Spaces:
Running
Running
darksakura
commited on
Commit
•
f82071f
0
Parent(s):
Duplicate from darksakura/l1
Browse filesThis view is limited to 50 files because it contains too many changes.
See raw diff
- .gitattributes +36 -0
- 12.mp3 +0 -0
- Eng_docs.md +109 -0
- LICENSE +407 -0
- README.md +14 -0
- app.cpython-310.pyc +0 -0
- cluster/__init__.py +29 -0
- cluster/train_cluster.py +89 -0
- configs/config-65.json +156 -0
- configs/config.json +159 -0
- cvec/checkpoint_best_legacy_500.pt +3 -0
- data_utils.py +184 -0
- diffusion/__init__.py +0 -0
- diffusion/__pycache__/__init__.cpython-38.pyc +0 -0
- diffusion/__pycache__/data_loaders.cpython-38.pyc +0 -0
- diffusion/__pycache__/diffusion.cpython-38.pyc +0 -0
- diffusion/__pycache__/dpm_solver_pytorch.cpython-38.pyc +0 -0
- diffusion/__pycache__/solver.cpython-38.pyc +0 -0
- diffusion/__pycache__/unit2mel.cpython-38.pyc +0 -0
- diffusion/__pycache__/vocoder.cpython-38.pyc +0 -0
- diffusion/__pycache__/wavenet.cpython-38.pyc +0 -0
- diffusion/data_loaders.py +284 -0
- diffusion/diffusion.py +317 -0
- diffusion/diffusion_onnx.py +612 -0
- diffusion/dpm_solver_pytorch.py +1201 -0
- diffusion/how to export onnx.md +4 -0
- diffusion/infer_gt_mel.py +74 -0
- diffusion/logger/__init__.py +0 -0
- diffusion/logger/__pycache__/__init__.cpython-38.pyc +0 -0
- diffusion/logger/__pycache__/saver.cpython-38.pyc +0 -0
- diffusion/logger/__pycache__/utils.cpython-38.pyc +0 -0
- diffusion/logger/saver.py +150 -0
- diffusion/logger/utils.py +126 -0
- diffusion/onnx_export.py +226 -0
- diffusion/solver.py +195 -0
- diffusion/unit2mel.py +147 -0
- diffusion/vocoder.py +94 -0
- diffusion/wavenet.py +108 -0
- filelists/test.txt +4 -0
- filelists/train.txt +15 -0
- filelists/val.txt +4 -0
- flask_api.py +60 -0
- hubert/__init__.py +0 -0
- hubert/checkpoint_best_legacy_500.pt +3 -0
- hubert/hubert_model.py +222 -0
- hubert/hubert_model_onnx.py +217 -0
- hubert/put_hubert_ckpt_here +0 -0
- inference/__init__.py +0 -0
- inference/chunks_temp.json +1 -0
- inference/infer_tool.py +533 -0
.gitattributes
ADDED
@@ -0,0 +1,36 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
*.7z filter=lfs diff=lfs merge=lfs -text
|
2 |
+
*.arrow filter=lfs diff=lfs merge=lfs -text
|
3 |
+
*.bin filter=lfs diff=lfs merge=lfs -text
|
4 |
+
*.bz2 filter=lfs diff=lfs merge=lfs -text
|
5 |
+
*.ckpt filter=lfs diff=lfs merge=lfs -text
|
6 |
+
*.ftz filter=lfs diff=lfs merge=lfs -text
|
7 |
+
*.gz filter=lfs diff=lfs merge=lfs -text
|
8 |
+
*.h5 filter=lfs diff=lfs merge=lfs -text
|
9 |
+
*.joblib filter=lfs diff=lfs merge=lfs -text
|
10 |
+
*.lfs.* filter=lfs diff=lfs merge=lfs -text
|
11 |
+
*.mlmodel filter=lfs diff=lfs merge=lfs -text
|
12 |
+
*.model filter=lfs diff=lfs merge=lfs -text
|
13 |
+
*.msgpack filter=lfs diff=lfs merge=lfs -text
|
14 |
+
*.npy filter=lfs diff=lfs merge=lfs -text
|
15 |
+
*.npz filter=lfs diff=lfs merge=lfs -text
|
16 |
+
*.onnx filter=lfs diff=lfs merge=lfs -text
|
17 |
+
*.ot filter=lfs diff=lfs merge=lfs -text
|
18 |
+
*.parquet filter=lfs diff=lfs merge=lfs -text
|
19 |
+
*.pb filter=lfs diff=lfs merge=lfs -text
|
20 |
+
*.pickle filter=lfs diff=lfs merge=lfs -text
|
21 |
+
*.pkl filter=lfs diff=lfs merge=lfs -text
|
22 |
+
*.pt filter=lfs diff=lfs merge=lfs -text
|
23 |
+
*.pth filter=lfs diff=lfs merge=lfs -text
|
24 |
+
*.rar filter=lfs diff=lfs merge=lfs -text
|
25 |
+
*.safetensors filter=lfs diff=lfs merge=lfs -text
|
26 |
+
saved_model/**/* filter=lfs diff=lfs merge=lfs -text
|
27 |
+
*.tar.* filter=lfs diff=lfs merge=lfs -text
|
28 |
+
*.tflite filter=lfs diff=lfs merge=lfs -text
|
29 |
+
*.tgz filter=lfs diff=lfs merge=lfs -text
|
30 |
+
*.wasm filter=lfs diff=lfs merge=lfs -text
|
31 |
+
*.xz filter=lfs diff=lfs merge=lfs -text
|
32 |
+
*.zip filter=lfs diff=lfs merge=lfs -text
|
33 |
+
*.zst filter=lfs diff=lfs merge=lfs -text
|
34 |
+
*tfevents* filter=lfs diff=lfs merge=lfs -text
|
35 |
+
Summertime.wav filter=lfs diff=lfs merge=lfs -text
|
36 |
+
pretrain/nsf_hifigan/model filter=lfs diff=lfs merge=lfs -text
|
12.mp3
ADDED
Binary file (206 kB). View file
|
|
Eng_docs.md
ADDED
@@ -0,0 +1,109 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# SoftVC VITS Singing Voice Conversion
|
2 |
+
|
3 |
+
## Updates
|
4 |
+
> According to incomplete statistics, it seems that training with multiple speakers may lead to **worsened leaking of voice timbre**. It is not recommended to train models with more than 5 speakers. The current suggestion is to try to train models with only a single speaker if you want to achieve a voice timbre that is more similar to the target.
|
5 |
+
> Fixed the issue with unwanted staccato, improving audio quality by a decent amount.\
|
6 |
+
> The 2.0 version has been moved to the 2.0 branch.\
|
7 |
+
> Version 3.0 uses the code structure of FreeVC, which isn't compatible with older versions.\
|
8 |
+
> Compared to [DiffSVC](https://github.com/prophesier/diff-svc) , diffsvc performs much better when the training data is of extremely high quality, but this repository may perform better on datasets with lower quality. Additionally, this repository is much faster in terms of inference speed compared to diffsvc.
|
9 |
+
|
10 |
+
## Model Overview
|
11 |
+
A singing voice coversion (SVC) model, using the SoftVC encoder to extract features from the input audio, sent into VITS along with the F0 to replace the original input to acheive a voice conversion effect. Additionally, changing the vocoder to [NSF HiFiGAN](https://github.com/openvpi/DiffSinger/tree/refactor/modules/nsf_hifigan) to fix the issue with unwanted staccato.
|
12 |
+
|
13 |
+
## Notice
|
14 |
+
+ The current branch is the 32kHz version, which requires less vram during inferencing, as well as faster inferencing speeds, and datasets for said branch take up less disk space. Thus the 32 kHz branch is recommended for use.
|
15 |
+
+ If you want to train 48 kHz variant models, switch to the [main branch](https://github.com/innnky/so-vits-svc/tree/main).
|
16 |
+
|
17 |
+
|
18 |
+
## Required models
|
19 |
+
+ soft vc hubert:[hubert-soft-0d54a1f4.pt](https://github.com/bshall/hubert/releases/download/v0.1/hubert-soft-0d54a1f4.pt)
|
20 |
+
+ Place under `hubert`.
|
21 |
+
+ Pretrained models [G_0.pth](https://huggingface.co/innnky/sovits_pretrained/resolve/main/G_0.pth) and [D_0.pth](https://huggingface.co/innnky/sovits_pretrained/resolve/main/D_0.pth)
|
22 |
+
+ Place under `logs/32k`.
|
23 |
+
+ Pretrained models are required, because from experiments, training from scratch can be rather unpredictable to say the least, and training with a pretrained model can greatly improve training speeds.
|
24 |
+
+ The pretrained model includes云灏, 即霜, 辉宇·星AI, 派蒙, and 绫地宁宁, covering the common ranges of both male and female voices, and so it can be seen as a rather universal pretrained model.
|
25 |
+
+ The pretrained model exludes the `optimizer speaker_embedding` section, rendering it only usable for pretraining and incapable of inferencing with.
|
26 |
+
```shell
|
27 |
+
# For simple downloading.
|
28 |
+
# hubert
|
29 |
+
wget -P hubert/ https://github.com/bshall/hubert/releases/download/v0.1/hubert-soft-0d54a1f4.pt
|
30 |
+
# G&D pretrained models
|
31 |
+
wget -P logs/32k/ https://huggingface.co/innnky/sovits_pretrained/resolve/main/G_0.pth
|
32 |
+
wget -P logs/32k/ https://huggingface.co/innnky/sovits_pretrained/resolve/main/D_0.pth
|
33 |
+
|
34 |
+
```
|
35 |
+
|
36 |
+
## Colab notebook script for dataset creation and training.
|
37 |
+
[colab training notebook](https://colab.research.google.com/drive/1rCUOOVG7-XQlVZuWRAj5IpGrMM8t07pE?usp=sharing)
|
38 |
+
|
39 |
+
## Dataset preparation
|
40 |
+
All that is required is that the data be put under the `dataset_raw` folder in the structure format provided below.
|
41 |
+
```shell
|
42 |
+
dataset_raw
|
43 |
+
├───speaker0
|
44 |
+
│ ├───xxx1-xxx1.wav
|
45 |
+
│ ├───...
|
46 |
+
│ └───Lxx-0xx8.wav
|
47 |
+
└───speaker1
|
48 |
+
├───xx2-0xxx2.wav
|
49 |
+
├───...
|
50 |
+
└───xxx7-xxx007.wav
|
51 |
+
```
|
52 |
+
|
53 |
+
## Data pre-processing.
|
54 |
+
1. Resample to 32khz
|
55 |
+
|
56 |
+
```shell
|
57 |
+
python resample.py
|
58 |
+
```
|
59 |
+
2. Automatically sort out training set, validation set, test set, and automatically generate configuration files.
|
60 |
+
```shell
|
61 |
+
python preprocess_flist_config.py
|
62 |
+
# Notice.
|
63 |
+
# The n_speakers value in the config will be set automatically according to the amount of speakers in the dataset.
|
64 |
+
# To reserve space for additionally added speakers in the dataset, the n_speakers value will be be set to twice the actual amount.
|
65 |
+
# If you want even more space for adding more data, you can edit the n_speakers value in the config after runing this step.
|
66 |
+
# This can not be changed after training starts.
|
67 |
+
```
|
68 |
+
3. Generate hubert and F0 features/
|
69 |
+
```shell
|
70 |
+
python preprocess_hubert_f0.py
|
71 |
+
```
|
72 |
+
After running the step above, the `dataset` folder will contain all the pre-processed data, you can delete the `dataset_raw` folder after that.
|
73 |
+
|
74 |
+
## Training.
|
75 |
+
```shell
|
76 |
+
python train.py -c configs/config.json -m 32k
|
77 |
+
```
|
78 |
+
|
79 |
+
## Inferencing.
|
80 |
+
|
81 |
+
Use [inference_main.py](inference_main.py)
|
82 |
+
+ Edit `model_path` to your newest checkpoint.
|
83 |
+
+ Place the input audio under the `raw` folder.
|
84 |
+
+ Change `clean_names` to the output file name.
|
85 |
+
+ Use `trans` to edit the pitch shifting amount (semitones).
|
86 |
+
+ Change `spk_list` to the speaker name.
|
87 |
+
|
88 |
+
## Onnx Exporting.
|
89 |
+
### **When exporting Onnx, please make sure you re-clone the whole repository!!!**
|
90 |
+
Use [onnx_export.py](onnx_export.py)
|
91 |
+
+ Create a new folder called `checkpoints`.
|
92 |
+
+ Create a project folder in `checkpoints` folder with the desired name for your project, let's use `myproject` as example. Folder structure looks like `./checkpoints/myproject`.
|
93 |
+
+ Rename your model to `model.pth`, rename your config file to `config.json` then move them into `myproject` folder.
|
94 |
+
+ Modify [onnx_export.py](onnx_export.py) where `path = "NyaruTaffy"`, change `NyaruTaffy` to your project name, here it will be `path = "myproject"`.
|
95 |
+
+ Run [onnx_export.py](onnx_export.py)
|
96 |
+
+ Once it finished, a `model.onnx` will be generated in `myproject` folder, that's the model you just exported.
|
97 |
+
+ Notice: if you want to export a 48K model, please follow the instruction below or use `model_onnx_48k.py` directly.
|
98 |
+
+ Open [model_onnx.py](model_onnx.py) and change `hps={"sampling_rate": 32000...}` to `hps={"sampling_rate": 48000}` in class `SynthesizerTrn`.
|
99 |
+
+ Open [nvSTFT](/vdecoder/hifigan/nvSTFT.py) and replace all `32000` with `48000`
|
100 |
+
### Onnx Model UI Support
|
101 |
+
+ [MoeSS](https://github.com/NaruseMioShirakana/MoeSS)
|
102 |
+
+ All training function and transformation are removed, only if they are all removed you are actually using Onnx.
|
103 |
+
|
104 |
+
## Gradio (WebUI)
|
105 |
+
Use [sovits_gradio.py](sovits_gradio.py) to run Gradio WebUI
|
106 |
+
+ Create a new folder called `checkpoints`.
|
107 |
+
+ Create a project folder in `checkpoints` folder with the desired name for your project, let's use `myproject` as example. Folder structure looks like `./checkpoints/myproject`.
|
108 |
+
+ Rename your model to `model.pth`, rename your config file to `config.json` then move them into `myproject` folder.
|
109 |
+
+ Run [sovits_gradio.py](sovits_gradio.py)
|
LICENSE
ADDED
@@ -0,0 +1,407 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
Attribution-NonCommercial 4.0 International
|
2 |
+
|
3 |
+
=======================================================================
|
4 |
+
|
5 |
+
Creative Commons Corporation ("Creative Commons") is not a law firm and
|
6 |
+
does not provide legal services or legal advice. Distribution of
|
7 |
+
Creative Commons public licenses does not create a lawyer-client or
|
8 |
+
other relationship. Creative Commons makes its licenses and related
|
9 |
+
information available on an "as-is" basis. Creative Commons gives no
|
10 |
+
warranties regarding its licenses, any material licensed under their
|
11 |
+
terms and conditions, or any related information. Creative Commons
|
12 |
+
disclaims all liability for damages resulting from their use to the
|
13 |
+
fullest extent possible.
|
14 |
+
|
15 |
+
Using Creative Commons Public Licenses
|
16 |
+
|
17 |
+
Creative Commons public licenses provide a standard set of terms and
|
18 |
+
conditions that creators and other rights holders may use to share
|
19 |
+
original works of authorship and other material subject to copyright
|
20 |
+
and certain other rights specified in the public license below. The
|
21 |
+
following considerations are for informational purposes only, are not
|
22 |
+
exhaustive, and do not form part of our licenses.
|
23 |
+
|
24 |
+
Considerations for licensors: Our public licenses are
|
25 |
+
intended for use by those authorized to give the public
|
26 |
+
permission to use material in ways otherwise restricted by
|
27 |
+
copyright and certain other rights. Our licenses are
|
28 |
+
irrevocable. Licensors should read and understand the terms
|
29 |
+
and conditions of the license they choose before applying it.
|
30 |
+
Licensors should also secure all rights necessary before
|
31 |
+
applying our licenses so that the public can reuse the
|
32 |
+
material as expected. Licensors should clearly mark any
|
33 |
+
material not subject to the license. This includes other CC-
|
34 |
+
licensed material, or material used under an exception or
|
35 |
+
limitation to copyright. More considerations for licensors:
|
36 |
+
wiki.creativecommons.org/Considerations_for_licensors
|
37 |
+
|
38 |
+
Considerations for the public: By using one of our public
|
39 |
+
licenses, a licensor grants the public permission to use the
|
40 |
+
licensed material under specified terms and conditions. If
|
41 |
+
the licensor's permission is not necessary for any reason--for
|
42 |
+
example, because of any applicable exception or limitation to
|
43 |
+
copyright--then that use is not regulated by the license. Our
|
44 |
+
licenses grant only permissions under copyright and certain
|
45 |
+
other rights that a licensor has authority to grant. Use of
|
46 |
+
the licensed material may still be restricted for other
|
47 |
+
reasons, including because others have copyright or other
|
48 |
+
rights in the material. A licensor may make special requests,
|
49 |
+
such as asking that all changes be marked or described.
|
50 |
+
Although not required by our licenses, you are encouraged to
|
51 |
+
respect those requests where reasonable. More considerations
|
52 |
+
for the public:
|
53 |
+
wiki.creativecommons.org/Considerations_for_licensees
|
54 |
+
|
55 |
+
=======================================================================
|
56 |
+
|
57 |
+
Creative Commons Attribution-NonCommercial 4.0 International Public
|
58 |
+
License
|
59 |
+
|
60 |
+
By exercising the Licensed Rights (defined below), You accept and agree
|
61 |
+
to be bound by the terms and conditions of this Creative Commons
|
62 |
+
Attribution-NonCommercial 4.0 International Public License ("Public
|
63 |
+
License"). To the extent this Public License may be interpreted as a
|
64 |
+
contract, You are granted the Licensed Rights in consideration of Your
|
65 |
+
acceptance of these terms and conditions, and the Licensor grants You
|
66 |
+
such rights in consideration of benefits the Licensor receives from
|
67 |
+
making the Licensed Material available under these terms and
|
68 |
+
conditions.
|
69 |
+
|
70 |
+
|
71 |
+
Section 1 -- Definitions.
|
72 |
+
|
73 |
+
a. Adapted Material means material subject to Copyright and Similar
|
74 |
+
Rights that is derived from or based upon the Licensed Material
|
75 |
+
and in which the Licensed Material is translated, altered,
|
76 |
+
arranged, transformed, or otherwise modified in a manner requiring
|
77 |
+
permission under the Copyright and Similar Rights held by the
|
78 |
+
Licensor. For purposes of this Public License, where the Licensed
|
79 |
+
Material is a musical work, performance, or sound recording,
|
80 |
+
Adapted Material is always produced where the Licensed Material is
|
81 |
+
synched in timed relation with a moving image.
|
82 |
+
|
83 |
+
b. Adapter's License means the license You apply to Your Copyright
|
84 |
+
and Similar Rights in Your contributions to Adapted Material in
|
85 |
+
accordance with the terms and conditions of this Public License.
|
86 |
+
|
87 |
+
c. Copyright and Similar Rights means copyright and/or similar rights
|
88 |
+
closely related to copyright including, without limitation,
|
89 |
+
performance, broadcast, sound recording, and Sui Generis Database
|
90 |
+
Rights, without regard to how the rights are labeled or
|
91 |
+
categorized. For purposes of this Public License, the rights
|
92 |
+
specified in Section 2(b)(1)-(2) are not Copyright and Similar
|
93 |
+
Rights.
|
94 |
+
d. Effective Technological Measures means those measures that, in the
|
95 |
+
absence of proper authority, may not be circumvented under laws
|
96 |
+
fulfilling obligations under Article 11 of the WIPO Copyright
|
97 |
+
Treaty adopted on December 20, 1996, and/or similar international
|
98 |
+
agreements.
|
99 |
+
|
100 |
+
e. Exceptions and Limitations means fair use, fair dealing, and/or
|
101 |
+
any other exception or limitation to Copyright and Similar Rights
|
102 |
+
that applies to Your use of the Licensed Material.
|
103 |
+
|
104 |
+
f. Licensed Material means the artistic or literary work, database,
|
105 |
+
or other material to which the Licensor applied this Public
|
106 |
+
License.
|
107 |
+
|
108 |
+
g. Licensed Rights means the rights granted to You subject to the
|
109 |
+
terms and conditions of this Public License, which are limited to
|
110 |
+
all Copyright and Similar Rights that apply to Your use of the
|
111 |
+
Licensed Material and that the Licensor has authority to license.
|
112 |
+
|
113 |
+
h. Licensor means the individual(s) or entity(ies) granting rights
|
114 |
+
under this Public License.
|
115 |
+
|
116 |
+
i. NonCommercial means not primarily intended for or directed towards
|
117 |
+
commercial advantage or monetary compensation. For purposes of
|
118 |
+
this Public License, the exchange of the Licensed Material for
|
119 |
+
other material subject to Copyright and Similar Rights by digital
|
120 |
+
file-sharing or similar means is NonCommercial provided there is
|
121 |
+
no payment of monetary compensation in connection with the
|
122 |
+
exchange.
|
123 |
+
|
124 |
+
j. Share means to provide material to the public by any means or
|
125 |
+
process that requires permission under the Licensed Rights, such
|
126 |
+
as reproduction, public display, public performance, distribution,
|
127 |
+
dissemination, communication, or importation, and to make material
|
128 |
+
available to the public including in ways that members of the
|
129 |
+
public may access the material from a place and at a time
|
130 |
+
individually chosen by them.
|
131 |
+
|
132 |
+
k. Sui Generis Database Rights means rights other than copyright
|
133 |
+
resulting from Directive 96/9/EC of the European Parliament and of
|
134 |
+
the Council of 11 March 1996 on the legal protection of databases,
|
135 |
+
as amended and/or succeeded, as well as other essentially
|
136 |
+
equivalent rights anywhere in the world.
|
137 |
+
|
138 |
+
l. You means the individual or entity exercising the Licensed Rights
|
139 |
+
under this Public License. Your has a corresponding meaning.
|
140 |
+
|
141 |
+
|
142 |
+
Section 2 -- Scope.
|
143 |
+
|
144 |
+
a. License grant.
|
145 |
+
|
146 |
+
1. Subject to the terms and conditions of this Public License,
|
147 |
+
the Licensor hereby grants You a worldwide, royalty-free,
|
148 |
+
non-sublicensable, non-exclusive, irrevocable license to
|
149 |
+
exercise the Licensed Rights in the Licensed Material to:
|
150 |
+
|
151 |
+
a. reproduce and Share the Licensed Material, in whole or
|
152 |
+
in part, for NonCommercial purposes only; and
|
153 |
+
|
154 |
+
b. produce, reproduce, and Share Adapted Material for
|
155 |
+
NonCommercial purposes only.
|
156 |
+
|
157 |
+
2. Exceptions and Limitations. For the avoidance of doubt, where
|
158 |
+
Exceptions and Limitations apply to Your use, this Public
|
159 |
+
License does not apply, and You do not need to comply with
|
160 |
+
its terms and conditions.
|
161 |
+
|
162 |
+
3. Term. The term of this Public License is specified in Section
|
163 |
+
6(a).
|
164 |
+
|
165 |
+
4. Media and formats; technical modifications allowed. The
|
166 |
+
Licensor authorizes You to exercise the Licensed Rights in
|
167 |
+
all media and formats whether now known or hereafter created,
|
168 |
+
and to make technical modifications necessary to do so. The
|
169 |
+
Licensor waives and/or agrees not to assert any right or
|
170 |
+
authority to forbid You from making technical modifications
|
171 |
+
necessary to exercise the Licensed Rights, including
|
172 |
+
technical modifications necessary to circumvent Effective
|
173 |
+
Technological Measures. For purposes of this Public License,
|
174 |
+
simply making modifications authorized by this Section 2(a)
|
175 |
+
(4) never produces Adapted Material.
|
176 |
+
|
177 |
+
5. Downstream recipients.
|
178 |
+
|
179 |
+
a. Offer from the Licensor -- Licensed Material. Every
|
180 |
+
recipient of the Licensed Material automatically
|
181 |
+
receives an offer from the Licensor to exercise the
|
182 |
+
Licensed Rights under the terms and conditions of this
|
183 |
+
Public License.
|
184 |
+
|
185 |
+
b. No downstream restrictions. You may not offer or impose
|
186 |
+
any additional or different terms or conditions on, or
|
187 |
+
apply any Effective Technological Measures to, the
|
188 |
+
Licensed Material if doing so restricts exercise of the
|
189 |
+
Licensed Rights by any recipient of the Licensed
|
190 |
+
Material.
|
191 |
+
|
192 |
+
6. No endorsement. Nothing in this Public License constitutes or
|
193 |
+
may be construed as permission to assert or imply that You
|
194 |
+
are, or that Your use of the Licensed Material is, connected
|
195 |
+
with, or sponsored, endorsed, or granted official status by,
|
196 |
+
the Licensor or others designated to receive attribution as
|
197 |
+
provided in Section 3(a)(1)(A)(i).
|
198 |
+
|
199 |
+
b. Other rights.
|
200 |
+
|
201 |
+
1. Moral rights, such as the right of integrity, are not
|
202 |
+
licensed under this Public License, nor are publicity,
|
203 |
+
privacy, and/or other similar personality rights; however, to
|
204 |
+
the extent possible, the Licensor waives and/or agrees not to
|
205 |
+
assert any such rights held by the Licensor to the limited
|
206 |
+
extent necessary to allow You to exercise the Licensed
|
207 |
+
Rights, but not otherwise.
|
208 |
+
|
209 |
+
2. Patent and trademark rights are not licensed under this
|
210 |
+
Public License.
|
211 |
+
|
212 |
+
3. To the extent possible, the Licensor waives any right to
|
213 |
+
collect royalties from You for the exercise of the Licensed
|
214 |
+
Rights, whether directly or through a collecting society
|
215 |
+
under any voluntary or waivable statutory or compulsory
|
216 |
+
licensing scheme. In all other cases the Licensor expressly
|
217 |
+
reserves any right to collect such royalties, including when
|
218 |
+
the Licensed Material is used other than for NonCommercial
|
219 |
+
purposes.
|
220 |
+
|
221 |
+
|
222 |
+
Section 3 -- License Conditions.
|
223 |
+
|
224 |
+
Your exercise of the Licensed Rights is expressly made subject to the
|
225 |
+
following conditions.
|
226 |
+
|
227 |
+
a. Attribution.
|
228 |
+
|
229 |
+
1. If You Share the Licensed Material (including in modified
|
230 |
+
form), You must:
|
231 |
+
|
232 |
+
a. retain the following if it is supplied by the Licensor
|
233 |
+
with the Licensed Material:
|
234 |
+
|
235 |
+
i. identification of the creator(s) of the Licensed
|
236 |
+
Material and any others designated to receive
|
237 |
+
attribution, in any reasonable manner requested by
|
238 |
+
the Licensor (including by pseudonym if
|
239 |
+
designated);
|
240 |
+
|
241 |
+
ii. a copyright notice;
|
242 |
+
|
243 |
+
iii. a notice that refers to this Public License;
|
244 |
+
|
245 |
+
iv. a notice that refers to the disclaimer of
|
246 |
+
warranties;
|
247 |
+
|
248 |
+
v. a URI or hyperlink to the Licensed Material to the
|
249 |
+
extent reasonably practicable;
|
250 |
+
|
251 |
+
b. indicate if You modified the Licensed Material and
|
252 |
+
retain an indication of any previous modifications; and
|
253 |
+
|
254 |
+
c. indicate the Licensed Material is licensed under this
|
255 |
+
Public License, and include the text of, or the URI or
|
256 |
+
hyperlink to, this Public License.
|
257 |
+
|
258 |
+
2. You may satisfy the conditions in Section 3(a)(1) in any
|
259 |
+
reasonable manner based on the medium, means, and context in
|
260 |
+
which You Share the Licensed Material. For example, it may be
|
261 |
+
reasonable to satisfy the conditions by providing a URI or
|
262 |
+
hyperlink to a resource that includes the required
|
263 |
+
information.
|
264 |
+
|
265 |
+
3. If requested by the Licensor, You must remove any of the
|
266 |
+
information required by Section 3(a)(1)(A) to the extent
|
267 |
+
reasonably practicable.
|
268 |
+
|
269 |
+
4. If You Share Adapted Material You produce, the Adapter's
|
270 |
+
License You apply must not prevent recipients of the Adapted
|
271 |
+
Material from complying with this Public License.
|
272 |
+
|
273 |
+
|
274 |
+
Section 4 -- Sui Generis Database Rights.
|
275 |
+
|
276 |
+
Where the Licensed Rights include Sui Generis Database Rights that
|
277 |
+
apply to Your use of the Licensed Material:
|
278 |
+
|
279 |
+
a. for the avoidance of doubt, Section 2(a)(1) grants You the right
|
280 |
+
to extract, reuse, reproduce, and Share all or a substantial
|
281 |
+
portion of the contents of the database for NonCommercial purposes
|
282 |
+
only;
|
283 |
+
|
284 |
+
b. if You include all or a substantial portion of the database
|
285 |
+
contents in a database in which You have Sui Generis Database
|
286 |
+
Rights, then the database in which You have Sui Generis Database
|
287 |
+
Rights (but not its individual contents) is Adapted Material; and
|
288 |
+
|
289 |
+
c. You must comply with the conditions in Section 3(a) if You Share
|
290 |
+
all or a substantial portion of the contents of the database.
|
291 |
+
|
292 |
+
For the avoidance of doubt, this Section 4 supplements and does not
|
293 |
+
replace Your obligations under this Public License where the Licensed
|
294 |
+
Rights include other Copyright and Similar Rights.
|
295 |
+
|
296 |
+
|
297 |
+
Section 5 -- Disclaimer of Warranties and Limitation of Liability.
|
298 |
+
|
299 |
+
a. UNLESS OTHERWISE SEPARATELY UNDERTAKEN BY THE LICENSOR, TO THE
|
300 |
+
EXTENT POSSIBLE, THE LICENSOR OFFERS THE LICENSED MATERIAL AS-IS
|
301 |
+
AND AS-AVAILABLE, AND MAKES NO REPRESENTATIONS OR WARRANTIES OF
|
302 |
+
ANY KIND CONCERNING THE LICENSED MATERIAL, WHETHER EXPRESS,
|
303 |
+
IMPLIED, STATUTORY, OR OTHER. THIS INCLUDES, WITHOUT LIMITATION,
|
304 |
+
WARRANTIES OF TITLE, MERCHANTABILITY, FITNESS FOR A PARTICULAR
|
305 |
+
PURPOSE, NON-INFRINGEMENT, ABSENCE OF LATENT OR OTHER DEFECTS,
|
306 |
+
ACCURACY, OR THE PRESENCE OR ABSENCE OF ERRORS, WHETHER OR NOT
|
307 |
+
KNOWN OR DISCOVERABLE. WHERE DISCLAIMERS OF WARRANTIES ARE NOT
|
308 |
+
ALLOWED IN FULL OR IN PART, THIS DISCLAIMER MAY NOT APPLY TO YOU.
|
309 |
+
|
310 |
+
b. TO THE EXTENT POSSIBLE, IN NO EVENT WILL THE LICENSOR BE LIABLE
|
311 |
+
TO YOU ON ANY LEGAL THEORY (INCLUDING, WITHOUT LIMITATION,
|
312 |
+
NEGLIGENCE) OR OTHERWISE FOR ANY DIRECT, SPECIAL, INDIRECT,
|
313 |
+
INCIDENTAL, CONSEQUENTIAL, PUNITIVE, EXEMPLARY, OR OTHER LOSSES,
|
314 |
+
COSTS, EXPENSES, OR DAMAGES ARISING OUT OF THIS PUBLIC LICENSE OR
|
315 |
+
USE OF THE LICENSED MATERIAL, EVEN IF THE LICENSOR HAS BEEN
|
316 |
+
ADVISED OF THE POSSIBILITY OF SUCH LOSSES, COSTS, EXPENSES, OR
|
317 |
+
DAMAGES. WHERE A LIMITATION OF LIABILITY IS NOT ALLOWED IN FULL OR
|
318 |
+
IN PART, THIS LIMITATION MAY NOT APPLY TO YOU.
|
319 |
+
|
320 |
+
c. The disclaimer of warranties and limitation of liability provided
|
321 |
+
above shall be interpreted in a manner that, to the extent
|
322 |
+
possible, most closely approximates an absolute disclaimer and
|
323 |
+
waiver of all liability.
|
324 |
+
|
325 |
+
|
326 |
+
Section 6 -- Term and Termination.
|
327 |
+
|
328 |
+
a. This Public License applies for the term of the Copyright and
|
329 |
+
Similar Rights licensed here. However, if You fail to comply with
|
330 |
+
this Public License, then Your rights under this Public License
|
331 |
+
terminate automatically.
|
332 |
+
|
333 |
+
b. Where Your right to use the Licensed Material has terminated under
|
334 |
+
Section 6(a), it reinstates:
|
335 |
+
|
336 |
+
1. automatically as of the date the violation is cured, provided
|
337 |
+
it is cured within 30 days of Your discovery of the
|
338 |
+
violation; or
|
339 |
+
|
340 |
+
2. upon express reinstatement by the Licensor.
|
341 |
+
|
342 |
+
For the avoidance of doubt, this Section 6(b) does not affect any
|
343 |
+
right the Licensor may have to seek remedies for Your violations
|
344 |
+
of this Public License.
|
345 |
+
|
346 |
+
c. For the avoidance of doubt, the Licensor may also offer the
|
347 |
+
Licensed Material under separate terms or conditions or stop
|
348 |
+
distributing the Licensed Material at any time; however, doing so
|
349 |
+
will not terminate this Public License.
|
350 |
+
|
351 |
+
d. Sections 1, 5, 6, 7, and 8 survive termination of this Public
|
352 |
+
License.
|
353 |
+
|
354 |
+
|
355 |
+
Section 7 -- Other Terms and Conditions.
|
356 |
+
|
357 |
+
a. The Licensor shall not be bound by any additional or different
|
358 |
+
terms or conditions communicated by You unless expressly agreed.
|
359 |
+
|
360 |
+
b. Any arrangements, understandings, or agreements regarding the
|
361 |
+
Licensed Material not stated herein are separate from and
|
362 |
+
independent of the terms and conditions of this Public License.
|
363 |
+
|
364 |
+
|
365 |
+
Section 8 -- Interpretation.
|
366 |
+
|
367 |
+
a. For the avoidance of doubt, this Public License does not, and
|
368 |
+
shall not be interpreted to, reduce, limit, restrict, or impose
|
369 |
+
conditions on any use of the Licensed Material that could lawfully
|
370 |
+
be made without permission under this Public License.
|
371 |
+
|
372 |
+
b. To the extent possible, if any provision of this Public License is
|
373 |
+
deemed unenforceable, it shall be automatically reformed to the
|
374 |
+
minimum extent necessary to make it enforceable. If the provision
|
375 |
+
cannot be reformed, it shall be severed from this Public License
|
376 |
+
without affecting the enforceability of the remaining terms and
|
377 |
+
conditions.
|
378 |
+
|
379 |
+
c. No term or condition of this Public License will be waived and no
|
380 |
+
failure to comply consented to unless expressly agreed to by the
|
381 |
+
Licensor.
|
382 |
+
|
383 |
+
d. Nothing in this Public License constitutes or may be interpreted
|
384 |
+
as a limitation upon, or waiver of, any privileges and immunities
|
385 |
+
that apply to the Licensor or You, including from the legal
|
386 |
+
processes of any jurisdiction or authority.
|
387 |
+
|
388 |
+
=======================================================================
|
389 |
+
|
390 |
+
Creative Commons is not a party to its public
|
391 |
+
licenses. Notwithstanding, Creative Commons may elect to apply one of
|
392 |
+
its public licenses to material it publishes and in those instances
|
393 |
+
will be considered the “Licensor.” The text of the Creative Commons
|
394 |
+
public licenses is dedicated to the public domain under the CC0 Public
|
395 |
+
Domain Dedication. Except for the limited purpose of indicating that
|
396 |
+
material is shared under a Creative Commons public license or as
|
397 |
+
otherwise permitted by the Creative Commons policies published at
|
398 |
+
creativecommons.org/policies, Creative Commons does not authorize the
|
399 |
+
use of the trademark "Creative Commons" or any other trademark or logo
|
400 |
+
of Creative Commons without its prior written consent including,
|
401 |
+
without limitation, in connection with any unauthorized modifications
|
402 |
+
to any of its public licenses or any other arrangements,
|
403 |
+
understandings, or agreements concerning use of licensed material. For
|
404 |
+
the avoidance of doubt, this paragraph does not form part of the
|
405 |
+
public licenses.
|
406 |
+
|
407 |
+
Creative Commons may be contacted at creativecommons.org.
|
README.md
ADDED
@@ -0,0 +1,14 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
---
|
2 |
+
title: 🏆
|
3 |
+
emoji: 🏆
|
4 |
+
colorFrom: gray
|
5 |
+
colorTo: pink
|
6 |
+
sdk: gradio
|
7 |
+
sdk_version: 3.18.0
|
8 |
+
app_file: app.cpython-310.pyc
|
9 |
+
pinned: false
|
10 |
+
license: mit
|
11 |
+
duplicated_from: darksakura/l1
|
12 |
+
---
|
13 |
+
|
14 |
+
Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
|
app.cpython-310.pyc
ADDED
Binary file (12.4 kB). View file
|
|
cluster/__init__.py
ADDED
@@ -0,0 +1,29 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import numpy as np
|
2 |
+
import torch
|
3 |
+
from sklearn.cluster import KMeans
|
4 |
+
|
5 |
+
def get_cluster_model(ckpt_path):
|
6 |
+
checkpoint = torch.load(ckpt_path)
|
7 |
+
kmeans_dict = {}
|
8 |
+
for spk, ckpt in checkpoint.items():
|
9 |
+
km = KMeans(ckpt["n_features_in_"])
|
10 |
+
km.__dict__["n_features_in_"] = ckpt["n_features_in_"]
|
11 |
+
km.__dict__["_n_threads"] = ckpt["_n_threads"]
|
12 |
+
km.__dict__["cluster_centers_"] = ckpt["cluster_centers_"]
|
13 |
+
kmeans_dict[spk] = km
|
14 |
+
return kmeans_dict
|
15 |
+
|
16 |
+
def get_cluster_result(model, x, speaker):
|
17 |
+
"""
|
18 |
+
x: np.array [t, 256]
|
19 |
+
return cluster class result
|
20 |
+
"""
|
21 |
+
return model[speaker].predict(x)
|
22 |
+
|
23 |
+
def get_cluster_center_result(model, x,speaker):
|
24 |
+
"""x: np.array [t, 256]"""
|
25 |
+
predict = model[speaker].predict(x)
|
26 |
+
return model[speaker].cluster_centers_[predict]
|
27 |
+
|
28 |
+
def get_center(model, x,speaker):
|
29 |
+
return model[speaker].cluster_centers_[x]
|
cluster/train_cluster.py
ADDED
@@ -0,0 +1,89 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
from glob import glob
|
3 |
+
from pathlib import Path
|
4 |
+
import torch
|
5 |
+
import logging
|
6 |
+
import argparse
|
7 |
+
import torch
|
8 |
+
import numpy as np
|
9 |
+
from sklearn.cluster import KMeans, MiniBatchKMeans
|
10 |
+
import tqdm
|
11 |
+
logging.basicConfig(level=logging.INFO)
|
12 |
+
logger = logging.getLogger(__name__)
|
13 |
+
import time
|
14 |
+
import random
|
15 |
+
|
16 |
+
def train_cluster(in_dir, n_clusters, use_minibatch=True, verbose=False):
|
17 |
+
|
18 |
+
logger.info(f"Loading features from {in_dir}")
|
19 |
+
features = []
|
20 |
+
nums = 0
|
21 |
+
for path in tqdm.tqdm(in_dir.glob("*.soft.pt")):
|
22 |
+
features.append(torch.load(path).squeeze(0).numpy().T)
|
23 |
+
# print(features[-1].shape)
|
24 |
+
features = np.concatenate(features, axis=0)
|
25 |
+
print(nums, features.nbytes/ 1024**2, "MB , shape:",features.shape, features.dtype)
|
26 |
+
features = features.astype(np.float32)
|
27 |
+
logger.info(f"Clustering features of shape: {features.shape}")
|
28 |
+
t = time.time()
|
29 |
+
if use_minibatch:
|
30 |
+
kmeans = MiniBatchKMeans(n_clusters=n_clusters,verbose=verbose, batch_size=4096, max_iter=80).fit(features)
|
31 |
+
else:
|
32 |
+
kmeans = KMeans(n_clusters=n_clusters,verbose=verbose).fit(features)
|
33 |
+
print(time.time()-t, "s")
|
34 |
+
|
35 |
+
x = {
|
36 |
+
"n_features_in_": kmeans.n_features_in_,
|
37 |
+
"_n_threads": kmeans._n_threads,
|
38 |
+
"cluster_centers_": kmeans.cluster_centers_,
|
39 |
+
}
|
40 |
+
print("end")
|
41 |
+
|
42 |
+
return x
|
43 |
+
|
44 |
+
|
45 |
+
if __name__ == "__main__":
|
46 |
+
|
47 |
+
parser = argparse.ArgumentParser()
|
48 |
+
parser.add_argument('--dataset', type=Path, default="./dataset/44k",
|
49 |
+
help='path of training data directory')
|
50 |
+
parser.add_argument('--output', type=Path, default="logs/44k",
|
51 |
+
help='path of model output directory')
|
52 |
+
|
53 |
+
args = parser.parse_args()
|
54 |
+
|
55 |
+
checkpoint_dir = args.output
|
56 |
+
dataset = args.dataset
|
57 |
+
n_clusters = 10000
|
58 |
+
|
59 |
+
ckpt = {}
|
60 |
+
for spk in os.listdir(dataset):
|
61 |
+
if os.path.isdir(dataset/spk):
|
62 |
+
print(f"train kmeans for {spk}...")
|
63 |
+
in_dir = dataset/spk
|
64 |
+
x = train_cluster(in_dir, n_clusters, verbose=False)
|
65 |
+
ckpt[spk] = x
|
66 |
+
|
67 |
+
checkpoint_path = checkpoint_dir / f"kmeans_{n_clusters}.pt"
|
68 |
+
checkpoint_path.parent.mkdir(exist_ok=True, parents=True)
|
69 |
+
torch.save(
|
70 |
+
ckpt,
|
71 |
+
checkpoint_path,
|
72 |
+
)
|
73 |
+
|
74 |
+
|
75 |
+
# import cluster
|
76 |
+
# for spk in tqdm.tqdm(os.listdir("dataset")):
|
77 |
+
# if os.path.isdir(f"dataset/{spk}"):
|
78 |
+
# print(f"start kmeans inference for {spk}...")
|
79 |
+
# for feature_path in tqdm.tqdm(glob(f"dataset/{spk}/*.discrete.npy", recursive=True)):
|
80 |
+
# mel_path = feature_path.replace(".discrete.npy",".mel.npy")
|
81 |
+
# mel_spectrogram = np.load(mel_path)
|
82 |
+
# feature_len = mel_spectrogram.shape[-1]
|
83 |
+
# c = np.load(feature_path)
|
84 |
+
# c = utils.tools.repeat_expand_2d(torch.FloatTensor(c), feature_len).numpy()
|
85 |
+
# feature = c.T
|
86 |
+
# feature_class = cluster.get_cluster_result(feature, spk)
|
87 |
+
# np.save(feature_path.replace(".discrete.npy", ".discrete_class.npy"), feature_class)
|
88 |
+
|
89 |
+
|
configs/config-65.json
ADDED
@@ -0,0 +1,156 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"train": {
|
3 |
+
"log_interval": 200,
|
4 |
+
"eval_interval": 1600,
|
5 |
+
"seed": 1234,
|
6 |
+
"epochs": 10000,
|
7 |
+
"learning_rate": 0.0001,
|
8 |
+
"betas": [
|
9 |
+
0.8,
|
10 |
+
0.99
|
11 |
+
],
|
12 |
+
"eps": 1e-09,
|
13 |
+
"batch_size": 1,
|
14 |
+
"fp16_run": false,
|
15 |
+
"lr_decay": 0.999875,
|
16 |
+
"segment_size": 10240,
|
17 |
+
"init_lr_ratio": 1,
|
18 |
+
"warmup_epochs": 0,
|
19 |
+
"c_mel": 45,
|
20 |
+
"c_kl": 1.0,
|
21 |
+
"use_sr": true,
|
22 |
+
"max_speclen": 512,
|
23 |
+
"port": "8001",
|
24 |
+
"keep_ckpts": 53,
|
25 |
+
"all_in_mem": false
|
26 |
+
},
|
27 |
+
"data": {
|
28 |
+
"training_files": "filelists/train.txt",
|
29 |
+
"validation_files": "filelists/val.txt",
|
30 |
+
"max_wav_value": 32768.0,
|
31 |
+
"sampling_rate": 44100,
|
32 |
+
"filter_length": 2048,
|
33 |
+
"hop_length": 512,
|
34 |
+
"win_length": 2048,
|
35 |
+
"n_mel_channels": 80,
|
36 |
+
"mel_fmin": 0.0,
|
37 |
+
"mel_fmax": 22050
|
38 |
+
},
|
39 |
+
"model": {
|
40 |
+
"inter_channels": 192,
|
41 |
+
"hidden_channels": 192,
|
42 |
+
"filter_channels": 768,
|
43 |
+
"n_heads": 2,
|
44 |
+
"n_layers": 6,
|
45 |
+
"kernel_size": 3,
|
46 |
+
"p_dropout": 0.1,
|
47 |
+
"resblock": "1",
|
48 |
+
"resblock_kernel_sizes": [
|
49 |
+
3,
|
50 |
+
7,
|
51 |
+
11
|
52 |
+
],
|
53 |
+
"resblock_dilation_sizes": [
|
54 |
+
[
|
55 |
+
1,
|
56 |
+
3,
|
57 |
+
5
|
58 |
+
],
|
59 |
+
[
|
60 |
+
1,
|
61 |
+
3,
|
62 |
+
5
|
63 |
+
],
|
64 |
+
[
|
65 |
+
1,
|
66 |
+
3,
|
67 |
+
5
|
68 |
+
]
|
69 |
+
],
|
70 |
+
"upsample_rates": [
|
71 |
+
8,
|
72 |
+
8,
|
73 |
+
2,
|
74 |
+
2,
|
75 |
+
2
|
76 |
+
],
|
77 |
+
"upsample_initial_channel": 512,
|
78 |
+
"upsample_kernel_sizes": [
|
79 |
+
16,
|
80 |
+
16,
|
81 |
+
4,
|
82 |
+
4,
|
83 |
+
4
|
84 |
+
],
|
85 |
+
"n_layers_q": 3,
|
86 |
+
"use_spectral_norm": false,
|
87 |
+
"gin_channels": 256,
|
88 |
+
"ssl_dim": 256,
|
89 |
+
"n_speakers": 63
|
90 |
+
},
|
91 |
+
"spk": {
|
92 |
+
"AKIMOTO_MANATSU": 0,
|
93 |
+
"ENDO_SAKURA": 1,
|
94 |
+
"ETO_MISA": 2,
|
95 |
+
"HARUKA_KUROMI": 3,
|
96 |
+
"HAYAKAWA_SEIRA": 4,
|
97 |
+
"HIGUCHI_HINA": 5,
|
98 |
+
"HORI_MIONA": 6,
|
99 |
+
"HOSHINO_MINAMI": 7,
|
100 |
+
"ICHINOSE_MIKU": 8,
|
101 |
+
"IKEDA_TERESA": 9,
|
102 |
+
"IKUTA_ERIKA": 10,
|
103 |
+
"INOUE_NAGI": 11,
|
104 |
+
"INOUE_SAYURI": 12,
|
105 |
+
"IOKI_MAO": 13,
|
106 |
+
"ITO_JUNNA": 14,
|
107 |
+
"ITO_RIRIA": 15,
|
108 |
+
"IWAMOTO_RENKA": 16,
|
109 |
+
"KAKEHASHI_SAYAKA": 17,
|
110 |
+
"KAKI_HARUKA": 18,
|
111 |
+
"KANAGAWA_SAYA": 19,
|
112 |
+
"KAWAGO_HINA": 20,
|
113 |
+
"KAWASAKI_SAKURA": 21,
|
114 |
+
"KITAGAWA_YURI": 22,
|
115 |
+
"KITANO_HINAKO": 23,
|
116 |
+
"KUBO_SHIORI": 24,
|
117 |
+
"MATSUMURA_SAYURI": 25,
|
118 |
+
"MIYU_MATSUO": 26,
|
119 |
+
"MUKAI_HAZUKI": 27,
|
120 |
+
"NAKAMURA_RENO": 28,
|
121 |
+
"NAKANISHI_ARUNO": 29,
|
122 |
+
"NAO_YUMIKI": 30,
|
123 |
+
"NISHINO_NANASE": 31,
|
124 |
+
"NOUJO_AMI": 32,
|
125 |
+
"OGAWA_AYA": 33,
|
126 |
+
"OKUDA_IROHA": 34,
|
127 |
+
"OZONO_MOMOKO": 35,
|
128 |
+
"RIKA_SATO": 36,
|
129 |
+
"RUNA_HAYASHI": 37,
|
130 |
+
"SAGARA_IORI": 38,
|
131 |
+
"SAITO_ASUKA": 39,
|
132 |
+
"SAKAGUCHI_TAMAMI": 40,
|
133 |
+
"SAKURAI_REIKA": 41,
|
134 |
+
"SASAKI_KOTOKO": 42,
|
135 |
+
"SATO_KAEDE": 43,
|
136 |
+
"SATO_YUURI": 44,
|
137 |
+
"SHIBATA_YUNA": 45,
|
138 |
+
"SHINUCHI_MAI": 46,
|
139 |
+
"SHIRAISHI_MAI": 47,
|
140 |
+
"SUGAWARA_SATSUKI": 48,
|
141 |
+
"SUZUKI_AYANE": 49,
|
142 |
+
"TAKAYAMA_KAZUMI": 50,
|
143 |
+
"TAMURA_MAYU": 51,
|
144 |
+
"TERADA_RANZE": 52,
|
145 |
+
"TOMISATO_NAO": 53,
|
146 |
+
"TSUTSUI_AYAME": 54,
|
147 |
+
"UMEZAWA_MINAMI": 55,
|
148 |
+
"WADA_MAAYA": 56,
|
149 |
+
"WAKATSUKI_YUMI": 57,
|
150 |
+
"WATANABE_MIRIA": 58,
|
151 |
+
"YAKUBO_MIO": 59,
|
152 |
+
"YAMASHITA_MIZUKI": 60,
|
153 |
+
"YAMAZAKI_RENA": 61,
|
154 |
+
"YODA_YUUKI": 62
|
155 |
+
}
|
156 |
+
}
|
configs/config.json
ADDED
@@ -0,0 +1,159 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"train": {
|
3 |
+
"log_interval": 200,
|
4 |
+
"eval_interval": 1800,
|
5 |
+
"seed": 1234,
|
6 |
+
"epochs": 10000,
|
7 |
+
"learning_rate": 0.0001,
|
8 |
+
"betas": [
|
9 |
+
0.8,
|
10 |
+
0.99
|
11 |
+
],
|
12 |
+
"eps": 1e-09,
|
13 |
+
"batch_size": 32,
|
14 |
+
"fp16_run": false,
|
15 |
+
"lr_decay": 0.999875,
|
16 |
+
"segment_size": 10240,
|
17 |
+
"init_lr_ratio": 1,
|
18 |
+
"warmup_epochs": 0,
|
19 |
+
"c_mel": 45,
|
20 |
+
"c_kl": 1.0,
|
21 |
+
"use_sr": true,
|
22 |
+
"max_speclen": 512,
|
23 |
+
"port": "8001",
|
24 |
+
"keep_ckpts": 40,
|
25 |
+
"all_in_mem": false,
|
26 |
+
"vol_aug": true
|
27 |
+
},
|
28 |
+
"data": {
|
29 |
+
"training_files": "filelists/train.txt",
|
30 |
+
"validation_files": "filelists/val.txt",
|
31 |
+
"max_wav_value": 32768.0,
|
32 |
+
"sampling_rate": 44100,
|
33 |
+
"filter_length": 2048,
|
34 |
+
"hop_length": 512,
|
35 |
+
"win_length": 2048,
|
36 |
+
"n_mel_channels": 80,
|
37 |
+
"mel_fmin": 0.0,
|
38 |
+
"mel_fmax": 22050
|
39 |
+
},
|
40 |
+
"model": {
|
41 |
+
"inter_channels": 192,
|
42 |
+
"hidden_channels": 192,
|
43 |
+
"filter_channels": 768,
|
44 |
+
"n_heads": 2,
|
45 |
+
"n_layers": 6,
|
46 |
+
"kernel_size": 3,
|
47 |
+
"p_dropout": 0.1,
|
48 |
+
"resblock": "1",
|
49 |
+
"resblock_kernel_sizes": [
|
50 |
+
3,
|
51 |
+
7,
|
52 |
+
11
|
53 |
+
],
|
54 |
+
"resblock_dilation_sizes": [
|
55 |
+
[
|
56 |
+
1,
|
57 |
+
3,
|
58 |
+
5
|
59 |
+
],
|
60 |
+
[
|
61 |
+
1,
|
62 |
+
3,
|
63 |
+
5
|
64 |
+
],
|
65 |
+
[
|
66 |
+
1,
|
67 |
+
3,
|
68 |
+
5
|
69 |
+
]
|
70 |
+
],
|
71 |
+
"upsample_rates": [
|
72 |
+
8,
|
73 |
+
8,
|
74 |
+
2,
|
75 |
+
2,
|
76 |
+
2
|
77 |
+
],
|
78 |
+
"upsample_initial_channel": 512,
|
79 |
+
"upsample_kernel_sizes": [
|
80 |
+
16,
|
81 |
+
16,
|
82 |
+
4,
|
83 |
+
4,
|
84 |
+
4
|
85 |
+
],
|
86 |
+
"n_layers_q": 3,
|
87 |
+
"use_spectral_norm": false,
|
88 |
+
"gin_channels": 768,
|
89 |
+
"ssl_dim": 768,
|
90 |
+
"n_speakers": 62,
|
91 |
+
"speech_encoder": "vec768l12",
|
92 |
+
"speaker_embedding": false,
|
93 |
+
"vol_embedding": true
|
94 |
+
},
|
95 |
+
"spk": {
|
96 |
+
"AKIMOTO_MANATSU": 0,
|
97 |
+
"ENDO_SAKURA": 1,
|
98 |
+
"ETO_MISA": 2,
|
99 |
+
"HARUKA_KUROMI": 3,
|
100 |
+
"HASHIMOTO_NANAMI": 4,
|
101 |
+
"HAYAKAWA_SEIRA": 5,
|
102 |
+
"HIGUCHI_HINA": 6,
|
103 |
+
"HORI_MIONA": 7,
|
104 |
+
"HOSHINO_MINAMI": 8,
|
105 |
+
"ICHINOSE_MIKU": 9,
|
106 |
+
"IKEDA_TERESA": 10,
|
107 |
+
"IKUTA_ERIKA": 11,
|
108 |
+
"INOUE_NAGI": 12,
|
109 |
+
"INOUE_SAYURI": 13,
|
110 |
+
"IOKI_MAO": 14,
|
111 |
+
"ITO_JUNNA": 15,
|
112 |
+
"ITO_RIRIA": 16,
|
113 |
+
"IWAMOTO_RENKA": 17,
|
114 |
+
"KAKEHASHI_SAYAKA": 18,
|
115 |
+
"KAKI_HARUKA": 19,
|
116 |
+
"KANAGAWA_SAYA": 20,
|
117 |
+
"KAWAGO_HINA": 21,
|
118 |
+
"KAWASAKI_SAKURA": 22,
|
119 |
+
"KITAGAWA_YURI": 23,
|
120 |
+
"KITANO_HINAKO": 24,
|
121 |
+
"KUBO_SHIORI": 25,
|
122 |
+
"MATSUMURA_SAYURI": 26,
|
123 |
+
"MIYU_MATSUO": 27,
|
124 |
+
"MUKAI_HAZUKI": 28,
|
125 |
+
"NAKAMURA_RENO": 29,
|
126 |
+
"NAKANISHI_ARUNO": 30,
|
127 |
+
"NAO_YUMIKI": 31,
|
128 |
+
"NISHINO_NANASE": 32,
|
129 |
+
"OGAWA_AYA": 33,
|
130 |
+
"OKUDA_IROHA": 34,
|
131 |
+
"OZONO_MOMOKO": 35,
|
132 |
+
"RIKA_SATO": 36,
|
133 |
+
"RUNA_HAYASHI": 37,
|
134 |
+
"SAITO_ASUKA": 38,
|
135 |
+
"SAKAGUCHI_TAMAMI": 39,
|
136 |
+
"SAKURAI_REIKA": 40,
|
137 |
+
"SASAKI_KOTOKO": 41,
|
138 |
+
"SATO_KAEDE": 42,
|
139 |
+
"SATO_YUURI": 43,
|
140 |
+
"SEIMIYA_REI": 44,
|
141 |
+
"SHIBATA_YUNA": 45,
|
142 |
+
"SHINUCHI_MAI": 46,
|
143 |
+
"SHIRAISHI_MAI": 47,
|
144 |
+
"SUGAWARA_SATSUKI": 48,
|
145 |
+
"SUZUKI_AYANE": 49,
|
146 |
+
"TAKAYAMA_KAZUMI": 50,
|
147 |
+
"TAMURA_MAYU": 51,
|
148 |
+
"TERADA_RANZE": 52,
|
149 |
+
"TOMISATO_NAO": 53,
|
150 |
+
"TSUTSUI_AYAME": 54,
|
151 |
+
"UMEZAWA_MINAMI": 55,
|
152 |
+
"WAKATSUKI_YUMI": 56,
|
153 |
+
"WATANABE_MIRIA": 57,
|
154 |
+
"YAKUBO_MIO": 58,
|
155 |
+
"YAMASHITA_MIZUKI": 59,
|
156 |
+
"YAMAZAKI_RENA": 60,
|
157 |
+
"YODA_YUUKI": 61
|
158 |
+
}
|
159 |
+
}
|
cvec/checkpoint_best_legacy_500.pt
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:294a2e8c98136070a999e040ec98dfa5a99b88a7938181c56cc2ab0e2f6ce0e8
|
3 |
+
size 48501067
|
data_utils.py
ADDED
@@ -0,0 +1,184 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import time
|
2 |
+
import os
|
3 |
+
import random
|
4 |
+
import numpy as np
|
5 |
+
import torch
|
6 |
+
import torch.utils.data
|
7 |
+
|
8 |
+
import modules.commons as commons
|
9 |
+
import utils
|
10 |
+
from modules.mel_processing import spectrogram_torch, spec_to_mel_torch, spectrogram_torch
|
11 |
+
from utils import load_wav_to_torch, load_filepaths_and_text
|
12 |
+
|
13 |
+
# import h5py
|
14 |
+
|
15 |
+
|
16 |
+
"""Multi speaker version"""
|
17 |
+
|
18 |
+
|
19 |
+
class TextAudioSpeakerLoader(torch.utils.data.Dataset):
|
20 |
+
"""
|
21 |
+
1) loads audio, speaker_id, text pairs
|
22 |
+
2) normalizes text and converts them to sequences of integers
|
23 |
+
3) computes spectrograms from audio files.
|
24 |
+
"""
|
25 |
+
|
26 |
+
def __init__(self, audiopaths, hparams, all_in_mem: bool = False, vol_aug: bool = True):
|
27 |
+
self.audiopaths = load_filepaths_and_text(audiopaths)
|
28 |
+
self.hparams = hparams
|
29 |
+
self.max_wav_value = hparams.data.max_wav_value
|
30 |
+
self.sampling_rate = hparams.data.sampling_rate
|
31 |
+
self.filter_length = hparams.data.filter_length
|
32 |
+
self.hop_length = hparams.data.hop_length
|
33 |
+
self.win_length = hparams.data.win_length
|
34 |
+
self.sampling_rate = hparams.data.sampling_rate
|
35 |
+
self.use_sr = hparams.train.use_sr
|
36 |
+
self.spec_len = hparams.train.max_speclen
|
37 |
+
self.spk_map = hparams.spk
|
38 |
+
self.vol_emb = hparams.model.vol_embedding
|
39 |
+
self.vol_aug = hparams.train.vol_aug and vol_aug
|
40 |
+
random.seed(1234)
|
41 |
+
random.shuffle(self.audiopaths)
|
42 |
+
|
43 |
+
self.all_in_mem = all_in_mem
|
44 |
+
if self.all_in_mem:
|
45 |
+
self.cache = [self.get_audio(p[0]) for p in self.audiopaths]
|
46 |
+
|
47 |
+
def get_audio(self, filename):
|
48 |
+
filename = filename.replace("\\", "/")
|
49 |
+
audio, sampling_rate = load_wav_to_torch(filename)
|
50 |
+
if sampling_rate != self.sampling_rate:
|
51 |
+
raise ValueError("{} SR doesn't match target {} SR".format(
|
52 |
+
sampling_rate, self.sampling_rate))
|
53 |
+
audio_norm = audio / self.max_wav_value
|
54 |
+
audio_norm = audio_norm.unsqueeze(0)
|
55 |
+
spec_filename = filename.replace(".wav", ".spec.pt")
|
56 |
+
|
57 |
+
# Ideally, all data generated after Mar 25 should have .spec.pt
|
58 |
+
if os.path.exists(spec_filename):
|
59 |
+
spec = torch.load(spec_filename)
|
60 |
+
else:
|
61 |
+
spec = spectrogram_torch(audio_norm, self.filter_length,
|
62 |
+
self.sampling_rate, self.hop_length, self.win_length,
|
63 |
+
center=False)
|
64 |
+
spec = torch.squeeze(spec, 0)
|
65 |
+
torch.save(spec, spec_filename)
|
66 |
+
|
67 |
+
spk = filename.split("/")[-2]
|
68 |
+
spk = torch.LongTensor([self.spk_map[spk]])
|
69 |
+
|
70 |
+
f0, uv = np.load(filename + ".f0.npy",allow_pickle=True)
|
71 |
+
|
72 |
+
f0 = torch.FloatTensor(np.array(f0,dtype=float))
|
73 |
+
uv = torch.FloatTensor(np.array(uv,dtype=float))
|
74 |
+
|
75 |
+
c = torch.load(filename+ ".soft.pt")
|
76 |
+
c = utils.repeat_expand_2d(c.squeeze(0), f0.shape[0])
|
77 |
+
if self.vol_emb:
|
78 |
+
volume_path = filename + ".vol.npy"
|
79 |
+
volume = np.load(volume_path)
|
80 |
+
volume = torch.from_numpy(volume).float()
|
81 |
+
else:
|
82 |
+
volume = None
|
83 |
+
|
84 |
+
lmin = min(c.size(-1), spec.size(-1))
|
85 |
+
assert abs(c.size(-1) - spec.size(-1)) < 3, (c.size(-1), spec.size(-1), f0.shape, filename)
|
86 |
+
assert abs(audio_norm.shape[1]-lmin * self.hop_length) < 3 * self.hop_length
|
87 |
+
spec, c, f0, uv = spec[:, :lmin], c[:, :lmin], f0[:lmin], uv[:lmin]
|
88 |
+
audio_norm = audio_norm[:, :lmin * self.hop_length]
|
89 |
+
if volume!= None:
|
90 |
+
volume = volume[:lmin]
|
91 |
+
return c, f0, spec, audio_norm, spk, uv, volume
|
92 |
+
|
93 |
+
def random_slice(self, c, f0, spec, audio_norm, spk, uv, volume):
|
94 |
+
# if spec.shape[1] < 30:
|
95 |
+
# print("skip too short audio:", filename)
|
96 |
+
# return None
|
97 |
+
|
98 |
+
if random.choice([True, False]) and self.vol_aug and volume!=None:
|
99 |
+
max_amp = float(torch.max(torch.abs(audio_norm))) + 1e-5
|
100 |
+
max_shift = min(1, np.log10(1/max_amp))
|
101 |
+
log10_vol_shift = random.uniform(-1, max_shift)
|
102 |
+
audio_norm = audio_norm * (10 ** log10_vol_shift)
|
103 |
+
volume = volume * (10 ** log10_vol_shift)
|
104 |
+
spec = spectrogram_torch(audio_norm,
|
105 |
+
self.hparams.data.filter_length,
|
106 |
+
self.hparams.data.sampling_rate,
|
107 |
+
self.hparams.data.hop_length,
|
108 |
+
self.hparams.data.win_length,
|
109 |
+
center=False)[0]
|
110 |
+
|
111 |
+
if spec.shape[1] > 800:
|
112 |
+
start = random.randint(0, spec.shape[1]-800)
|
113 |
+
end = start + 790
|
114 |
+
spec, c, f0, uv = spec[:, start:end], c[:, start:end], f0[start:end], uv[start:end]
|
115 |
+
audio_norm = audio_norm[:, start * self.hop_length : end * self.hop_length]
|
116 |
+
if volume !=None:
|
117 |
+
volume = volume[start:end]
|
118 |
+
return c, f0, spec, audio_norm, spk, uv,volume
|
119 |
+
|
120 |
+
def __getitem__(self, index):
|
121 |
+
if self.all_in_mem:
|
122 |
+
return self.random_slice(*self.cache[index])
|
123 |
+
else:
|
124 |
+
return self.random_slice(*self.get_audio(self.audiopaths[index][0]))
|
125 |
+
|
126 |
+
def __len__(self):
|
127 |
+
return len(self.audiopaths)
|
128 |
+
|
129 |
+
|
130 |
+
class TextAudioCollate:
|
131 |
+
|
132 |
+
def __call__(self, batch):
|
133 |
+
batch = [b for b in batch if b is not None]
|
134 |
+
|
135 |
+
input_lengths, ids_sorted_decreasing = torch.sort(
|
136 |
+
torch.LongTensor([x[0].shape[1] for x in batch]),
|
137 |
+
dim=0, descending=True)
|
138 |
+
|
139 |
+
max_c_len = max([x[0].size(1) for x in batch])
|
140 |
+
max_wav_len = max([x[3].size(1) for x in batch])
|
141 |
+
|
142 |
+
lengths = torch.LongTensor(len(batch))
|
143 |
+
|
144 |
+
c_padded = torch.FloatTensor(len(batch), batch[0][0].shape[0], max_c_len)
|
145 |
+
f0_padded = torch.FloatTensor(len(batch), max_c_len)
|
146 |
+
spec_padded = torch.FloatTensor(len(batch), batch[0][2].shape[0], max_c_len)
|
147 |
+
wav_padded = torch.FloatTensor(len(batch), 1, max_wav_len)
|
148 |
+
spkids = torch.LongTensor(len(batch), 1)
|
149 |
+
uv_padded = torch.FloatTensor(len(batch), max_c_len)
|
150 |
+
volume_padded = torch.FloatTensor(len(batch), max_c_len)
|
151 |
+
|
152 |
+
c_padded.zero_()
|
153 |
+
spec_padded.zero_()
|
154 |
+
f0_padded.zero_()
|
155 |
+
wav_padded.zero_()
|
156 |
+
uv_padded.zero_()
|
157 |
+
volume_padded.zero_()
|
158 |
+
|
159 |
+
for i in range(len(ids_sorted_decreasing)):
|
160 |
+
row = batch[ids_sorted_decreasing[i]]
|
161 |
+
|
162 |
+
c = row[0]
|
163 |
+
c_padded[i, :, :c.size(1)] = c
|
164 |
+
lengths[i] = c.size(1)
|
165 |
+
|
166 |
+
f0 = row[1]
|
167 |
+
f0_padded[i, :f0.size(0)] = f0
|
168 |
+
|
169 |
+
spec = row[2]
|
170 |
+
spec_padded[i, :, :spec.size(1)] = spec
|
171 |
+
|
172 |
+
wav = row[3]
|
173 |
+
wav_padded[i, :, :wav.size(1)] = wav
|
174 |
+
|
175 |
+
spkids[i, 0] = row[4]
|
176 |
+
|
177 |
+
uv = row[5]
|
178 |
+
uv_padded[i, :uv.size(0)] = uv
|
179 |
+
volume = row[6]
|
180 |
+
if volume != None:
|
181 |
+
volume_padded[i, :volume.size(0)] = volume
|
182 |
+
else :
|
183 |
+
volume_padded = None
|
184 |
+
return c_padded, f0_padded, spec_padded, wav_padded, spkids, lengths, uv_padded, volume_padded
|
diffusion/__init__.py
ADDED
File without changes
|
diffusion/__pycache__/__init__.cpython-38.pyc
ADDED
Binary file (131 Bytes). View file
|
|
diffusion/__pycache__/data_loaders.cpython-38.pyc
ADDED
Binary file (5.7 kB). View file
|
|
diffusion/__pycache__/diffusion.cpython-38.pyc
ADDED
Binary file (10.1 kB). View file
|
|
diffusion/__pycache__/dpm_solver_pytorch.cpython-38.pyc
ADDED
Binary file (50.4 kB). View file
|
|
diffusion/__pycache__/solver.cpython-38.pyc
ADDED
Binary file (4.04 kB). View file
|
|
diffusion/__pycache__/unit2mel.cpython-38.pyc
ADDED
Binary file (4.43 kB). View file
|
|
diffusion/__pycache__/vocoder.cpython-38.pyc
ADDED
Binary file (3.5 kB). View file
|
|
diffusion/__pycache__/wavenet.cpython-38.pyc
ADDED
Binary file (3.84 kB). View file
|
|
diffusion/data_loaders.py
ADDED
@@ -0,0 +1,284 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
import random
|
3 |
+
import re
|
4 |
+
import numpy as np
|
5 |
+
import librosa
|
6 |
+
import torch
|
7 |
+
import random
|
8 |
+
from utils import repeat_expand_2d
|
9 |
+
from tqdm import tqdm
|
10 |
+
from torch.utils.data import Dataset
|
11 |
+
|
12 |
+
def traverse_dir(
|
13 |
+
root_dir,
|
14 |
+
extensions,
|
15 |
+
amount=None,
|
16 |
+
str_include=None,
|
17 |
+
str_exclude=None,
|
18 |
+
is_pure=False,
|
19 |
+
is_sort=False,
|
20 |
+
is_ext=True):
|
21 |
+
|
22 |
+
file_list = []
|
23 |
+
cnt = 0
|
24 |
+
for root, _, files in os.walk(root_dir):
|
25 |
+
for file in files:
|
26 |
+
if any([file.endswith(f".{ext}") for ext in extensions]):
|
27 |
+
# path
|
28 |
+
mix_path = os.path.join(root, file)
|
29 |
+
pure_path = mix_path[len(root_dir)+1:] if is_pure else mix_path
|
30 |
+
|
31 |
+
# amount
|
32 |
+
if (amount is not None) and (cnt == amount):
|
33 |
+
if is_sort:
|
34 |
+
file_list.sort()
|
35 |
+
return file_list
|
36 |
+
|
37 |
+
# check string
|
38 |
+
if (str_include is not None) and (str_include not in pure_path):
|
39 |
+
continue
|
40 |
+
if (str_exclude is not None) and (str_exclude in pure_path):
|
41 |
+
continue
|
42 |
+
|
43 |
+
if not is_ext:
|
44 |
+
ext = pure_path.split('.')[-1]
|
45 |
+
pure_path = pure_path[:-(len(ext)+1)]
|
46 |
+
file_list.append(pure_path)
|
47 |
+
cnt += 1
|
48 |
+
if is_sort:
|
49 |
+
file_list.sort()
|
50 |
+
return file_list
|
51 |
+
|
52 |
+
|
53 |
+
def get_data_loaders(args, whole_audio=False):
|
54 |
+
data_train = AudioDataset(
|
55 |
+
filelists = args.data.training_files,
|
56 |
+
waveform_sec=args.data.duration,
|
57 |
+
hop_size=args.data.block_size,
|
58 |
+
sample_rate=args.data.sampling_rate,
|
59 |
+
load_all_data=args.train.cache_all_data,
|
60 |
+
whole_audio=whole_audio,
|
61 |
+
extensions=args.data.extensions,
|
62 |
+
n_spk=args.model.n_spk,
|
63 |
+
spk=args.spk,
|
64 |
+
device=args.train.cache_device,
|
65 |
+
fp16=args.train.cache_fp16,
|
66 |
+
use_aug=True)
|
67 |
+
loader_train = torch.utils.data.DataLoader(
|
68 |
+
data_train ,
|
69 |
+
batch_size=args.train.batch_size if not whole_audio else 1,
|
70 |
+
shuffle=True,
|
71 |
+
num_workers=args.train.num_workers if args.train.cache_device=='cpu' else 0,
|
72 |
+
persistent_workers=(args.train.num_workers > 0) if args.train.cache_device=='cpu' else False,
|
73 |
+
pin_memory=True if args.train.cache_device=='cpu' else False
|
74 |
+
)
|
75 |
+
data_valid = AudioDataset(
|
76 |
+
filelists = args.data.validation_files,
|
77 |
+
waveform_sec=args.data.duration,
|
78 |
+
hop_size=args.data.block_size,
|
79 |
+
sample_rate=args.data.sampling_rate,
|
80 |
+
load_all_data=args.train.cache_all_data,
|
81 |
+
whole_audio=True,
|
82 |
+
spk=args.spk,
|
83 |
+
extensions=args.data.extensions,
|
84 |
+
n_spk=args.model.n_spk)
|
85 |
+
loader_valid = torch.utils.data.DataLoader(
|
86 |
+
data_valid,
|
87 |
+
batch_size=1,
|
88 |
+
shuffle=False,
|
89 |
+
num_workers=0,
|
90 |
+
pin_memory=True
|
91 |
+
)
|
92 |
+
return loader_train, loader_valid
|
93 |
+
|
94 |
+
|
95 |
+
class AudioDataset(Dataset):
|
96 |
+
def __init__(
|
97 |
+
self,
|
98 |
+
filelists,
|
99 |
+
waveform_sec,
|
100 |
+
hop_size,
|
101 |
+
sample_rate,
|
102 |
+
spk,
|
103 |
+
load_all_data=True,
|
104 |
+
whole_audio=False,
|
105 |
+
extensions=['wav'],
|
106 |
+
n_spk=1,
|
107 |
+
device='cpu',
|
108 |
+
fp16=False,
|
109 |
+
use_aug=False,
|
110 |
+
):
|
111 |
+
super().__init__()
|
112 |
+
|
113 |
+
self.waveform_sec = waveform_sec
|
114 |
+
self.sample_rate = sample_rate
|
115 |
+
self.hop_size = hop_size
|
116 |
+
self.filelists = filelists
|
117 |
+
self.whole_audio = whole_audio
|
118 |
+
self.use_aug = use_aug
|
119 |
+
self.data_buffer={}
|
120 |
+
self.pitch_aug_dict = {}
|
121 |
+
# np.load(os.path.join(self.path_root, 'pitch_aug_dict.npy'), allow_pickle=True).item()
|
122 |
+
if load_all_data:
|
123 |
+
print('Load all the data filelists:', filelists)
|
124 |
+
else:
|
125 |
+
print('Load the f0, volume data filelists:', filelists)
|
126 |
+
with open(filelists,"r") as f:
|
127 |
+
self.paths = f.read().splitlines()
|
128 |
+
for name_ext in tqdm(self.paths, total=len(self.paths)):
|
129 |
+
name = os.path.splitext(name_ext)[0]
|
130 |
+
path_audio = name_ext
|
131 |
+
duration = librosa.get_duration(filename = path_audio, sr = self.sample_rate)
|
132 |
+
|
133 |
+
path_f0 = name_ext + ".f0.npy"
|
134 |
+
f0,_ = np.load(path_f0,allow_pickle=True)
|
135 |
+
f0 = torch.from_numpy(np.array(f0,dtype=float)).float().unsqueeze(-1).to(device)
|
136 |
+
|
137 |
+
path_volume = name_ext + ".vol.npy"
|
138 |
+
volume = np.load(path_volume)
|
139 |
+
volume = torch.from_numpy(volume).float().unsqueeze(-1).to(device)
|
140 |
+
|
141 |
+
path_augvol = name_ext + ".aug_vol.npy"
|
142 |
+
aug_vol = np.load(path_augvol)
|
143 |
+
aug_vol = torch.from_numpy(aug_vol).float().unsqueeze(-1).to(device)
|
144 |
+
|
145 |
+
if n_spk is not None and n_spk > 1:
|
146 |
+
spk_name = name_ext.split("/")[-2]
|
147 |
+
spk_id = spk[spk_name] if spk_name in spk else 0
|
148 |
+
if spk_id < 0 or spk_id >= n_spk:
|
149 |
+
raise ValueError(' [x] Muiti-speaker traing error : spk_id must be a positive integer from 0 to n_spk-1 ')
|
150 |
+
else:
|
151 |
+
spk_id = 0
|
152 |
+
spk_id = torch.LongTensor(np.array([spk_id])).to(device)
|
153 |
+
|
154 |
+
if load_all_data:
|
155 |
+
'''
|
156 |
+
audio, sr = librosa.load(path_audio, sr=self.sample_rate)
|
157 |
+
if len(audio.shape) > 1:
|
158 |
+
audio = librosa.to_mono(audio)
|
159 |
+
audio = torch.from_numpy(audio).to(device)
|
160 |
+
'''
|
161 |
+
path_mel = name_ext + ".mel.npy"
|
162 |
+
mel = np.load(path_mel)
|
163 |
+
mel = torch.from_numpy(mel).to(device)
|
164 |
+
|
165 |
+
path_augmel = name_ext + ".aug_mel.npy"
|
166 |
+
aug_mel,keyshift = np.load(path_augmel, allow_pickle=True)
|
167 |
+
aug_mel = np.array(aug_mel,dtype=float)
|
168 |
+
aug_mel = torch.from_numpy(aug_mel).to(device)
|
169 |
+
self.pitch_aug_dict[name_ext] = keyshift
|
170 |
+
|
171 |
+
path_units = name_ext + ".soft.pt"
|
172 |
+
units = torch.load(path_units).to(device)
|
173 |
+
units = units[0]
|
174 |
+
units = repeat_expand_2d(units,f0.size(0)).transpose(0,1)
|
175 |
+
|
176 |
+
if fp16:
|
177 |
+
mel = mel.half()
|
178 |
+
aug_mel = aug_mel.half()
|
179 |
+
units = units.half()
|
180 |
+
|
181 |
+
self.data_buffer[name_ext] = {
|
182 |
+
'duration': duration,
|
183 |
+
'mel': mel,
|
184 |
+
'aug_mel': aug_mel,
|
185 |
+
'units': units,
|
186 |
+
'f0': f0,
|
187 |
+
'volume': volume,
|
188 |
+
'aug_vol': aug_vol,
|
189 |
+
'spk_id': spk_id
|
190 |
+
}
|
191 |
+
else:
|
192 |
+
path_augmel = name_ext + ".aug_mel.npy"
|
193 |
+
aug_mel,keyshift = np.load(path_augmel, allow_pickle=True)
|
194 |
+
self.pitch_aug_dict[name_ext] = keyshift
|
195 |
+
self.data_buffer[name_ext] = {
|
196 |
+
'duration': duration,
|
197 |
+
'f0': f0,
|
198 |
+
'volume': volume,
|
199 |
+
'aug_vol': aug_vol,
|
200 |
+
'spk_id': spk_id
|
201 |
+
}
|
202 |
+
|
203 |
+
|
204 |
+
def __getitem__(self, file_idx):
|
205 |
+
name_ext = self.paths[file_idx]
|
206 |
+
data_buffer = self.data_buffer[name_ext]
|
207 |
+
# check duration. if too short, then skip
|
208 |
+
if data_buffer['duration'] < (self.waveform_sec + 0.1):
|
209 |
+
return self.__getitem__( (file_idx + 1) % len(self.paths))
|
210 |
+
|
211 |
+
# get item
|
212 |
+
return self.get_data(name_ext, data_buffer)
|
213 |
+
|
214 |
+
def get_data(self, name_ext, data_buffer):
|
215 |
+
name = os.path.splitext(name_ext)[0]
|
216 |
+
frame_resolution = self.hop_size / self.sample_rate
|
217 |
+
duration = data_buffer['duration']
|
218 |
+
waveform_sec = duration if self.whole_audio else self.waveform_sec
|
219 |
+
|
220 |
+
# load audio
|
221 |
+
idx_from = 0 if self.whole_audio else random.uniform(0, duration - waveform_sec - 0.1)
|
222 |
+
start_frame = int(idx_from / frame_resolution)
|
223 |
+
units_frame_len = int(waveform_sec / frame_resolution)
|
224 |
+
aug_flag = random.choice([True, False]) and self.use_aug
|
225 |
+
'''
|
226 |
+
audio = data_buffer.get('audio')
|
227 |
+
if audio is None:
|
228 |
+
path_audio = os.path.join(self.path_root, 'audio', name) + '.wav'
|
229 |
+
audio, sr = librosa.load(
|
230 |
+
path_audio,
|
231 |
+
sr = self.sample_rate,
|
232 |
+
offset = start_frame * frame_resolution,
|
233 |
+
duration = waveform_sec)
|
234 |
+
if len(audio.shape) > 1:
|
235 |
+
audio = librosa.to_mono(audio)
|
236 |
+
# clip audio into N seconds
|
237 |
+
audio = audio[ : audio.shape[-1] // self.hop_size * self.hop_size]
|
238 |
+
audio = torch.from_numpy(audio).float()
|
239 |
+
else:
|
240 |
+
audio = audio[start_frame * self.hop_size : (start_frame + units_frame_len) * self.hop_size]
|
241 |
+
'''
|
242 |
+
# load mel
|
243 |
+
mel_key = 'aug_mel' if aug_flag else 'mel'
|
244 |
+
mel = data_buffer.get(mel_key)
|
245 |
+
if mel is None:
|
246 |
+
mel = name_ext + ".mel.npy"
|
247 |
+
mel = np.load(mel)
|
248 |
+
mel = mel[start_frame : start_frame + units_frame_len]
|
249 |
+
mel = torch.from_numpy(mel).float()
|
250 |
+
else:
|
251 |
+
mel = mel[start_frame : start_frame + units_frame_len]
|
252 |
+
|
253 |
+
# load f0
|
254 |
+
f0 = data_buffer.get('f0')
|
255 |
+
aug_shift = 0
|
256 |
+
if aug_flag:
|
257 |
+
aug_shift = self.pitch_aug_dict[name_ext]
|
258 |
+
f0_frames = 2 ** (aug_shift / 12) * f0[start_frame : start_frame + units_frame_len]
|
259 |
+
|
260 |
+
# load units
|
261 |
+
units = data_buffer.get('units')
|
262 |
+
if units is None:
|
263 |
+
path_units = name_ext + ".soft.pt"
|
264 |
+
units = torch.load(path_units)
|
265 |
+
units = units[0]
|
266 |
+
units = repeat_expand_2d(units,f0.size(0)).transpose(0,1)
|
267 |
+
|
268 |
+
units = units[start_frame : start_frame + units_frame_len]
|
269 |
+
|
270 |
+
# load volume
|
271 |
+
vol_key = 'aug_vol' if aug_flag else 'volume'
|
272 |
+
volume = data_buffer.get(vol_key)
|
273 |
+
volume_frames = volume[start_frame : start_frame + units_frame_len]
|
274 |
+
|
275 |
+
# load spk_id
|
276 |
+
spk_id = data_buffer.get('spk_id')
|
277 |
+
|
278 |
+
# load shift
|
279 |
+
aug_shift = torch.from_numpy(np.array([[aug_shift]])).float()
|
280 |
+
|
281 |
+
return dict(mel=mel, f0=f0_frames, volume=volume_frames, units=units, spk_id=spk_id, aug_shift=aug_shift, name=name, name_ext=name_ext)
|
282 |
+
|
283 |
+
def __len__(self):
|
284 |
+
return len(self.paths)
|
diffusion/diffusion.py
ADDED
@@ -0,0 +1,317 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from collections import deque
|
2 |
+
from functools import partial
|
3 |
+
from inspect import isfunction
|
4 |
+
import torch.nn.functional as F
|
5 |
+
import librosa.sequence
|
6 |
+
import numpy as np
|
7 |
+
import torch
|
8 |
+
from torch import nn
|
9 |
+
from tqdm import tqdm
|
10 |
+
|
11 |
+
|
12 |
+
def exists(x):
|
13 |
+
return x is not None
|
14 |
+
|
15 |
+
|
16 |
+
def default(val, d):
|
17 |
+
if exists(val):
|
18 |
+
return val
|
19 |
+
return d() if isfunction(d) else d
|
20 |
+
|
21 |
+
|
22 |
+
def extract(a, t, x_shape):
|
23 |
+
b, *_ = t.shape
|
24 |
+
out = a.gather(-1, t)
|
25 |
+
return out.reshape(b, *((1,) * (len(x_shape) - 1)))
|
26 |
+
|
27 |
+
|
28 |
+
def noise_like(shape, device, repeat=False):
|
29 |
+
repeat_noise = lambda: torch.randn((1, *shape[1:]), device=device).repeat(shape[0], *((1,) * (len(shape) - 1)))
|
30 |
+
noise = lambda: torch.randn(shape, device=device)
|
31 |
+
return repeat_noise() if repeat else noise()
|
32 |
+
|
33 |
+
|
34 |
+
def linear_beta_schedule(timesteps, max_beta=0.02):
|
35 |
+
"""
|
36 |
+
linear schedule
|
37 |
+
"""
|
38 |
+
betas = np.linspace(1e-4, max_beta, timesteps)
|
39 |
+
return betas
|
40 |
+
|
41 |
+
|
42 |
+
def cosine_beta_schedule(timesteps, s=0.008):
|
43 |
+
"""
|
44 |
+
cosine schedule
|
45 |
+
as proposed in https://openreview.net/forum?id=-NEXDKk8gZ
|
46 |
+
"""
|
47 |
+
steps = timesteps + 1
|
48 |
+
x = np.linspace(0, steps, steps)
|
49 |
+
alphas_cumprod = np.cos(((x / steps) + s) / (1 + s) * np.pi * 0.5) ** 2
|
50 |
+
alphas_cumprod = alphas_cumprod / alphas_cumprod[0]
|
51 |
+
betas = 1 - (alphas_cumprod[1:] / alphas_cumprod[:-1])
|
52 |
+
return np.clip(betas, a_min=0, a_max=0.999)
|
53 |
+
|
54 |
+
|
55 |
+
beta_schedule = {
|
56 |
+
"cosine": cosine_beta_schedule,
|
57 |
+
"linear": linear_beta_schedule,
|
58 |
+
}
|
59 |
+
|
60 |
+
|
61 |
+
class GaussianDiffusion(nn.Module):
|
62 |
+
def __init__(self,
|
63 |
+
denoise_fn,
|
64 |
+
out_dims=128,
|
65 |
+
timesteps=1000,
|
66 |
+
k_step=1000,
|
67 |
+
max_beta=0.02,
|
68 |
+
spec_min=-12,
|
69 |
+
spec_max=2):
|
70 |
+
super().__init__()
|
71 |
+
self.denoise_fn = denoise_fn
|
72 |
+
self.out_dims = out_dims
|
73 |
+
betas = beta_schedule['linear'](timesteps, max_beta=max_beta)
|
74 |
+
|
75 |
+
alphas = 1. - betas
|
76 |
+
alphas_cumprod = np.cumprod(alphas, axis=0)
|
77 |
+
alphas_cumprod_prev = np.append(1., alphas_cumprod[:-1])
|
78 |
+
|
79 |
+
timesteps, = betas.shape
|
80 |
+
self.num_timesteps = int(timesteps)
|
81 |
+
self.k_step = k_step
|
82 |
+
|
83 |
+
self.noise_list = deque(maxlen=4)
|
84 |
+
|
85 |
+
to_torch = partial(torch.tensor, dtype=torch.float32)
|
86 |
+
|
87 |
+
self.register_buffer('betas', to_torch(betas))
|
88 |
+
self.register_buffer('alphas_cumprod', to_torch(alphas_cumprod))
|
89 |
+
self.register_buffer('alphas_cumprod_prev', to_torch(alphas_cumprod_prev))
|
90 |
+
|
91 |
+
# calculations for diffusion q(x_t | x_{t-1}) and others
|
92 |
+
self.register_buffer('sqrt_alphas_cumprod', to_torch(np.sqrt(alphas_cumprod)))
|
93 |
+
self.register_buffer('sqrt_one_minus_alphas_cumprod', to_torch(np.sqrt(1. - alphas_cumprod)))
|
94 |
+
self.register_buffer('log_one_minus_alphas_cumprod', to_torch(np.log(1. - alphas_cumprod)))
|
95 |
+
self.register_buffer('sqrt_recip_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod)))
|
96 |
+
self.register_buffer('sqrt_recipm1_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod - 1)))
|
97 |
+
|
98 |
+
# calculations for posterior q(x_{t-1} | x_t, x_0)
|
99 |
+
posterior_variance = betas * (1. - alphas_cumprod_prev) / (1. - alphas_cumprod)
|
100 |
+
# above: equal to 1. / (1. / (1. - alpha_cumprod_tm1) + alpha_t / beta_t)
|
101 |
+
self.register_buffer('posterior_variance', to_torch(posterior_variance))
|
102 |
+
# below: log calculation clipped because the posterior variance is 0 at the beginning of the diffusion chain
|
103 |
+
self.register_buffer('posterior_log_variance_clipped', to_torch(np.log(np.maximum(posterior_variance, 1e-20))))
|
104 |
+
self.register_buffer('posterior_mean_coef1', to_torch(
|
105 |
+
betas * np.sqrt(alphas_cumprod_prev) / (1. - alphas_cumprod)))
|
106 |
+
self.register_buffer('posterior_mean_coef2', to_torch(
|
107 |
+
(1. - alphas_cumprod_prev) * np.sqrt(alphas) / (1. - alphas_cumprod)))
|
108 |
+
|
109 |
+
self.register_buffer('spec_min', torch.FloatTensor([spec_min])[None, None, :out_dims])
|
110 |
+
self.register_buffer('spec_max', torch.FloatTensor([spec_max])[None, None, :out_dims])
|
111 |
+
|
112 |
+
def q_mean_variance(self, x_start, t):
|
113 |
+
mean = extract(self.sqrt_alphas_cumprod, t, x_start.shape) * x_start
|
114 |
+
variance = extract(1. - self.alphas_cumprod, t, x_start.shape)
|
115 |
+
log_variance = extract(self.log_one_minus_alphas_cumprod, t, x_start.shape)
|
116 |
+
return mean, variance, log_variance
|
117 |
+
|
118 |
+
def predict_start_from_noise(self, x_t, t, noise):
|
119 |
+
return (
|
120 |
+
extract(self.sqrt_recip_alphas_cumprod, t, x_t.shape) * x_t -
|
121 |
+
extract(self.sqrt_recipm1_alphas_cumprod, t, x_t.shape) * noise
|
122 |
+
)
|
123 |
+
|
124 |
+
def q_posterior(self, x_start, x_t, t):
|
125 |
+
posterior_mean = (
|
126 |
+
extract(self.posterior_mean_coef1, t, x_t.shape) * x_start +
|
127 |
+
extract(self.posterior_mean_coef2, t, x_t.shape) * x_t
|
128 |
+
)
|
129 |
+
posterior_variance = extract(self.posterior_variance, t, x_t.shape)
|
130 |
+
posterior_log_variance_clipped = extract(self.posterior_log_variance_clipped, t, x_t.shape)
|
131 |
+
return posterior_mean, posterior_variance, posterior_log_variance_clipped
|
132 |
+
|
133 |
+
def p_mean_variance(self, x, t, cond):
|
134 |
+
noise_pred = self.denoise_fn(x, t, cond=cond)
|
135 |
+
x_recon = self.predict_start_from_noise(x, t=t, noise=noise_pred)
|
136 |
+
|
137 |
+
x_recon.clamp_(-1., 1.)
|
138 |
+
|
139 |
+
model_mean, posterior_variance, posterior_log_variance = self.q_posterior(x_start=x_recon, x_t=x, t=t)
|
140 |
+
return model_mean, posterior_variance, posterior_log_variance
|
141 |
+
|
142 |
+
@torch.no_grad()
|
143 |
+
def p_sample(self, x, t, cond, clip_denoised=True, repeat_noise=False):
|
144 |
+
b, *_, device = *x.shape, x.device
|
145 |
+
model_mean, _, model_log_variance = self.p_mean_variance(x=x, t=t, cond=cond)
|
146 |
+
noise = noise_like(x.shape, device, repeat_noise)
|
147 |
+
# no noise when t == 0
|
148 |
+
nonzero_mask = (1 - (t == 0).float()).reshape(b, *((1,) * (len(x.shape) - 1)))
|
149 |
+
return model_mean + nonzero_mask * (0.5 * model_log_variance).exp() * noise
|
150 |
+
|
151 |
+
@torch.no_grad()
|
152 |
+
def p_sample_plms(self, x, t, interval, cond, clip_denoised=True, repeat_noise=False):
|
153 |
+
"""
|
154 |
+
Use the PLMS method from
|
155 |
+
[Pseudo Numerical Methods for Diffusion Models on Manifolds](https://arxiv.org/abs/2202.09778).
|
156 |
+
"""
|
157 |
+
|
158 |
+
def get_x_pred(x, noise_t, t):
|
159 |
+
a_t = extract(self.alphas_cumprod, t, x.shape)
|
160 |
+
a_prev = extract(self.alphas_cumprod, torch.max(t - interval, torch.zeros_like(t)), x.shape)
|
161 |
+
a_t_sq, a_prev_sq = a_t.sqrt(), a_prev.sqrt()
|
162 |
+
|
163 |
+
x_delta = (a_prev - a_t) * ((1 / (a_t_sq * (a_t_sq + a_prev_sq))) * x - 1 / (
|
164 |
+
a_t_sq * (((1 - a_prev) * a_t).sqrt() + ((1 - a_t) * a_prev).sqrt())) * noise_t)
|
165 |
+
x_pred = x + x_delta
|
166 |
+
|
167 |
+
return x_pred
|
168 |
+
|
169 |
+
noise_list = self.noise_list
|
170 |
+
noise_pred = self.denoise_fn(x, t, cond=cond)
|
171 |
+
|
172 |
+
if len(noise_list) == 0:
|
173 |
+
x_pred = get_x_pred(x, noise_pred, t)
|
174 |
+
noise_pred_prev = self.denoise_fn(x_pred, max(t - interval, 0), cond=cond)
|
175 |
+
noise_pred_prime = (noise_pred + noise_pred_prev) / 2
|
176 |
+
elif len(noise_list) == 1:
|
177 |
+
noise_pred_prime = (3 * noise_pred - noise_list[-1]) / 2
|
178 |
+
elif len(noise_list) == 2:
|
179 |
+
noise_pred_prime = (23 * noise_pred - 16 * noise_list[-1] + 5 * noise_list[-2]) / 12
|
180 |
+
else:
|
181 |
+
noise_pred_prime = (55 * noise_pred - 59 * noise_list[-1] + 37 * noise_list[-2] - 9 * noise_list[-3]) / 24
|
182 |
+
|
183 |
+
x_prev = get_x_pred(x, noise_pred_prime, t)
|
184 |
+
noise_list.append(noise_pred)
|
185 |
+
|
186 |
+
return x_prev
|
187 |
+
|
188 |
+
def q_sample(self, x_start, t, noise=None):
|
189 |
+
noise = default(noise, lambda: torch.randn_like(x_start))
|
190 |
+
return (
|
191 |
+
extract(self.sqrt_alphas_cumprod, t, x_start.shape) * x_start +
|
192 |
+
extract(self.sqrt_one_minus_alphas_cumprod, t, x_start.shape) * noise
|
193 |
+
)
|
194 |
+
|
195 |
+
def p_losses(self, x_start, t, cond, noise=None, loss_type='l2'):
|
196 |
+
noise = default(noise, lambda: torch.randn_like(x_start))
|
197 |
+
|
198 |
+
x_noisy = self.q_sample(x_start=x_start, t=t, noise=noise)
|
199 |
+
x_recon = self.denoise_fn(x_noisy, t, cond)
|
200 |
+
|
201 |
+
if loss_type == 'l1':
|
202 |
+
loss = (noise - x_recon).abs().mean()
|
203 |
+
elif loss_type == 'l2':
|
204 |
+
loss = F.mse_loss(noise, x_recon)
|
205 |
+
else:
|
206 |
+
raise NotImplementedError()
|
207 |
+
|
208 |
+
return loss
|
209 |
+
|
210 |
+
def forward(self,
|
211 |
+
condition,
|
212 |
+
gt_spec=None,
|
213 |
+
infer=True,
|
214 |
+
infer_speedup=10,
|
215 |
+
method='dpm-solver',
|
216 |
+
k_step=300,
|
217 |
+
use_tqdm=True):
|
218 |
+
"""
|
219 |
+
conditioning diffusion, use fastspeech2 encoder output as the condition
|
220 |
+
"""
|
221 |
+
cond = condition.transpose(1, 2)
|
222 |
+
b, device = condition.shape[0], condition.device
|
223 |
+
|
224 |
+
if not infer:
|
225 |
+
spec = self.norm_spec(gt_spec)
|
226 |
+
t = torch.randint(0, self.k_step, (b,), device=device).long()
|
227 |
+
norm_spec = spec.transpose(1, 2)[:, None, :, :] # [B, 1, M, T]
|
228 |
+
return self.p_losses(norm_spec, t, cond=cond)
|
229 |
+
else:
|
230 |
+
shape = (cond.shape[0], 1, self.out_dims, cond.shape[2])
|
231 |
+
|
232 |
+
if gt_spec is None:
|
233 |
+
t = self.k_step
|
234 |
+
x = torch.randn(shape, device=device)
|
235 |
+
else:
|
236 |
+
t = k_step
|
237 |
+
norm_spec = self.norm_spec(gt_spec)
|
238 |
+
norm_spec = norm_spec.transpose(1, 2)[:, None, :, :]
|
239 |
+
x = self.q_sample(x_start=norm_spec, t=torch.tensor([t - 1], device=device).long())
|
240 |
+
|
241 |
+
if method is not None and infer_speedup > 1:
|
242 |
+
if method == 'dpm-solver':
|
243 |
+
from .dpm_solver_pytorch import NoiseScheduleVP, model_wrapper, DPM_Solver
|
244 |
+
# 1. Define the noise schedule.
|
245 |
+
noise_schedule = NoiseScheduleVP(schedule='discrete', betas=self.betas[:t])
|
246 |
+
|
247 |
+
# 2. Convert your discrete-time `model` to the continuous-time
|
248 |
+
# noise prediction model. Here is an example for a diffusion model
|
249 |
+
# `model` with the noise prediction type ("noise") .
|
250 |
+
def my_wrapper(fn):
|
251 |
+
def wrapped(x, t, **kwargs):
|
252 |
+
ret = fn(x, t, **kwargs)
|
253 |
+
if use_tqdm:
|
254 |
+
self.bar.update(1)
|
255 |
+
return ret
|
256 |
+
|
257 |
+
return wrapped
|
258 |
+
|
259 |
+
model_fn = model_wrapper(
|
260 |
+
my_wrapper(self.denoise_fn),
|
261 |
+
noise_schedule,
|
262 |
+
model_type="noise", # or "x_start" or "v" or "score"
|
263 |
+
model_kwargs={"cond": cond}
|
264 |
+
)
|
265 |
+
|
266 |
+
# 3. Define dpm-solver and sample by singlestep DPM-Solver.
|
267 |
+
# (We recommend singlestep DPM-Solver for unconditional sampling)
|
268 |
+
# You can adjust the `steps` to balance the computation
|
269 |
+
# costs and the sample quality.
|
270 |
+
dpm_solver = DPM_Solver(model_fn, noise_schedule)
|
271 |
+
|
272 |
+
steps = t // infer_speedup
|
273 |
+
if use_tqdm:
|
274 |
+
self.bar = tqdm(desc="sample time step", total=steps)
|
275 |
+
x = dpm_solver.sample(
|
276 |
+
x,
|
277 |
+
steps=steps,
|
278 |
+
order=3,
|
279 |
+
skip_type="time_uniform",
|
280 |
+
method="singlestep",
|
281 |
+
)
|
282 |
+
if use_tqdm:
|
283 |
+
self.bar.close()
|
284 |
+
elif method == 'pndm':
|
285 |
+
self.noise_list = deque(maxlen=4)
|
286 |
+
if use_tqdm:
|
287 |
+
for i in tqdm(
|
288 |
+
reversed(range(0, t, infer_speedup)), desc='sample time step',
|
289 |
+
total=t // infer_speedup,
|
290 |
+
):
|
291 |
+
x = self.p_sample_plms(
|
292 |
+
x, torch.full((b,), i, device=device, dtype=torch.long),
|
293 |
+
infer_speedup, cond=cond
|
294 |
+
)
|
295 |
+
else:
|
296 |
+
for i in reversed(range(0, t, infer_speedup)):
|
297 |
+
x = self.p_sample_plms(
|
298 |
+
x, torch.full((b,), i, device=device, dtype=torch.long),
|
299 |
+
infer_speedup, cond=cond
|
300 |
+
)
|
301 |
+
else:
|
302 |
+
raise NotImplementedError(method)
|
303 |
+
else:
|
304 |
+
if use_tqdm:
|
305 |
+
for i in tqdm(reversed(range(0, t)), desc='sample time step', total=t):
|
306 |
+
x = self.p_sample(x, torch.full((b,), i, device=device, dtype=torch.long), cond)
|
307 |
+
else:
|
308 |
+
for i in reversed(range(0, t)):
|
309 |
+
x = self.p_sample(x, torch.full((b,), i, device=device, dtype=torch.long), cond)
|
310 |
+
x = x.squeeze(1).transpose(1, 2) # [B, T, M]
|
311 |
+
return self.denorm_spec(x)
|
312 |
+
|
313 |
+
def norm_spec(self, x):
|
314 |
+
return (x - self.spec_min) / (self.spec_max - self.spec_min) * 2 - 1
|
315 |
+
|
316 |
+
def denorm_spec(self, x):
|
317 |
+
return (x + 1) / 2 * (self.spec_max - self.spec_min) + self.spec_min
|
diffusion/diffusion_onnx.py
ADDED
@@ -0,0 +1,612 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from collections import deque
|
2 |
+
from functools import partial
|
3 |
+
from inspect import isfunction
|
4 |
+
import torch.nn.functional as F
|
5 |
+
import librosa.sequence
|
6 |
+
import numpy as np
|
7 |
+
from torch.nn import Conv1d
|
8 |
+
from torch.nn import Mish
|
9 |
+
import torch
|
10 |
+
from torch import nn
|
11 |
+
from tqdm import tqdm
|
12 |
+
import math
|
13 |
+
|
14 |
+
|
15 |
+
def exists(x):
|
16 |
+
return x is not None
|
17 |
+
|
18 |
+
|
19 |
+
def default(val, d):
|
20 |
+
if exists(val):
|
21 |
+
return val
|
22 |
+
return d() if isfunction(d) else d
|
23 |
+
|
24 |
+
|
25 |
+
def extract(a, t):
|
26 |
+
return a[t].reshape((1, 1, 1, 1))
|
27 |
+
|
28 |
+
|
29 |
+
def noise_like(shape, device, repeat=False):
|
30 |
+
repeat_noise = lambda: torch.randn((1, *shape[1:]), device=device).repeat(shape[0], *((1,) * (len(shape) - 1)))
|
31 |
+
noise = lambda: torch.randn(shape, device=device)
|
32 |
+
return repeat_noise() if repeat else noise()
|
33 |
+
|
34 |
+
|
35 |
+
def linear_beta_schedule(timesteps, max_beta=0.02):
|
36 |
+
"""
|
37 |
+
linear schedule
|
38 |
+
"""
|
39 |
+
betas = np.linspace(1e-4, max_beta, timesteps)
|
40 |
+
return betas
|
41 |
+
|
42 |
+
|
43 |
+
def cosine_beta_schedule(timesteps, s=0.008):
|
44 |
+
"""
|
45 |
+
cosine schedule
|
46 |
+
as proposed in https://openreview.net/forum?id=-NEXDKk8gZ
|
47 |
+
"""
|
48 |
+
steps = timesteps + 1
|
49 |
+
x = np.linspace(0, steps, steps)
|
50 |
+
alphas_cumprod = np.cos(((x / steps) + s) / (1 + s) * np.pi * 0.5) ** 2
|
51 |
+
alphas_cumprod = alphas_cumprod / alphas_cumprod[0]
|
52 |
+
betas = 1 - (alphas_cumprod[1:] / alphas_cumprod[:-1])
|
53 |
+
return np.clip(betas, a_min=0, a_max=0.999)
|
54 |
+
|
55 |
+
|
56 |
+
beta_schedule = {
|
57 |
+
"cosine": cosine_beta_schedule,
|
58 |
+
"linear": linear_beta_schedule,
|
59 |
+
}
|
60 |
+
|
61 |
+
|
62 |
+
def extract_1(a, t):
|
63 |
+
return a[t].reshape((1, 1, 1, 1))
|
64 |
+
|
65 |
+
|
66 |
+
def predict_stage0(noise_pred, noise_pred_prev):
|
67 |
+
return (noise_pred + noise_pred_prev) / 2
|
68 |
+
|
69 |
+
|
70 |
+
def predict_stage1(noise_pred, noise_list):
|
71 |
+
return (noise_pred * 3
|
72 |
+
- noise_list[-1]) / 2
|
73 |
+
|
74 |
+
|
75 |
+
def predict_stage2(noise_pred, noise_list):
|
76 |
+
return (noise_pred * 23
|
77 |
+
- noise_list[-1] * 16
|
78 |
+
+ noise_list[-2] * 5) / 12
|
79 |
+
|
80 |
+
|
81 |
+
def predict_stage3(noise_pred, noise_list):
|
82 |
+
return (noise_pred * 55
|
83 |
+
- noise_list[-1] * 59
|
84 |
+
+ noise_list[-2] * 37
|
85 |
+
- noise_list[-3] * 9) / 24
|
86 |
+
|
87 |
+
|
88 |
+
class SinusoidalPosEmb(nn.Module):
|
89 |
+
def __init__(self, dim):
|
90 |
+
super().__init__()
|
91 |
+
self.dim = dim
|
92 |
+
self.half_dim = dim // 2
|
93 |
+
self.emb = 9.21034037 / (self.half_dim - 1)
|
94 |
+
self.emb = torch.exp(torch.arange(self.half_dim) * torch.tensor(-self.emb)).unsqueeze(0)
|
95 |
+
self.emb = self.emb.cpu()
|
96 |
+
|
97 |
+
def forward(self, x):
|
98 |
+
emb = self.emb * x
|
99 |
+
emb = torch.cat((emb.sin(), emb.cos()), dim=-1)
|
100 |
+
return emb
|
101 |
+
|
102 |
+
|
103 |
+
class ResidualBlock(nn.Module):
|
104 |
+
def __init__(self, encoder_hidden, residual_channels, dilation):
|
105 |
+
super().__init__()
|
106 |
+
self.residual_channels = residual_channels
|
107 |
+
self.dilated_conv = Conv1d(residual_channels, 2 * residual_channels, 3, padding=dilation, dilation=dilation)
|
108 |
+
self.diffusion_projection = nn.Linear(residual_channels, residual_channels)
|
109 |
+
self.conditioner_projection = Conv1d(encoder_hidden, 2 * residual_channels, 1)
|
110 |
+
self.output_projection = Conv1d(residual_channels, 2 * residual_channels, 1)
|
111 |
+
|
112 |
+
def forward(self, x, conditioner, diffusion_step):
|
113 |
+
diffusion_step = self.diffusion_projection(diffusion_step).unsqueeze(-1)
|
114 |
+
conditioner = self.conditioner_projection(conditioner)
|
115 |
+
y = x + diffusion_step
|
116 |
+
y = self.dilated_conv(y) + conditioner
|
117 |
+
|
118 |
+
gate, filter_1 = torch.split(y, [self.residual_channels, self.residual_channels], dim=1)
|
119 |
+
|
120 |
+
y = torch.sigmoid(gate) * torch.tanh(filter_1)
|
121 |
+
y = self.output_projection(y)
|
122 |
+
|
123 |
+
residual, skip = torch.split(y, [self.residual_channels, self.residual_channels], dim=1)
|
124 |
+
|
125 |
+
return (x + residual) / 1.41421356, skip
|
126 |
+
|
127 |
+
|
128 |
+
class DiffNet(nn.Module):
|
129 |
+
def __init__(self, in_dims, n_layers, n_chans, n_hidden):
|
130 |
+
super().__init__()
|
131 |
+
self.encoder_hidden = n_hidden
|
132 |
+
self.residual_layers = n_layers
|
133 |
+
self.residual_channels = n_chans
|
134 |
+
self.input_projection = Conv1d(in_dims, self.residual_channels, 1)
|
135 |
+
self.diffusion_embedding = SinusoidalPosEmb(self.residual_channels)
|
136 |
+
dim = self.residual_channels
|
137 |
+
self.mlp = nn.Sequential(
|
138 |
+
nn.Linear(dim, dim * 4),
|
139 |
+
Mish(),
|
140 |
+
nn.Linear(dim * 4, dim)
|
141 |
+
)
|
142 |
+
self.residual_layers = nn.ModuleList([
|
143 |
+
ResidualBlock(self.encoder_hidden, self.residual_channels, 1)
|
144 |
+
for i in range(self.residual_layers)
|
145 |
+
])
|
146 |
+
self.skip_projection = Conv1d(self.residual_channels, self.residual_channels, 1)
|
147 |
+
self.output_projection = Conv1d(self.residual_channels, in_dims, 1)
|
148 |
+
nn.init.zeros_(self.output_projection.weight)
|
149 |
+
|
150 |
+
def forward(self, spec, diffusion_step, cond):
|
151 |
+
x = spec.squeeze(0)
|
152 |
+
x = self.input_projection(x) # x [B, residual_channel, T]
|
153 |
+
x = F.relu(x)
|
154 |
+
# skip = torch.randn_like(x)
|
155 |
+
diffusion_step = diffusion_step.float()
|
156 |
+
diffusion_step = self.diffusion_embedding(diffusion_step)
|
157 |
+
diffusion_step = self.mlp(diffusion_step)
|
158 |
+
|
159 |
+
x, skip = self.residual_layers[0](x, cond, diffusion_step)
|
160 |
+
# noinspection PyTypeChecker
|
161 |
+
for layer in self.residual_layers[1:]:
|
162 |
+
x, skip_connection = layer.forward(x, cond, diffusion_step)
|
163 |
+
skip = skip + skip_connection
|
164 |
+
x = skip / math.sqrt(len(self.residual_layers))
|
165 |
+
x = self.skip_projection(x)
|
166 |
+
x = F.relu(x)
|
167 |
+
x = self.output_projection(x) # [B, 80, T]
|
168 |
+
return x.unsqueeze(1)
|
169 |
+
|
170 |
+
|
171 |
+
class AfterDiffusion(nn.Module):
|
172 |
+
def __init__(self, spec_max, spec_min, v_type='a'):
|
173 |
+
super().__init__()
|
174 |
+
self.spec_max = spec_max
|
175 |
+
self.spec_min = spec_min
|
176 |
+
self.type = v_type
|
177 |
+
|
178 |
+
def forward(self, x):
|
179 |
+
x = x.squeeze(1).permute(0, 2, 1)
|
180 |
+
mel_out = (x + 1) / 2 * (self.spec_max - self.spec_min) + self.spec_min
|
181 |
+
if self.type == 'nsf-hifigan-log10':
|
182 |
+
mel_out = mel_out * 0.434294
|
183 |
+
return mel_out.transpose(2, 1)
|
184 |
+
|
185 |
+
|
186 |
+
class Pred(nn.Module):
|
187 |
+
def __init__(self, alphas_cumprod):
|
188 |
+
super().__init__()
|
189 |
+
self.alphas_cumprod = alphas_cumprod
|
190 |
+
|
191 |
+
def forward(self, x_1, noise_t, t_1, t_prev):
|
192 |
+
a_t = extract(self.alphas_cumprod, t_1).cpu()
|
193 |
+
a_prev = extract(self.alphas_cumprod, t_prev).cpu()
|
194 |
+
a_t_sq, a_prev_sq = a_t.sqrt().cpu(), a_prev.sqrt().cpu()
|
195 |
+
x_delta = (a_prev - a_t) * ((1 / (a_t_sq * (a_t_sq + a_prev_sq))) * x_1 - 1 / (
|
196 |
+
a_t_sq * (((1 - a_prev) * a_t).sqrt() + ((1 - a_t) * a_prev).sqrt())) * noise_t)
|
197 |
+
x_pred = x_1 + x_delta.cpu()
|
198 |
+
|
199 |
+
return x_pred
|
200 |
+
|
201 |
+
|
202 |
+
class GaussianDiffusion(nn.Module):
|
203 |
+
def __init__(self,
|
204 |
+
out_dims=128,
|
205 |
+
n_layers=20,
|
206 |
+
n_chans=384,
|
207 |
+
n_hidden=256,
|
208 |
+
timesteps=1000,
|
209 |
+
k_step=1000,
|
210 |
+
max_beta=0.02,
|
211 |
+
spec_min=-12,
|
212 |
+
spec_max=2):
|
213 |
+
super().__init__()
|
214 |
+
self.denoise_fn = DiffNet(out_dims, n_layers, n_chans, n_hidden)
|
215 |
+
self.out_dims = out_dims
|
216 |
+
self.mel_bins = out_dims
|
217 |
+
self.n_hidden = n_hidden
|
218 |
+
betas = beta_schedule['linear'](timesteps, max_beta=max_beta)
|
219 |
+
|
220 |
+
alphas = 1. - betas
|
221 |
+
alphas_cumprod = np.cumprod(alphas, axis=0)
|
222 |
+
alphas_cumprod_prev = np.append(1., alphas_cumprod[:-1])
|
223 |
+
timesteps, = betas.shape
|
224 |
+
self.num_timesteps = int(timesteps)
|
225 |
+
self.k_step = k_step
|
226 |
+
|
227 |
+
self.noise_list = deque(maxlen=4)
|
228 |
+
|
229 |
+
to_torch = partial(torch.tensor, dtype=torch.float32)
|
230 |
+
|
231 |
+
self.register_buffer('betas', to_torch(betas))
|
232 |
+
self.register_buffer('alphas_cumprod', to_torch(alphas_cumprod))
|
233 |
+
self.register_buffer('alphas_cumprod_prev', to_torch(alphas_cumprod_prev))
|
234 |
+
|
235 |
+
# calculations for diffusion q(x_t | x_{t-1}) and others
|
236 |
+
self.register_buffer('sqrt_alphas_cumprod', to_torch(np.sqrt(alphas_cumprod)))
|
237 |
+
self.register_buffer('sqrt_one_minus_alphas_cumprod', to_torch(np.sqrt(1. - alphas_cumprod)))
|
238 |
+
self.register_buffer('log_one_minus_alphas_cumprod', to_torch(np.log(1. - alphas_cumprod)))
|
239 |
+
self.register_buffer('sqrt_recip_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod)))
|
240 |
+
self.register_buffer('sqrt_recipm1_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod - 1)))
|
241 |
+
|
242 |
+
# calculations for posterior q(x_{t-1} | x_t, x_0)
|
243 |
+
posterior_variance = betas * (1. - alphas_cumprod_prev) / (1. - alphas_cumprod)
|
244 |
+
# above: equal to 1. / (1. / (1. - alpha_cumprod_tm1) + alpha_t / beta_t)
|
245 |
+
self.register_buffer('posterior_variance', to_torch(posterior_variance))
|
246 |
+
# below: log calculation clipped because the posterior variance is 0 at the beginning of the diffusion chain
|
247 |
+
self.register_buffer('posterior_log_variance_clipped', to_torch(np.log(np.maximum(posterior_variance, 1e-20))))
|
248 |
+
self.register_buffer('posterior_mean_coef1', to_torch(
|
249 |
+
betas * np.sqrt(alphas_cumprod_prev) / (1. - alphas_cumprod)))
|
250 |
+
self.register_buffer('posterior_mean_coef2', to_torch(
|
251 |
+
(1. - alphas_cumprod_prev) * np.sqrt(alphas) / (1. - alphas_cumprod)))
|
252 |
+
|
253 |
+
self.register_buffer('spec_min', torch.FloatTensor([spec_min])[None, None, :out_dims])
|
254 |
+
self.register_buffer('spec_max', torch.FloatTensor([spec_max])[None, None, :out_dims])
|
255 |
+
self.ad = AfterDiffusion(self.spec_max, self.spec_min)
|
256 |
+
self.xp = Pred(self.alphas_cumprod)
|
257 |
+
|
258 |
+
def q_mean_variance(self, x_start, t):
|
259 |
+
mean = extract(self.sqrt_alphas_cumprod, t, x_start.shape) * x_start
|
260 |
+
variance = extract(1. - self.alphas_cumprod, t, x_start.shape)
|
261 |
+
log_variance = extract(self.log_one_minus_alphas_cumprod, t, x_start.shape)
|
262 |
+
return mean, variance, log_variance
|
263 |
+
|
264 |
+
def predict_start_from_noise(self, x_t, t, noise):
|
265 |
+
return (
|
266 |
+
extract(self.sqrt_recip_alphas_cumprod, t, x_t.shape) * x_t -
|
267 |
+
extract(self.sqrt_recipm1_alphas_cumprod, t, x_t.shape) * noise
|
268 |
+
)
|
269 |
+
|
270 |
+
def q_posterior(self, x_start, x_t, t):
|
271 |
+
posterior_mean = (
|
272 |
+
extract(self.posterior_mean_coef1, t, x_t.shape) * x_start +
|
273 |
+
extract(self.posterior_mean_coef2, t, x_t.shape) * x_t
|
274 |
+
)
|
275 |
+
posterior_variance = extract(self.posterior_variance, t, x_t.shape)
|
276 |
+
posterior_log_variance_clipped = extract(self.posterior_log_variance_clipped, t, x_t.shape)
|
277 |
+
return posterior_mean, posterior_variance, posterior_log_variance_clipped
|
278 |
+
|
279 |
+
def p_mean_variance(self, x, t, cond):
|
280 |
+
noise_pred = self.denoise_fn(x, t, cond=cond)
|
281 |
+
x_recon = self.predict_start_from_noise(x, t=t, noise=noise_pred)
|
282 |
+
|
283 |
+
x_recon.clamp_(-1., 1.)
|
284 |
+
|
285 |
+
model_mean, posterior_variance, posterior_log_variance = self.q_posterior(x_start=x_recon, x_t=x, t=t)
|
286 |
+
return model_mean, posterior_variance, posterior_log_variance
|
287 |
+
|
288 |
+
@torch.no_grad()
|
289 |
+
def p_sample(self, x, t, cond, clip_denoised=True, repeat_noise=False):
|
290 |
+
b, *_, device = *x.shape, x.device
|
291 |
+
model_mean, _, model_log_variance = self.p_mean_variance(x=x, t=t, cond=cond)
|
292 |
+
noise = noise_like(x.shape, device, repeat_noise)
|
293 |
+
# no noise when t == 0
|
294 |
+
nonzero_mask = (1 - (t == 0).float()).reshape(b, *((1,) * (len(x.shape) - 1)))
|
295 |
+
return model_mean + nonzero_mask * (0.5 * model_log_variance).exp() * noise
|
296 |
+
|
297 |
+
@torch.no_grad()
|
298 |
+
def p_sample_plms(self, x, t, interval, cond, clip_denoised=True, repeat_noise=False):
|
299 |
+
"""
|
300 |
+
Use the PLMS method from
|
301 |
+
[Pseudo Numerical Methods for Diffusion Models on Manifolds](https://arxiv.org/abs/2202.09778).
|
302 |
+
"""
|
303 |
+
|
304 |
+
def get_x_pred(x, noise_t, t):
|
305 |
+
a_t = extract(self.alphas_cumprod, t)
|
306 |
+
a_prev = extract(self.alphas_cumprod, torch.max(t - interval, torch.zeros_like(t)))
|
307 |
+
a_t_sq, a_prev_sq = a_t.sqrt(), a_prev.sqrt()
|
308 |
+
|
309 |
+
x_delta = (a_prev - a_t) * ((1 / (a_t_sq * (a_t_sq + a_prev_sq))) * x - 1 / (
|
310 |
+
a_t_sq * (((1 - a_prev) * a_t).sqrt() + ((1 - a_t) * a_prev).sqrt())) * noise_t)
|
311 |
+
x_pred = x + x_delta
|
312 |
+
|
313 |
+
return x_pred
|
314 |
+
|
315 |
+
noise_list = self.noise_list
|
316 |
+
noise_pred = self.denoise_fn(x, t, cond=cond)
|
317 |
+
|
318 |
+
if len(noise_list) == 0:
|
319 |
+
x_pred = get_x_pred(x, noise_pred, t)
|
320 |
+
noise_pred_prev = self.denoise_fn(x_pred, max(t - interval, 0), cond=cond)
|
321 |
+
noise_pred_prime = (noise_pred + noise_pred_prev) / 2
|
322 |
+
elif len(noise_list) == 1:
|
323 |
+
noise_pred_prime = (3 * noise_pred - noise_list[-1]) / 2
|
324 |
+
elif len(noise_list) == 2:
|
325 |
+
noise_pred_prime = (23 * noise_pred - 16 * noise_list[-1] + 5 * noise_list[-2]) / 12
|
326 |
+
else:
|
327 |
+
noise_pred_prime = (55 * noise_pred - 59 * noise_list[-1] + 37 * noise_list[-2] - 9 * noise_list[-3]) / 24
|
328 |
+
|
329 |
+
x_prev = get_x_pred(x, noise_pred_prime, t)
|
330 |
+
noise_list.append(noise_pred)
|
331 |
+
|
332 |
+
return x_prev
|
333 |
+
|
334 |
+
def q_sample(self, x_start, t, noise=None):
|
335 |
+
noise = default(noise, lambda: torch.randn_like(x_start))
|
336 |
+
return (
|
337 |
+
extract(self.sqrt_alphas_cumprod, t, x_start.shape) * x_start +
|
338 |
+
extract(self.sqrt_one_minus_alphas_cumprod, t, x_start.shape) * noise
|
339 |
+
)
|
340 |
+
|
341 |
+
def p_losses(self, x_start, t, cond, noise=None, loss_type='l2'):
|
342 |
+
noise = default(noise, lambda: torch.randn_like(x_start))
|
343 |
+
|
344 |
+
x_noisy = self.q_sample(x_start=x_start, t=t, noise=noise)
|
345 |
+
x_recon = self.denoise_fn(x_noisy, t, cond)
|
346 |
+
|
347 |
+
if loss_type == 'l1':
|
348 |
+
loss = (noise - x_recon).abs().mean()
|
349 |
+
elif loss_type == 'l2':
|
350 |
+
loss = F.mse_loss(noise, x_recon)
|
351 |
+
else:
|
352 |
+
raise NotImplementedError()
|
353 |
+
|
354 |
+
return loss
|
355 |
+
|
356 |
+
def org_forward(self,
|
357 |
+
condition,
|
358 |
+
init_noise=None,
|
359 |
+
gt_spec=None,
|
360 |
+
infer=True,
|
361 |
+
infer_speedup=100,
|
362 |
+
method='pndm',
|
363 |
+
k_step=1000,
|
364 |
+
use_tqdm=True):
|
365 |
+
"""
|
366 |
+
conditioning diffusion, use fastspeech2 encoder output as the condition
|
367 |
+
"""
|
368 |
+
cond = condition
|
369 |
+
b, device = condition.shape[0], condition.device
|
370 |
+
if not infer:
|
371 |
+
spec = self.norm_spec(gt_spec)
|
372 |
+
t = torch.randint(0, self.k_step, (b,), device=device).long()
|
373 |
+
norm_spec = spec.transpose(1, 2)[:, None, :, :] # [B, 1, M, T]
|
374 |
+
return self.p_losses(norm_spec, t, cond=cond)
|
375 |
+
else:
|
376 |
+
shape = (cond.shape[0], 1, self.out_dims, cond.shape[2])
|
377 |
+
|
378 |
+
if gt_spec is None:
|
379 |
+
t = self.k_step
|
380 |
+
if init_noise is None:
|
381 |
+
x = torch.randn(shape, device=device)
|
382 |
+
else:
|
383 |
+
x = init_noise
|
384 |
+
else:
|
385 |
+
t = k_step
|
386 |
+
norm_spec = self.norm_spec(gt_spec)
|
387 |
+
norm_spec = norm_spec.transpose(1, 2)[:, None, :, :]
|
388 |
+
x = self.q_sample(x_start=norm_spec, t=torch.tensor([t - 1], device=device).long())
|
389 |
+
|
390 |
+
if method is not None and infer_speedup > 1:
|
391 |
+
if method == 'dpm-solver':
|
392 |
+
from .dpm_solver_pytorch import NoiseScheduleVP, model_wrapper, DPM_Solver
|
393 |
+
# 1. Define the noise schedule.
|
394 |
+
noise_schedule = NoiseScheduleVP(schedule='discrete', betas=self.betas[:t])
|
395 |
+
|
396 |
+
# 2. Convert your discrete-time `model` to the continuous-time
|
397 |
+
# noise prediction model. Here is an example for a diffusion model
|
398 |
+
# `model` with the noise prediction type ("noise") .
|
399 |
+
def my_wrapper(fn):
|
400 |
+
def wrapped(x, t, **kwargs):
|
401 |
+
ret = fn(x, t, **kwargs)
|
402 |
+
if use_tqdm:
|
403 |
+
self.bar.update(1)
|
404 |
+
return ret
|
405 |
+
|
406 |
+
return wrapped
|
407 |
+
|
408 |
+
model_fn = model_wrapper(
|
409 |
+
my_wrapper(self.denoise_fn),
|
410 |
+
noise_schedule,
|
411 |
+
model_type="noise", # or "x_start" or "v" or "score"
|
412 |
+
model_kwargs={"cond": cond}
|
413 |
+
)
|
414 |
+
|
415 |
+
# 3. Define dpm-solver and sample by singlestep DPM-Solver.
|
416 |
+
# (We recommend singlestep DPM-Solver for unconditional sampling)
|
417 |
+
# You can adjust the `steps` to balance the computation
|
418 |
+
# costs and the sample quality.
|
419 |
+
dpm_solver = DPM_Solver(model_fn, noise_schedule)
|
420 |
+
|
421 |
+
steps = t // infer_speedup
|
422 |
+
if use_tqdm:
|
423 |
+
self.bar = tqdm(desc="sample time step", total=steps)
|
424 |
+
x = dpm_solver.sample(
|
425 |
+
x,
|
426 |
+
steps=steps,
|
427 |
+
order=3,
|
428 |
+
skip_type="time_uniform",
|
429 |
+
method="singlestep",
|
430 |
+
)
|
431 |
+
if use_tqdm:
|
432 |
+
self.bar.close()
|
433 |
+
elif method == 'pndm':
|
434 |
+
self.noise_list = deque(maxlen=4)
|
435 |
+
if use_tqdm:
|
436 |
+
for i in tqdm(
|
437 |
+
reversed(range(0, t, infer_speedup)), desc='sample time step',
|
438 |
+
total=t // infer_speedup,
|
439 |
+
):
|
440 |
+
x = self.p_sample_plms(
|
441 |
+
x, torch.full((b,), i, device=device, dtype=torch.long),
|
442 |
+
infer_speedup, cond=cond
|
443 |
+
)
|
444 |
+
else:
|
445 |
+
for i in reversed(range(0, t, infer_speedup)):
|
446 |
+
x = self.p_sample_plms(
|
447 |
+
x, torch.full((b,), i, device=device, dtype=torch.long),
|
448 |
+
infer_speedup, cond=cond
|
449 |
+
)
|
450 |
+
else:
|
451 |
+
raise NotImplementedError(method)
|
452 |
+
else:
|
453 |
+
if use_tqdm:
|
454 |
+
for i in tqdm(reversed(range(0, t)), desc='sample time step', total=t):
|
455 |
+
x = self.p_sample(x, torch.full((b,), i, device=device, dtype=torch.long), cond)
|
456 |
+
else:
|
457 |
+
for i in reversed(range(0, t)):
|
458 |
+
x = self.p_sample(x, torch.full((b,), i, device=device, dtype=torch.long), cond)
|
459 |
+
x = x.squeeze(1).transpose(1, 2) # [B, T, M]
|
460 |
+
return self.denorm_spec(x).transpose(2, 1)
|
461 |
+
|
462 |
+
def norm_spec(self, x):
|
463 |
+
return (x - self.spec_min) / (self.spec_max - self.spec_min) * 2 - 1
|
464 |
+
|
465 |
+
def denorm_spec(self, x):
|
466 |
+
return (x + 1) / 2 * (self.spec_max - self.spec_min) + self.spec_min
|
467 |
+
|
468 |
+
def get_x_pred(self, x_1, noise_t, t_1, t_prev):
|
469 |
+
a_t = extract(self.alphas_cumprod, t_1)
|
470 |
+
a_prev = extract(self.alphas_cumprod, t_prev)
|
471 |
+
a_t_sq, a_prev_sq = a_t.sqrt(), a_prev.sqrt()
|
472 |
+
x_delta = (a_prev - a_t) * ((1 / (a_t_sq * (a_t_sq + a_prev_sq))) * x_1 - 1 / (
|
473 |
+
a_t_sq * (((1 - a_prev) * a_t).sqrt() + ((1 - a_t) * a_prev).sqrt())) * noise_t)
|
474 |
+
x_pred = x_1 + x_delta
|
475 |
+
return x_pred
|
476 |
+
|
477 |
+
def OnnxExport(self, project_name=None, init_noise=None, hidden_channels=256, export_denoise=True, export_pred=True, export_after=True):
|
478 |
+
cond = torch.randn([1, self.n_hidden, 10]).cpu()
|
479 |
+
if init_noise is None:
|
480 |
+
x = torch.randn((1, 1, self.mel_bins, cond.shape[2]), dtype=torch.float32).cpu()
|
481 |
+
else:
|
482 |
+
x = init_noise
|
483 |
+
pndms = 100
|
484 |
+
|
485 |
+
org_y_x = self.org_forward(cond, init_noise=x)
|
486 |
+
|
487 |
+
device = cond.device
|
488 |
+
n_frames = cond.shape[2]
|
489 |
+
step_range = torch.arange(0, self.k_step, pndms, dtype=torch.long, device=device).flip(0)
|
490 |
+
plms_noise_stage = torch.tensor(0, dtype=torch.long, device=device)
|
491 |
+
noise_list = torch.zeros((0, 1, 1, self.mel_bins, n_frames), device=device)
|
492 |
+
|
493 |
+
ot = step_range[0]
|
494 |
+
ot_1 = torch.full((1,), ot, device=device, dtype=torch.long)
|
495 |
+
if export_denoise:
|
496 |
+
torch.onnx.export(
|
497 |
+
self.denoise_fn,
|
498 |
+
(x.cpu(), ot_1.cpu(), cond.cpu()),
|
499 |
+
f"{project_name}_denoise.onnx",
|
500 |
+
input_names=["noise", "time", "condition"],
|
501 |
+
output_names=["noise_pred"],
|
502 |
+
dynamic_axes={
|
503 |
+
"noise": [3],
|
504 |
+
"condition": [2]
|
505 |
+
},
|
506 |
+
opset_version=16
|
507 |
+
)
|
508 |
+
|
509 |
+
for t in step_range:
|
510 |
+
t_1 = torch.full((1,), t, device=device, dtype=torch.long)
|
511 |
+
noise_pred = self.denoise_fn(x, t_1, cond)
|
512 |
+
t_prev = t_1 - pndms
|
513 |
+
t_prev = t_prev * (t_prev > 0)
|
514 |
+
if plms_noise_stage == 0:
|
515 |
+
if export_pred:
|
516 |
+
torch.onnx.export(
|
517 |
+
self.xp,
|
518 |
+
(x.cpu(), noise_pred.cpu(), t_1.cpu(), t_prev.cpu()),
|
519 |
+
f"{project_name}_pred.onnx",
|
520 |
+
input_names=["noise", "noise_pred", "time", "time_prev"],
|
521 |
+
output_names=["noise_pred_o"],
|
522 |
+
dynamic_axes={
|
523 |
+
"noise": [3],
|
524 |
+
"noise_pred": [3]
|
525 |
+
},
|
526 |
+
opset_version=16
|
527 |
+
)
|
528 |
+
|
529 |
+
x_pred = self.get_x_pred(x, noise_pred, t_1, t_prev)
|
530 |
+
noise_pred_prev = self.denoise_fn(x_pred, t_prev, cond=cond)
|
531 |
+
noise_pred_prime = predict_stage0(noise_pred, noise_pred_prev)
|
532 |
+
|
533 |
+
elif plms_noise_stage == 1:
|
534 |
+
noise_pred_prime = predict_stage1(noise_pred, noise_list)
|
535 |
+
|
536 |
+
elif plms_noise_stage == 2:
|
537 |
+
noise_pred_prime = predict_stage2(noise_pred, noise_list)
|
538 |
+
|
539 |
+
else:
|
540 |
+
noise_pred_prime = predict_stage3(noise_pred, noise_list)
|
541 |
+
|
542 |
+
noise_pred = noise_pred.unsqueeze(0)
|
543 |
+
|
544 |
+
if plms_noise_stage < 3:
|
545 |
+
noise_list = torch.cat((noise_list, noise_pred), dim=0)
|
546 |
+
plms_noise_stage = plms_noise_stage + 1
|
547 |
+
|
548 |
+
else:
|
549 |
+
noise_list = torch.cat((noise_list[-2:], noise_pred), dim=0)
|
550 |
+
|
551 |
+
x = self.get_x_pred(x, noise_pred_prime, t_1, t_prev)
|
552 |
+
if export_after:
|
553 |
+
torch.onnx.export(
|
554 |
+
self.ad,
|
555 |
+
x.cpu(),
|
556 |
+
f"{project_name}_after.onnx",
|
557 |
+
input_names=["x"],
|
558 |
+
output_names=["mel_out"],
|
559 |
+
dynamic_axes={
|
560 |
+
"x": [3]
|
561 |
+
},
|
562 |
+
opset_version=16
|
563 |
+
)
|
564 |
+
x = self.ad(x)
|
565 |
+
|
566 |
+
print((x == org_y_x).all())
|
567 |
+
return x
|
568 |
+
|
569 |
+
def forward(self, condition=None, init_noise=None, pndms=None, k_step=None):
|
570 |
+
cond = condition
|
571 |
+
x = init_noise
|
572 |
+
|
573 |
+
device = cond.device
|
574 |
+
n_frames = cond.shape[2]
|
575 |
+
step_range = torch.arange(0, k_step.item(), pndms.item(), dtype=torch.long, device=device).flip(0)
|
576 |
+
plms_noise_stage = torch.tensor(0, dtype=torch.long, device=device)
|
577 |
+
noise_list = torch.zeros((0, 1, 1, self.mel_bins, n_frames), device=device)
|
578 |
+
|
579 |
+
ot = step_range[0]
|
580 |
+
ot_1 = torch.full((1,), ot, device=device, dtype=torch.long)
|
581 |
+
|
582 |
+
for t in step_range:
|
583 |
+
t_1 = torch.full((1,), t, device=device, dtype=torch.long)
|
584 |
+
noise_pred = self.denoise_fn(x, t_1, cond)
|
585 |
+
t_prev = t_1 - pndms
|
586 |
+
t_prev = t_prev * (t_prev > 0)
|
587 |
+
if plms_noise_stage == 0:
|
588 |
+
x_pred = self.get_x_pred(x, noise_pred, t_1, t_prev)
|
589 |
+
noise_pred_prev = self.denoise_fn(x_pred, t_prev, cond=cond)
|
590 |
+
noise_pred_prime = predict_stage0(noise_pred, noise_pred_prev)
|
591 |
+
|
592 |
+
elif plms_noise_stage == 1:
|
593 |
+
noise_pred_prime = predict_stage1(noise_pred, noise_list)
|
594 |
+
|
595 |
+
elif plms_noise_stage == 2:
|
596 |
+
noise_pred_prime = predict_stage2(noise_pred, noise_list)
|
597 |
+
|
598 |
+
else:
|
599 |
+
noise_pred_prime = predict_stage3(noise_pred, noise_list)
|
600 |
+
|
601 |
+
noise_pred = noise_pred.unsqueeze(0)
|
602 |
+
|
603 |
+
if plms_noise_stage < 3:
|
604 |
+
noise_list = torch.cat((noise_list, noise_pred), dim=0)
|
605 |
+
plms_noise_stage = plms_noise_stage + 1
|
606 |
+
|
607 |
+
else:
|
608 |
+
noise_list = torch.cat((noise_list[-2:], noise_pred), dim=0)
|
609 |
+
|
610 |
+
x = self.get_x_pred(x, noise_pred_prime, t_1, t_prev)
|
611 |
+
x = self.ad(x)
|
612 |
+
return x
|
diffusion/dpm_solver_pytorch.py
ADDED
@@ -0,0 +1,1201 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import math
|
2 |
+
|
3 |
+
import torch
|
4 |
+
|
5 |
+
|
6 |
+
class NoiseScheduleVP:
|
7 |
+
def __init__(
|
8 |
+
self,
|
9 |
+
schedule='discrete',
|
10 |
+
betas=None,
|
11 |
+
alphas_cumprod=None,
|
12 |
+
continuous_beta_0=0.1,
|
13 |
+
continuous_beta_1=20.,
|
14 |
+
):
|
15 |
+
"""Create a wrapper class for the forward SDE (VP type).
|
16 |
+
|
17 |
+
***
|
18 |
+
Update: We support discrete-time diffusion models by implementing a picewise linear interpolation for log_alpha_t.
|
19 |
+
We recommend to use schedule='discrete' for the discrete-time diffusion models, especially for high-resolution images.
|
20 |
+
***
|
21 |
+
|
22 |
+
The forward SDE ensures that the condition distribution q_{t|0}(x_t | x_0) = N ( alpha_t * x_0, sigma_t^2 * I ).
|
23 |
+
We further define lambda_t = log(alpha_t) - log(sigma_t), which is the half-logSNR (described in the DPM-Solver paper).
|
24 |
+
Therefore, we implement the functions for computing alpha_t, sigma_t and lambda_t. For t in [0, T], we have:
|
25 |
+
|
26 |
+
log_alpha_t = self.marginal_log_mean_coeff(t)
|
27 |
+
sigma_t = self.marginal_std(t)
|
28 |
+
lambda_t = self.marginal_lambda(t)
|
29 |
+
|
30 |
+
Moreover, as lambda(t) is an invertible function, we also support its inverse function:
|
31 |
+
|
32 |
+
t = self.inverse_lambda(lambda_t)
|
33 |
+
|
34 |
+
===============================================================
|
35 |
+
|
36 |
+
We support both discrete-time DPMs (trained on n = 0, 1, ..., N-1) and continuous-time DPMs (trained on t in [t_0, T]).
|
37 |
+
|
38 |
+
1. For discrete-time DPMs:
|
39 |
+
|
40 |
+
For discrete-time DPMs trained on n = 0, 1, ..., N-1, we convert the discrete steps to continuous time steps by:
|
41 |
+
t_i = (i + 1) / N
|
42 |
+
e.g. for N = 1000, we have t_0 = 1e-3 and T = t_{N-1} = 1.
|
43 |
+
We solve the corresponding diffusion ODE from time T = 1 to time t_0 = 1e-3.
|
44 |
+
|
45 |
+
Args:
|
46 |
+
betas: A `torch.Tensor`. The beta array for the discrete-time DPM. (See the original DDPM paper for details)
|
47 |
+
alphas_cumprod: A `torch.Tensor`. The cumprod alphas for the discrete-time DPM. (See the original DDPM paper for details)
|
48 |
+
|
49 |
+
Note that we always have alphas_cumprod = cumprod(betas). Therefore, we only need to set one of `betas` and `alphas_cumprod`.
|
50 |
+
|
51 |
+
**Important**: Please pay special attention for the args for `alphas_cumprod`:
|
52 |
+
The `alphas_cumprod` is the \hat{alpha_n} arrays in the notations of DDPM. Specifically, DDPMs assume that
|
53 |
+
q_{t_n | 0}(x_{t_n} | x_0) = N ( \sqrt{\hat{alpha_n}} * x_0, (1 - \hat{alpha_n}) * I ).
|
54 |
+
Therefore, the notation \hat{alpha_n} is different from the notation alpha_t in DPM-Solver. In fact, we have
|
55 |
+
alpha_{t_n} = \sqrt{\hat{alpha_n}},
|
56 |
+
and
|
57 |
+
log(alpha_{t_n}) = 0.5 * log(\hat{alpha_n}).
|
58 |
+
|
59 |
+
|
60 |
+
2. For continuous-time DPMs:
|
61 |
+
|
62 |
+
We support two types of VPSDEs: linear (DDPM) and cosine (improved-DDPM). The hyperparameters for the noise
|
63 |
+
schedule are the default settings in DDPM and improved-DDPM:
|
64 |
+
|
65 |
+
Args:
|
66 |
+
beta_min: A `float` number. The smallest beta for the linear schedule.
|
67 |
+
beta_max: A `float` number. The largest beta for the linear schedule.
|
68 |
+
cosine_s: A `float` number. The hyperparameter in the cosine schedule.
|
69 |
+
cosine_beta_max: A `float` number. The hyperparameter in the cosine schedule.
|
70 |
+
T: A `float` number. The ending time of the forward process.
|
71 |
+
|
72 |
+
===============================================================
|
73 |
+
|
74 |
+
Args:
|
75 |
+
schedule: A `str`. The noise schedule of the forward SDE. 'discrete' for discrete-time DPMs,
|
76 |
+
'linear' or 'cosine' for continuous-time DPMs.
|
77 |
+
Returns:
|
78 |
+
A wrapper object of the forward SDE (VP type).
|
79 |
+
|
80 |
+
===============================================================
|
81 |
+
|
82 |
+
Example:
|
83 |
+
|
84 |
+
# For discrete-time DPMs, given betas (the beta array for n = 0, 1, ..., N - 1):
|
85 |
+
>>> ns = NoiseScheduleVP('discrete', betas=betas)
|
86 |
+
|
87 |
+
# For discrete-time DPMs, given alphas_cumprod (the \hat{alpha_n} array for n = 0, 1, ..., N - 1):
|
88 |
+
>>> ns = NoiseScheduleVP('discrete', alphas_cumprod=alphas_cumprod)
|
89 |
+
|
90 |
+
# For continuous-time DPMs (VPSDE), linear schedule:
|
91 |
+
>>> ns = NoiseScheduleVP('linear', continuous_beta_0=0.1, continuous_beta_1=20.)
|
92 |
+
|
93 |
+
"""
|
94 |
+
|
95 |
+
if schedule not in ['discrete', 'linear', 'cosine']:
|
96 |
+
raise ValueError(
|
97 |
+
"Unsupported noise schedule {}. The schedule needs to be 'discrete' or 'linear' or 'cosine'".format(
|
98 |
+
schedule))
|
99 |
+
|
100 |
+
self.schedule = schedule
|
101 |
+
if schedule == 'discrete':
|
102 |
+
if betas is not None:
|
103 |
+
log_alphas = 0.5 * torch.log(1 - betas).cumsum(dim=0)
|
104 |
+
else:
|
105 |
+
assert alphas_cumprod is not None
|
106 |
+
log_alphas = 0.5 * torch.log(alphas_cumprod)
|
107 |
+
self.total_N = len(log_alphas)
|
108 |
+
self.T = 1.
|
109 |
+
self.t_array = torch.linspace(0., 1., self.total_N + 1)[1:].reshape((1, -1))
|
110 |
+
self.log_alpha_array = log_alphas.reshape((1, -1,))
|
111 |
+
else:
|
112 |
+
self.total_N = 1000
|
113 |
+
self.beta_0 = continuous_beta_0
|
114 |
+
self.beta_1 = continuous_beta_1
|
115 |
+
self.cosine_s = 0.008
|
116 |
+
self.cosine_beta_max = 999.
|
117 |
+
self.cosine_t_max = math.atan(self.cosine_beta_max * (1. + self.cosine_s) / math.pi) * 2. * (
|
118 |
+
1. + self.cosine_s) / math.pi - self.cosine_s
|
119 |
+
self.cosine_log_alpha_0 = math.log(math.cos(self.cosine_s / (1. + self.cosine_s) * math.pi / 2.))
|
120 |
+
self.schedule = schedule
|
121 |
+
if schedule == 'cosine':
|
122 |
+
# For the cosine schedule, T = 1 will have numerical issues. So we manually set the ending time T.
|
123 |
+
# Note that T = 0.9946 may be not the optimal setting. However, we find it works well.
|
124 |
+
self.T = 0.9946
|
125 |
+
else:
|
126 |
+
self.T = 1.
|
127 |
+
|
128 |
+
def marginal_log_mean_coeff(self, t):
|
129 |
+
"""
|
130 |
+
Compute log(alpha_t) of a given continuous-time label t in [0, T].
|
131 |
+
"""
|
132 |
+
if self.schedule == 'discrete':
|
133 |
+
return interpolate_fn(t.reshape((-1, 1)), self.t_array.to(t.device),
|
134 |
+
self.log_alpha_array.to(t.device)).reshape((-1))
|
135 |
+
elif self.schedule == 'linear':
|
136 |
+
return -0.25 * t ** 2 * (self.beta_1 - self.beta_0) - 0.5 * t * self.beta_0
|
137 |
+
elif self.schedule == 'cosine':
|
138 |
+
log_alpha_fn = lambda s: torch.log(torch.cos((s + self.cosine_s) / (1. + self.cosine_s) * math.pi / 2.))
|
139 |
+
log_alpha_t = log_alpha_fn(t) - self.cosine_log_alpha_0
|
140 |
+
return log_alpha_t
|
141 |
+
|
142 |
+
def marginal_alpha(self, t):
|
143 |
+
"""
|
144 |
+
Compute alpha_t of a given continuous-time label t in [0, T].
|
145 |
+
"""
|
146 |
+
return torch.exp(self.marginal_log_mean_coeff(t))
|
147 |
+
|
148 |
+
def marginal_std(self, t):
|
149 |
+
"""
|
150 |
+
Compute sigma_t of a given continuous-time label t in [0, T].
|
151 |
+
"""
|
152 |
+
return torch.sqrt(1. - torch.exp(2. * self.marginal_log_mean_coeff(t)))
|
153 |
+
|
154 |
+
def marginal_lambda(self, t):
|
155 |
+
"""
|
156 |
+
Compute lambda_t = log(alpha_t) - log(sigma_t) of a given continuous-time label t in [0, T].
|
157 |
+
"""
|
158 |
+
log_mean_coeff = self.marginal_log_mean_coeff(t)
|
159 |
+
log_std = 0.5 * torch.log(1. - torch.exp(2. * log_mean_coeff))
|
160 |
+
return log_mean_coeff - log_std
|
161 |
+
|
162 |
+
def inverse_lambda(self, lamb):
|
163 |
+
"""
|
164 |
+
Compute the continuous-time label t in [0, T] of a given half-logSNR lambda_t.
|
165 |
+
"""
|
166 |
+
if self.schedule == 'linear':
|
167 |
+
tmp = 2. * (self.beta_1 - self.beta_0) * torch.logaddexp(-2. * lamb, torch.zeros((1,)).to(lamb))
|
168 |
+
Delta = self.beta_0 ** 2 + tmp
|
169 |
+
return tmp / (torch.sqrt(Delta) + self.beta_0) / (self.beta_1 - self.beta_0)
|
170 |
+
elif self.schedule == 'discrete':
|
171 |
+
log_alpha = -0.5 * torch.logaddexp(torch.zeros((1,)).to(lamb.device), -2. * lamb)
|
172 |
+
t = interpolate_fn(log_alpha.reshape((-1, 1)), torch.flip(self.log_alpha_array.to(lamb.device), [1]),
|
173 |
+
torch.flip(self.t_array.to(lamb.device), [1]))
|
174 |
+
return t.reshape((-1,))
|
175 |
+
else:
|
176 |
+
log_alpha = -0.5 * torch.logaddexp(-2. * lamb, torch.zeros((1,)).to(lamb))
|
177 |
+
t_fn = lambda log_alpha_t: torch.arccos(torch.exp(log_alpha_t + self.cosine_log_alpha_0)) * 2. * (
|
178 |
+
1. + self.cosine_s) / math.pi - self.cosine_s
|
179 |
+
t = t_fn(log_alpha)
|
180 |
+
return t
|
181 |
+
|
182 |
+
|
183 |
+
def model_wrapper(
|
184 |
+
model,
|
185 |
+
noise_schedule,
|
186 |
+
model_type="noise",
|
187 |
+
model_kwargs={},
|
188 |
+
guidance_type="uncond",
|
189 |
+
condition=None,
|
190 |
+
unconditional_condition=None,
|
191 |
+
guidance_scale=1.,
|
192 |
+
classifier_fn=None,
|
193 |
+
classifier_kwargs={},
|
194 |
+
):
|
195 |
+
"""Create a wrapper function for the noise prediction model.
|
196 |
+
|
197 |
+
DPM-Solver needs to solve the continuous-time diffusion ODEs. For DPMs trained on discrete-time labels, we need to
|
198 |
+
firstly wrap the model function to a noise prediction model that accepts the continuous time as the input.
|
199 |
+
|
200 |
+
We support four types of the diffusion model by setting `model_type`:
|
201 |
+
|
202 |
+
1. "noise": noise prediction model. (Trained by predicting noise).
|
203 |
+
|
204 |
+
2. "x_start": data prediction model. (Trained by predicting the data x_0 at time 0).
|
205 |
+
|
206 |
+
3. "v": velocity prediction model. (Trained by predicting the velocity).
|
207 |
+
The "v" prediction is derivation detailed in Appendix D of [1], and is used in Imagen-Video [2].
|
208 |
+
|
209 |
+
[1] Salimans, Tim, and Jonathan Ho. "Progressive distillation for fast sampling of diffusion models."
|
210 |
+
arXiv preprint arXiv:2202.00512 (2022).
|
211 |
+
[2] Ho, Jonathan, et al. "Imagen Video: High Definition Video Generation with Diffusion Models."
|
212 |
+
arXiv preprint arXiv:2210.02303 (2022).
|
213 |
+
|
214 |
+
4. "score": marginal score function. (Trained by denoising score matching).
|
215 |
+
Note that the score function and the noise prediction model follows a simple relationship:
|
216 |
+
```
|
217 |
+
noise(x_t, t) = -sigma_t * score(x_t, t)
|
218 |
+
```
|
219 |
+
|
220 |
+
We support three types of guided sampling by DPMs by setting `guidance_type`:
|
221 |
+
1. "uncond": unconditional sampling by DPMs.
|
222 |
+
The input `model` has the following format:
|
223 |
+
``
|
224 |
+
model(x, t_input, **model_kwargs) -> noise | x_start | v | score
|
225 |
+
``
|
226 |
+
|
227 |
+
2. "classifier": classifier guidance sampling [3] by DPMs and another classifier.
|
228 |
+
The input `model` has the following format:
|
229 |
+
``
|
230 |
+
model(x, t_input, **model_kwargs) -> noise | x_start | v | score
|
231 |
+
``
|
232 |
+
|
233 |
+
The input `classifier_fn` has the following format:
|
234 |
+
``
|
235 |
+
classifier_fn(x, t_input, cond, **classifier_kwargs) -> logits(x, t_input, cond)
|
236 |
+
``
|
237 |
+
|
238 |
+
[3] P. Dhariwal and A. Q. Nichol, "Diffusion models beat GANs on image synthesis,"
|
239 |
+
in Advances in Neural Information Processing Systems, vol. 34, 2021, pp. 8780-8794.
|
240 |
+
|
241 |
+
3. "classifier-free": classifier-free guidance sampling by conditional DPMs.
|
242 |
+
The input `model` has the following format:
|
243 |
+
``
|
244 |
+
model(x, t_input, cond, **model_kwargs) -> noise | x_start | v | score
|
245 |
+
``
|
246 |
+
And if cond == `unconditional_condition`, the model output is the unconditional DPM output.
|
247 |
+
|
248 |
+
[4] Ho, Jonathan, and Tim Salimans. "Classifier-free diffusion guidance."
|
249 |
+
arXiv preprint arXiv:2207.12598 (2022).
|
250 |
+
|
251 |
+
|
252 |
+
The `t_input` is the time label of the model, which may be discrete-time labels (i.e. 0 to 999)
|
253 |
+
or continuous-time labels (i.e. epsilon to T).
|
254 |
+
|
255 |
+
We wrap the model function to accept only `x` and `t_continuous` as inputs, and outputs the predicted noise:
|
256 |
+
``
|
257 |
+
def model_fn(x, t_continuous) -> noise:
|
258 |
+
t_input = get_model_input_time(t_continuous)
|
259 |
+
return noise_pred(model, x, t_input, **model_kwargs)
|
260 |
+
``
|
261 |
+
where `t_continuous` is the continuous time labels (i.e. epsilon to T). And we use `model_fn` for DPM-Solver.
|
262 |
+
|
263 |
+
===============================================================
|
264 |
+
|
265 |
+
Args:
|
266 |
+
model: A diffusion model with the corresponding format described above.
|
267 |
+
noise_schedule: A noise schedule object, such as NoiseScheduleVP.
|
268 |
+
model_type: A `str`. The parameterization type of the diffusion model.
|
269 |
+
"noise" or "x_start" or "v" or "score".
|
270 |
+
model_kwargs: A `dict`. A dict for the other inputs of the model function.
|
271 |
+
guidance_type: A `str`. The type of the guidance for sampling.
|
272 |
+
"uncond" or "classifier" or "classifier-free".
|
273 |
+
condition: A pytorch tensor. The condition for the guided sampling.
|
274 |
+
Only used for "classifier" or "classifier-free" guidance type.
|
275 |
+
unconditional_condition: A pytorch tensor. The condition for the unconditional sampling.
|
276 |
+
Only used for "classifier-free" guidance type.
|
277 |
+
guidance_scale: A `float`. The scale for the guided sampling.
|
278 |
+
classifier_fn: A classifier function. Only used for the classifier guidance.
|
279 |
+
classifier_kwargs: A `dict`. A dict for the other inputs of the classifier function.
|
280 |
+
Returns:
|
281 |
+
A noise prediction model that accepts the noised data and the continuous time as the inputs.
|
282 |
+
"""
|
283 |
+
|
284 |
+
def get_model_input_time(t_continuous):
|
285 |
+
"""
|
286 |
+
Convert the continuous-time `t_continuous` (in [epsilon, T]) to the model input time.
|
287 |
+
For discrete-time DPMs, we convert `t_continuous` in [1 / N, 1] to `t_input` in [0, 1000 * (N - 1) / N].
|
288 |
+
For continuous-time DPMs, we just use `t_continuous`.
|
289 |
+
"""
|
290 |
+
if noise_schedule.schedule == 'discrete':
|
291 |
+
return (t_continuous - 1. / noise_schedule.total_N) * noise_schedule.total_N
|
292 |
+
else:
|
293 |
+
return t_continuous
|
294 |
+
|
295 |
+
def noise_pred_fn(x, t_continuous, cond=None):
|
296 |
+
if t_continuous.reshape((-1,)).shape[0] == 1:
|
297 |
+
t_continuous = t_continuous.expand((x.shape[0]))
|
298 |
+
t_input = get_model_input_time(t_continuous)
|
299 |
+
if cond is None:
|
300 |
+
output = model(x, t_input, **model_kwargs)
|
301 |
+
else:
|
302 |
+
output = model(x, t_input, cond, **model_kwargs)
|
303 |
+
if model_type == "noise":
|
304 |
+
return output
|
305 |
+
elif model_type == "x_start":
|
306 |
+
alpha_t, sigma_t = noise_schedule.marginal_alpha(t_continuous), noise_schedule.marginal_std(t_continuous)
|
307 |
+
dims = x.dim()
|
308 |
+
return (x - expand_dims(alpha_t, dims) * output) / expand_dims(sigma_t, dims)
|
309 |
+
elif model_type == "v":
|
310 |
+
alpha_t, sigma_t = noise_schedule.marginal_alpha(t_continuous), noise_schedule.marginal_std(t_continuous)
|
311 |
+
dims = x.dim()
|
312 |
+
return expand_dims(alpha_t, dims) * output + expand_dims(sigma_t, dims) * x
|
313 |
+
elif model_type == "score":
|
314 |
+
sigma_t = noise_schedule.marginal_std(t_continuous)
|
315 |
+
dims = x.dim()
|
316 |
+
return -expand_dims(sigma_t, dims) * output
|
317 |
+
|
318 |
+
def cond_grad_fn(x, t_input):
|
319 |
+
"""
|
320 |
+
Compute the gradient of the classifier, i.e. nabla_{x} log p_t(cond | x_t).
|
321 |
+
"""
|
322 |
+
with torch.enable_grad():
|
323 |
+
x_in = x.detach().requires_grad_(True)
|
324 |
+
log_prob = classifier_fn(x_in, t_input, condition, **classifier_kwargs)
|
325 |
+
return torch.autograd.grad(log_prob.sum(), x_in)[0]
|
326 |
+
|
327 |
+
def model_fn(x, t_continuous):
|
328 |
+
"""
|
329 |
+
The noise predicition model function that is used for DPM-Solver.
|
330 |
+
"""
|
331 |
+
if t_continuous.reshape((-1,)).shape[0] == 1:
|
332 |
+
t_continuous = t_continuous.expand((x.shape[0]))
|
333 |
+
if guidance_type == "uncond":
|
334 |
+
return noise_pred_fn(x, t_continuous)
|
335 |
+
elif guidance_type == "classifier":
|
336 |
+
assert classifier_fn is not None
|
337 |
+
t_input = get_model_input_time(t_continuous)
|
338 |
+
cond_grad = cond_grad_fn(x, t_input)
|
339 |
+
sigma_t = noise_schedule.marginal_std(t_continuous)
|
340 |
+
noise = noise_pred_fn(x, t_continuous)
|
341 |
+
return noise - guidance_scale * expand_dims(sigma_t, dims=cond_grad.dim()) * cond_grad
|
342 |
+
elif guidance_type == "classifier-free":
|
343 |
+
if guidance_scale == 1. or unconditional_condition is None:
|
344 |
+
return noise_pred_fn(x, t_continuous, cond=condition)
|
345 |
+
else:
|
346 |
+
x_in = torch.cat([x] * 2)
|
347 |
+
t_in = torch.cat([t_continuous] * 2)
|
348 |
+
c_in = torch.cat([unconditional_condition, condition])
|
349 |
+
noise_uncond, noise = noise_pred_fn(x_in, t_in, cond=c_in).chunk(2)
|
350 |
+
return noise_uncond + guidance_scale * (noise - noise_uncond)
|
351 |
+
|
352 |
+
assert model_type in ["noise", "x_start", "v"]
|
353 |
+
assert guidance_type in ["uncond", "classifier", "classifier-free"]
|
354 |
+
return model_fn
|
355 |
+
|
356 |
+
|
357 |
+
class DPM_Solver:
|
358 |
+
def __init__(self, model_fn, noise_schedule, predict_x0=False, thresholding=False, max_val=1.):
|
359 |
+
"""Construct a DPM-Solver.
|
360 |
+
|
361 |
+
We support both the noise prediction model ("predicting epsilon") and the data prediction model ("predicting x0").
|
362 |
+
If `predict_x0` is False, we use the solver for the noise prediction model (DPM-Solver).
|
363 |
+
If `predict_x0` is True, we use the solver for the data prediction model (DPM-Solver++).
|
364 |
+
In such case, we further support the "dynamic thresholding" in [1] when `thresholding` is True.
|
365 |
+
The "dynamic thresholding" can greatly improve the sample quality for pixel-space DPMs with large guidance scales.
|
366 |
+
|
367 |
+
Args:
|
368 |
+
model_fn: A noise prediction model function which accepts the continuous-time input (t in [epsilon, T]):
|
369 |
+
``
|
370 |
+
def model_fn(x, t_continuous):
|
371 |
+
return noise
|
372 |
+
``
|
373 |
+
noise_schedule: A noise schedule object, such as NoiseScheduleVP.
|
374 |
+
predict_x0: A `bool`. If true, use the data prediction model; else, use the noise prediction model.
|
375 |
+
thresholding: A `bool`. Valid when `predict_x0` is True. Whether to use the "dynamic thresholding" in [1].
|
376 |
+
max_val: A `float`. Valid when both `predict_x0` and `thresholding` are True. The max value for thresholding.
|
377 |
+
|
378 |
+
[1] Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022b.
|
379 |
+
"""
|
380 |
+
self.model = model_fn
|
381 |
+
self.noise_schedule = noise_schedule
|
382 |
+
self.predict_x0 = predict_x0
|
383 |
+
self.thresholding = thresholding
|
384 |
+
self.max_val = max_val
|
385 |
+
|
386 |
+
def noise_prediction_fn(self, x, t):
|
387 |
+
"""
|
388 |
+
Return the noise prediction model.
|
389 |
+
"""
|
390 |
+
return self.model(x, t)
|
391 |
+
|
392 |
+
def data_prediction_fn(self, x, t):
|
393 |
+
"""
|
394 |
+
Return the data prediction model (with thresholding).
|
395 |
+
"""
|
396 |
+
noise = self.noise_prediction_fn(x, t)
|
397 |
+
dims = x.dim()
|
398 |
+
alpha_t, sigma_t = self.noise_schedule.marginal_alpha(t), self.noise_schedule.marginal_std(t)
|
399 |
+
x0 = (x - expand_dims(sigma_t, dims) * noise) / expand_dims(alpha_t, dims)
|
400 |
+
if self.thresholding:
|
401 |
+
p = 0.995 # A hyperparameter in the paper of "Imagen" [1].
|
402 |
+
s = torch.quantile(torch.abs(x0).reshape((x0.shape[0], -1)), p, dim=1)
|
403 |
+
s = expand_dims(torch.maximum(s, self.max_val * torch.ones_like(s).to(s.device)), dims)
|
404 |
+
x0 = torch.clamp(x0, -s, s) / s
|
405 |
+
return x0
|
406 |
+
|
407 |
+
def model_fn(self, x, t):
|
408 |
+
"""
|
409 |
+
Convert the model to the noise prediction model or the data prediction model.
|
410 |
+
"""
|
411 |
+
if self.predict_x0:
|
412 |
+
return self.data_prediction_fn(x, t)
|
413 |
+
else:
|
414 |
+
return self.noise_prediction_fn(x, t)
|
415 |
+
|
416 |
+
def get_time_steps(self, skip_type, t_T, t_0, N, device):
|
417 |
+
"""Compute the intermediate time steps for sampling.
|
418 |
+
|
419 |
+
Args:
|
420 |
+
skip_type: A `str`. The type for the spacing of the time steps. We support three types:
|
421 |
+
- 'logSNR': uniform logSNR for the time steps.
|
422 |
+
- 'time_uniform': uniform time for the time steps. (**Recommended for high-resolutional data**.)
|
423 |
+
- 'time_quadratic': quadratic time for the time steps. (Used in DDIM for low-resolutional data.)
|
424 |
+
t_T: A `float`. The starting time of the sampling (default is T).
|
425 |
+
t_0: A `float`. The ending time of the sampling (default is epsilon).
|
426 |
+
N: A `int`. The total number of the spacing of the time steps.
|
427 |
+
device: A torch device.
|
428 |
+
Returns:
|
429 |
+
A pytorch tensor of the time steps, with the shape (N + 1,).
|
430 |
+
"""
|
431 |
+
if skip_type == 'logSNR':
|
432 |
+
lambda_T = self.noise_schedule.marginal_lambda(torch.tensor(t_T).to(device))
|
433 |
+
lambda_0 = self.noise_schedule.marginal_lambda(torch.tensor(t_0).to(device))
|
434 |
+
logSNR_steps = torch.linspace(lambda_T.cpu().item(), lambda_0.cpu().item(), N + 1).to(device)
|
435 |
+
return self.noise_schedule.inverse_lambda(logSNR_steps)
|
436 |
+
elif skip_type == 'time_uniform':
|
437 |
+
return torch.linspace(t_T, t_0, N + 1).to(device)
|
438 |
+
elif skip_type == 'time_quadratic':
|
439 |
+
t_order = 2
|
440 |
+
t = torch.linspace(t_T ** (1. / t_order), t_0 ** (1. / t_order), N + 1).pow(t_order).to(device)
|
441 |
+
return t
|
442 |
+
else:
|
443 |
+
raise ValueError(
|
444 |
+
"Unsupported skip_type {}, need to be 'logSNR' or 'time_uniform' or 'time_quadratic'".format(skip_type))
|
445 |
+
|
446 |
+
def get_orders_and_timesteps_for_singlestep_solver(self, steps, order, skip_type, t_T, t_0, device):
|
447 |
+
"""
|
448 |
+
Get the order of each step for sampling by the singlestep DPM-Solver.
|
449 |
+
|
450 |
+
We combine both DPM-Solver-1,2,3 to use all the function evaluations, which is named as "DPM-Solver-fast".
|
451 |
+
Given a fixed number of function evaluations by `steps`, the sampling procedure by DPM-Solver-fast is:
|
452 |
+
- If order == 1:
|
453 |
+
We take `steps` of DPM-Solver-1 (i.e. DDIM).
|
454 |
+
- If order == 2:
|
455 |
+
- Denote K = (steps // 2). We take K or (K + 1) intermediate time steps for sampling.
|
456 |
+
- If steps % 2 == 0, we use K steps of DPM-Solver-2.
|
457 |
+
- If steps % 2 == 1, we use K steps of DPM-Solver-2 and 1 step of DPM-Solver-1.
|
458 |
+
- If order == 3:
|
459 |
+
- Denote K = (steps // 3 + 1). We take K intermediate time steps for sampling.
|
460 |
+
- If steps % 3 == 0, we use (K - 2) steps of DPM-Solver-3, and 1 step of DPM-Solver-2 and 1 step of DPM-Solver-1.
|
461 |
+
- If steps % 3 == 1, we use (K - 1) steps of DPM-Solver-3 and 1 step of DPM-Solver-1.
|
462 |
+
- If steps % 3 == 2, we use (K - 1) steps of DPM-Solver-3 and 1 step of DPM-Solver-2.
|
463 |
+
|
464 |
+
============================================
|
465 |
+
Args:
|
466 |
+
order: A `int`. The max order for the solver (2 or 3).
|
467 |
+
steps: A `int`. The total number of function evaluations (NFE).
|
468 |
+
skip_type: A `str`. The type for the spacing of the time steps. We support three types:
|
469 |
+
- 'logSNR': uniform logSNR for the time steps.
|
470 |
+
- 'time_uniform': uniform time for the time steps. (**Recommended for high-resolutional data**.)
|
471 |
+
- 'time_quadratic': quadratic time for the time steps. (Used in DDIM for low-resolutional data.)
|
472 |
+
t_T: A `float`. The starting time of the sampling (default is T).
|
473 |
+
t_0: A `float`. The ending time of the sampling (default is epsilon).
|
474 |
+
device: A torch device.
|
475 |
+
Returns:
|
476 |
+
orders: A list of the solver order of each step.
|
477 |
+
"""
|
478 |
+
if order == 3:
|
479 |
+
K = steps // 3 + 1
|
480 |
+
if steps % 3 == 0:
|
481 |
+
orders = [3, ] * (K - 2) + [2, 1]
|
482 |
+
elif steps % 3 == 1:
|
483 |
+
orders = [3, ] * (K - 1) + [1]
|
484 |
+
else:
|
485 |
+
orders = [3, ] * (K - 1) + [2]
|
486 |
+
elif order == 2:
|
487 |
+
if steps % 2 == 0:
|
488 |
+
K = steps // 2
|
489 |
+
orders = [2, ] * K
|
490 |
+
else:
|
491 |
+
K = steps // 2 + 1
|
492 |
+
orders = [2, ] * (K - 1) + [1]
|
493 |
+
elif order == 1:
|
494 |
+
K = 1
|
495 |
+
orders = [1, ] * steps
|
496 |
+
else:
|
497 |
+
raise ValueError("'order' must be '1' or '2' or '3'.")
|
498 |
+
if skip_type == 'logSNR':
|
499 |
+
# To reproduce the results in DPM-Solver paper
|
500 |
+
timesteps_outer = self.get_time_steps(skip_type, t_T, t_0, K, device)
|
501 |
+
else:
|
502 |
+
timesteps_outer = self.get_time_steps(skip_type, t_T, t_0, steps, device)[
|
503 |
+
torch.cumsum(torch.tensor([0, ] + orders), dim=0).to(device)]
|
504 |
+
return timesteps_outer, orders
|
505 |
+
|
506 |
+
def denoise_fn(self, x, s):
|
507 |
+
"""
|
508 |
+
Denoise at the final step, which is equivalent to solve the ODE from lambda_s to infty by first-order discretization.
|
509 |
+
"""
|
510 |
+
return self.data_prediction_fn(x, s)
|
511 |
+
|
512 |
+
def dpm_solver_first_update(self, x, s, t, model_s=None, return_intermediate=False):
|
513 |
+
"""
|
514 |
+
DPM-Solver-1 (equivalent to DDIM) from time `s` to time `t`.
|
515 |
+
|
516 |
+
Args:
|
517 |
+
x: A pytorch tensor. The initial value at time `s`.
|
518 |
+
s: A pytorch tensor. The starting time, with the shape (x.shape[0],).
|
519 |
+
t: A pytorch tensor. The ending time, with the shape (x.shape[0],).
|
520 |
+
model_s: A pytorch tensor. The model function evaluated at time `s`.
|
521 |
+
If `model_s` is None, we evaluate the model by `x` and `s`; otherwise we directly use it.
|
522 |
+
return_intermediate: A `bool`. If true, also return the model value at time `s`.
|
523 |
+
Returns:
|
524 |
+
x_t: A pytorch tensor. The approximated solution at time `t`.
|
525 |
+
"""
|
526 |
+
ns = self.noise_schedule
|
527 |
+
dims = x.dim()
|
528 |
+
lambda_s, lambda_t = ns.marginal_lambda(s), ns.marginal_lambda(t)
|
529 |
+
h = lambda_t - lambda_s
|
530 |
+
log_alpha_s, log_alpha_t = ns.marginal_log_mean_coeff(s), ns.marginal_log_mean_coeff(t)
|
531 |
+
sigma_s, sigma_t = ns.marginal_std(s), ns.marginal_std(t)
|
532 |
+
alpha_t = torch.exp(log_alpha_t)
|
533 |
+
|
534 |
+
if self.predict_x0:
|
535 |
+
phi_1 = torch.expm1(-h)
|
536 |
+
if model_s is None:
|
537 |
+
model_s = self.model_fn(x, s)
|
538 |
+
x_t = (
|
539 |
+
expand_dims(sigma_t / sigma_s, dims) * x
|
540 |
+
- expand_dims(alpha_t * phi_1, dims) * model_s
|
541 |
+
)
|
542 |
+
if return_intermediate:
|
543 |
+
return x_t, {'model_s': model_s}
|
544 |
+
else:
|
545 |
+
return x_t
|
546 |
+
else:
|
547 |
+
phi_1 = torch.expm1(h)
|
548 |
+
if model_s is None:
|
549 |
+
model_s = self.model_fn(x, s)
|
550 |
+
x_t = (
|
551 |
+
expand_dims(torch.exp(log_alpha_t - log_alpha_s), dims) * x
|
552 |
+
- expand_dims(sigma_t * phi_1, dims) * model_s
|
553 |
+
)
|
554 |
+
if return_intermediate:
|
555 |
+
return x_t, {'model_s': model_s}
|
556 |
+
else:
|
557 |
+
return x_t
|
558 |
+
|
559 |
+
def singlestep_dpm_solver_second_update(self, x, s, t, r1=0.5, model_s=None, return_intermediate=False,
|
560 |
+
solver_type='dpm_solver'):
|
561 |
+
"""
|
562 |
+
Singlestep solver DPM-Solver-2 from time `s` to time `t`.
|
563 |
+
|
564 |
+
Args:
|
565 |
+
x: A pytorch tensor. The initial value at time `s`.
|
566 |
+
s: A pytorch tensor. The starting time, with the shape (x.shape[0],).
|
567 |
+
t: A pytorch tensor. The ending time, with the shape (x.shape[0],).
|
568 |
+
r1: A `float`. The hyperparameter of the second-order solver.
|
569 |
+
model_s: A pytorch tensor. The model function evaluated at time `s`.
|
570 |
+
If `model_s` is None, we evaluate the model by `x` and `s`; otherwise we directly use it.
|
571 |
+
return_intermediate: A `bool`. If true, also return the model value at time `s` and `s1` (the intermediate time).
|
572 |
+
solver_type: either 'dpm_solver' or 'taylor'. The type for the high-order solvers.
|
573 |
+
The type slightly impacts the performance. We recommend to use 'dpm_solver' type.
|
574 |
+
Returns:
|
575 |
+
x_t: A pytorch tensor. The approximated solution at time `t`.
|
576 |
+
"""
|
577 |
+
if solver_type not in ['dpm_solver', 'taylor']:
|
578 |
+
raise ValueError("'solver_type' must be either 'dpm_solver' or 'taylor', got {}".format(solver_type))
|
579 |
+
if r1 is None:
|
580 |
+
r1 = 0.5
|
581 |
+
ns = self.noise_schedule
|
582 |
+
dims = x.dim()
|
583 |
+
lambda_s, lambda_t = ns.marginal_lambda(s), ns.marginal_lambda(t)
|
584 |
+
h = lambda_t - lambda_s
|
585 |
+
lambda_s1 = lambda_s + r1 * h
|
586 |
+
s1 = ns.inverse_lambda(lambda_s1)
|
587 |
+
log_alpha_s, log_alpha_s1, log_alpha_t = ns.marginal_log_mean_coeff(s), ns.marginal_log_mean_coeff(
|
588 |
+
s1), ns.marginal_log_mean_coeff(t)
|
589 |
+
sigma_s, sigma_s1, sigma_t = ns.marginal_std(s), ns.marginal_std(s1), ns.marginal_std(t)
|
590 |
+
alpha_s1, alpha_t = torch.exp(log_alpha_s1), torch.exp(log_alpha_t)
|
591 |
+
|
592 |
+
if self.predict_x0:
|
593 |
+
phi_11 = torch.expm1(-r1 * h)
|
594 |
+
phi_1 = torch.expm1(-h)
|
595 |
+
|
596 |
+
if model_s is None:
|
597 |
+
model_s = self.model_fn(x, s)
|
598 |
+
x_s1 = (
|
599 |
+
expand_dims(sigma_s1 / sigma_s, dims) * x
|
600 |
+
- expand_dims(alpha_s1 * phi_11, dims) * model_s
|
601 |
+
)
|
602 |
+
model_s1 = self.model_fn(x_s1, s1)
|
603 |
+
if solver_type == 'dpm_solver':
|
604 |
+
x_t = (
|
605 |
+
expand_dims(sigma_t / sigma_s, dims) * x
|
606 |
+
- expand_dims(alpha_t * phi_1, dims) * model_s
|
607 |
+
- (0.5 / r1) * expand_dims(alpha_t * phi_1, dims) * (model_s1 - model_s)
|
608 |
+
)
|
609 |
+
elif solver_type == 'taylor':
|
610 |
+
x_t = (
|
611 |
+
expand_dims(sigma_t / sigma_s, dims) * x
|
612 |
+
- expand_dims(alpha_t * phi_1, dims) * model_s
|
613 |
+
+ (1. / r1) * expand_dims(alpha_t * ((torch.exp(-h) - 1.) / h + 1.), dims) * (
|
614 |
+
model_s1 - model_s)
|
615 |
+
)
|
616 |
+
else:
|
617 |
+
phi_11 = torch.expm1(r1 * h)
|
618 |
+
phi_1 = torch.expm1(h)
|
619 |
+
|
620 |
+
if model_s is None:
|
621 |
+
model_s = self.model_fn(x, s)
|
622 |
+
x_s1 = (
|
623 |
+
expand_dims(torch.exp(log_alpha_s1 - log_alpha_s), dims) * x
|
624 |
+
- expand_dims(sigma_s1 * phi_11, dims) * model_s
|
625 |
+
)
|
626 |
+
model_s1 = self.model_fn(x_s1, s1)
|
627 |
+
if solver_type == 'dpm_solver':
|
628 |
+
x_t = (
|
629 |
+
expand_dims(torch.exp(log_alpha_t - log_alpha_s), dims) * x
|
630 |
+
- expand_dims(sigma_t * phi_1, dims) * model_s
|
631 |
+
- (0.5 / r1) * expand_dims(sigma_t * phi_1, dims) * (model_s1 - model_s)
|
632 |
+
)
|
633 |
+
elif solver_type == 'taylor':
|
634 |
+
x_t = (
|
635 |
+
expand_dims(torch.exp(log_alpha_t - log_alpha_s), dims) * x
|
636 |
+
- expand_dims(sigma_t * phi_1, dims) * model_s
|
637 |
+
- (1. / r1) * expand_dims(sigma_t * ((torch.exp(h) - 1.) / h - 1.), dims) * (model_s1 - model_s)
|
638 |
+
)
|
639 |
+
if return_intermediate:
|
640 |
+
return x_t, {'model_s': model_s, 'model_s1': model_s1}
|
641 |
+
else:
|
642 |
+
return x_t
|
643 |
+
|
644 |
+
def singlestep_dpm_solver_third_update(self, x, s, t, r1=1. / 3., r2=2. / 3., model_s=None, model_s1=None,
|
645 |
+
return_intermediate=False, solver_type='dpm_solver'):
|
646 |
+
"""
|
647 |
+
Singlestep solver DPM-Solver-3 from time `s` to time `t`.
|
648 |
+
|
649 |
+
Args:
|
650 |
+
x: A pytorch tensor. The initial value at time `s`.
|
651 |
+
s: A pytorch tensor. The starting time, with the shape (x.shape[0],).
|
652 |
+
t: A pytorch tensor. The ending time, with the shape (x.shape[0],).
|
653 |
+
r1: A `float`. The hyperparameter of the third-order solver.
|
654 |
+
r2: A `float`. The hyperparameter of the third-order solver.
|
655 |
+
model_s: A pytorch tensor. The model function evaluated at time `s`.
|
656 |
+
If `model_s` is None, we evaluate the model by `x` and `s`; otherwise we directly use it.
|
657 |
+
model_s1: A pytorch tensor. The model function evaluated at time `s1` (the intermediate time given by `r1`).
|
658 |
+
If `model_s1` is None, we evaluate the model at `s1`; otherwise we directly use it.
|
659 |
+
return_intermediate: A `bool`. If true, also return the model value at time `s`, `s1` and `s2` (the intermediate times).
|
660 |
+
solver_type: either 'dpm_solver' or 'taylor'. The type for the high-order solvers.
|
661 |
+
The type slightly impacts the performance. We recommend to use 'dpm_solver' type.
|
662 |
+
Returns:
|
663 |
+
x_t: A pytorch tensor. The approximated solution at time `t`.
|
664 |
+
"""
|
665 |
+
if solver_type not in ['dpm_solver', 'taylor']:
|
666 |
+
raise ValueError("'solver_type' must be either 'dpm_solver' or 'taylor', got {}".format(solver_type))
|
667 |
+
if r1 is None:
|
668 |
+
r1 = 1. / 3.
|
669 |
+
if r2 is None:
|
670 |
+
r2 = 2. / 3.
|
671 |
+
ns = self.noise_schedule
|
672 |
+
dims = x.dim()
|
673 |
+
lambda_s, lambda_t = ns.marginal_lambda(s), ns.marginal_lambda(t)
|
674 |
+
h = lambda_t - lambda_s
|
675 |
+
lambda_s1 = lambda_s + r1 * h
|
676 |
+
lambda_s2 = lambda_s + r2 * h
|
677 |
+
s1 = ns.inverse_lambda(lambda_s1)
|
678 |
+
s2 = ns.inverse_lambda(lambda_s2)
|
679 |
+
log_alpha_s, log_alpha_s1, log_alpha_s2, log_alpha_t = ns.marginal_log_mean_coeff(
|
680 |
+
s), ns.marginal_log_mean_coeff(s1), ns.marginal_log_mean_coeff(s2), ns.marginal_log_mean_coeff(t)
|
681 |
+
sigma_s, sigma_s1, sigma_s2, sigma_t = ns.marginal_std(s), ns.marginal_std(s1), ns.marginal_std(
|
682 |
+
s2), ns.marginal_std(t)
|
683 |
+
alpha_s1, alpha_s2, alpha_t = torch.exp(log_alpha_s1), torch.exp(log_alpha_s2), torch.exp(log_alpha_t)
|
684 |
+
|
685 |
+
if self.predict_x0:
|
686 |
+
phi_11 = torch.expm1(-r1 * h)
|
687 |
+
phi_12 = torch.expm1(-r2 * h)
|
688 |
+
phi_1 = torch.expm1(-h)
|
689 |
+
phi_22 = torch.expm1(-r2 * h) / (r2 * h) + 1.
|
690 |
+
phi_2 = phi_1 / h + 1.
|
691 |
+
phi_3 = phi_2 / h - 0.5
|
692 |
+
|
693 |
+
if model_s is None:
|
694 |
+
model_s = self.model_fn(x, s)
|
695 |
+
if model_s1 is None:
|
696 |
+
x_s1 = (
|
697 |
+
expand_dims(sigma_s1 / sigma_s, dims) * x
|
698 |
+
- expand_dims(alpha_s1 * phi_11, dims) * model_s
|
699 |
+
)
|
700 |
+
model_s1 = self.model_fn(x_s1, s1)
|
701 |
+
x_s2 = (
|
702 |
+
expand_dims(sigma_s2 / sigma_s, dims) * x
|
703 |
+
- expand_dims(alpha_s2 * phi_12, dims) * model_s
|
704 |
+
+ r2 / r1 * expand_dims(alpha_s2 * phi_22, dims) * (model_s1 - model_s)
|
705 |
+
)
|
706 |
+
model_s2 = self.model_fn(x_s2, s2)
|
707 |
+
if solver_type == 'dpm_solver':
|
708 |
+
x_t = (
|
709 |
+
expand_dims(sigma_t / sigma_s, dims) * x
|
710 |
+
- expand_dims(alpha_t * phi_1, dims) * model_s
|
711 |
+
+ (1. / r2) * expand_dims(alpha_t * phi_2, dims) * (model_s2 - model_s)
|
712 |
+
)
|
713 |
+
elif solver_type == 'taylor':
|
714 |
+
D1_0 = (1. / r1) * (model_s1 - model_s)
|
715 |
+
D1_1 = (1. / r2) * (model_s2 - model_s)
|
716 |
+
D1 = (r2 * D1_0 - r1 * D1_1) / (r2 - r1)
|
717 |
+
D2 = 2. * (D1_1 - D1_0) / (r2 - r1)
|
718 |
+
x_t = (
|
719 |
+
expand_dims(sigma_t / sigma_s, dims) * x
|
720 |
+
- expand_dims(alpha_t * phi_1, dims) * model_s
|
721 |
+
+ expand_dims(alpha_t * phi_2, dims) * D1
|
722 |
+
- expand_dims(alpha_t * phi_3, dims) * D2
|
723 |
+
)
|
724 |
+
else:
|
725 |
+
phi_11 = torch.expm1(r1 * h)
|
726 |
+
phi_12 = torch.expm1(r2 * h)
|
727 |
+
phi_1 = torch.expm1(h)
|
728 |
+
phi_22 = torch.expm1(r2 * h) / (r2 * h) - 1.
|
729 |
+
phi_2 = phi_1 / h - 1.
|
730 |
+
phi_3 = phi_2 / h - 0.5
|
731 |
+
|
732 |
+
if model_s is None:
|
733 |
+
model_s = self.model_fn(x, s)
|
734 |
+
if model_s1 is None:
|
735 |
+
x_s1 = (
|
736 |
+
expand_dims(torch.exp(log_alpha_s1 - log_alpha_s), dims) * x
|
737 |
+
- expand_dims(sigma_s1 * phi_11, dims) * model_s
|
738 |
+
)
|
739 |
+
model_s1 = self.model_fn(x_s1, s1)
|
740 |
+
x_s2 = (
|
741 |
+
expand_dims(torch.exp(log_alpha_s2 - log_alpha_s), dims) * x
|
742 |
+
- expand_dims(sigma_s2 * phi_12, dims) * model_s
|
743 |
+
- r2 / r1 * expand_dims(sigma_s2 * phi_22, dims) * (model_s1 - model_s)
|
744 |
+
)
|
745 |
+
model_s2 = self.model_fn(x_s2, s2)
|
746 |
+
if solver_type == 'dpm_solver':
|
747 |
+
x_t = (
|
748 |
+
expand_dims(torch.exp(log_alpha_t - log_alpha_s), dims) * x
|
749 |
+
- expand_dims(sigma_t * phi_1, dims) * model_s
|
750 |
+
- (1. / r2) * expand_dims(sigma_t * phi_2, dims) * (model_s2 - model_s)
|
751 |
+
)
|
752 |
+
elif solver_type == 'taylor':
|
753 |
+
D1_0 = (1. / r1) * (model_s1 - model_s)
|
754 |
+
D1_1 = (1. / r2) * (model_s2 - model_s)
|
755 |
+
D1 = (r2 * D1_0 - r1 * D1_1) / (r2 - r1)
|
756 |
+
D2 = 2. * (D1_1 - D1_0) / (r2 - r1)
|
757 |
+
x_t = (
|
758 |
+
expand_dims(torch.exp(log_alpha_t - log_alpha_s), dims) * x
|
759 |
+
- expand_dims(sigma_t * phi_1, dims) * model_s
|
760 |
+
- expand_dims(sigma_t * phi_2, dims) * D1
|
761 |
+
- expand_dims(sigma_t * phi_3, dims) * D2
|
762 |
+
)
|
763 |
+
|
764 |
+
if return_intermediate:
|
765 |
+
return x_t, {'model_s': model_s, 'model_s1': model_s1, 'model_s2': model_s2}
|
766 |
+
else:
|
767 |
+
return x_t
|
768 |
+
|
769 |
+
def multistep_dpm_solver_second_update(self, x, model_prev_list, t_prev_list, t, solver_type="dpm_solver"):
|
770 |
+
"""
|
771 |
+
Multistep solver DPM-Solver-2 from time `t_prev_list[-1]` to time `t`.
|
772 |
+
|
773 |
+
Args:
|
774 |
+
x: A pytorch tensor. The initial value at time `s`.
|
775 |
+
model_prev_list: A list of pytorch tensor. The previous computed model values.
|
776 |
+
t_prev_list: A list of pytorch tensor. The previous times, each time has the shape (x.shape[0],)
|
777 |
+
t: A pytorch tensor. The ending time, with the shape (x.shape[0],).
|
778 |
+
solver_type: either 'dpm_solver' or 'taylor'. The type for the high-order solvers.
|
779 |
+
The type slightly impacts the performance. We recommend to use 'dpm_solver' type.
|
780 |
+
Returns:
|
781 |
+
x_t: A pytorch tensor. The approximated solution at time `t`.
|
782 |
+
"""
|
783 |
+
if solver_type not in ['dpm_solver', 'taylor']:
|
784 |
+
raise ValueError("'solver_type' must be either 'dpm_solver' or 'taylor', got {}".format(solver_type))
|
785 |
+
ns = self.noise_schedule
|
786 |
+
dims = x.dim()
|
787 |
+
model_prev_1, model_prev_0 = model_prev_list
|
788 |
+
t_prev_1, t_prev_0 = t_prev_list
|
789 |
+
lambda_prev_1, lambda_prev_0, lambda_t = ns.marginal_lambda(t_prev_1), ns.marginal_lambda(
|
790 |
+
t_prev_0), ns.marginal_lambda(t)
|
791 |
+
log_alpha_prev_0, log_alpha_t = ns.marginal_log_mean_coeff(t_prev_0), ns.marginal_log_mean_coeff(t)
|
792 |
+
sigma_prev_0, sigma_t = ns.marginal_std(t_prev_0), ns.marginal_std(t)
|
793 |
+
alpha_t = torch.exp(log_alpha_t)
|
794 |
+
|
795 |
+
h_0 = lambda_prev_0 - lambda_prev_1
|
796 |
+
h = lambda_t - lambda_prev_0
|
797 |
+
r0 = h_0 / h
|
798 |
+
D1_0 = expand_dims(1. / r0, dims) * (model_prev_0 - model_prev_1)
|
799 |
+
if self.predict_x0:
|
800 |
+
if solver_type == 'dpm_solver':
|
801 |
+
x_t = (
|
802 |
+
expand_dims(sigma_t / sigma_prev_0, dims) * x
|
803 |
+
- expand_dims(alpha_t * (torch.exp(-h) - 1.), dims) * model_prev_0
|
804 |
+
- 0.5 * expand_dims(alpha_t * (torch.exp(-h) - 1.), dims) * D1_0
|
805 |
+
)
|
806 |
+
elif solver_type == 'taylor':
|
807 |
+
x_t = (
|
808 |
+
expand_dims(sigma_t / sigma_prev_0, dims) * x
|
809 |
+
- expand_dims(alpha_t * (torch.exp(-h) - 1.), dims) * model_prev_0
|
810 |
+
+ expand_dims(alpha_t * ((torch.exp(-h) - 1.) / h + 1.), dims) * D1_0
|
811 |
+
)
|
812 |
+
else:
|
813 |
+
if solver_type == 'dpm_solver':
|
814 |
+
x_t = (
|
815 |
+
expand_dims(torch.exp(log_alpha_t - log_alpha_prev_0), dims) * x
|
816 |
+
- expand_dims(sigma_t * (torch.exp(h) - 1.), dims) * model_prev_0
|
817 |
+
- 0.5 * expand_dims(sigma_t * (torch.exp(h) - 1.), dims) * D1_0
|
818 |
+
)
|
819 |
+
elif solver_type == 'taylor':
|
820 |
+
x_t = (
|
821 |
+
expand_dims(torch.exp(log_alpha_t - log_alpha_prev_0), dims) * x
|
822 |
+
- expand_dims(sigma_t * (torch.exp(h) - 1.), dims) * model_prev_0
|
823 |
+
- expand_dims(sigma_t * ((torch.exp(h) - 1.) / h - 1.), dims) * D1_0
|
824 |
+
)
|
825 |
+
return x_t
|
826 |
+
|
827 |
+
def multistep_dpm_solver_third_update(self, x, model_prev_list, t_prev_list, t, solver_type='dpm_solver'):
|
828 |
+
"""
|
829 |
+
Multistep solver DPM-Solver-3 from time `t_prev_list[-1]` to time `t`.
|
830 |
+
|
831 |
+
Args:
|
832 |
+
x: A pytorch tensor. The initial value at time `s`.
|
833 |
+
model_prev_list: A list of pytorch tensor. The previous computed model values.
|
834 |
+
t_prev_list: A list of pytorch tensor. The previous times, each time has the shape (x.shape[0],)
|
835 |
+
t: A pytorch tensor. The ending time, with the shape (x.shape[0],).
|
836 |
+
solver_type: either 'dpm_solver' or 'taylor'. The type for the high-order solvers.
|
837 |
+
The type slightly impacts the performance. We recommend to use 'dpm_solver' type.
|
838 |
+
Returns:
|
839 |
+
x_t: A pytorch tensor. The approximated solution at time `t`.
|
840 |
+
"""
|
841 |
+
ns = self.noise_schedule
|
842 |
+
dims = x.dim()
|
843 |
+
model_prev_2, model_prev_1, model_prev_0 = model_prev_list
|
844 |
+
t_prev_2, t_prev_1, t_prev_0 = t_prev_list
|
845 |
+
lambda_prev_2, lambda_prev_1, lambda_prev_0, lambda_t = ns.marginal_lambda(t_prev_2), ns.marginal_lambda(
|
846 |
+
t_prev_1), ns.marginal_lambda(t_prev_0), ns.marginal_lambda(t)
|
847 |
+
log_alpha_prev_0, log_alpha_t = ns.marginal_log_mean_coeff(t_prev_0), ns.marginal_log_mean_coeff(t)
|
848 |
+
sigma_prev_0, sigma_t = ns.marginal_std(t_prev_0), ns.marginal_std(t)
|
849 |
+
alpha_t = torch.exp(log_alpha_t)
|
850 |
+
|
851 |
+
h_1 = lambda_prev_1 - lambda_prev_2
|
852 |
+
h_0 = lambda_prev_0 - lambda_prev_1
|
853 |
+
h = lambda_t - lambda_prev_0
|
854 |
+
r0, r1 = h_0 / h, h_1 / h
|
855 |
+
D1_0 = expand_dims(1. / r0, dims) * (model_prev_0 - model_prev_1)
|
856 |
+
D1_1 = expand_dims(1. / r1, dims) * (model_prev_1 - model_prev_2)
|
857 |
+
D1 = D1_0 + expand_dims(r0 / (r0 + r1), dims) * (D1_0 - D1_1)
|
858 |
+
D2 = expand_dims(1. / (r0 + r1), dims) * (D1_0 - D1_1)
|
859 |
+
if self.predict_x0:
|
860 |
+
x_t = (
|
861 |
+
expand_dims(sigma_t / sigma_prev_0, dims) * x
|
862 |
+
- expand_dims(alpha_t * (torch.exp(-h) - 1.), dims) * model_prev_0
|
863 |
+
+ expand_dims(alpha_t * ((torch.exp(-h) - 1.) / h + 1.), dims) * D1
|
864 |
+
- expand_dims(alpha_t * ((torch.exp(-h) - 1. + h) / h ** 2 - 0.5), dims) * D2
|
865 |
+
)
|
866 |
+
else:
|
867 |
+
x_t = (
|
868 |
+
expand_dims(torch.exp(log_alpha_t - log_alpha_prev_0), dims) * x
|
869 |
+
- expand_dims(sigma_t * (torch.exp(h) - 1.), dims) * model_prev_0
|
870 |
+
- expand_dims(sigma_t * ((torch.exp(h) - 1.) / h - 1.), dims) * D1
|
871 |
+
- expand_dims(sigma_t * ((torch.exp(h) - 1. - h) / h ** 2 - 0.5), dims) * D2
|
872 |
+
)
|
873 |
+
return x_t
|
874 |
+
|
875 |
+
def singlestep_dpm_solver_update(self, x, s, t, order, return_intermediate=False, solver_type='dpm_solver', r1=None,
|
876 |
+
r2=None):
|
877 |
+
"""
|
878 |
+
Singlestep DPM-Solver with the order `order` from time `s` to time `t`.
|
879 |
+
|
880 |
+
Args:
|
881 |
+
x: A pytorch tensor. The initial value at time `s`.
|
882 |
+
s: A pytorch tensor. The starting time, with the shape (x.shape[0],).
|
883 |
+
t: A pytorch tensor. The ending time, with the shape (x.shape[0],).
|
884 |
+
order: A `int`. The order of DPM-Solver. We only support order == 1 or 2 or 3.
|
885 |
+
return_intermediate: A `bool`. If true, also return the model value at time `s`, `s1` and `s2` (the intermediate times).
|
886 |
+
solver_type: either 'dpm_solver' or 'taylor'. The type for the high-order solvers.
|
887 |
+
The type slightly impacts the performance. We recommend to use 'dpm_solver' type.
|
888 |
+
r1: A `float`. The hyperparameter of the second-order or third-order solver.
|
889 |
+
r2: A `float`. The hyperparameter of the third-order solver.
|
890 |
+
Returns:
|
891 |
+
x_t: A pytorch tensor. The approximated solution at time `t`.
|
892 |
+
"""
|
893 |
+
if order == 1:
|
894 |
+
return self.dpm_solver_first_update(x, s, t, return_intermediate=return_intermediate)
|
895 |
+
elif order == 2:
|
896 |
+
return self.singlestep_dpm_solver_second_update(x, s, t, return_intermediate=return_intermediate,
|
897 |
+
solver_type=solver_type, r1=r1)
|
898 |
+
elif order == 3:
|
899 |
+
return self.singlestep_dpm_solver_third_update(x, s, t, return_intermediate=return_intermediate,
|
900 |
+
solver_type=solver_type, r1=r1, r2=r2)
|
901 |
+
else:
|
902 |
+
raise ValueError("Solver order must be 1 or 2 or 3, got {}".format(order))
|
903 |
+
|
904 |
+
def multistep_dpm_solver_update(self, x, model_prev_list, t_prev_list, t, order, solver_type='dpm_solver'):
|
905 |
+
"""
|
906 |
+
Multistep DPM-Solver with the order `order` from time `t_prev_list[-1]` to time `t`.
|
907 |
+
|
908 |
+
Args:
|
909 |
+
x: A pytorch tensor. The initial value at time `s`.
|
910 |
+
model_prev_list: A list of pytorch tensor. The previous computed model values.
|
911 |
+
t_prev_list: A list of pytorch tensor. The previous times, each time has the shape (x.shape[0],)
|
912 |
+
t: A pytorch tensor. The ending time, with the shape (x.shape[0],).
|
913 |
+
order: A `int`. The order of DPM-Solver. We only support order == 1 or 2 or 3.
|
914 |
+
solver_type: either 'dpm_solver' or 'taylor'. The type for the high-order solvers.
|
915 |
+
The type slightly impacts the performance. We recommend to use 'dpm_solver' type.
|
916 |
+
Returns:
|
917 |
+
x_t: A pytorch tensor. The approximated solution at time `t`.
|
918 |
+
"""
|
919 |
+
if order == 1:
|
920 |
+
return self.dpm_solver_first_update(x, t_prev_list[-1], t, model_s=model_prev_list[-1])
|
921 |
+
elif order == 2:
|
922 |
+
return self.multistep_dpm_solver_second_update(x, model_prev_list, t_prev_list, t, solver_type=solver_type)
|
923 |
+
elif order == 3:
|
924 |
+
return self.multistep_dpm_solver_third_update(x, model_prev_list, t_prev_list, t, solver_type=solver_type)
|
925 |
+
else:
|
926 |
+
raise ValueError("Solver order must be 1 or 2 or 3, got {}".format(order))
|
927 |
+
|
928 |
+
def dpm_solver_adaptive(self, x, order, t_T, t_0, h_init=0.05, atol=0.0078, rtol=0.05, theta=0.9, t_err=1e-5,
|
929 |
+
solver_type='dpm_solver'):
|
930 |
+
"""
|
931 |
+
The adaptive step size solver based on singlestep DPM-Solver.
|
932 |
+
|
933 |
+
Args:
|
934 |
+
x: A pytorch tensor. The initial value at time `t_T`.
|
935 |
+
order: A `int`. The (higher) order of the solver. We only support order == 2 or 3.
|
936 |
+
t_T: A `float`. The starting time of the sampling (default is T).
|
937 |
+
t_0: A `float`. The ending time of the sampling (default is epsilon).
|
938 |
+
h_init: A `float`. The initial step size (for logSNR).
|
939 |
+
atol: A `float`. The absolute tolerance of the solver. For image data, the default setting is 0.0078, followed [1].
|
940 |
+
rtol: A `float`. The relative tolerance of the solver. The default setting is 0.05.
|
941 |
+
theta: A `float`. The safety hyperparameter for adapting the step size. The default setting is 0.9, followed [1].
|
942 |
+
t_err: A `float`. The tolerance for the time. We solve the diffusion ODE until the absolute error between the
|
943 |
+
current time and `t_0` is less than `t_err`. The default setting is 1e-5.
|
944 |
+
solver_type: either 'dpm_solver' or 'taylor'. The type for the high-order solvers.
|
945 |
+
The type slightly impacts the performance. We recommend to use 'dpm_solver' type.
|
946 |
+
Returns:
|
947 |
+
x_0: A pytorch tensor. The approximated solution at time `t_0`.
|
948 |
+
|
949 |
+
[1] A. Jolicoeur-Martineau, K. Li, R. Piché-Taillefer, T. Kachman, and I. Mitliagkas, "Gotta go fast when generating data with score-based models," arXiv preprint arXiv:2105.14080, 2021.
|
950 |
+
"""
|
951 |
+
ns = self.noise_schedule
|
952 |
+
s = t_T * torch.ones((x.shape[0],)).to(x)
|
953 |
+
lambda_s = ns.marginal_lambda(s)
|
954 |
+
lambda_0 = ns.marginal_lambda(t_0 * torch.ones_like(s).to(x))
|
955 |
+
h = h_init * torch.ones_like(s).to(x)
|
956 |
+
x_prev = x
|
957 |
+
nfe = 0
|
958 |
+
if order == 2:
|
959 |
+
r1 = 0.5
|
960 |
+
lower_update = lambda x, s, t: self.dpm_solver_first_update(x, s, t, return_intermediate=True)
|
961 |
+
higher_update = lambda x, s, t, **kwargs: self.singlestep_dpm_solver_second_update(x, s, t, r1=r1,
|
962 |
+
solver_type=solver_type,
|
963 |
+
**kwargs)
|
964 |
+
elif order == 3:
|
965 |
+
r1, r2 = 1. / 3., 2. / 3.
|
966 |
+
lower_update = lambda x, s, t: self.singlestep_dpm_solver_second_update(x, s, t, r1=r1,
|
967 |
+
return_intermediate=True,
|
968 |
+
solver_type=solver_type)
|
969 |
+
higher_update = lambda x, s, t, **kwargs: self.singlestep_dpm_solver_third_update(x, s, t, r1=r1, r2=r2,
|
970 |
+
solver_type=solver_type,
|
971 |
+
**kwargs)
|
972 |
+
else:
|
973 |
+
raise ValueError("For adaptive step size solver, order must be 2 or 3, got {}".format(order))
|
974 |
+
while torch.abs((s - t_0)).mean() > t_err:
|
975 |
+
t = ns.inverse_lambda(lambda_s + h)
|
976 |
+
x_lower, lower_noise_kwargs = lower_update(x, s, t)
|
977 |
+
x_higher = higher_update(x, s, t, **lower_noise_kwargs)
|
978 |
+
delta = torch.max(torch.ones_like(x).to(x) * atol, rtol * torch.max(torch.abs(x_lower), torch.abs(x_prev)))
|
979 |
+
norm_fn = lambda v: torch.sqrt(torch.square(v.reshape((v.shape[0], -1))).mean(dim=-1, keepdim=True))
|
980 |
+
E = norm_fn((x_higher - x_lower) / delta).max()
|
981 |
+
if torch.all(E <= 1.):
|
982 |
+
x = x_higher
|
983 |
+
s = t
|
984 |
+
x_prev = x_lower
|
985 |
+
lambda_s = ns.marginal_lambda(s)
|
986 |
+
h = torch.min(theta * h * torch.float_power(E, -1. / order).float(), lambda_0 - lambda_s)
|
987 |
+
nfe += order
|
988 |
+
print('adaptive solver nfe', nfe)
|
989 |
+
return x
|
990 |
+
|
991 |
+
def sample(self, x, steps=20, t_start=None, t_end=None, order=3, skip_type='time_uniform',
|
992 |
+
method='singlestep', denoise=False, solver_type='dpm_solver', atol=0.0078,
|
993 |
+
rtol=0.05,
|
994 |
+
):
|
995 |
+
"""
|
996 |
+
Compute the sample at time `t_end` by DPM-Solver, given the initial `x` at time `t_start`.
|
997 |
+
|
998 |
+
=====================================================
|
999 |
+
|
1000 |
+
We support the following algorithms for both noise prediction model and data prediction model:
|
1001 |
+
- 'singlestep':
|
1002 |
+
Singlestep DPM-Solver (i.e. "DPM-Solver-fast" in the paper), which combines different orders of singlestep DPM-Solver.
|
1003 |
+
We combine all the singlestep solvers with order <= `order` to use up all the function evaluations (steps).
|
1004 |
+
The total number of function evaluations (NFE) == `steps`.
|
1005 |
+
Given a fixed NFE == `steps`, the sampling procedure is:
|
1006 |
+
- If `order` == 1:
|
1007 |
+
- Denote K = steps. We use K steps of DPM-Solver-1 (i.e. DDIM).
|
1008 |
+
- If `order` == 2:
|
1009 |
+
- Denote K = (steps // 2) + (steps % 2). We take K intermediate time steps for sampling.
|
1010 |
+
- If steps % 2 == 0, we use K steps of singlestep DPM-Solver-2.
|
1011 |
+
- If steps % 2 == 1, we use (K - 1) steps of singlestep DPM-Solver-2 and 1 step of DPM-Solver-1.
|
1012 |
+
- If `order` == 3:
|
1013 |
+
- Denote K = (steps // 3 + 1). We take K intermediate time steps for sampling.
|
1014 |
+
- If steps % 3 == 0, we use (K - 2) steps of singlestep DPM-Solver-3, and 1 step of singlestep DPM-Solver-2 and 1 step of DPM-Solver-1.
|
1015 |
+
- If steps % 3 == 1, we use (K - 1) steps of singlestep DPM-Solver-3 and 1 step of DPM-Solver-1.
|
1016 |
+
- If steps % 3 == 2, we use (K - 1) steps of singlestep DPM-Solver-3 and 1 step of singlestep DPM-Solver-2.
|
1017 |
+
- 'multistep':
|
1018 |
+
Multistep DPM-Solver with the order of `order`. The total number of function evaluations (NFE) == `steps`.
|
1019 |
+
We initialize the first `order` values by lower order multistep solvers.
|
1020 |
+
Given a fixed NFE == `steps`, the sampling procedure is:
|
1021 |
+
Denote K = steps.
|
1022 |
+
- If `order` == 1:
|
1023 |
+
- We use K steps of DPM-Solver-1 (i.e. DDIM).
|
1024 |
+
- If `order` == 2:
|
1025 |
+
- We firstly use 1 step of DPM-Solver-1, then use (K - 1) step of multistep DPM-Solver-2.
|
1026 |
+
- If `order` == 3:
|
1027 |
+
- We firstly use 1 step of DPM-Solver-1, then 1 step of multistep DPM-Solver-2, then (K - 2) step of multistep DPM-Solver-3.
|
1028 |
+
- 'singlestep_fixed':
|
1029 |
+
Fixed order singlestep DPM-Solver (i.e. DPM-Solver-1 or singlestep DPM-Solver-2 or singlestep DPM-Solver-3).
|
1030 |
+
We use singlestep DPM-Solver-`order` for `order`=1 or 2 or 3, with total [`steps` // `order`] * `order` NFE.
|
1031 |
+
- 'adaptive':
|
1032 |
+
Adaptive step size DPM-Solver (i.e. "DPM-Solver-12" and "DPM-Solver-23" in the paper).
|
1033 |
+
We ignore `steps` and use adaptive step size DPM-Solver with a higher order of `order`.
|
1034 |
+
You can adjust the absolute tolerance `atol` and the relative tolerance `rtol` to balance the computatation costs
|
1035 |
+
(NFE) and the sample quality.
|
1036 |
+
- If `order` == 2, we use DPM-Solver-12 which combines DPM-Solver-1 and singlestep DPM-Solver-2.
|
1037 |
+
- If `order` == 3, we use DPM-Solver-23 which combines singlestep DPM-Solver-2 and singlestep DPM-Solver-3.
|
1038 |
+
|
1039 |
+
=====================================================
|
1040 |
+
|
1041 |
+
Some advices for choosing the algorithm:
|
1042 |
+
- For **unconditional sampling** or **guided sampling with small guidance scale** by DPMs:
|
1043 |
+
Use singlestep DPM-Solver ("DPM-Solver-fast" in the paper) with `order = 3`.
|
1044 |
+
e.g.
|
1045 |
+
>>> dpm_solver = DPM_Solver(model_fn, noise_schedule, predict_x0=False)
|
1046 |
+
>>> x_sample = dpm_solver.sample(x, steps=steps, t_start=t_start, t_end=t_end, order=3,
|
1047 |
+
skip_type='time_uniform', method='singlestep')
|
1048 |
+
- For **guided sampling with large guidance scale** by DPMs:
|
1049 |
+
Use multistep DPM-Solver with `predict_x0 = True` and `order = 2`.
|
1050 |
+
e.g.
|
1051 |
+
>>> dpm_solver = DPM_Solver(model_fn, noise_schedule, predict_x0=True)
|
1052 |
+
>>> x_sample = dpm_solver.sample(x, steps=steps, t_start=t_start, t_end=t_end, order=2,
|
1053 |
+
skip_type='time_uniform', method='multistep')
|
1054 |
+
|
1055 |
+
We support three types of `skip_type`:
|
1056 |
+
- 'logSNR': uniform logSNR for the time steps. **Recommended for low-resolutional images**
|
1057 |
+
- 'time_uniform': uniform time for the time steps. **Recommended for high-resolutional images**.
|
1058 |
+
- 'time_quadratic': quadratic time for the time steps.
|
1059 |
+
|
1060 |
+
=====================================================
|
1061 |
+
Args:
|
1062 |
+
x: A pytorch tensor. The initial value at time `t_start`
|
1063 |
+
e.g. if `t_start` == T, then `x` is a sample from the standard normal distribution.
|
1064 |
+
steps: A `int`. The total number of function evaluations (NFE).
|
1065 |
+
t_start: A `float`. The starting time of the sampling.
|
1066 |
+
If `T` is None, we use self.noise_schedule.T (default is 1.0).
|
1067 |
+
t_end: A `float`. The ending time of the sampling.
|
1068 |
+
If `t_end` is None, we use 1. / self.noise_schedule.total_N.
|
1069 |
+
e.g. if total_N == 1000, we have `t_end` == 1e-3.
|
1070 |
+
For discrete-time DPMs:
|
1071 |
+
- We recommend `t_end` == 1. / self.noise_schedule.total_N.
|
1072 |
+
For continuous-time DPMs:
|
1073 |
+
- We recommend `t_end` == 1e-3 when `steps` <= 15; and `t_end` == 1e-4 when `steps` > 15.
|
1074 |
+
order: A `int`. The order of DPM-Solver.
|
1075 |
+
skip_type: A `str`. The type for the spacing of the time steps. 'time_uniform' or 'logSNR' or 'time_quadratic'.
|
1076 |
+
method: A `str`. The method for sampling. 'singlestep' or 'multistep' or 'singlestep_fixed' or 'adaptive'.
|
1077 |
+
denoise: A `bool`. Whether to denoise at the final step. Default is False.
|
1078 |
+
If `denoise` is True, the total NFE is (`steps` + 1).
|
1079 |
+
solver_type: A `str`. The taylor expansion type for the solver. `dpm_solver` or `taylor`. We recommend `dpm_solver`.
|
1080 |
+
atol: A `float`. The absolute tolerance of the adaptive step size solver. Valid when `method` == 'adaptive'.
|
1081 |
+
rtol: A `float`. The relative tolerance of the adaptive step size solver. Valid when `method` == 'adaptive'.
|
1082 |
+
Returns:
|
1083 |
+
x_end: A pytorch tensor. The approximated solution at time `t_end`.
|
1084 |
+
|
1085 |
+
"""
|
1086 |
+
t_0 = 1. / self.noise_schedule.total_N if t_end is None else t_end
|
1087 |
+
t_T = self.noise_schedule.T if t_start is None else t_start
|
1088 |
+
device = x.device
|
1089 |
+
if method == 'adaptive':
|
1090 |
+
with torch.no_grad():
|
1091 |
+
x = self.dpm_solver_adaptive(x, order=order, t_T=t_T, t_0=t_0, atol=atol, rtol=rtol,
|
1092 |
+
solver_type=solver_type)
|
1093 |
+
elif method == 'multistep':
|
1094 |
+
assert steps >= order
|
1095 |
+
timesteps = self.get_time_steps(skip_type=skip_type, t_T=t_T, t_0=t_0, N=steps, device=device)
|
1096 |
+
assert timesteps.shape[0] - 1 == steps
|
1097 |
+
with torch.no_grad():
|
1098 |
+
vec_t = timesteps[0].expand((x.shape[0]))
|
1099 |
+
model_prev_list = [self.model_fn(x, vec_t)]
|
1100 |
+
t_prev_list = [vec_t]
|
1101 |
+
# Init the first `order` values by lower order multistep DPM-Solver.
|
1102 |
+
for init_order in range(1, order):
|
1103 |
+
vec_t = timesteps[init_order].expand(x.shape[0])
|
1104 |
+
x = self.multistep_dpm_solver_update(x, model_prev_list, t_prev_list, vec_t, init_order,
|
1105 |
+
solver_type=solver_type)
|
1106 |
+
model_prev_list.append(self.model_fn(x, vec_t))
|
1107 |
+
t_prev_list.append(vec_t)
|
1108 |
+
# Compute the remaining values by `order`-th order multistep DPM-Solver.
|
1109 |
+
for step in range(order, steps + 1):
|
1110 |
+
vec_t = timesteps[step].expand(x.shape[0])
|
1111 |
+
x = self.multistep_dpm_solver_update(x, model_prev_list, t_prev_list, vec_t, order,
|
1112 |
+
solver_type=solver_type)
|
1113 |
+
for i in range(order - 1):
|
1114 |
+
t_prev_list[i] = t_prev_list[i + 1]
|
1115 |
+
model_prev_list[i] = model_prev_list[i + 1]
|
1116 |
+
t_prev_list[-1] = vec_t
|
1117 |
+
# We do not need to evaluate the final model value.
|
1118 |
+
if step < steps:
|
1119 |
+
model_prev_list[-1] = self.model_fn(x, vec_t)
|
1120 |
+
elif method in ['singlestep', 'singlestep_fixed']:
|
1121 |
+
if method == 'singlestep':
|
1122 |
+
timesteps_outer, orders = self.get_orders_and_timesteps_for_singlestep_solver(steps=steps, order=order,
|
1123 |
+
skip_type=skip_type,
|
1124 |
+
t_T=t_T, t_0=t_0,
|
1125 |
+
device=device)
|
1126 |
+
elif method == 'singlestep_fixed':
|
1127 |
+
K = steps // order
|
1128 |
+
orders = [order, ] * K
|
1129 |
+
timesteps_outer = self.get_time_steps(skip_type=skip_type, t_T=t_T, t_0=t_0, N=K, device=device)
|
1130 |
+
for i, order in enumerate(orders):
|
1131 |
+
t_T_inner, t_0_inner = timesteps_outer[i], timesteps_outer[i + 1]
|
1132 |
+
timesteps_inner = self.get_time_steps(skip_type=skip_type, t_T=t_T_inner.item(), t_0=t_0_inner.item(),
|
1133 |
+
N=order, device=device)
|
1134 |
+
lambda_inner = self.noise_schedule.marginal_lambda(timesteps_inner)
|
1135 |
+
vec_s, vec_t = t_T_inner.repeat(x.shape[0]), t_0_inner.repeat(x.shape[0])
|
1136 |
+
h = lambda_inner[-1] - lambda_inner[0]
|
1137 |
+
r1 = None if order <= 1 else (lambda_inner[1] - lambda_inner[0]) / h
|
1138 |
+
r2 = None if order <= 2 else (lambda_inner[2] - lambda_inner[0]) / h
|
1139 |
+
x = self.singlestep_dpm_solver_update(x, vec_s, vec_t, order, solver_type=solver_type, r1=r1, r2=r2)
|
1140 |
+
if denoise:
|
1141 |
+
x = self.denoise_fn(x, torch.ones((x.shape[0],)).to(device) * t_0)
|
1142 |
+
return x
|
1143 |
+
|
1144 |
+
|
1145 |
+
#############################################################
|
1146 |
+
# other utility functions
|
1147 |
+
#############################################################
|
1148 |
+
|
1149 |
+
def interpolate_fn(x, xp, yp):
|
1150 |
+
"""
|
1151 |
+
A piecewise linear function y = f(x), using xp and yp as keypoints.
|
1152 |
+
We implement f(x) in a differentiable way (i.e. applicable for autograd).
|
1153 |
+
The function f(x) is well-defined for all x-axis. (For x beyond the bounds of xp, we use the outmost points of xp to define the linear function.)
|
1154 |
+
|
1155 |
+
Args:
|
1156 |
+
x: PyTorch tensor with shape [N, C], where N is the batch size, C is the number of channels (we use C = 1 for DPM-Solver).
|
1157 |
+
xp: PyTorch tensor with shape [C, K], where K is the number of keypoints.
|
1158 |
+
yp: PyTorch tensor with shape [C, K].
|
1159 |
+
Returns:
|
1160 |
+
The function values f(x), with shape [N, C].
|
1161 |
+
"""
|
1162 |
+
N, K = x.shape[0], xp.shape[1]
|
1163 |
+
all_x = torch.cat([x.unsqueeze(2), xp.unsqueeze(0).repeat((N, 1, 1))], dim=2)
|
1164 |
+
sorted_all_x, x_indices = torch.sort(all_x, dim=2)
|
1165 |
+
x_idx = torch.argmin(x_indices, dim=2)
|
1166 |
+
cand_start_idx = x_idx - 1
|
1167 |
+
start_idx = torch.where(
|
1168 |
+
torch.eq(x_idx, 0),
|
1169 |
+
torch.tensor(1, device=x.device),
|
1170 |
+
torch.where(
|
1171 |
+
torch.eq(x_idx, K), torch.tensor(K - 2, device=x.device), cand_start_idx,
|
1172 |
+
),
|
1173 |
+
)
|
1174 |
+
end_idx = torch.where(torch.eq(start_idx, cand_start_idx), start_idx + 2, start_idx + 1)
|
1175 |
+
start_x = torch.gather(sorted_all_x, dim=2, index=start_idx.unsqueeze(2)).squeeze(2)
|
1176 |
+
end_x = torch.gather(sorted_all_x, dim=2, index=end_idx.unsqueeze(2)).squeeze(2)
|
1177 |
+
start_idx2 = torch.where(
|
1178 |
+
torch.eq(x_idx, 0),
|
1179 |
+
torch.tensor(0, device=x.device),
|
1180 |
+
torch.where(
|
1181 |
+
torch.eq(x_idx, K), torch.tensor(K - 2, device=x.device), cand_start_idx,
|
1182 |
+
),
|
1183 |
+
)
|
1184 |
+
y_positions_expanded = yp.unsqueeze(0).expand(N, -1, -1)
|
1185 |
+
start_y = torch.gather(y_positions_expanded, dim=2, index=start_idx2.unsqueeze(2)).squeeze(2)
|
1186 |
+
end_y = torch.gather(y_positions_expanded, dim=2, index=(start_idx2 + 1).unsqueeze(2)).squeeze(2)
|
1187 |
+
cand = start_y + (x - start_x) * (end_y - start_y) / (end_x - start_x)
|
1188 |
+
return cand
|
1189 |
+
|
1190 |
+
|
1191 |
+
def expand_dims(v, dims):
|
1192 |
+
"""
|
1193 |
+
Expand the tensor `v` to the dim `dims`.
|
1194 |
+
|
1195 |
+
Args:
|
1196 |
+
`v`: a PyTorch tensor with shape [N].
|
1197 |
+
`dim`: a `int`.
|
1198 |
+
Returns:
|
1199 |
+
a PyTorch tensor with shape [N, 1, 1, ..., 1] and the total dimension is `dims`.
|
1200 |
+
"""
|
1201 |
+
return v[(...,) + (None,) * (dims - 1)]
|
diffusion/how to export onnx.md
ADDED
@@ -0,0 +1,4 @@
|
|
|
|
|
|
|
|
|
|
|
1 |
+
- Open [onnx_export](onnx_export.py)
|
2 |
+
- project_name = "dddsp" change "project_name" to your project name
|
3 |
+
- model_path = f'{project_name}/model_500000.pt' change "model_path" to your model path
|
4 |
+
- Run
|
diffusion/infer_gt_mel.py
ADDED
@@ -0,0 +1,74 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import numpy as np
|
2 |
+
import torch
|
3 |
+
import torch.nn.functional as F
|
4 |
+
from diffusion.unit2mel import load_model_vocoder
|
5 |
+
|
6 |
+
|
7 |
+
class DiffGtMel:
|
8 |
+
def __init__(self, project_path=None, device=None):
|
9 |
+
self.project_path = project_path
|
10 |
+
if device is not None:
|
11 |
+
self.device = device
|
12 |
+
else:
|
13 |
+
self.device = 'cuda' if torch.cuda.is_available() else 'cpu'
|
14 |
+
self.model = None
|
15 |
+
self.vocoder = None
|
16 |
+
self.args = None
|
17 |
+
|
18 |
+
def flush_model(self, project_path, ddsp_config=None):
|
19 |
+
if (self.model is None) or (project_path != self.project_path):
|
20 |
+
model, vocoder, args = load_model_vocoder(project_path, device=self.device)
|
21 |
+
if self.check_args(ddsp_config, args):
|
22 |
+
self.model = model
|
23 |
+
self.vocoder = vocoder
|
24 |
+
self.args = args
|
25 |
+
|
26 |
+
def check_args(self, args1, args2):
|
27 |
+
if args1.data.block_size != args2.data.block_size:
|
28 |
+
raise ValueError("DDSP与DIFF模型的block_size不一致")
|
29 |
+
if args1.data.sampling_rate != args2.data.sampling_rate:
|
30 |
+
raise ValueError("DDSP与DIFF模型的sampling_rate不一致")
|
31 |
+
if args1.data.encoder != args2.data.encoder:
|
32 |
+
raise ValueError("DDSP与DIFF模型的encoder不一致")
|
33 |
+
return True
|
34 |
+
|
35 |
+
def __call__(self, audio, f0, hubert, volume, acc=1, spk_id=1, k_step=0, method='pndm',
|
36 |
+
spk_mix_dict=None, start_frame=0):
|
37 |
+
input_mel = self.vocoder.extract(audio, self.args.data.sampling_rate)
|
38 |
+
out_mel = self.model(
|
39 |
+
hubert,
|
40 |
+
f0,
|
41 |
+
volume,
|
42 |
+
spk_id=spk_id,
|
43 |
+
spk_mix_dict=spk_mix_dict,
|
44 |
+
gt_spec=input_mel,
|
45 |
+
infer=True,
|
46 |
+
infer_speedup=acc,
|
47 |
+
method=method,
|
48 |
+
k_step=k_step,
|
49 |
+
use_tqdm=False)
|
50 |
+
if start_frame > 0:
|
51 |
+
out_mel = out_mel[:, start_frame:, :]
|
52 |
+
f0 = f0[:, start_frame:, :]
|
53 |
+
output = self.vocoder.infer(out_mel, f0)
|
54 |
+
if start_frame > 0:
|
55 |
+
output = F.pad(output, (start_frame * self.vocoder.vocoder_hop_size, 0))
|
56 |
+
return output
|
57 |
+
|
58 |
+
def infer(self, audio, f0, hubert, volume, acc=1, spk_id=1, k_step=0, method='pndm', silence_front=0,
|
59 |
+
use_silence=False, spk_mix_dict=None):
|
60 |
+
start_frame = int(silence_front * self.vocoder.vocoder_sample_rate / self.vocoder.vocoder_hop_size)
|
61 |
+
if use_silence:
|
62 |
+
audio = audio[:, start_frame * self.vocoder.vocoder_hop_size:]
|
63 |
+
f0 = f0[:, start_frame:, :]
|
64 |
+
hubert = hubert[:, start_frame:, :]
|
65 |
+
volume = volume[:, start_frame:, :]
|
66 |
+
_start_frame = 0
|
67 |
+
else:
|
68 |
+
_start_frame = start_frame
|
69 |
+
audio = self.__call__(audio, f0, hubert, volume, acc=acc, spk_id=spk_id, k_step=k_step,
|
70 |
+
method=method, spk_mix_dict=spk_mix_dict, start_frame=_start_frame)
|
71 |
+
if use_silence:
|
72 |
+
if start_frame > 0:
|
73 |
+
audio = F.pad(audio, (start_frame * self.vocoder.vocoder_hop_size, 0))
|
74 |
+
return audio
|
diffusion/logger/__init__.py
ADDED
File without changes
|
diffusion/logger/__pycache__/__init__.cpython-38.pyc
ADDED
Binary file (138 Bytes). View file
|
|
diffusion/logger/__pycache__/saver.cpython-38.pyc
ADDED
Binary file (4 kB). View file
|
|
diffusion/logger/__pycache__/utils.cpython-38.pyc
ADDED
Binary file (3.8 kB). View file
|
|
diffusion/logger/saver.py
ADDED
@@ -0,0 +1,150 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
'''
|
2 |
+
author: wayn391@mastertones
|
3 |
+
'''
|
4 |
+
|
5 |
+
import os
|
6 |
+
import json
|
7 |
+
import time
|
8 |
+
import yaml
|
9 |
+
import datetime
|
10 |
+
import torch
|
11 |
+
import matplotlib.pyplot as plt
|
12 |
+
from . import utils
|
13 |
+
from torch.utils.tensorboard import SummaryWriter
|
14 |
+
|
15 |
+
class Saver(object):
|
16 |
+
def __init__(
|
17 |
+
self,
|
18 |
+
args,
|
19 |
+
initial_global_step=-1):
|
20 |
+
|
21 |
+
self.expdir = args.env.expdir
|
22 |
+
self.sample_rate = args.data.sampling_rate
|
23 |
+
|
24 |
+
# cold start
|
25 |
+
self.global_step = initial_global_step
|
26 |
+
self.init_time = time.time()
|
27 |
+
self.last_time = time.time()
|
28 |
+
|
29 |
+
# makedirs
|
30 |
+
os.makedirs(self.expdir, exist_ok=True)
|
31 |
+
|
32 |
+
# path
|
33 |
+
self.path_log_info = os.path.join(self.expdir, 'log_info.txt')
|
34 |
+
|
35 |
+
# ckpt
|
36 |
+
os.makedirs(self.expdir, exist_ok=True)
|
37 |
+
|
38 |
+
# writer
|
39 |
+
self.writer = SummaryWriter(os.path.join(self.expdir, 'logs'))
|
40 |
+
|
41 |
+
# save config
|
42 |
+
path_config = os.path.join(self.expdir, 'config.yaml')
|
43 |
+
with open(path_config, "w") as out_config:
|
44 |
+
yaml.dump(dict(args), out_config)
|
45 |
+
|
46 |
+
|
47 |
+
def log_info(self, msg):
|
48 |
+
'''log method'''
|
49 |
+
if isinstance(msg, dict):
|
50 |
+
msg_list = []
|
51 |
+
for k, v in msg.items():
|
52 |
+
tmp_str = ''
|
53 |
+
if isinstance(v, int):
|
54 |
+
tmp_str = '{}: {:,}'.format(k, v)
|
55 |
+
else:
|
56 |
+
tmp_str = '{}: {}'.format(k, v)
|
57 |
+
|
58 |
+
msg_list.append(tmp_str)
|
59 |
+
msg_str = '\n'.join(msg_list)
|
60 |
+
else:
|
61 |
+
msg_str = msg
|
62 |
+
|
63 |
+
# dsplay
|
64 |
+
print(msg_str)
|
65 |
+
|
66 |
+
# save
|
67 |
+
with open(self.path_log_info, 'a') as fp:
|
68 |
+
fp.write(msg_str+'\n')
|
69 |
+
|
70 |
+
def log_value(self, dict):
|
71 |
+
for k, v in dict.items():
|
72 |
+
self.writer.add_scalar(k, v, self.global_step)
|
73 |
+
|
74 |
+
def log_spec(self, name, spec, spec_out, vmin=-14, vmax=3.5):
|
75 |
+
spec_cat = torch.cat([(spec_out - spec).abs() + vmin, spec, spec_out], -1)
|
76 |
+
spec = spec_cat[0]
|
77 |
+
if isinstance(spec, torch.Tensor):
|
78 |
+
spec = spec.cpu().numpy()
|
79 |
+
fig = plt.figure(figsize=(12, 9))
|
80 |
+
plt.pcolor(spec.T, vmin=vmin, vmax=vmax)
|
81 |
+
plt.tight_layout()
|
82 |
+
self.writer.add_figure(name, fig, self.global_step)
|
83 |
+
|
84 |
+
def log_audio(self, dict):
|
85 |
+
for k, v in dict.items():
|
86 |
+
self.writer.add_audio(k, v, global_step=self.global_step, sample_rate=self.sample_rate)
|
87 |
+
|
88 |
+
def get_interval_time(self, update=True):
|
89 |
+
cur_time = time.time()
|
90 |
+
time_interval = cur_time - self.last_time
|
91 |
+
if update:
|
92 |
+
self.last_time = cur_time
|
93 |
+
return time_interval
|
94 |
+
|
95 |
+
def get_total_time(self, to_str=True):
|
96 |
+
total_time = time.time() - self.init_time
|
97 |
+
if to_str:
|
98 |
+
total_time = str(datetime.timedelta(
|
99 |
+
seconds=total_time))[:-5]
|
100 |
+
return total_time
|
101 |
+
|
102 |
+
def save_model(
|
103 |
+
self,
|
104 |
+
model,
|
105 |
+
optimizer,
|
106 |
+
name='model',
|
107 |
+
postfix='',
|
108 |
+
to_json=False):
|
109 |
+
# path
|
110 |
+
if postfix:
|
111 |
+
postfix = '_' + postfix
|
112 |
+
path_pt = os.path.join(
|
113 |
+
self.expdir , name+postfix+'.pt')
|
114 |
+
|
115 |
+
# check
|
116 |
+
print(' [*] model checkpoint saved: {}'.format(path_pt))
|
117 |
+
|
118 |
+
# save
|
119 |
+
if optimizer is not None:
|
120 |
+
torch.save({
|
121 |
+
'global_step': self.global_step,
|
122 |
+
'model': model.state_dict(),
|
123 |
+
'optimizer': optimizer.state_dict()}, path_pt)
|
124 |
+
else:
|
125 |
+
torch.save({
|
126 |
+
'global_step': self.global_step,
|
127 |
+
'model': model.state_dict()}, path_pt)
|
128 |
+
|
129 |
+
# to json
|
130 |
+
if to_json:
|
131 |
+
path_json = os.path.join(
|
132 |
+
self.expdir , name+'.json')
|
133 |
+
utils.to_json(path_params, path_json)
|
134 |
+
|
135 |
+
def delete_model(self, name='model', postfix=''):
|
136 |
+
# path
|
137 |
+
if postfix:
|
138 |
+
postfix = '_' + postfix
|
139 |
+
path_pt = os.path.join(
|
140 |
+
self.expdir , name+postfix+'.pt')
|
141 |
+
|
142 |
+
# delete
|
143 |
+
if os.path.exists(path_pt):
|
144 |
+
os.remove(path_pt)
|
145 |
+
print(' [*] model checkpoint deleted: {}'.format(path_pt))
|
146 |
+
|
147 |
+
def global_step_increment(self):
|
148 |
+
self.global_step += 1
|
149 |
+
|
150 |
+
|
diffusion/logger/utils.py
ADDED
@@ -0,0 +1,126 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
import yaml
|
3 |
+
import json
|
4 |
+
import pickle
|
5 |
+
import torch
|
6 |
+
|
7 |
+
def traverse_dir(
|
8 |
+
root_dir,
|
9 |
+
extensions,
|
10 |
+
amount=None,
|
11 |
+
str_include=None,
|
12 |
+
str_exclude=None,
|
13 |
+
is_pure=False,
|
14 |
+
is_sort=False,
|
15 |
+
is_ext=True):
|
16 |
+
|
17 |
+
file_list = []
|
18 |
+
cnt = 0
|
19 |
+
for root, _, files in os.walk(root_dir):
|
20 |
+
for file in files:
|
21 |
+
if any([file.endswith(f".{ext}") for ext in extensions]):
|
22 |
+
# path
|
23 |
+
mix_path = os.path.join(root, file)
|
24 |
+
pure_path = mix_path[len(root_dir)+1:] if is_pure else mix_path
|
25 |
+
|
26 |
+
# amount
|
27 |
+
if (amount is not None) and (cnt == amount):
|
28 |
+
if is_sort:
|
29 |
+
file_list.sort()
|
30 |
+
return file_list
|
31 |
+
|
32 |
+
# check string
|
33 |
+
if (str_include is not None) and (str_include not in pure_path):
|
34 |
+
continue
|
35 |
+
if (str_exclude is not None) and (str_exclude in pure_path):
|
36 |
+
continue
|
37 |
+
|
38 |
+
if not is_ext:
|
39 |
+
ext = pure_path.split('.')[-1]
|
40 |
+
pure_path = pure_path[:-(len(ext)+1)]
|
41 |
+
file_list.append(pure_path)
|
42 |
+
cnt += 1
|
43 |
+
if is_sort:
|
44 |
+
file_list.sort()
|
45 |
+
return file_list
|
46 |
+
|
47 |
+
|
48 |
+
|
49 |
+
class DotDict(dict):
|
50 |
+
def __getattr__(*args):
|
51 |
+
val = dict.get(*args)
|
52 |
+
return DotDict(val) if type(val) is dict else val
|
53 |
+
|
54 |
+
__setattr__ = dict.__setitem__
|
55 |
+
__delattr__ = dict.__delitem__
|
56 |
+
|
57 |
+
|
58 |
+
def get_network_paras_amount(model_dict):
|
59 |
+
info = dict()
|
60 |
+
for model_name, model in model_dict.items():
|
61 |
+
# all_params = sum(p.numel() for p in model.parameters())
|
62 |
+
trainable_params = sum(p.numel() for p in model.parameters() if p.requires_grad)
|
63 |
+
|
64 |
+
info[model_name] = trainable_params
|
65 |
+
return info
|
66 |
+
|
67 |
+
|
68 |
+
def load_config(path_config):
|
69 |
+
with open(path_config, "r") as config:
|
70 |
+
args = yaml.safe_load(config)
|
71 |
+
args = DotDict(args)
|
72 |
+
# print(args)
|
73 |
+
return args
|
74 |
+
|
75 |
+
def save_config(path_config,config):
|
76 |
+
config = dict(config)
|
77 |
+
with open(path_config, "w") as f:
|
78 |
+
yaml.dump(config, f)
|
79 |
+
|
80 |
+
def to_json(path_params, path_json):
|
81 |
+
params = torch.load(path_params, map_location=torch.device('cpu'))
|
82 |
+
raw_state_dict = {}
|
83 |
+
for k, v in params.items():
|
84 |
+
val = v.flatten().numpy().tolist()
|
85 |
+
raw_state_dict[k] = val
|
86 |
+
|
87 |
+
with open(path_json, 'w') as outfile:
|
88 |
+
json.dump(raw_state_dict, outfile,indent= "\t")
|
89 |
+
|
90 |
+
|
91 |
+
def convert_tensor_to_numpy(tensor, is_squeeze=True):
|
92 |
+
if is_squeeze:
|
93 |
+
tensor = tensor.squeeze()
|
94 |
+
if tensor.requires_grad:
|
95 |
+
tensor = tensor.detach()
|
96 |
+
if tensor.is_cuda:
|
97 |
+
tensor = tensor.cpu()
|
98 |
+
return tensor.numpy()
|
99 |
+
|
100 |
+
|
101 |
+
def load_model(
|
102 |
+
expdir,
|
103 |
+
model,
|
104 |
+
optimizer,
|
105 |
+
name='model',
|
106 |
+
postfix='',
|
107 |
+
device='cpu'):
|
108 |
+
if postfix == '':
|
109 |
+
postfix = '_' + postfix
|
110 |
+
path = os.path.join(expdir, name+postfix)
|
111 |
+
path_pt = traverse_dir(expdir, ['pt'], is_ext=False)
|
112 |
+
global_step = 0
|
113 |
+
if len(path_pt) > 0:
|
114 |
+
steps = [s[len(path):] for s in path_pt]
|
115 |
+
maxstep = max([int(s) if s.isdigit() else 0 for s in steps])
|
116 |
+
if maxstep >= 0:
|
117 |
+
path_pt = path+str(maxstep)+'.pt'
|
118 |
+
else:
|
119 |
+
path_pt = path+'best.pt'
|
120 |
+
print(' [*] restoring model from', path_pt)
|
121 |
+
ckpt = torch.load(path_pt, map_location=torch.device(device))
|
122 |
+
global_step = ckpt['global_step']
|
123 |
+
model.load_state_dict(ckpt['model'], strict=False)
|
124 |
+
if ckpt.get('optimizer') != None:
|
125 |
+
optimizer.load_state_dict(ckpt['optimizer'])
|
126 |
+
return global_step, model, optimizer
|
diffusion/onnx_export.py
ADDED
@@ -0,0 +1,226 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from diffusion_onnx import GaussianDiffusion
|
2 |
+
import os
|
3 |
+
import yaml
|
4 |
+
import torch
|
5 |
+
import torch.nn as nn
|
6 |
+
import numpy as np
|
7 |
+
from wavenet import WaveNet
|
8 |
+
import torch.nn.functional as F
|
9 |
+
import diffusion
|
10 |
+
|
11 |
+
class DotDict(dict):
|
12 |
+
def __getattr__(*args):
|
13 |
+
val = dict.get(*args)
|
14 |
+
return DotDict(val) if type(val) is dict else val
|
15 |
+
|
16 |
+
__setattr__ = dict.__setitem__
|
17 |
+
__delattr__ = dict.__delitem__
|
18 |
+
|
19 |
+
|
20 |
+
def load_model_vocoder(
|
21 |
+
model_path,
|
22 |
+
device='cpu'):
|
23 |
+
config_file = os.path.join(os.path.split(model_path)[0], 'config.yaml')
|
24 |
+
with open(config_file, "r") as config:
|
25 |
+
args = yaml.safe_load(config)
|
26 |
+
args = DotDict(args)
|
27 |
+
|
28 |
+
# load model
|
29 |
+
model = Unit2Mel(
|
30 |
+
args.data.encoder_out_channels,
|
31 |
+
args.model.n_spk,
|
32 |
+
args.model.use_pitch_aug,
|
33 |
+
128,
|
34 |
+
args.model.n_layers,
|
35 |
+
args.model.n_chans,
|
36 |
+
args.model.n_hidden)
|
37 |
+
|
38 |
+
print(' [Loading] ' + model_path)
|
39 |
+
ckpt = torch.load(model_path, map_location=torch.device(device))
|
40 |
+
model.to(device)
|
41 |
+
model.load_state_dict(ckpt['model'])
|
42 |
+
model.eval()
|
43 |
+
return model, args
|
44 |
+
|
45 |
+
|
46 |
+
class Unit2Mel(nn.Module):
|
47 |
+
def __init__(
|
48 |
+
self,
|
49 |
+
input_channel,
|
50 |
+
n_spk,
|
51 |
+
use_pitch_aug=False,
|
52 |
+
out_dims=128,
|
53 |
+
n_layers=20,
|
54 |
+
n_chans=384,
|
55 |
+
n_hidden=256):
|
56 |
+
super().__init__()
|
57 |
+
self.unit_embed = nn.Linear(input_channel, n_hidden)
|
58 |
+
self.f0_embed = nn.Linear(1, n_hidden)
|
59 |
+
self.volume_embed = nn.Linear(1, n_hidden)
|
60 |
+
if use_pitch_aug:
|
61 |
+
self.aug_shift_embed = nn.Linear(1, n_hidden, bias=False)
|
62 |
+
else:
|
63 |
+
self.aug_shift_embed = None
|
64 |
+
self.n_spk = n_spk
|
65 |
+
if n_spk is not None and n_spk > 1:
|
66 |
+
self.spk_embed = nn.Embedding(n_spk, n_hidden)
|
67 |
+
|
68 |
+
# diffusion
|
69 |
+
self.decoder = GaussianDiffusion(out_dims, n_layers, n_chans, n_hidden)
|
70 |
+
self.hidden_size = n_hidden
|
71 |
+
self.speaker_map = torch.zeros((self.n_spk,1,1,n_hidden))
|
72 |
+
|
73 |
+
|
74 |
+
|
75 |
+
def forward(self, units, mel2ph, f0, volume, g = None):
|
76 |
+
|
77 |
+
'''
|
78 |
+
input:
|
79 |
+
B x n_frames x n_unit
|
80 |
+
return:
|
81 |
+
dict of B x n_frames x feat
|
82 |
+
'''
|
83 |
+
|
84 |
+
decoder_inp = F.pad(units, [0, 0, 1, 0])
|
85 |
+
mel2ph_ = mel2ph.unsqueeze(2).repeat([1, 1, units.shape[-1]])
|
86 |
+
units = torch.gather(decoder_inp, 1, mel2ph_) # [B, T, H]
|
87 |
+
|
88 |
+
x = self.unit_embed(units) + self.f0_embed((1 + f0.unsqueeze(-1) / 700).log()) + self.volume_embed(volume.unsqueeze(-1))
|
89 |
+
|
90 |
+
if self.n_spk is not None and self.n_spk > 1: # [N, S] * [S, B, 1, H]
|
91 |
+
g = g.reshape((g.shape[0], g.shape[1], 1, 1, 1)) # [N, S, B, 1, 1]
|
92 |
+
g = g * self.speaker_map # [N, S, B, 1, H]
|
93 |
+
g = torch.sum(g, dim=1) # [N, 1, B, 1, H]
|
94 |
+
g = g.transpose(0, -1).transpose(0, -2).squeeze(0) # [B, H, N]
|
95 |
+
x = x.transpose(1, 2) + g
|
96 |
+
return x
|
97 |
+
else:
|
98 |
+
return x.transpose(1, 2)
|
99 |
+
|
100 |
+
|
101 |
+
def init_spkembed(self, units, f0, volume, spk_id = None, spk_mix_dict = None, aug_shift = None,
|
102 |
+
gt_spec=None, infer=True, infer_speedup=10, method='dpm-solver', k_step=300, use_tqdm=True):
|
103 |
+
|
104 |
+
'''
|
105 |
+
input:
|
106 |
+
B x n_frames x n_unit
|
107 |
+
return:
|
108 |
+
dict of B x n_frames x feat
|
109 |
+
'''
|
110 |
+
x = self.unit_embed(units) + self.f0_embed((1+ f0 / 700).log()) + self.volume_embed(volume)
|
111 |
+
if self.n_spk is not None and self.n_spk > 1:
|
112 |
+
if spk_mix_dict is not None:
|
113 |
+
spk_embed_mix = torch.zeros((1,1,self.hidden_size))
|
114 |
+
for k, v in spk_mix_dict.items():
|
115 |
+
spk_id_torch = torch.LongTensor(np.array([[k]])).to(units.device)
|
116 |
+
spk_embeddd = self.spk_embed(spk_id_torch)
|
117 |
+
self.speaker_map[k] = spk_embeddd
|
118 |
+
spk_embed_mix = spk_embed_mix + v * spk_embeddd
|
119 |
+
x = x + spk_embed_mix
|
120 |
+
else:
|
121 |
+
x = x + self.spk_embed(spk_id - 1)
|
122 |
+
self.speaker_map = self.speaker_map.unsqueeze(0)
|
123 |
+
self.speaker_map = self.speaker_map.detach()
|
124 |
+
return x.transpose(1, 2)
|
125 |
+
|
126 |
+
def OnnxExport(self, project_name=None, init_noise=None, export_encoder=True, export_denoise=True, export_pred=True, export_after=True):
|
127 |
+
hubert_hidden_size = 768
|
128 |
+
n_frames = 100
|
129 |
+
hubert = torch.randn((1, n_frames, hubert_hidden_size))
|
130 |
+
mel2ph = torch.arange(end=n_frames).unsqueeze(0).long()
|
131 |
+
f0 = torch.randn((1, n_frames))
|
132 |
+
volume = torch.randn((1, n_frames))
|
133 |
+
spk_mix = []
|
134 |
+
spks = {}
|
135 |
+
if self.n_spk is not None and self.n_spk > 1:
|
136 |
+
for i in range(self.n_spk):
|
137 |
+
spk_mix.append(1.0/float(self.n_spk))
|
138 |
+
spks.update({i:1.0/float(self.n_spk)})
|
139 |
+
spk_mix = torch.tensor(spk_mix)
|
140 |
+
spk_mix = spk_mix.repeat(n_frames, 1)
|
141 |
+
orgouttt = self.init_spkembed(hubert, f0.unsqueeze(-1), volume.unsqueeze(-1), spk_mix_dict=spks)
|
142 |
+
outtt = self.forward(hubert, mel2ph, f0, volume, spk_mix)
|
143 |
+
if export_encoder:
|
144 |
+
torch.onnx.export(
|
145 |
+
self,
|
146 |
+
(hubert, mel2ph, f0, volume, spk_mix),
|
147 |
+
f"{project_name}_encoder.onnx",
|
148 |
+
input_names=["hubert", "mel2ph", "f0", "volume", "spk_mix"],
|
149 |
+
output_names=["mel_pred"],
|
150 |
+
dynamic_axes={
|
151 |
+
"hubert": [1],
|
152 |
+
"f0": [1],
|
153 |
+
"volume": [1],
|
154 |
+
"mel2ph": [1],
|
155 |
+
"spk_mix": [0],
|
156 |
+
},
|
157 |
+
opset_version=16
|
158 |
+
)
|
159 |
+
|
160 |
+
self.decoder.OnnxExport(project_name, init_noise=init_noise, export_denoise=export_denoise, export_pred=export_pred, export_after=export_after)
|
161 |
+
|
162 |
+
def ExportOnnx(self, project_name=None):
|
163 |
+
hubert_hidden_size = 768
|
164 |
+
n_frames = 100
|
165 |
+
hubert = torch.randn((1, n_frames, hubert_hidden_size))
|
166 |
+
mel2ph = torch.arange(end=n_frames).unsqueeze(0).long()
|
167 |
+
f0 = torch.randn((1, n_frames))
|
168 |
+
volume = torch.randn((1, n_frames))
|
169 |
+
spk_mix = []
|
170 |
+
spks = {}
|
171 |
+
if self.n_spk is not None and self.n_spk > 1:
|
172 |
+
for i in range(self.n_spk):
|
173 |
+
spk_mix.append(1.0/float(self.n_spk))
|
174 |
+
spks.update({i:1.0/float(self.n_spk)})
|
175 |
+
spk_mix = torch.tensor(spk_mix)
|
176 |
+
orgouttt = self.orgforward(hubert, f0.unsqueeze(-1), volume.unsqueeze(-1), spk_mix_dict=spks)
|
177 |
+
outtt = self.forward(hubert, mel2ph, f0, volume, spk_mix)
|
178 |
+
|
179 |
+
torch.onnx.export(
|
180 |
+
self,
|
181 |
+
(hubert, mel2ph, f0, volume, spk_mix),
|
182 |
+
f"{project_name}_encoder.onnx",
|
183 |
+
input_names=["hubert", "mel2ph", "f0", "volume", "spk_mix"],
|
184 |
+
output_names=["mel_pred"],
|
185 |
+
dynamic_axes={
|
186 |
+
"hubert": [1],
|
187 |
+
"f0": [1],
|
188 |
+
"volume": [1],
|
189 |
+
"mel2ph": [1]
|
190 |
+
},
|
191 |
+
opset_version=16
|
192 |
+
)
|
193 |
+
|
194 |
+
condition = torch.randn(1,self.decoder.n_hidden,n_frames)
|
195 |
+
noise = torch.randn((1, 1, self.decoder.mel_bins, condition.shape[2]), dtype=torch.float32)
|
196 |
+
pndm_speedup = torch.LongTensor([100])
|
197 |
+
K_steps = torch.LongTensor([1000])
|
198 |
+
self.decoder = torch.jit.script(self.decoder)
|
199 |
+
self.decoder(condition, noise, pndm_speedup, K_steps)
|
200 |
+
|
201 |
+
torch.onnx.export(
|
202 |
+
self.decoder,
|
203 |
+
(condition, noise, pndm_speedup, K_steps),
|
204 |
+
f"{project_name}_diffusion.onnx",
|
205 |
+
input_names=["condition", "noise", "pndm_speedup", "K_steps"],
|
206 |
+
output_names=["mel"],
|
207 |
+
dynamic_axes={
|
208 |
+
"condition": [2],
|
209 |
+
"noise": [3],
|
210 |
+
},
|
211 |
+
opset_version=16
|
212 |
+
)
|
213 |
+
|
214 |
+
|
215 |
+
if __name__ == "__main__":
|
216 |
+
project_name = "dddsp"
|
217 |
+
model_path = f'{project_name}/model_500000.pt'
|
218 |
+
|
219 |
+
model, _ = load_model_vocoder(model_path)
|
220 |
+
|
221 |
+
# 分开Diffusion导出(需要使用MoeSS/MoeVoiceStudio或者自己编写Pndm/Dpm采样)
|
222 |
+
model.OnnxExport(project_name, export_encoder=True, export_denoise=True, export_pred=True, export_after=True)
|
223 |
+
|
224 |
+
# 合并Diffusion导出(Encoder和Diffusion分开,直接将Encoder的结果和初始噪声输入Diffusion即可)
|
225 |
+
# model.ExportOnnx(project_name)
|
226 |
+
|
diffusion/solver.py
ADDED
@@ -0,0 +1,195 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
import time
|
3 |
+
import numpy as np
|
4 |
+
import torch
|
5 |
+
import librosa
|
6 |
+
from diffusion.logger.saver import Saver
|
7 |
+
from diffusion.logger import utils
|
8 |
+
from torch import autocast
|
9 |
+
from torch.cuda.amp import GradScaler
|
10 |
+
|
11 |
+
def test(args, model, vocoder, loader_test, saver):
|
12 |
+
print(' [*] testing...')
|
13 |
+
model.eval()
|
14 |
+
|
15 |
+
# losses
|
16 |
+
test_loss = 0.
|
17 |
+
|
18 |
+
# intialization
|
19 |
+
num_batches = len(loader_test)
|
20 |
+
rtf_all = []
|
21 |
+
|
22 |
+
# run
|
23 |
+
with torch.no_grad():
|
24 |
+
for bidx, data in enumerate(loader_test):
|
25 |
+
fn = data['name'][0].split("/")[-1]
|
26 |
+
speaker = data['name'][0].split("/")[-2]
|
27 |
+
print('--------')
|
28 |
+
print('{}/{} - {}'.format(bidx, num_batches, fn))
|
29 |
+
|
30 |
+
# unpack data
|
31 |
+
for k in data.keys():
|
32 |
+
if not k.startswith('name'):
|
33 |
+
data[k] = data[k].to(args.device)
|
34 |
+
print('>>', data['name'][0])
|
35 |
+
|
36 |
+
# forward
|
37 |
+
st_time = time.time()
|
38 |
+
mel = model(
|
39 |
+
data['units'],
|
40 |
+
data['f0'],
|
41 |
+
data['volume'],
|
42 |
+
data['spk_id'],
|
43 |
+
gt_spec=None,
|
44 |
+
infer=True,
|
45 |
+
infer_speedup=args.infer.speedup,
|
46 |
+
method=args.infer.method)
|
47 |
+
signal = vocoder.infer(mel, data['f0'])
|
48 |
+
ed_time = time.time()
|
49 |
+
|
50 |
+
# RTF
|
51 |
+
run_time = ed_time - st_time
|
52 |
+
song_time = signal.shape[-1] / args.data.sampling_rate
|
53 |
+
rtf = run_time / song_time
|
54 |
+
print('RTF: {} | {} / {}'.format(rtf, run_time, song_time))
|
55 |
+
rtf_all.append(rtf)
|
56 |
+
|
57 |
+
# loss
|
58 |
+
for i in range(args.train.batch_size):
|
59 |
+
loss = model(
|
60 |
+
data['units'],
|
61 |
+
data['f0'],
|
62 |
+
data['volume'],
|
63 |
+
data['spk_id'],
|
64 |
+
gt_spec=data['mel'],
|
65 |
+
infer=False)
|
66 |
+
test_loss += loss.item()
|
67 |
+
|
68 |
+
# log mel
|
69 |
+
saver.log_spec(f"{speaker}_{fn}.wav", data['mel'], mel)
|
70 |
+
|
71 |
+
# log audi
|
72 |
+
path_audio = data['name_ext'][0]
|
73 |
+
audio, sr = librosa.load(path_audio, sr=args.data.sampling_rate)
|
74 |
+
if len(audio.shape) > 1:
|
75 |
+
audio = librosa.to_mono(audio)
|
76 |
+
audio = torch.from_numpy(audio).unsqueeze(0).to(signal)
|
77 |
+
saver.log_audio({f"{speaker}_{fn}_gt.wav": audio,f"{speaker}_{fn}_pred.wav": signal})
|
78 |
+
# report
|
79 |
+
test_loss /= args.train.batch_size
|
80 |
+
test_loss /= num_batches
|
81 |
+
|
82 |
+
# check
|
83 |
+
print(' [test_loss] test_loss:', test_loss)
|
84 |
+
print(' Real Time Factor', np.mean(rtf_all))
|
85 |
+
return test_loss
|
86 |
+
|
87 |
+
|
88 |
+
def train(args, initial_global_step, model, optimizer, scheduler, vocoder, loader_train, loader_test):
|
89 |
+
# saver
|
90 |
+
saver = Saver(args, initial_global_step=initial_global_step)
|
91 |
+
|
92 |
+
# model size
|
93 |
+
params_count = utils.get_network_paras_amount({'model': model})
|
94 |
+
saver.log_info('--- model size ---')
|
95 |
+
saver.log_info(params_count)
|
96 |
+
|
97 |
+
# run
|
98 |
+
num_batches = len(loader_train)
|
99 |
+
model.train()
|
100 |
+
saver.log_info('======= start training =======')
|
101 |
+
scaler = GradScaler()
|
102 |
+
if args.train.amp_dtype == 'fp32':
|
103 |
+
dtype = torch.float32
|
104 |
+
elif args.train.amp_dtype == 'fp16':
|
105 |
+
dtype = torch.float16
|
106 |
+
elif args.train.amp_dtype == 'bf16':
|
107 |
+
dtype = torch.bfloat16
|
108 |
+
else:
|
109 |
+
raise ValueError(' [x] Unknown amp_dtype: ' + args.train.amp_dtype)
|
110 |
+
saver.log_info("epoch|batch_idx/num_batches|output_dir|batch/s|lr|time|step")
|
111 |
+
for epoch in range(args.train.epochs):
|
112 |
+
for batch_idx, data in enumerate(loader_train):
|
113 |
+
saver.global_step_increment()
|
114 |
+
optimizer.zero_grad()
|
115 |
+
|
116 |
+
# unpack data
|
117 |
+
for k in data.keys():
|
118 |
+
if not k.startswith('name'):
|
119 |
+
data[k] = data[k].to(args.device)
|
120 |
+
|
121 |
+
# forward
|
122 |
+
if dtype == torch.float32:
|
123 |
+
loss = model(data['units'].float(), data['f0'], data['volume'], data['spk_id'],
|
124 |
+
aug_shift = data['aug_shift'], gt_spec=data['mel'].float(), infer=False)
|
125 |
+
else:
|
126 |
+
with autocast(device_type=args.device, dtype=dtype):
|
127 |
+
loss = model(data['units'], data['f0'], data['volume'], data['spk_id'],
|
128 |
+
aug_shift = data['aug_shift'], gt_spec=data['mel'], infer=False)
|
129 |
+
|
130 |
+
# handle nan loss
|
131 |
+
if torch.isnan(loss):
|
132 |
+
raise ValueError(' [x] nan loss ')
|
133 |
+
else:
|
134 |
+
# backpropagate
|
135 |
+
if dtype == torch.float32:
|
136 |
+
loss.backward()
|
137 |
+
optimizer.step()
|
138 |
+
else:
|
139 |
+
scaler.scale(loss).backward()
|
140 |
+
scaler.step(optimizer)
|
141 |
+
scaler.update()
|
142 |
+
scheduler.step()
|
143 |
+
|
144 |
+
# log loss
|
145 |
+
if saver.global_step % args.train.interval_log == 0:
|
146 |
+
current_lr = optimizer.param_groups[0]['lr']
|
147 |
+
saver.log_info(
|
148 |
+
'epoch: {} | {:3d}/{:3d} | {} | batch/s: {:.2f} | lr: {:.6} | loss: {:.3f} | time: {} | step: {}'.format(
|
149 |
+
epoch,
|
150 |
+
batch_idx,
|
151 |
+
num_batches,
|
152 |
+
args.env.expdir,
|
153 |
+
args.train.interval_log/saver.get_interval_time(),
|
154 |
+
current_lr,
|
155 |
+
loss.item(),
|
156 |
+
saver.get_total_time(),
|
157 |
+
saver.global_step
|
158 |
+
)
|
159 |
+
)
|
160 |
+
|
161 |
+
saver.log_value({
|
162 |
+
'train/loss': loss.item()
|
163 |
+
})
|
164 |
+
|
165 |
+
saver.log_value({
|
166 |
+
'train/lr': current_lr
|
167 |
+
})
|
168 |
+
|
169 |
+
# validation
|
170 |
+
if saver.global_step % args.train.interval_val == 0:
|
171 |
+
optimizer_save = optimizer if args.train.save_opt else None
|
172 |
+
|
173 |
+
# save latest
|
174 |
+
saver.save_model(model, optimizer_save, postfix=f'{saver.global_step}')
|
175 |
+
last_val_step = saver.global_step - args.train.interval_val
|
176 |
+
if last_val_step % args.train.interval_force_save != 0:
|
177 |
+
saver.delete_model(postfix=f'{last_val_step}')
|
178 |
+
|
179 |
+
# run testing set
|
180 |
+
test_loss = test(args, model, vocoder, loader_test, saver)
|
181 |
+
|
182 |
+
# log loss
|
183 |
+
saver.log_info(
|
184 |
+
' --- <validation> --- \nloss: {:.3f}. '.format(
|
185 |
+
test_loss,
|
186 |
+
)
|
187 |
+
)
|
188 |
+
|
189 |
+
saver.log_value({
|
190 |
+
'validation/loss': test_loss
|
191 |
+
})
|
192 |
+
|
193 |
+
model.train()
|
194 |
+
|
195 |
+
|
diffusion/unit2mel.py
ADDED
@@ -0,0 +1,147 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
import yaml
|
3 |
+
import torch
|
4 |
+
import torch.nn as nn
|
5 |
+
import numpy as np
|
6 |
+
from .diffusion import GaussianDiffusion
|
7 |
+
from .wavenet import WaveNet
|
8 |
+
from .vocoder import Vocoder
|
9 |
+
|
10 |
+
class DotDict(dict):
|
11 |
+
def __getattr__(*args):
|
12 |
+
val = dict.get(*args)
|
13 |
+
return DotDict(val) if type(val) is dict else val
|
14 |
+
|
15 |
+
__setattr__ = dict.__setitem__
|
16 |
+
__delattr__ = dict.__delitem__
|
17 |
+
|
18 |
+
|
19 |
+
def load_model_vocoder(
|
20 |
+
model_path,
|
21 |
+
device='cpu',
|
22 |
+
config_path = None
|
23 |
+
):
|
24 |
+
if config_path is None: config_file = os.path.join(os.path.split(model_path)[0], 'config.yaml')
|
25 |
+
else: config_file = config_path
|
26 |
+
|
27 |
+
with open(config_file, "r") as config:
|
28 |
+
args = yaml.safe_load(config)
|
29 |
+
args = DotDict(args)
|
30 |
+
|
31 |
+
# load vocoder
|
32 |
+
vocoder = Vocoder(args.vocoder.type, args.vocoder.ckpt, device=device)
|
33 |
+
|
34 |
+
# load model
|
35 |
+
model = Unit2Mel(
|
36 |
+
args.data.encoder_out_channels,
|
37 |
+
args.model.n_spk,
|
38 |
+
args.model.use_pitch_aug,
|
39 |
+
vocoder.dimension,
|
40 |
+
args.model.n_layers,
|
41 |
+
args.model.n_chans,
|
42 |
+
args.model.n_hidden)
|
43 |
+
|
44 |
+
print(' [Loading] ' + model_path)
|
45 |
+
ckpt = torch.load(model_path, map_location=torch.device(device))
|
46 |
+
model.to(device)
|
47 |
+
model.load_state_dict(ckpt['model'])
|
48 |
+
model.eval()
|
49 |
+
return model, vocoder, args
|
50 |
+
|
51 |
+
|
52 |
+
class Unit2Mel(nn.Module):
|
53 |
+
def __init__(
|
54 |
+
self,
|
55 |
+
input_channel,
|
56 |
+
n_spk,
|
57 |
+
use_pitch_aug=False,
|
58 |
+
out_dims=128,
|
59 |
+
n_layers=20,
|
60 |
+
n_chans=384,
|
61 |
+
n_hidden=256):
|
62 |
+
super().__init__()
|
63 |
+
self.unit_embed = nn.Linear(input_channel, n_hidden)
|
64 |
+
self.f0_embed = nn.Linear(1, n_hidden)
|
65 |
+
self.volume_embed = nn.Linear(1, n_hidden)
|
66 |
+
if use_pitch_aug:
|
67 |
+
self.aug_shift_embed = nn.Linear(1, n_hidden, bias=False)
|
68 |
+
else:
|
69 |
+
self.aug_shift_embed = None
|
70 |
+
self.n_spk = n_spk
|
71 |
+
if n_spk is not None and n_spk > 1:
|
72 |
+
self.spk_embed = nn.Embedding(n_spk, n_hidden)
|
73 |
+
|
74 |
+
self.n_hidden = n_hidden
|
75 |
+
# diffusion
|
76 |
+
self.decoder = GaussianDiffusion(WaveNet(out_dims, n_layers, n_chans, n_hidden), out_dims=out_dims)
|
77 |
+
self.input_channel = input_channel
|
78 |
+
|
79 |
+
def init_spkembed(self, units, f0, volume, spk_id = None, spk_mix_dict = None, aug_shift = None,
|
80 |
+
gt_spec=None, infer=True, infer_speedup=10, method='dpm-solver', k_step=300, use_tqdm=True):
|
81 |
+
|
82 |
+
'''
|
83 |
+
input:
|
84 |
+
B x n_frames x n_unit
|
85 |
+
return:
|
86 |
+
dict of B x n_frames x feat
|
87 |
+
'''
|
88 |
+
x = self.unit_embed(units) + self.f0_embed((1+ f0 / 700).log()) + self.volume_embed(volume)
|
89 |
+
if self.n_spk is not None and self.n_spk > 1:
|
90 |
+
if spk_mix_dict is not None:
|
91 |
+
spk_embed_mix = torch.zeros((1,1,self.hidden_size))
|
92 |
+
for k, v in spk_mix_dict.items():
|
93 |
+
spk_id_torch = torch.LongTensor(np.array([[k]])).to(units.device)
|
94 |
+
spk_embeddd = self.spk_embed(spk_id_torch)
|
95 |
+
self.speaker_map[k] = spk_embeddd
|
96 |
+
spk_embed_mix = spk_embed_mix + v * spk_embeddd
|
97 |
+
x = x + spk_embed_mix
|
98 |
+
else:
|
99 |
+
x = x + self.spk_embed(spk_id - 1)
|
100 |
+
self.speaker_map = self.speaker_map.unsqueeze(0)
|
101 |
+
self.speaker_map = self.speaker_map.detach()
|
102 |
+
return x.transpose(1, 2)
|
103 |
+
|
104 |
+
def init_spkmix(self, n_spk):
|
105 |
+
self.speaker_map = torch.zeros((n_spk,1,1,self.n_hidden))
|
106 |
+
hubert_hidden_size = self.input_channel
|
107 |
+
n_frames = 10
|
108 |
+
hubert = torch.randn((1, n_frames, hubert_hidden_size))
|
109 |
+
mel2ph = torch.arange(end=n_frames).unsqueeze(0).long()
|
110 |
+
f0 = torch.randn((1, n_frames))
|
111 |
+
volume = torch.randn((1, n_frames))
|
112 |
+
spks = {}
|
113 |
+
for i in range(n_spk):
|
114 |
+
spks.update({i:1.0/float(self.n_spk)})
|
115 |
+
orgouttt = self.init_spkembed(hubert, f0.unsqueeze(-1), volume.unsqueeze(-1), spk_mix_dict=spks)
|
116 |
+
|
117 |
+
def forward(self, units, f0, volume, spk_id = None, spk_mix_dict = None, aug_shift = None,
|
118 |
+
gt_spec=None, infer=True, infer_speedup=10, method='dpm-solver', k_step=300, use_tqdm=True):
|
119 |
+
|
120 |
+
'''
|
121 |
+
input:
|
122 |
+
B x n_frames x n_unit
|
123 |
+
return:
|
124 |
+
dict of B x n_frames x feat
|
125 |
+
'''
|
126 |
+
|
127 |
+
x = self.unit_embed(units) + self.f0_embed((1+ f0 / 700).log()) + self.volume_embed(volume)
|
128 |
+
if self.n_spk is not None and self.n_spk > 1:
|
129 |
+
if spk_mix_dict is not None:
|
130 |
+
for k, v in spk_mix_dict.items():
|
131 |
+
spk_id_torch = torch.LongTensor(np.array([[k]])).to(units.device)
|
132 |
+
x = x + v * self.spk_embed(spk_id_torch)
|
133 |
+
else:
|
134 |
+
if spk_id.shape[1] > 1:
|
135 |
+
g = spk_id.reshape((spk_id.shape[0], spk_id.shape[1], 1, 1, 1)) # [N, S, B, 1, 1]
|
136 |
+
g = g * self.speaker_map # [N, S, B, 1, H]
|
137 |
+
g = torch.sum(g, dim=1) # [N, 1, B, 1, H]
|
138 |
+
g = g.transpose(0, -1).transpose(0, -2).squeeze(0) # [B, H, N]
|
139 |
+
x = x + g
|
140 |
+
else:
|
141 |
+
x = x + self.spk_embed(spk_id)
|
142 |
+
if self.aug_shift_embed is not None and aug_shift is not None:
|
143 |
+
x = x + self.aug_shift_embed(aug_shift / 5)
|
144 |
+
x = self.decoder(x, gt_spec=gt_spec, infer=infer, infer_speedup=infer_speedup, method=method, k_step=k_step, use_tqdm=use_tqdm)
|
145 |
+
|
146 |
+
return x
|
147 |
+
|
diffusion/vocoder.py
ADDED
@@ -0,0 +1,94 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
from vdecoder.nsf_hifigan.nvSTFT import STFT
|
3 |
+
from vdecoder.nsf_hifigan.models import load_model,load_config
|
4 |
+
from torchaudio.transforms import Resample
|
5 |
+
|
6 |
+
|
7 |
+
class Vocoder:
|
8 |
+
def __init__(self, vocoder_type, vocoder_ckpt, device = None):
|
9 |
+
if device is None:
|
10 |
+
device = 'cuda' if torch.cuda.is_available() else 'cpu'
|
11 |
+
self.device = device
|
12 |
+
|
13 |
+
if vocoder_type == 'nsf-hifigan':
|
14 |
+
self.vocoder = NsfHifiGAN(vocoder_ckpt, device = device)
|
15 |
+
elif vocoder_type == 'nsf-hifigan-log10':
|
16 |
+
self.vocoder = NsfHifiGANLog10(vocoder_ckpt, device = device)
|
17 |
+
else:
|
18 |
+
raise ValueError(f" [x] Unknown vocoder: {vocoder_type}")
|
19 |
+
|
20 |
+
self.resample_kernel = {}
|
21 |
+
self.vocoder_sample_rate = self.vocoder.sample_rate()
|
22 |
+
self.vocoder_hop_size = self.vocoder.hop_size()
|
23 |
+
self.dimension = self.vocoder.dimension()
|
24 |
+
|
25 |
+
def extract(self, audio, sample_rate, keyshift=0):
|
26 |
+
|
27 |
+
# resample
|
28 |
+
if sample_rate == self.vocoder_sample_rate:
|
29 |
+
audio_res = audio
|
30 |
+
else:
|
31 |
+
key_str = str(sample_rate)
|
32 |
+
if key_str not in self.resample_kernel:
|
33 |
+
self.resample_kernel[key_str] = Resample(sample_rate, self.vocoder_sample_rate, lowpass_filter_width = 128).to(self.device)
|
34 |
+
audio_res = self.resample_kernel[key_str](audio)
|
35 |
+
|
36 |
+
# extract
|
37 |
+
mel = self.vocoder.extract(audio_res, keyshift=keyshift) # B, n_frames, bins
|
38 |
+
return mel
|
39 |
+
|
40 |
+
def infer(self, mel, f0):
|
41 |
+
f0 = f0[:,:mel.size(1),0] # B, n_frames
|
42 |
+
audio = self.vocoder(mel, f0)
|
43 |
+
return audio
|
44 |
+
|
45 |
+
|
46 |
+
class NsfHifiGAN(torch.nn.Module):
|
47 |
+
def __init__(self, model_path, device=None):
|
48 |
+
super().__init__()
|
49 |
+
if device is None:
|
50 |
+
device = 'cuda' if torch.cuda.is_available() else 'cpu'
|
51 |
+
self.device = device
|
52 |
+
self.model_path = model_path
|
53 |
+
self.model = None
|
54 |
+
self.h = load_config(model_path)
|
55 |
+
self.stft = STFT(
|
56 |
+
self.h.sampling_rate,
|
57 |
+
self.h.num_mels,
|
58 |
+
self.h.n_fft,
|
59 |
+
self.h.win_size,
|
60 |
+
self.h.hop_size,
|
61 |
+
self.h.fmin,
|
62 |
+
self.h.fmax)
|
63 |
+
|
64 |
+
def sample_rate(self):
|
65 |
+
return self.h.sampling_rate
|
66 |
+
|
67 |
+
def hop_size(self):
|
68 |
+
return self.h.hop_size
|
69 |
+
|
70 |
+
def dimension(self):
|
71 |
+
return self.h.num_mels
|
72 |
+
|
73 |
+
def extract(self, audio, keyshift=0):
|
74 |
+
mel = self.stft.get_mel(audio, keyshift=keyshift).transpose(1, 2) # B, n_frames, bins
|
75 |
+
return mel
|
76 |
+
|
77 |
+
def forward(self, mel, f0):
|
78 |
+
if self.model is None:
|
79 |
+
print('| Load HifiGAN: ', self.model_path)
|
80 |
+
self.model, self.h = load_model(self.model_path, device=self.device)
|
81 |
+
with torch.no_grad():
|
82 |
+
c = mel.transpose(1, 2)
|
83 |
+
audio = self.model(c, f0)
|
84 |
+
return audio
|
85 |
+
|
86 |
+
class NsfHifiGANLog10(NsfHifiGAN):
|
87 |
+
def forward(self, mel, f0):
|
88 |
+
if self.model is None:
|
89 |
+
print('| Load HifiGAN: ', self.model_path)
|
90 |
+
self.model, self.h = load_model(self.model_path, device=self.device)
|
91 |
+
with torch.no_grad():
|
92 |
+
c = 0.434294 * mel.transpose(1, 2)
|
93 |
+
audio = self.model(c, f0)
|
94 |
+
return audio
|
diffusion/wavenet.py
ADDED
@@ -0,0 +1,108 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import math
|
2 |
+
from math import sqrt
|
3 |
+
|
4 |
+
import torch
|
5 |
+
import torch.nn as nn
|
6 |
+
import torch.nn.functional as F
|
7 |
+
from torch.nn import Mish
|
8 |
+
|
9 |
+
|
10 |
+
class Conv1d(torch.nn.Conv1d):
|
11 |
+
def __init__(self, *args, **kwargs):
|
12 |
+
super().__init__(*args, **kwargs)
|
13 |
+
nn.init.kaiming_normal_(self.weight)
|
14 |
+
|
15 |
+
|
16 |
+
class SinusoidalPosEmb(nn.Module):
|
17 |
+
def __init__(self, dim):
|
18 |
+
super().__init__()
|
19 |
+
self.dim = dim
|
20 |
+
|
21 |
+
def forward(self, x):
|
22 |
+
device = x.device
|
23 |
+
half_dim = self.dim // 2
|
24 |
+
emb = math.log(10000) / (half_dim - 1)
|
25 |
+
emb = torch.exp(torch.arange(half_dim, device=device) * -emb)
|
26 |
+
emb = x[:, None] * emb[None, :]
|
27 |
+
emb = torch.cat((emb.sin(), emb.cos()), dim=-1)
|
28 |
+
return emb
|
29 |
+
|
30 |
+
|
31 |
+
class ResidualBlock(nn.Module):
|
32 |
+
def __init__(self, encoder_hidden, residual_channels, dilation):
|
33 |
+
super().__init__()
|
34 |
+
self.residual_channels = residual_channels
|
35 |
+
self.dilated_conv = nn.Conv1d(
|
36 |
+
residual_channels,
|
37 |
+
2 * residual_channels,
|
38 |
+
kernel_size=3,
|
39 |
+
padding=dilation,
|
40 |
+
dilation=dilation
|
41 |
+
)
|
42 |
+
self.diffusion_projection = nn.Linear(residual_channels, residual_channels)
|
43 |
+
self.conditioner_projection = nn.Conv1d(encoder_hidden, 2 * residual_channels, 1)
|
44 |
+
self.output_projection = nn.Conv1d(residual_channels, 2 * residual_channels, 1)
|
45 |
+
|
46 |
+
def forward(self, x, conditioner, diffusion_step):
|
47 |
+
diffusion_step = self.diffusion_projection(diffusion_step).unsqueeze(-1)
|
48 |
+
conditioner = self.conditioner_projection(conditioner)
|
49 |
+
y = x + diffusion_step
|
50 |
+
|
51 |
+
y = self.dilated_conv(y) + conditioner
|
52 |
+
|
53 |
+
# Using torch.split instead of torch.chunk to avoid using onnx::Slice
|
54 |
+
gate, filter = torch.split(y, [self.residual_channels, self.residual_channels], dim=1)
|
55 |
+
y = torch.sigmoid(gate) * torch.tanh(filter)
|
56 |
+
|
57 |
+
y = self.output_projection(y)
|
58 |
+
|
59 |
+
# Using torch.split instead of torch.chunk to avoid using onnx::Slice
|
60 |
+
residual, skip = torch.split(y, [self.residual_channels, self.residual_channels], dim=1)
|
61 |
+
return (x + residual) / math.sqrt(2.0), skip
|
62 |
+
|
63 |
+
|
64 |
+
class WaveNet(nn.Module):
|
65 |
+
def __init__(self, in_dims=128, n_layers=20, n_chans=384, n_hidden=256):
|
66 |
+
super().__init__()
|
67 |
+
self.input_projection = Conv1d(in_dims, n_chans, 1)
|
68 |
+
self.diffusion_embedding = SinusoidalPosEmb(n_chans)
|
69 |
+
self.mlp = nn.Sequential(
|
70 |
+
nn.Linear(n_chans, n_chans * 4),
|
71 |
+
Mish(),
|
72 |
+
nn.Linear(n_chans * 4, n_chans)
|
73 |
+
)
|
74 |
+
self.residual_layers = nn.ModuleList([
|
75 |
+
ResidualBlock(
|
76 |
+
encoder_hidden=n_hidden,
|
77 |
+
residual_channels=n_chans,
|
78 |
+
dilation=1
|
79 |
+
)
|
80 |
+
for i in range(n_layers)
|
81 |
+
])
|
82 |
+
self.skip_projection = Conv1d(n_chans, n_chans, 1)
|
83 |
+
self.output_projection = Conv1d(n_chans, in_dims, 1)
|
84 |
+
nn.init.zeros_(self.output_projection.weight)
|
85 |
+
|
86 |
+
def forward(self, spec, diffusion_step, cond):
|
87 |
+
"""
|
88 |
+
:param spec: [B, 1, M, T]
|
89 |
+
:param diffusion_step: [B, 1]
|
90 |
+
:param cond: [B, M, T]
|
91 |
+
:return:
|
92 |
+
"""
|
93 |
+
x = spec.squeeze(1)
|
94 |
+
x = self.input_projection(x) # [B, residual_channel, T]
|
95 |
+
|
96 |
+
x = F.relu(x)
|
97 |
+
diffusion_step = self.diffusion_embedding(diffusion_step)
|
98 |
+
diffusion_step = self.mlp(diffusion_step)
|
99 |
+
skip = []
|
100 |
+
for layer in self.residual_layers:
|
101 |
+
x, skip_connection = layer(x, cond, diffusion_step)
|
102 |
+
skip.append(skip_connection)
|
103 |
+
|
104 |
+
x = torch.sum(torch.stack(skip), dim=0) / sqrt(len(self.residual_layers))
|
105 |
+
x = self.skip_projection(x)
|
106 |
+
x = F.relu(x)
|
107 |
+
x = self.output_projection(x) # [B, mel_bins, T]
|
108 |
+
return x[:, None, :, :]
|
filelists/test.txt
ADDED
@@ -0,0 +1,4 @@
|
|
|
|
|
|
|
|
|
|
|
1 |
+
./dataset/44k/taffy/000562.wav
|
2 |
+
./dataset/44k/nyaru/000011.wav
|
3 |
+
./dataset/44k/nyaru/000008.wav
|
4 |
+
./dataset/44k/taffy/000563.wav
|
filelists/train.txt
ADDED
@@ -0,0 +1,15 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
./dataset/44k/taffy/000549.wav
|
2 |
+
./dataset/44k/nyaru/000004.wav
|
3 |
+
./dataset/44k/nyaru/000006.wav
|
4 |
+
./dataset/44k/taffy/000551.wav
|
5 |
+
./dataset/44k/nyaru/000009.wav
|
6 |
+
./dataset/44k/taffy/000561.wav
|
7 |
+
./dataset/44k/nyaru/000001.wav
|
8 |
+
./dataset/44k/taffy/000553.wav
|
9 |
+
./dataset/44k/nyaru/000002.wav
|
10 |
+
./dataset/44k/taffy/000560.wav
|
11 |
+
./dataset/44k/taffy/000557.wav
|
12 |
+
./dataset/44k/nyaru/000005.wav
|
13 |
+
./dataset/44k/taffy/000554.wav
|
14 |
+
./dataset/44k/taffy/000550.wav
|
15 |
+
./dataset/44k/taffy/000559.wav
|
filelists/val.txt
ADDED
@@ -0,0 +1,4 @@
|
|
|
|
|
|
|
|
|
|
|
1 |
+
./dataset/44k/nyaru/000003.wav
|
2 |
+
./dataset/44k/nyaru/000007.wav
|
3 |
+
./dataset/44k/taffy/000558.wav
|
4 |
+
./dataset/44k/taffy/000556.wav
|
flask_api.py
ADDED
@@ -0,0 +1,60 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import io
|
2 |
+
import logging
|
3 |
+
|
4 |
+
import soundfile
|
5 |
+
import torch
|
6 |
+
import torchaudio
|
7 |
+
from flask import Flask, request, send_file
|
8 |
+
from flask_cors import CORS
|
9 |
+
|
10 |
+
from inference.infer_tool import Svc, RealTimeVC
|
11 |
+
|
12 |
+
app = Flask(__name__)
|
13 |
+
|
14 |
+
CORS(app)
|
15 |
+
|
16 |
+
logging.getLogger('numba').setLevel(logging.WARNING)
|
17 |
+
|
18 |
+
|
19 |
+
@app.route("/voiceChangeModel", methods=["POST"])
|
20 |
+
def voice_change_model():
|
21 |
+
request_form = request.form
|
22 |
+
wave_file = request.files.get("sample", None)
|
23 |
+
# 变调信息
|
24 |
+
f_pitch_change = float(request_form.get("fPitchChange", 0))
|
25 |
+
# DAW所需的采样率
|
26 |
+
daw_sample = int(float(request_form.get("sampleRate", 0)))
|
27 |
+
speaker_id = int(float(request_form.get("sSpeakId", 0)))
|
28 |
+
# http获得wav文件并转换
|
29 |
+
input_wav_path = io.BytesIO(wave_file.read())
|
30 |
+
|
31 |
+
# 模型推理
|
32 |
+
if raw_infer:
|
33 |
+
# out_audio, out_sr = svc_model.infer(speaker_id, f_pitch_change, input_wav_path)
|
34 |
+
out_audio, out_sr = svc_model.infer(speaker_id, f_pitch_change, input_wav_path, cluster_infer_ratio=0,
|
35 |
+
auto_predict_f0=False, noice_scale=0.4, f0_filter=False)
|
36 |
+
tar_audio = torchaudio.functional.resample(out_audio, svc_model.target_sample, daw_sample)
|
37 |
+
else:
|
38 |
+
out_audio = svc.process(svc_model, speaker_id, f_pitch_change, input_wav_path, cluster_infer_ratio=0,
|
39 |
+
auto_predict_f0=False, noice_scale=0.4, f0_filter=False)
|
40 |
+
tar_audio = torchaudio.functional.resample(torch.from_numpy(out_audio), svc_model.target_sample, daw_sample)
|
41 |
+
# 返回音频
|
42 |
+
out_wav_path = io.BytesIO()
|
43 |
+
soundfile.write(out_wav_path, tar_audio.cpu().numpy(), daw_sample, format="wav")
|
44 |
+
out_wav_path.seek(0)
|
45 |
+
return send_file(out_wav_path, download_name="temp.wav", as_attachment=True)
|
46 |
+
|
47 |
+
|
48 |
+
if __name__ == '__main__':
|
49 |
+
# 启用则为直接切片合成,False为交叉淡化方式
|
50 |
+
# vst插件调整0.3-0.5s切片时间可以降低延迟,直接切片方法会有连接处爆音、交叉淡化会有轻微重叠声音
|
51 |
+
# 自行选择能接受的方法,或将vst最大切片时间调整为1s,此处设为Ture,延迟大音质稳定一些
|
52 |
+
raw_infer = True
|
53 |
+
# 每个模型和config是唯一对应的
|
54 |
+
model_name = "logs/32k/G_174000-Copy1.pth"
|
55 |
+
config_name = "configs/config.json"
|
56 |
+
cluster_model_path = "logs/44k/kmeans_10000.pt"
|
57 |
+
svc_model = Svc(model_name, config_name, cluster_model_path=cluster_model_path)
|
58 |
+
svc = RealTimeVC()
|
59 |
+
# 此处与vst插件对应,不建议更改
|
60 |
+
app.run(port=6842, host="0.0.0.0", debug=False, threaded=False)
|
hubert/__init__.py
ADDED
File without changes
|
hubert/checkpoint_best_legacy_500.pt
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:60d936ec5a566776fc392e69ad8b630d14eb588111233fe313436e200a7b187b
|
3 |
+
size 1330114945
|
hubert/hubert_model.py
ADDED
@@ -0,0 +1,222 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import copy
|
2 |
+
import random
|
3 |
+
from typing import Optional, Tuple
|
4 |
+
|
5 |
+
import torch
|
6 |
+
import torch.nn as nn
|
7 |
+
import torch.nn.functional as t_func
|
8 |
+
from torch.nn.modules.utils import consume_prefix_in_state_dict_if_present
|
9 |
+
|
10 |
+
|
11 |
+
class Hubert(nn.Module):
|
12 |
+
def __init__(self, num_label_embeddings: int = 100, mask: bool = True):
|
13 |
+
super().__init__()
|
14 |
+
self._mask = mask
|
15 |
+
self.feature_extractor = FeatureExtractor()
|
16 |
+
self.feature_projection = FeatureProjection()
|
17 |
+
self.positional_embedding = PositionalConvEmbedding()
|
18 |
+
self.norm = nn.LayerNorm(768)
|
19 |
+
self.dropout = nn.Dropout(0.1)
|
20 |
+
self.encoder = TransformerEncoder(
|
21 |
+
nn.TransformerEncoderLayer(
|
22 |
+
768, 12, 3072, activation="gelu", batch_first=True
|
23 |
+
),
|
24 |
+
12,
|
25 |
+
)
|
26 |
+
self.proj = nn.Linear(768, 256)
|
27 |
+
|
28 |
+
self.masked_spec_embed = nn.Parameter(torch.FloatTensor(768).uniform_())
|
29 |
+
self.label_embedding = nn.Embedding(num_label_embeddings, 256)
|
30 |
+
|
31 |
+
def mask(self, x: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
|
32 |
+
mask = None
|
33 |
+
if self.training and self._mask:
|
34 |
+
mask = _compute_mask((x.size(0), x.size(1)), 0.8, 10, x.device, 2)
|
35 |
+
x[mask] = self.masked_spec_embed.to(x.dtype)
|
36 |
+
return x, mask
|
37 |
+
|
38 |
+
def encode(
|
39 |
+
self, x: torch.Tensor, layer: Optional[int] = None
|
40 |
+
) -> Tuple[torch.Tensor, torch.Tensor]:
|
41 |
+
x = self.feature_extractor(x)
|
42 |
+
x = self.feature_projection(x.transpose(1, 2))
|
43 |
+
x, mask = self.mask(x)
|
44 |
+
x = x + self.positional_embedding(x)
|
45 |
+
x = self.dropout(self.norm(x))
|
46 |
+
x = self.encoder(x, output_layer=layer)
|
47 |
+
return x, mask
|
48 |
+
|
49 |
+
def logits(self, x: torch.Tensor) -> torch.Tensor:
|
50 |
+
logits = torch.cosine_similarity(
|
51 |
+
x.unsqueeze(2),
|
52 |
+
self.label_embedding.weight.unsqueeze(0).unsqueeze(0),
|
53 |
+
dim=-1,
|
54 |
+
)
|
55 |
+
return logits / 0.1
|
56 |
+
|
57 |
+
def forward(self, x: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
|
58 |
+
x, mask = self.encode(x)
|
59 |
+
x = self.proj(x)
|
60 |
+
logits = self.logits(x)
|
61 |
+
return logits, mask
|
62 |
+
|
63 |
+
|
64 |
+
class HubertSoft(Hubert):
|
65 |
+
def __init__(self):
|
66 |
+
super().__init__()
|
67 |
+
|
68 |
+
@torch.inference_mode()
|
69 |
+
def units(self, wav: torch.Tensor) -> torch.Tensor:
|
70 |
+
wav = t_func.pad(wav, ((400 - 320) // 2, (400 - 320) // 2))
|
71 |
+
x, _ = self.encode(wav)
|
72 |
+
return self.proj(x)
|
73 |
+
|
74 |
+
|
75 |
+
class FeatureExtractor(nn.Module):
|
76 |
+
def __init__(self):
|
77 |
+
super().__init__()
|
78 |
+
self.conv0 = nn.Conv1d(1, 512, 10, 5, bias=False)
|
79 |
+
self.norm0 = nn.GroupNorm(512, 512)
|
80 |
+
self.conv1 = nn.Conv1d(512, 512, 3, 2, bias=False)
|
81 |
+
self.conv2 = nn.Conv1d(512, 512, 3, 2, bias=False)
|
82 |
+
self.conv3 = nn.Conv1d(512, 512, 3, 2, bias=False)
|
83 |
+
self.conv4 = nn.Conv1d(512, 512, 3, 2, bias=False)
|
84 |
+
self.conv5 = nn.Conv1d(512, 512, 2, 2, bias=False)
|
85 |
+
self.conv6 = nn.Conv1d(512, 512, 2, 2, bias=False)
|
86 |
+
|
87 |
+
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
88 |
+
x = t_func.gelu(self.norm0(self.conv0(x)))
|
89 |
+
x = t_func.gelu(self.conv1(x))
|
90 |
+
x = t_func.gelu(self.conv2(x))
|
91 |
+
x = t_func.gelu(self.conv3(x))
|
92 |
+
x = t_func.gelu(self.conv4(x))
|
93 |
+
x = t_func.gelu(self.conv5(x))
|
94 |
+
x = t_func.gelu(self.conv6(x))
|
95 |
+
return x
|
96 |
+
|
97 |
+
|
98 |
+
class FeatureProjection(nn.Module):
|
99 |
+
def __init__(self):
|
100 |
+
super().__init__()
|
101 |
+
self.norm = nn.LayerNorm(512)
|
102 |
+
self.projection = nn.Linear(512, 768)
|
103 |
+
self.dropout = nn.Dropout(0.1)
|
104 |
+
|
105 |
+
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
106 |
+
x = self.norm(x)
|
107 |
+
x = self.projection(x)
|
108 |
+
x = self.dropout(x)
|
109 |
+
return x
|
110 |
+
|
111 |
+
|
112 |
+
class PositionalConvEmbedding(nn.Module):
|
113 |
+
def __init__(self):
|
114 |
+
super().__init__()
|
115 |
+
self.conv = nn.Conv1d(
|
116 |
+
768,
|
117 |
+
768,
|
118 |
+
kernel_size=128,
|
119 |
+
padding=128 // 2,
|
120 |
+
groups=16,
|
121 |
+
)
|
122 |
+
self.conv = nn.utils.weight_norm(self.conv, name="weight", dim=2)
|
123 |
+
|
124 |
+
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
125 |
+
x = self.conv(x.transpose(1, 2))
|
126 |
+
x = t_func.gelu(x[:, :, :-1])
|
127 |
+
return x.transpose(1, 2)
|
128 |
+
|
129 |
+
|
130 |
+
class TransformerEncoder(nn.Module):
|
131 |
+
def __init__(
|
132 |
+
self, encoder_layer: nn.TransformerEncoderLayer, num_layers: int
|
133 |
+
) -> None:
|
134 |
+
super(TransformerEncoder, self).__init__()
|
135 |
+
self.layers = nn.ModuleList(
|
136 |
+
[copy.deepcopy(encoder_layer) for _ in range(num_layers)]
|
137 |
+
)
|
138 |
+
self.num_layers = num_layers
|
139 |
+
|
140 |
+
def forward(
|
141 |
+
self,
|
142 |
+
src: torch.Tensor,
|
143 |
+
mask: torch.Tensor = None,
|
144 |
+
src_key_padding_mask: torch.Tensor = None,
|
145 |
+
output_layer: Optional[int] = None,
|
146 |
+
) -> torch.Tensor:
|
147 |
+
output = src
|
148 |
+
for layer in self.layers[:output_layer]:
|
149 |
+
output = layer(
|
150 |
+
output, src_mask=mask, src_key_padding_mask=src_key_padding_mask
|
151 |
+
)
|
152 |
+
return output
|
153 |
+
|
154 |
+
|
155 |
+
def _compute_mask(
|
156 |
+
shape: Tuple[int, int],
|
157 |
+
mask_prob: float,
|
158 |
+
mask_length: int,
|
159 |
+
device: torch.device,
|
160 |
+
min_masks: int = 0,
|
161 |
+
) -> torch.Tensor:
|
162 |
+
batch_size, sequence_length = shape
|
163 |
+
|
164 |
+
if mask_length < 1:
|
165 |
+
raise ValueError("`mask_length` has to be bigger than 0.")
|
166 |
+
|
167 |
+
if mask_length > sequence_length:
|
168 |
+
raise ValueError(
|
169 |
+
f"`mask_length` has to be smaller than `sequence_length`, but got `mask_length`: {mask_length} and `sequence_length`: {sequence_length}`"
|
170 |
+
)
|
171 |
+
|
172 |
+
# compute number of masked spans in batch
|
173 |
+
num_masked_spans = int(mask_prob * sequence_length / mask_length + random.random())
|
174 |
+
num_masked_spans = max(num_masked_spans, min_masks)
|
175 |
+
|
176 |
+
# make sure num masked indices <= sequence_length
|
177 |
+
if num_masked_spans * mask_length > sequence_length:
|
178 |
+
num_masked_spans = sequence_length // mask_length
|
179 |
+
|
180 |
+
# SpecAugment mask to fill
|
181 |
+
mask = torch.zeros((batch_size, sequence_length), device=device, dtype=torch.bool)
|
182 |
+
|
183 |
+
# uniform distribution to sample from, make sure that offset samples are < sequence_length
|
184 |
+
uniform_dist = torch.ones(
|
185 |
+
(batch_size, sequence_length - (mask_length - 1)), device=device
|
186 |
+
)
|
187 |
+
|
188 |
+
# get random indices to mask
|
189 |
+
mask_indices = torch.multinomial(uniform_dist, num_masked_spans)
|
190 |
+
|
191 |
+
# expand masked indices to masked spans
|
192 |
+
mask_indices = (
|
193 |
+
mask_indices.unsqueeze(dim=-1)
|
194 |
+
.expand((batch_size, num_masked_spans, mask_length))
|
195 |
+
.reshape(batch_size, num_masked_spans * mask_length)
|
196 |
+
)
|
197 |
+
offsets = (
|
198 |
+
torch.arange(mask_length, device=device)[None, None, :]
|
199 |
+
.expand((batch_size, num_masked_spans, mask_length))
|
200 |
+
.reshape(batch_size, num_masked_spans * mask_length)
|
201 |
+
)
|
202 |
+
mask_idxs = mask_indices + offsets
|
203 |
+
|
204 |
+
# scatter indices to mask
|
205 |
+
mask = mask.scatter(1, mask_idxs, True)
|
206 |
+
|
207 |
+
return mask
|
208 |
+
|
209 |
+
|
210 |
+
def hubert_soft(
|
211 |
+
path: str,
|
212 |
+
) -> HubertSoft:
|
213 |
+
r"""HuBERT-Soft from `"A Comparison of Discrete and Soft Speech Units for Improved Voice Conversion"`.
|
214 |
+
Args:
|
215 |
+
path (str): path of a pretrained model
|
216 |
+
"""
|
217 |
+
hubert = HubertSoft()
|
218 |
+
checkpoint = torch.load(path)
|
219 |
+
consume_prefix_in_state_dict_if_present(checkpoint, "module.")
|
220 |
+
hubert.load_state_dict(checkpoint)
|
221 |
+
hubert.eval()
|
222 |
+
return hubert
|
hubert/hubert_model_onnx.py
ADDED
@@ -0,0 +1,217 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import copy
|
2 |
+
import random
|
3 |
+
from typing import Optional, Tuple
|
4 |
+
|
5 |
+
import torch
|
6 |
+
import torch.nn as nn
|
7 |
+
import torch.nn.functional as t_func
|
8 |
+
from torch.nn.modules.utils import consume_prefix_in_state_dict_if_present
|
9 |
+
|
10 |
+
|
11 |
+
class Hubert(nn.Module):
|
12 |
+
def __init__(self, num_label_embeddings: int = 100, mask: bool = True):
|
13 |
+
super().__init__()
|
14 |
+
self._mask = mask
|
15 |
+
self.feature_extractor = FeatureExtractor()
|
16 |
+
self.feature_projection = FeatureProjection()
|
17 |
+
self.positional_embedding = PositionalConvEmbedding()
|
18 |
+
self.norm = nn.LayerNorm(768)
|
19 |
+
self.dropout = nn.Dropout(0.1)
|
20 |
+
self.encoder = TransformerEncoder(
|
21 |
+
nn.TransformerEncoderLayer(
|
22 |
+
768, 12, 3072, activation="gelu", batch_first=True
|
23 |
+
),
|
24 |
+
12,
|
25 |
+
)
|
26 |
+
self.proj = nn.Linear(768, 256)
|
27 |
+
|
28 |
+
self.masked_spec_embed = nn.Parameter(torch.FloatTensor(768).uniform_())
|
29 |
+
self.label_embedding = nn.Embedding(num_label_embeddings, 256)
|
30 |
+
|
31 |
+
def mask(self, x: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
|
32 |
+
mask = None
|
33 |
+
if self.training and self._mask:
|
34 |
+
mask = _compute_mask((x.size(0), x.size(1)), 0.8, 10, x.device, 2)
|
35 |
+
x[mask] = self.masked_spec_embed.to(x.dtype)
|
36 |
+
return x, mask
|
37 |
+
|
38 |
+
def encode(
|
39 |
+
self, x: torch.Tensor, layer: Optional[int] = None
|
40 |
+
) -> Tuple[torch.Tensor, torch.Tensor]:
|
41 |
+
x = self.feature_extractor(x)
|
42 |
+
x = self.feature_projection(x.transpose(1, 2))
|
43 |
+
x, mask = self.mask(x)
|
44 |
+
x = x + self.positional_embedding(x)
|
45 |
+
x = self.dropout(self.norm(x))
|
46 |
+
x = self.encoder(x, output_layer=layer)
|
47 |
+
return x, mask
|
48 |
+
|
49 |
+
def logits(self, x: torch.Tensor) -> torch.Tensor:
|
50 |
+
logits = torch.cosine_similarity(
|
51 |
+
x.unsqueeze(2),
|
52 |
+
self.label_embedding.weight.unsqueeze(0).unsqueeze(0),
|
53 |
+
dim=-1,
|
54 |
+
)
|
55 |
+
return logits / 0.1
|
56 |
+
|
57 |
+
|
58 |
+
class HubertSoft(Hubert):
|
59 |
+
def __init__(self):
|
60 |
+
super().__init__()
|
61 |
+
|
62 |
+
def units(self, wav: torch.Tensor) -> torch.Tensor:
|
63 |
+
wav = t_func.pad(wav, ((400 - 320) // 2, (400 - 320) // 2))
|
64 |
+
x, _ = self.encode(wav)
|
65 |
+
return self.proj(x)
|
66 |
+
|
67 |
+
def forward(self, x):
|
68 |
+
return self.units(x)
|
69 |
+
|
70 |
+
class FeatureExtractor(nn.Module):
|
71 |
+
def __init__(self):
|
72 |
+
super().__init__()
|
73 |
+
self.conv0 = nn.Conv1d(1, 512, 10, 5, bias=False)
|
74 |
+
self.norm0 = nn.GroupNorm(512, 512)
|
75 |
+
self.conv1 = nn.Conv1d(512, 512, 3, 2, bias=False)
|
76 |
+
self.conv2 = nn.Conv1d(512, 512, 3, 2, bias=False)
|
77 |
+
self.conv3 = nn.Conv1d(512, 512, 3, 2, bias=False)
|
78 |
+
self.conv4 = nn.Conv1d(512, 512, 3, 2, bias=False)
|
79 |
+
self.conv5 = nn.Conv1d(512, 512, 2, 2, bias=False)
|
80 |
+
self.conv6 = nn.Conv1d(512, 512, 2, 2, bias=False)
|
81 |
+
|
82 |
+
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
83 |
+
x = t_func.gelu(self.norm0(self.conv0(x)))
|
84 |
+
x = t_func.gelu(self.conv1(x))
|
85 |
+
x = t_func.gelu(self.conv2(x))
|
86 |
+
x = t_func.gelu(self.conv3(x))
|
87 |
+
x = t_func.gelu(self.conv4(x))
|
88 |
+
x = t_func.gelu(self.conv5(x))
|
89 |
+
x = t_func.gelu(self.conv6(x))
|
90 |
+
return x
|
91 |
+
|
92 |
+
|
93 |
+
class FeatureProjection(nn.Module):
|
94 |
+
def __init__(self):
|
95 |
+
super().__init__()
|
96 |
+
self.norm = nn.LayerNorm(512)
|
97 |
+
self.projection = nn.Linear(512, 768)
|
98 |
+
self.dropout = nn.Dropout(0.1)
|
99 |
+
|
100 |
+
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
101 |
+
x = self.norm(x)
|
102 |
+
x = self.projection(x)
|
103 |
+
x = self.dropout(x)
|
104 |
+
return x
|
105 |
+
|
106 |
+
|
107 |
+
class PositionalConvEmbedding(nn.Module):
|
108 |
+
def __init__(self):
|
109 |
+
super().__init__()
|
110 |
+
self.conv = nn.Conv1d(
|
111 |
+
768,
|
112 |
+
768,
|
113 |
+
kernel_size=128,
|
114 |
+
padding=128 // 2,
|
115 |
+
groups=16,
|
116 |
+
)
|
117 |
+
self.conv = nn.utils.weight_norm(self.conv, name="weight", dim=2)
|
118 |
+
|
119 |
+
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
120 |
+
x = self.conv(x.transpose(1, 2))
|
121 |
+
x = t_func.gelu(x[:, :, :-1])
|
122 |
+
return x.transpose(1, 2)
|
123 |
+
|
124 |
+
|
125 |
+
class TransformerEncoder(nn.Module):
|
126 |
+
def __init__(
|
127 |
+
self, encoder_layer: nn.TransformerEncoderLayer, num_layers: int
|
128 |
+
) -> None:
|
129 |
+
super(TransformerEncoder, self).__init__()
|
130 |
+
self.layers = nn.ModuleList(
|
131 |
+
[copy.deepcopy(encoder_layer) for _ in range(num_layers)]
|
132 |
+
)
|
133 |
+
self.num_layers = num_layers
|
134 |
+
|
135 |
+
def forward(
|
136 |
+
self,
|
137 |
+
src: torch.Tensor,
|
138 |
+
mask: torch.Tensor = None,
|
139 |
+
src_key_padding_mask: torch.Tensor = None,
|
140 |
+
output_layer: Optional[int] = None,
|
141 |
+
) -> torch.Tensor:
|
142 |
+
output = src
|
143 |
+
for layer in self.layers[:output_layer]:
|
144 |
+
output = layer(
|
145 |
+
output, src_mask=mask, src_key_padding_mask=src_key_padding_mask
|
146 |
+
)
|
147 |
+
return output
|
148 |
+
|
149 |
+
|
150 |
+
def _compute_mask(
|
151 |
+
shape: Tuple[int, int],
|
152 |
+
mask_prob: float,
|
153 |
+
mask_length: int,
|
154 |
+
device: torch.device,
|
155 |
+
min_masks: int = 0,
|
156 |
+
) -> torch.Tensor:
|
157 |
+
batch_size, sequence_length = shape
|
158 |
+
|
159 |
+
if mask_length < 1:
|
160 |
+
raise ValueError("`mask_length` has to be bigger than 0.")
|
161 |
+
|
162 |
+
if mask_length > sequence_length:
|
163 |
+
raise ValueError(
|
164 |
+
f"`mask_length` has to be smaller than `sequence_length`, but got `mask_length`: {mask_length} and `sequence_length`: {sequence_length}`"
|
165 |
+
)
|
166 |
+
|
167 |
+
# compute number of masked spans in batch
|
168 |
+
num_masked_spans = int(mask_prob * sequence_length / mask_length + random.random())
|
169 |
+
num_masked_spans = max(num_masked_spans, min_masks)
|
170 |
+
|
171 |
+
# make sure num masked indices <= sequence_length
|
172 |
+
if num_masked_spans * mask_length > sequence_length:
|
173 |
+
num_masked_spans = sequence_length // mask_length
|
174 |
+
|
175 |
+
# SpecAugment mask to fill
|
176 |
+
mask = torch.zeros((batch_size, sequence_length), device=device, dtype=torch.bool)
|
177 |
+
|
178 |
+
# uniform distribution to sample from, make sure that offset samples are < sequence_length
|
179 |
+
uniform_dist = torch.ones(
|
180 |
+
(batch_size, sequence_length - (mask_length - 1)), device=device
|
181 |
+
)
|
182 |
+
|
183 |
+
# get random indices to mask
|
184 |
+
mask_indices = torch.multinomial(uniform_dist, num_masked_spans)
|
185 |
+
|
186 |
+
# expand masked indices to masked spans
|
187 |
+
mask_indices = (
|
188 |
+
mask_indices.unsqueeze(dim=-1)
|
189 |
+
.expand((batch_size, num_masked_spans, mask_length))
|
190 |
+
.reshape(batch_size, num_masked_spans * mask_length)
|
191 |
+
)
|
192 |
+
offsets = (
|
193 |
+
torch.arange(mask_length, device=device)[None, None, :]
|
194 |
+
.expand((batch_size, num_masked_spans, mask_length))
|
195 |
+
.reshape(batch_size, num_masked_spans * mask_length)
|
196 |
+
)
|
197 |
+
mask_idxs = mask_indices + offsets
|
198 |
+
|
199 |
+
# scatter indices to mask
|
200 |
+
mask = mask.scatter(1, mask_idxs, True)
|
201 |
+
|
202 |
+
return mask
|
203 |
+
|
204 |
+
|
205 |
+
def hubert_soft(
|
206 |
+
path: str,
|
207 |
+
) -> HubertSoft:
|
208 |
+
r"""HuBERT-Soft from `"A Comparison of Discrete and Soft Speech Units for Improved Voice Conversion"`.
|
209 |
+
Args:
|
210 |
+
path (str): path of a pretrained model
|
211 |
+
"""
|
212 |
+
hubert = HubertSoft()
|
213 |
+
checkpoint = torch.load(path)
|
214 |
+
consume_prefix_in_state_dict_if_present(checkpoint, "module.")
|
215 |
+
hubert.load_state_dict(checkpoint)
|
216 |
+
hubert.eval()
|
217 |
+
return hubert
|
hubert/put_hubert_ckpt_here
ADDED
File without changes
|
inference/__init__.py
ADDED
File without changes
|
inference/chunks_temp.json
ADDED
@@ -0,0 +1 @@
|
|
|
|
|
1 |
+
{"info": "temp_dict"}
|
inference/infer_tool.py
ADDED
@@ -0,0 +1,533 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import hashlib
|
2 |
+
import io
|
3 |
+
import json
|
4 |
+
import logging
|
5 |
+
import os
|
6 |
+
import time
|
7 |
+
from pathlib import Path
|
8 |
+
from inference import slicer
|
9 |
+
import gc
|
10 |
+
|
11 |
+
import librosa
|
12 |
+
import numpy as np
|
13 |
+
# import onnxruntime
|
14 |
+
import soundfile
|
15 |
+
import torch
|
16 |
+
import torchaudio
|
17 |
+
|
18 |
+
import cluster
|
19 |
+
import utils
|
20 |
+
from models import SynthesizerTrn
|
21 |
+
import pickle
|
22 |
+
|
23 |
+
from diffusion.unit2mel import load_model_vocoder
|
24 |
+
import yaml
|
25 |
+
|
26 |
+
logging.getLogger('matplotlib').setLevel(logging.WARNING)
|
27 |
+
|
28 |
+
|
29 |
+
def read_temp(file_name):
|
30 |
+
if not os.path.exists(file_name):
|
31 |
+
with open(file_name, "w") as f:
|
32 |
+
f.write(json.dumps({"info": "temp_dict"}))
|
33 |
+
return {}
|
34 |
+
else:
|
35 |
+
try:
|
36 |
+
with open(file_name, "r") as f:
|
37 |
+
data = f.read()
|
38 |
+
data_dict = json.loads(data)
|
39 |
+
if os.path.getsize(file_name) > 50 * 1024 * 1024:
|
40 |
+
f_name = file_name.replace("\\", "/").split("/")[-1]
|
41 |
+
print(f"clean {f_name}")
|
42 |
+
for wav_hash in list(data_dict.keys()):
|
43 |
+
if int(time.time()) - int(data_dict[wav_hash]["time"]) > 14 * 24 * 3600:
|
44 |
+
del data_dict[wav_hash]
|
45 |
+
except Exception as e:
|
46 |
+
print(e)
|
47 |
+
print(f"{file_name} error,auto rebuild file")
|
48 |
+
data_dict = {"info": "temp_dict"}
|
49 |
+
return data_dict
|
50 |
+
|
51 |
+
|
52 |
+
def write_temp(file_name, data):
|
53 |
+
with open(file_name, "w") as f:
|
54 |
+
f.write(json.dumps(data))
|
55 |
+
|
56 |
+
|
57 |
+
def timeit(func):
|
58 |
+
def run(*args, **kwargs):
|
59 |
+
t = time.time()
|
60 |
+
res = func(*args, **kwargs)
|
61 |
+
print('executing \'%s\' costed %.3fs' % (func.__name__, time.time() - t))
|
62 |
+
return res
|
63 |
+
|
64 |
+
return run
|
65 |
+
|
66 |
+
|
67 |
+
def format_wav(audio_path):
|
68 |
+
if Path(audio_path).suffix == '.wav':
|
69 |
+
return
|
70 |
+
raw_audio, raw_sample_rate = librosa.load(audio_path, mono=True, sr=None)
|
71 |
+
soundfile.write(Path(audio_path).with_suffix(".wav"), raw_audio, raw_sample_rate)
|
72 |
+
|
73 |
+
|
74 |
+
def get_end_file(dir_path, end):
|
75 |
+
file_lists = []
|
76 |
+
for root, dirs, files in os.walk(dir_path):
|
77 |
+
files = [f for f in files if f[0] != '.']
|
78 |
+
dirs[:] = [d for d in dirs if d[0] != '.']
|
79 |
+
for f_file in files:
|
80 |
+
if f_file.endswith(end):
|
81 |
+
file_lists.append(os.path.join(root, f_file).replace("\\", "/"))
|
82 |
+
return file_lists
|
83 |
+
|
84 |
+
|
85 |
+
def get_md5(content):
|
86 |
+
return hashlib.new("md5", content).hexdigest()
|
87 |
+
|
88 |
+
def fill_a_to_b(a, b):
|
89 |
+
if len(a) < len(b):
|
90 |
+
for _ in range(0, len(b) - len(a)):
|
91 |
+
a.append(a[0])
|
92 |
+
|
93 |
+
def mkdir(paths: list):
|
94 |
+
for path in paths:
|
95 |
+
if not os.path.exists(path):
|
96 |
+
os.mkdir(path)
|
97 |
+
|
98 |
+
def pad_array(arr, target_length):
|
99 |
+
current_length = arr.shape[0]
|
100 |
+
if current_length >= target_length:
|
101 |
+
return arr
|
102 |
+
else:
|
103 |
+
pad_width = target_length - current_length
|
104 |
+
pad_left = pad_width // 2
|
105 |
+
pad_right = pad_width - pad_left
|
106 |
+
padded_arr = np.pad(arr, (pad_left, pad_right), 'constant', constant_values=(0, 0))
|
107 |
+
return padded_arr
|
108 |
+
|
109 |
+
def split_list_by_n(list_collection, n, pre=0):
|
110 |
+
for i in range(0, len(list_collection), n):
|
111 |
+
yield list_collection[i-pre if i-pre>=0 else i: i + n]
|
112 |
+
|
113 |
+
|
114 |
+
class F0FilterException(Exception):
|
115 |
+
pass
|
116 |
+
|
117 |
+
class Svc(object):
|
118 |
+
def __init__(self, net_g_path, config_path,
|
119 |
+
device=None,
|
120 |
+
cluster_model_path="logs/44k/kmeans_10000.pt",
|
121 |
+
nsf_hifigan_enhance = False,
|
122 |
+
diffusion_model_path="logs/44k/diffusion/model_0.pt",
|
123 |
+
diffusion_config_path="configs/diffusion.yaml",
|
124 |
+
shallow_diffusion = False,
|
125 |
+
only_diffusion = False,
|
126 |
+
spk_mix_enable = False,
|
127 |
+
feature_retrieval = False
|
128 |
+
):
|
129 |
+
self.net_g_path = net_g_path
|
130 |
+
self.only_diffusion = only_diffusion
|
131 |
+
self.shallow_diffusion = shallow_diffusion
|
132 |
+
self.feature_retrieval = feature_retrieval
|
133 |
+
if device is None:
|
134 |
+
self.dev = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
135 |
+
else:
|
136 |
+
self.dev = torch.device(device)
|
137 |
+
self.net_g_ms = None
|
138 |
+
if not self.only_diffusion:
|
139 |
+
self.hps_ms = utils.get_hparams_from_file(config_path)
|
140 |
+
self.target_sample = self.hps_ms.data.sampling_rate
|
141 |
+
self.hop_size = self.hps_ms.data.hop_length
|
142 |
+
self.spk2id = self.hps_ms.spk
|
143 |
+
try:
|
144 |
+
self.vol_embedding = self.hps_ms.model.vol_embedding
|
145 |
+
except Exception as e:
|
146 |
+
self.vol_embedding = False
|
147 |
+
try:
|
148 |
+
self.speech_encoder = self.hps_ms.model.speech_encoder
|
149 |
+
except Exception as e:
|
150 |
+
self.speech_encoder = 'vec768l12'
|
151 |
+
|
152 |
+
self.nsf_hifigan_enhance = nsf_hifigan_enhance
|
153 |
+
if self.shallow_diffusion or self.only_diffusion:
|
154 |
+
if os.path.exists(diffusion_model_path) and os.path.exists(diffusion_model_path):
|
155 |
+
self.diffusion_model,self.vocoder,self.diffusion_args = load_model_vocoder(diffusion_model_path,self.dev,config_path=diffusion_config_path)
|
156 |
+
if self.only_diffusion:
|
157 |
+
self.target_sample = self.diffusion_args.data.sampling_rate
|
158 |
+
self.hop_size = self.diffusion_args.data.block_size
|
159 |
+
self.spk2id = self.diffusion_args.spk
|
160 |
+
self.speech_encoder = self.diffusion_args.data.encoder
|
161 |
+
if spk_mix_enable:
|
162 |
+
self.diffusion_model.init_spkmix(len(self.spk2id))
|
163 |
+
else:
|
164 |
+
print("No diffusion model or config found. Shallow diffusion mode will False")
|
165 |
+
self.shallow_diffusion = self.only_diffusion = False
|
166 |
+
|
167 |
+
# load hubert and model
|
168 |
+
if not self.only_diffusion:
|
169 |
+
self.load_model(spk_mix_enable)
|
170 |
+
self.hubert_model = utils.get_speech_encoder(self.speech_encoder,device=self.dev)
|
171 |
+
self.volume_extractor = utils.Volume_Extractor(self.hop_size)
|
172 |
+
else:
|
173 |
+
self.hubert_model = utils.get_speech_encoder(self.diffusion_args.data.encoder,device=self.dev)
|
174 |
+
self.volume_extractor = utils.Volume_Extractor(self.diffusion_args.data.block_size)
|
175 |
+
|
176 |
+
if os.path.exists(cluster_model_path):
|
177 |
+
if self.feature_retrieval:
|
178 |
+
with open(cluster_model_path,"rb") as f:
|
179 |
+
self.cluster_model = pickle.load(f)
|
180 |
+
self.big_npy = None
|
181 |
+
self.now_spk_id = -1
|
182 |
+
else:
|
183 |
+
self.cluster_model = cluster.get_cluster_model(cluster_model_path)
|
184 |
+
else:
|
185 |
+
self.feature_retrieval=False
|
186 |
+
|
187 |
+
if self.shallow_diffusion : self.nsf_hifigan_enhance = False
|
188 |
+
if self.nsf_hifigan_enhance:
|
189 |
+
from modules.enhancer import Enhancer
|
190 |
+
self.enhancer = Enhancer('nsf-hifigan', 'pretrain/nsf_hifigan/model',device=self.dev)
|
191 |
+
|
192 |
+
def load_model(self, spk_mix_enable=False):
|
193 |
+
# get model configuration
|
194 |
+
self.net_g_ms = SynthesizerTrn(
|
195 |
+
self.hps_ms.data.filter_length // 2 + 1,
|
196 |
+
self.hps_ms.train.segment_size // self.hps_ms.data.hop_length,
|
197 |
+
**self.hps_ms.model)
|
198 |
+
_ = utils.load_checkpoint(self.net_g_path, self.net_g_ms, None)
|
199 |
+
if "half" in self.net_g_path and torch.cuda.is_available():
|
200 |
+
_ = self.net_g_ms.half().eval().to(self.dev)
|
201 |
+
else:
|
202 |
+
_ = self.net_g_ms.eval().to(self.dev)
|
203 |
+
if spk_mix_enable:
|
204 |
+
self.net_g_ms.EnableCharacterMix(len(self.spk2id), self.dev)
|
205 |
+
|
206 |
+
def get_unit_f0(self, wav, tran, cluster_infer_ratio, speaker, f0_filter ,f0_predictor,cr_threshold=0.05):
|
207 |
+
|
208 |
+
f0_predictor_object = utils.get_f0_predictor(f0_predictor,hop_length=self.hop_size,sampling_rate=self.target_sample,device=self.dev,threshold=cr_threshold)
|
209 |
+
|
210 |
+
f0, uv = f0_predictor_object.compute_f0_uv(wav)
|
211 |
+
if f0_filter and sum(f0) == 0:
|
212 |
+
raise F0FilterException("No voice detected")
|
213 |
+
f0 = torch.FloatTensor(f0).to(self.dev)
|
214 |
+
uv = torch.FloatTensor(uv).to(self.dev)
|
215 |
+
|
216 |
+
f0 = f0 * 2 ** (tran / 12)
|
217 |
+
f0 = f0.unsqueeze(0)
|
218 |
+
uv = uv.unsqueeze(0)
|
219 |
+
|
220 |
+
wav16k = librosa.resample(wav, orig_sr=self.target_sample, target_sr=16000)
|
221 |
+
wav16k = torch.from_numpy(wav16k).to(self.dev)
|
222 |
+
c = self.hubert_model.encoder(wav16k)
|
223 |
+
c = utils.repeat_expand_2d(c.squeeze(0), f0.shape[1])
|
224 |
+
|
225 |
+
if cluster_infer_ratio !=0:
|
226 |
+
if self.feature_retrieval:
|
227 |
+
speaker_id = self.spk2id.get(speaker)
|
228 |
+
if speaker_id is None:
|
229 |
+
raise RuntimeError("The name you entered is not in the speaker list!")
|
230 |
+
if not speaker_id and type(speaker) is int:
|
231 |
+
if len(self.spk2id.__dict__) >= speaker:
|
232 |
+
speaker_id = speaker
|
233 |
+
feature_index = self.cluster_model[speaker_id]
|
234 |
+
feat_np = c.transpose(0,1).cpu().numpy()
|
235 |
+
if self.big_npy is not None or self.now_spk_id != speaker_id:
|
236 |
+
self.big_npy = feature_index.reconstruct_n(0, feature_index.ntotal)
|
237 |
+
self.now_spk_id = speaker_id
|
238 |
+
print("starting feature retrieval...")
|
239 |
+
score, ix = feature_index.search(feat_np, k=8)
|
240 |
+
weight = np.square(1 / score)
|
241 |
+
weight /= weight.sum(axis=1, keepdims=True)
|
242 |
+
npy = np.sum(self.big_npy[ix] * np.expand_dims(weight, axis=2), axis=1)
|
243 |
+
c = cluster_infer_ratio * npy + (1 - cluster_infer_ratio) * feat_np
|
244 |
+
c = torch.FloatTensor(c).to(self.dev).transpose(0,1)
|
245 |
+
print("end feature retrieval...")
|
246 |
+
else:
|
247 |
+
cluster_c = cluster.get_cluster_center_result(self.cluster_model, c.cpu().numpy().T, speaker).T
|
248 |
+
cluster_c = torch.FloatTensor(cluster_c).to(self.dev)
|
249 |
+
c = cluster_infer_ratio * cluster_c + (1 - cluster_infer_ratio) * c
|
250 |
+
|
251 |
+
c = c.unsqueeze(0)
|
252 |
+
return c, f0, uv
|
253 |
+
|
254 |
+
def infer(self, speaker, tran, raw_path,
|
255 |
+
cluster_infer_ratio=0,
|
256 |
+
auto_predict_f0=False,
|
257 |
+
noice_scale=0.4,
|
258 |
+
f0_filter=False,
|
259 |
+
f0_predictor='pm',
|
260 |
+
enhancer_adaptive_key = 0,
|
261 |
+
cr_threshold = 0.05,
|
262 |
+
k_step = 100,
|
263 |
+
frame = 0,
|
264 |
+
spk_mix = False,
|
265 |
+
second_encoding = False,
|
266 |
+
loudness_envelope_adjustment = 1
|
267 |
+
):
|
268 |
+
wav, sr = librosa.load(raw_path, sr=self.target_sample)
|
269 |
+
if spk_mix:
|
270 |
+
c, f0, uv = self.get_unit_f0(wav, tran, 0, None, f0_filter,f0_predictor,cr_threshold=cr_threshold)
|
271 |
+
n_frames = f0.size(1)
|
272 |
+
sid = speaker[:, frame:frame+n_frames].transpose(0,1)
|
273 |
+
else:
|
274 |
+
speaker_id = self.spk2id.get(speaker)
|
275 |
+
if speaker_id is None:
|
276 |
+
raise RuntimeError("The name you entered is not in the speaker list!")
|
277 |
+
if not speaker_id and type(speaker) is int:
|
278 |
+
if len(self.spk2id.__dict__) >= speaker:
|
279 |
+
speaker_id = speaker
|
280 |
+
sid = torch.LongTensor([int(speaker_id)]).to(self.dev).unsqueeze(0)
|
281 |
+
c, f0, uv = self.get_unit_f0(wav, tran, cluster_infer_ratio, speaker, f0_filter,f0_predictor,cr_threshold=cr_threshold)
|
282 |
+
n_frames = f0.size(1)
|
283 |
+
if "half" in self.net_g_path and torch.cuda.is_available():
|
284 |
+
c = c.half()
|
285 |
+
with torch.no_grad():
|
286 |
+
start = time.time()
|
287 |
+
vol = None
|
288 |
+
if not self.only_diffusion:
|
289 |
+
vol = self.volume_extractor.extract(torch.FloatTensor(wav).to(self.dev)[None,:])[None,:].to(self.dev) if self.vol_embedding else None
|
290 |
+
audio,f0 = self.net_g_ms.infer(c, f0=f0, g=sid, uv=uv, predict_f0=auto_predict_f0, noice_scale=noice_scale,vol=vol)
|
291 |
+
audio = audio[0,0].data.float()
|
292 |
+
audio_mel = self.vocoder.extract(audio[None,:],self.target_sample) if self.shallow_diffusion else None
|
293 |
+
else:
|
294 |
+
audio = torch.FloatTensor(wav).to(self.dev)
|
295 |
+
audio_mel = None
|
296 |
+
if self.only_diffusion or self.shallow_diffusion:
|
297 |
+
vol = self.volume_extractor.extract(audio[None,:])[None,:,None].to(self.dev) if vol==None else vol[:,:,None]
|
298 |
+
if self.shallow_diffusion and second_encoding:
|
299 |
+
audio16k = librosa.resample(audio.detach().cpu().numpy(), orig_sr=self.target_sample, target_sr=16000)
|
300 |
+
audio16k = torch.from_numpy(audio16k).to(self.dev)
|
301 |
+
c = self.hubert_model.encoder(audio16k)
|
302 |
+
c = utils.repeat_expand_2d(c.squeeze(0), f0.shape[1])
|
303 |
+
f0 = f0[:,:,None]
|
304 |
+
c = c.transpose(-1,-2)
|
305 |
+
audio_mel = self.diffusion_model(
|
306 |
+
c,
|
307 |
+
f0,
|
308 |
+
vol,
|
309 |
+
spk_id = sid,
|
310 |
+
spk_mix_dict = None,
|
311 |
+
gt_spec=audio_mel,
|
312 |
+
infer=True,
|
313 |
+
infer_speedup=self.diffusion_args.infer.speedup,
|
314 |
+
method=self.diffusion_args.infer.method,
|
315 |
+
k_step=k_step)
|
316 |
+
audio = self.vocoder.infer(audio_mel, f0).squeeze()
|
317 |
+
if self.nsf_hifigan_enhance:
|
318 |
+
audio, _ = self.enhancer.enhance(
|
319 |
+
audio[None,:],
|
320 |
+
self.target_sample,
|
321 |
+
f0[:,:,None],
|
322 |
+
self.hps_ms.data.hop_length,
|
323 |
+
adaptive_key = enhancer_adaptive_key)
|
324 |
+
if loudness_envelope_adjustment != 1:
|
325 |
+
audio = utils.change_rms(wav,self.target_sample,audio,self.target_sample,loudness_envelope_adjustment)
|
326 |
+
use_time = time.time() - start
|
327 |
+
print("vits use time:{}".format(use_time))
|
328 |
+
return audio, audio.shape[-1], n_frames
|
329 |
+
|
330 |
+
def clear_empty(self):
|
331 |
+
# clean up vram
|
332 |
+
torch.cuda.empty_cache()
|
333 |
+
|
334 |
+
def unload_model(self):
|
335 |
+
# unload model
|
336 |
+
self.net_g_ms = self.net_g_ms.to("cpu")
|
337 |
+
del self.net_g_ms
|
338 |
+
if hasattr(self,"enhancer"):
|
339 |
+
self.enhancer.enhancer = self.enhancer.enhancer.to("cpu")
|
340 |
+
del self.enhancer.enhancer
|
341 |
+
del self.enhancer
|
342 |
+
gc.collect()
|
343 |
+
|
344 |
+
def slice_inference(self,
|
345 |
+
raw_audio_path,
|
346 |
+
spk,
|
347 |
+
tran,
|
348 |
+
slice_db,
|
349 |
+
cluster_infer_ratio,
|
350 |
+
auto_predict_f0,
|
351 |
+
noice_scale,
|
352 |
+
pad_seconds=0.5,
|
353 |
+
clip_seconds=0,
|
354 |
+
lg_num=0,
|
355 |
+
lgr_num =0.75,
|
356 |
+
f0_predictor='pm',
|
357 |
+
enhancer_adaptive_key = 0,
|
358 |
+
cr_threshold = 0.05,
|
359 |
+
k_step = 100,
|
360 |
+
use_spk_mix = False,
|
361 |
+
second_encoding = False,
|
362 |
+
loudness_envelope_adjustment = 1
|
363 |
+
):
|
364 |
+
if use_spk_mix:
|
365 |
+
if len(self.spk2id) == 1:
|
366 |
+
spk = self.spk2id.keys()[0]
|
367 |
+
use_spk_mix = False
|
368 |
+
wav_path = Path(raw_audio_path).with_suffix('.wav')
|
369 |
+
chunks = slicer.cut(wav_path, db_thresh=slice_db)
|
370 |
+
audio_data, audio_sr = slicer.chunks2audio(wav_path, chunks)
|
371 |
+
per_size = int(clip_seconds*audio_sr)
|
372 |
+
lg_size = int(lg_num*audio_sr)
|
373 |
+
lg_size_r = int(lg_size*lgr_num)
|
374 |
+
lg_size_c_l = (lg_size-lg_size_r)//2
|
375 |
+
lg_size_c_r = lg_size-lg_size_r-lg_size_c_l
|
376 |
+
lg = np.linspace(0,1,lg_size_r) if lg_size!=0 else 0
|
377 |
+
|
378 |
+
if use_spk_mix:
|
379 |
+
assert len(self.spk2id) == len(spk)
|
380 |
+
audio_length = 0
|
381 |
+
for (slice_tag, data) in audio_data:
|
382 |
+
aud_length = int(np.ceil(len(data) / audio_sr * self.target_sample))
|
383 |
+
if slice_tag:
|
384 |
+
audio_length += aud_length // self.hop_size
|
385 |
+
continue
|
386 |
+
if per_size != 0:
|
387 |
+
datas = split_list_by_n(data, per_size,lg_size)
|
388 |
+
else:
|
389 |
+
datas = [data]
|
390 |
+
for k,dat in enumerate(datas):
|
391 |
+
pad_len = int(audio_sr * pad_seconds)
|
392 |
+
per_length = int(np.ceil(len(dat) / audio_sr * self.target_sample))
|
393 |
+
a_length = per_length + 2 * pad_len
|
394 |
+
audio_length += a_length // self.hop_size
|
395 |
+
audio_length += len(audio_data)
|
396 |
+
spk_mix_tensor = torch.zeros(size=(len(spk), audio_length)).to(self.dev)
|
397 |
+
for i in range(len(spk)):
|
398 |
+
last_end = None
|
399 |
+
for mix in spk[i]:
|
400 |
+
if mix[3]<0. or mix[2]<0.:
|
401 |
+
raise RuntimeError("mix value must higer Than zero!")
|
402 |
+
begin = int(audio_length * mix[0])
|
403 |
+
end = int(audio_length * mix[1])
|
404 |
+
length = end - begin
|
405 |
+
if length<=0:
|
406 |
+
raise RuntimeError("begin Must lower Than end!")
|
407 |
+
step = (mix[3] - mix[2])/length
|
408 |
+
if last_end is not None:
|
409 |
+
if last_end != begin:
|
410 |
+
raise RuntimeError("[i]EndTime Must Equal [i+1]BeginTime!")
|
411 |
+
last_end = end
|
412 |
+
if step == 0.:
|
413 |
+
spk_mix_data = torch.zeros(length).to(self.dev) + mix[2]
|
414 |
+
else:
|
415 |
+
spk_mix_data = torch.arange(mix[2],mix[3],step).to(self.dev)
|
416 |
+
if(len(spk_mix_data)<length):
|
417 |
+
num_pad = length - len(spk_mix_data)
|
418 |
+
spk_mix_data = torch.nn.functional.pad(spk_mix_data, [0, num_pad], mode="reflect").to(self.dev)
|
419 |
+
spk_mix_tensor[i][begin:end] = spk_mix_data[:length]
|
420 |
+
|
421 |
+
spk_mix_ten = torch.sum(spk_mix_tensor,dim=0).unsqueeze(0).to(self.dev)
|
422 |
+
# spk_mix_tensor[0][spk_mix_ten<0.001] = 1.0
|
423 |
+
for i, x in enumerate(spk_mix_ten[0]):
|
424 |
+
if x == 0.0:
|
425 |
+
spk_mix_ten[0][i] = 1.0
|
426 |
+
spk_mix_tensor[:,i] = 1.0 / len(spk)
|
427 |
+
spk_mix_tensor = spk_mix_tensor / spk_mix_ten
|
428 |
+
if not ((torch.sum(spk_mix_tensor,dim=0) - 1.)<0.0001).all():
|
429 |
+
raise RuntimeError("sum(spk_mix_tensor) not equal 1")
|
430 |
+
spk = spk_mix_tensor
|
431 |
+
|
432 |
+
global_frame = 0
|
433 |
+
audio = []
|
434 |
+
for (slice_tag, data) in audio_data:
|
435 |
+
print(f'#=====segment start, {round(len(data) / audio_sr, 3)}s======')
|
436 |
+
# padd
|
437 |
+
length = int(np.ceil(len(data) / audio_sr * self.target_sample))
|
438 |
+
if slice_tag:
|
439 |
+
print('jump empty segment')
|
440 |
+
_audio = np.zeros(length)
|
441 |
+
audio.extend(list(pad_array(_audio, length)))
|
442 |
+
global_frame += length // self.hop_size
|
443 |
+
continue
|
444 |
+
if per_size != 0:
|
445 |
+
datas = split_list_by_n(data, per_size,lg_size)
|
446 |
+
else:
|
447 |
+
datas = [data]
|
448 |
+
for k,dat in enumerate(datas):
|
449 |
+
per_length = int(np.ceil(len(dat) / audio_sr * self.target_sample)) if clip_seconds!=0 else length
|
450 |
+
if clip_seconds!=0: print(f'###=====segment clip start, {round(len(dat) / audio_sr, 3)}s======')
|
451 |
+
# padd
|
452 |
+
pad_len = int(audio_sr * pad_seconds)
|
453 |
+
dat = np.concatenate([np.zeros([pad_len]), dat, np.zeros([pad_len])])
|
454 |
+
raw_path = io.BytesIO()
|
455 |
+
soundfile.write(raw_path, dat, audio_sr, format="wav")
|
456 |
+
raw_path.seek(0)
|
457 |
+
out_audio, out_sr, out_frame = self.infer(spk, tran, raw_path,
|
458 |
+
cluster_infer_ratio=cluster_infer_ratio,
|
459 |
+
auto_predict_f0=auto_predict_f0,
|
460 |
+
noice_scale=noice_scale,
|
461 |
+
f0_predictor = f0_predictor,
|
462 |
+
enhancer_adaptive_key = enhancer_adaptive_key,
|
463 |
+
cr_threshold = cr_threshold,
|
464 |
+
k_step = k_step,
|
465 |
+
frame = global_frame,
|
466 |
+
spk_mix = use_spk_mix,
|
467 |
+
second_encoding = second_encoding,
|
468 |
+
loudness_envelope_adjustment = loudness_envelope_adjustment
|
469 |
+
)
|
470 |
+
global_frame += out_frame
|
471 |
+
_audio = out_audio.cpu().numpy()
|
472 |
+
pad_len = int(self.target_sample * pad_seconds)
|
473 |
+
_audio = _audio[pad_len:-pad_len]
|
474 |
+
_audio = pad_array(_audio, per_length)
|
475 |
+
if lg_size!=0 and k!=0:
|
476 |
+
lg1 = audio[-(lg_size_r+lg_size_c_r):-lg_size_c_r] if lgr_num != 1 else audio[-lg_size:]
|
477 |
+
lg2 = _audio[lg_size_c_l:lg_size_c_l+lg_size_r] if lgr_num != 1 else _audio[0:lg_size]
|
478 |
+
lg_pre = lg1*(1-lg)+lg2*lg
|
479 |
+
audio = audio[0:-(lg_size_r+lg_size_c_r)] if lgr_num != 1 else audio[0:-lg_size]
|
480 |
+
audio.extend(lg_pre)
|
481 |
+
_audio = _audio[lg_size_c_l+lg_size_r:] if lgr_num != 1 else _audio[lg_size:]
|
482 |
+
audio.extend(list(_audio))
|
483 |
+
return np.array(audio)
|
484 |
+
|
485 |
+
class RealTimeVC:
|
486 |
+
def __init__(self):
|
487 |
+
self.last_chunk = None
|
488 |
+
self.last_o = None
|
489 |
+
self.chunk_len = 16000 # chunk length
|
490 |
+
self.pre_len = 3840 # cross fade length, multiples of 640
|
491 |
+
|
492 |
+
# Input and output are 1-dimensional numpy waveform arrays
|
493 |
+
|
494 |
+
def process(self, svc_model, speaker_id, f_pitch_change, input_wav_path,
|
495 |
+
cluster_infer_ratio=0,
|
496 |
+
auto_predict_f0=False,
|
497 |
+
noice_scale=0.4,
|
498 |
+
f0_filter=False):
|
499 |
+
|
500 |
+
import maad
|
501 |
+
audio, sr = torchaudio.load(input_wav_path)
|
502 |
+
audio = audio.cpu().numpy()[0]
|
503 |
+
temp_wav = io.BytesIO()
|
504 |
+
if self.last_chunk is None:
|
505 |
+
input_wav_path.seek(0)
|
506 |
+
|
507 |
+
audio, sr = svc_model.infer(speaker_id, f_pitch_change, input_wav_path,
|
508 |
+
cluster_infer_ratio=cluster_infer_ratio,
|
509 |
+
auto_predict_f0=auto_predict_f0,
|
510 |
+
noice_scale=noice_scale,
|
511 |
+
f0_filter=f0_filter)
|
512 |
+
|
513 |
+
audio = audio.cpu().numpy()
|
514 |
+
self.last_chunk = audio[-self.pre_len:]
|
515 |
+
self.last_o = audio
|
516 |
+
return audio[-self.chunk_len:]
|
517 |
+
else:
|
518 |
+
audio = np.concatenate([self.last_chunk, audio])
|
519 |
+
soundfile.write(temp_wav, audio, sr, format="wav")
|
520 |
+
temp_wav.seek(0)
|
521 |
+
|
522 |
+
audio, sr = svc_model.infer(speaker_id, f_pitch_change, temp_wav,
|
523 |
+
cluster_infer_ratio=cluster_infer_ratio,
|
524 |
+
auto_predict_f0=auto_predict_f0,
|
525 |
+
noice_scale=noice_scale,
|
526 |
+
f0_filter=f0_filter)
|
527 |
+
|
528 |
+
audio = audio.cpu().numpy()
|
529 |
+
ret = maad.util.crossfade(self.last_o, audio, self.pre_len)
|
530 |
+
self.last_chunk = audio[-self.pre_len:]
|
531 |
+
self.last_o = audio
|
532 |
+
return ret[self.chunk_len:2 * self.chunk_len]
|
533 |
+
|