Add training scripts and weights

.gitattributes

@@ -30,3 +30,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+*.nemo filter=lfs diff=lfs merge=lfs -text

README.md

@@ -0,0 +1,44 @@
+---
+language:
+- en
+tags:
+- esc
+datasets:
+- spgispeech
+---
+To reproduce this run, execute: 
+```python 
+#!/usr/bin/env bash
+CUDA_VISIBLE_DEVICES=0 python run_speech_recognition_rnnt.py \
+        --config_path="conf/conformer_transducer_bpe_xlarge.yaml" \
+        --model_name_or_path="stt_en_conformer_transducer_xlarge" \
+        --dataset_name="esc-benchmark/esc-datasets" \
+        --tokenizer_path="tokenizer" \
+        --vocab_size="1024" \
+        --max_steps="100000" \
+        --dataset_config_name="spgispeech" \
+        --output_dir="./" \
+        --run_name="conformer-rnnt-spgispeech" \
+        --wandb_project="rnnt" \
+        --per_device_train_batch_size="8" \
+        --per_device_eval_batch_size="4" \
+        --logging_steps="50" \
+        --learning_rate="1e-4" \
+        --warmup_steps="500" \
+        --save_strategy="steps" \
+        --save_steps="20000" \
+        --evaluation_strategy="steps" \
+        --eval_steps="20000" \
+        --report_to="wandb" \
+        --preprocessing_num_workers="4" \
+        --fused_batch_size="4" \
+        --length_column_name="input_lengths" \
+        --fuse_loss_wer \
+        --group_by_length \
+        --overwrite_output_dir \
+        --do_train \
+        --do_eval \
+        --do_predict \
+        --use_auth_token
+
+```

models/__init__.py

@@ -0,0 +1 @@
+from .modeling_rnnt import RNNTBPEModel

models/modeling_rnnt.py

@@ -0,0 +1,126 @@
+from dataclasses import dataclass
+from typing import Optional
+
+import torch
+from nemo.collections.asr.models import EncDecRNNTBPEModel
+from omegaconf import DictConfig
+from transformers.utils import ModelOutput
+
+
+@dataclass
+class RNNTOutput(ModelOutput):
+    """
+    Base class for RNNT outputs.
+    """
+
+    loss: Optional[torch.FloatTensor] = None
+    wer: Optional[float] = None
+    wer_num: Optional[float] = None
+    wer_denom: Optional[float] = None
+
+
+# Adapted from https://github.com/NVIDIA/NeMo/blob/66c7677cd4a68d78965d4905dd1febbf5385dff3/nemo/collections/asr/models/rnnt_bpe_models.py#L33
+class RNNTBPEModel(EncDecRNNTBPEModel):
+    def __init__(self, cfg: DictConfig):
+        super().__init__(cfg=cfg, trainer=None)
+
+    def encoding(
+            self, input_signal=None, input_signal_length=None, processed_signal=None, processed_signal_length=None
+    ):
+        """
+        Forward pass of the acoustic model. Note that for RNNT Models, the forward pass of the model is a 3 step process,
+        and this method only performs the first step - forward of the acoustic model.
+
+        Please refer to the `forward` in order to see the full `forward` step for training - which
+        performs the forward of the acoustic model, the prediction network and then the joint network.
+        Finally, it computes the loss and possibly compute the detokenized text via the `decoding` step.
+
+        Please refer to the `validation_step` in order to see the full `forward` step for inference - which
+        performs the forward of the acoustic model, the prediction network and then the joint network.
+        Finally, it computes the decoded tokens via the `decoding` step and possibly compute the batch metrics.
+
+        Args:
+            input_signal: Tensor that represents a batch of raw audio signals,
+                of shape [B, T]. T here represents timesteps, with 1 second of audio represented as
+                `self.sample_rate` number of floating point values.
+            input_signal_length: Vector of length B, that contains the individual lengths of the audio
+                sequences.
+            processed_signal: Tensor that represents a batch of processed audio signals,
+                of shape (B, D, T) that has undergone processing via some DALI preprocessor.
+            processed_signal_length: Vector of length B, that contains the individual lengths of the
+                processed audio sequences.
+
+        Returns:
+            A tuple of 2 elements -
+            1) The log probabilities tensor of shape [B, T, D].
+            2) The lengths of the acoustic sequence after propagation through the encoder, of shape [B].
+        """
+        has_input_signal = input_signal is not None and input_signal_length is not None
+        has_processed_signal = processed_signal is not None and processed_signal_length is not None
+        if (has_input_signal ^ has_processed_signal) is False:
+            raise ValueError(
+                f"{self} Arguments ``input_signal`` and ``input_signal_length`` are mutually exclusive "
+                " with ``processed_signal`` and ``processed_signal_len`` arguments."
+            )
+
+        if not has_processed_signal:
+            processed_signal, processed_signal_length = self.preprocessor(
+                input_signal=input_signal, length=input_signal_length,
+            )
+
+        # Spec augment is not applied during evaluation/testing
+        if self.spec_augmentation is not None and self.training:
+            processed_signal = self.spec_augmentation(input_spec=processed_signal, length=processed_signal_length)
+
+        encoded, encoded_len = self.encoder(audio_signal=processed_signal, length=processed_signal_length)
+        return encoded, encoded_len
+
+    def forward(self, input_ids, input_lengths=None, labels=None, label_lengths=None):
+        # encoding() only performs encoder forward
+        encoded, encoded_len = self.encoding(input_signal=input_ids, input_signal_length=input_lengths)
+        del input_ids
+
+        # During training, loss must be computed, so decoder forward is necessary
+        decoder, target_length, states = self.decoder(targets=labels, target_length=label_lengths)
+
+        # If experimental fused Joint-Loss-WER is not used
+        if not self.joint.fuse_loss_wer:
+            # Compute full joint and loss
+            joint = self.joint(encoder_outputs=encoded, decoder_outputs=decoder)
+            loss_value = self.loss(
+                log_probs=joint, targets=labels, input_lengths=encoded_len, target_lengths=target_length
+            )
+            # Add auxiliary losses, if registered
+            loss_value = self.add_auxiliary_losses(loss_value)
+            wer = wer_num = wer_denom = None
+            if not self.training:
+                self.wer.update(encoded, encoded_len, labels, target_length)
+                wer, wer_num, wer_denom = self.wer.compute()
+                self.wer.reset()
+
+        else:
+            # If experimental fused Joint-Loss-WER is used
+            # Fused joint step
+            loss_value, wer, wer_num, wer_denom = self.joint(
+                encoder_outputs=encoded,
+                decoder_outputs=decoder,
+                encoder_lengths=encoded_len,
+                transcripts=labels,
+                transcript_lengths=label_lengths,
+                compute_wer=not self.training,
+            )
+            # Add auxiliary losses, if registered
+            loss_value = self.add_auxiliary_losses(loss_value)
+
+        return RNNTOutput(loss=loss_value, wer=wer, wer_num=wer_num, wer_denom=wer_denom)
+
+    def transcribe(self, input_ids, input_lengths=None, labels=None, label_lengths=None, return_hypotheses: bool = False, partial_hypothesis: Optional = None):
+        encoded, encoded_len = self.encoding(input_signal=input_ids, input_signal_length=input_lengths)
+        del input_ids
+        best_hyp, all_hyp = self.decoding.rnnt_decoder_predictions_tensor(
+            encoded,
+            encoded_len,
+            return_hypotheses=return_hypotheses,
+            partial_hypotheses=partial_hypothesis,
+        )
+        return best_hyp, all_hyp

run_speech_recognition_rnnt.py

@@ -0,0 +1,789 @@
+#!/usr/bin/env python
+# coding=utf-8
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Fine-tuning NVIDIA RNN-T models for speech recognition.
+"""
+# You can also adapt this script on your own sequence to sequence task. Pointers for this are left as comments.
+import copy
+import logging
+import os
+import sys
+from dataclasses import dataclass, field
+
+import wandb
+from torch.utils.data import Dataset
+from tqdm import tqdm
+import json
+from typing import Optional, Dict, Union, List, Any
+
+import numpy as np
+import torch
+
+from omegaconf import OmegaConf
+from models import RNNTBPEModel
+
+import datasets
+from datasets import DatasetDict, load_dataset, load_metric
+import transformers
+from transformers import (
+    HfArgumentParser,
+    Seq2SeqTrainingArguments,
+    set_seed,
+    Trainer,
+)
+from transformers.trainer_utils import get_last_checkpoint, is_main_process
+from transformers.utils import check_min_version
+from transformers.utils.versions import require_version
+
+from process_asr_text_tokenizer import __process_data as nemo_process_data, \
+    __build_document_from_manifests as nemo_build_document_from_manifests
+
+
+# Will error if the minimal version of Transformers is not installed. Remove at your own risks.
+check_min_version("4.17.0.dev0")
+
+require_version("datasets>=1.18.0", "To fix: pip install -r examples/pytorch/speech-recognition/requirements.txt")
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass
+class ModelArguments:
+    """
+    Arguments pertaining to which model/config/tokenizer we are going to fine-tune from.
+    """
+
+    config_path: str = field(
+        metadata={"help": "Path to pretrained model or model identifier from huggingface.co/models."},
+    )
+    model_name_or_path: Optional[str] = field(
+        default=None,
+        metadata={"help": "Path to pretrained model or model identifier from NVIDIA NeMo NGC."}
+    )
+    pretrained_model_name_or_path: Optional[str] = field(
+        default=None,
+        metadata={"help": "Path to local pretrained model or model identifier."}
+    )
+    cache_dir: Optional[str] = field(
+        default=None,
+        metadata={"help": "Where to store the pretrained models downloaded from huggingface.co or NVIDIA NeMo NGC."},
+    )
+    use_auth_token: bool = field(
+        default=False,
+        metadata={
+            "help": "Will use the token generated when running `transformers-cli login` (necessary to use this script "
+                    "with private models)."
+        },
+    )
+    manifest_path: str = field(
+        default="data",
+        metadata={
+            "help": "Manifest path."
+        },
+    )
+    tokenizer_path: str = field(
+        default="tokenizers",
+        metadata={
+            "help": "Tokenizer path."
+        },
+    )
+    vocab_size: int = field(
+        default=1024,
+        metadata={"help": "Tokenizer vocab size."}
+    )
+    tokenizer_type: str = field(
+        default="spe",
+        metadata={
+            "help": "Can be either spe or wpe. spe refers to the Google sentencepiece library tokenizer."
+                    "wpe refers to the HuggingFace BERT Word Piece tokenizer."
+        },
+    )
+    spe_type: str = field(
+        default="bpe",
+        metadata={
+            "help": "Type of the SentencePiece model. Can be `bpe`, `unigram`, `char` or `word`."
+                    "Used only if `tokenizer_type` == `spe`"
+        },
+    )
+    cutoff_freq: str = field(
+        default=0.001,
+        metadata={"help": "Drop the least frequent chars from the train set when building the tokenizer."}
+    )
+    fuse_loss_wer: bool = field(
+        default=True,
+        metadata={
+            "help": "Whether to fuse the computation of prediction net + joint net + loss + WER calculation to be run "
+                    "on sub-batches of size `fused_batch_size`"
+        }
+    )
+    fused_batch_size: int = field(
+        default=8,
+        metadata={
+            "help": "`fused_batch_size` is the actual batch size of the prediction net, joint net and transducer loss."
+                    "Using small values here will preserve a lot of memory during training, but will make training slower as well."
+                    "An optimal ratio of fused_batch_size : per_device_train_batch_size is 1:1."
+                    "However, to preserve memory, this ratio can be 1:8 or even 1:16."
+        }
+    )
+    final_decoding_strategy: str = field(
+        default="greedy_batch",
+        metadata={
+            "help": "Decoding strategy for final eval/prediction steps. One of: [`greedy`, `greedy_batch`, `beam`, "
+                    "`tsd`, `alsd`]."
+        }
+    )
+    final_num_beams: int = field(
+        default=1,
+        metadata={
+            "help": "Number of beams for final eval/prediction steps. Increase beam size for better scores, "
+                    "but it will take much longer for transcription!"
+        }
+    )
+
+
+@dataclass
+class DataTrainingArguments:
+    """
+    Arguments pertaining to what data we are going to input our model for training and eval.
+    """
+
+    dataset_name: str = field(
+        default=None, metadata={"help": "The name of the dataset to use (via the datasets library)."}
+    )
+    dataset_config_name: Optional[str] = field(
+        default=None, metadata={"help": "The configuration name of the dataset to use (via the datasets library)."}
+    )
+    text_column: Optional[str] = field(
+        default=None,
+        metadata={"help": "The name of the column in the datasets containing the full texts (for summarization)."},
+    )
+    dataset_cache_dir: Optional[str] = field(
+        default=None, metadata={"help": "Path to cache directory for saving and loading datasets"}
+    )
+    overwrite_cache: bool = field(
+        default=False, metadata={"help": "Overwrite the cached training and evaluation sets"}
+    )
+    preprocessing_num_workers: Optional[int] = field(
+        default=None,
+        metadata={"help": "The number of processes to use for the preprocessing."},
+    )
+    max_train_samples: Optional[int] = field(
+        default=None,
+        metadata={
+            "help": "For debugging purposes or quicker training, truncate the number of training examples to this "
+                    "value if set."
+        },
+    )
+    max_eval_samples: Optional[int] = field(
+        default=None,
+        metadata={
+            "help": "For debugging purposes or quicker training, truncate the number of evaluation examples to this "
+                    "value if set."
+        },
+    )
+    max_predict_samples: Optional[int] = field(
+        default=None,
+        metadata={
+            "help": "For debugging purposes or quicker training, truncate the number of test examples to this "
+                    "value if set."
+        },
+    )
+    audio_column_name: str = field(
+        default="audio",
+        metadata={"help": "The name of the dataset column containing the audio data. Defaults to 'audio'"},
+    )
+    text_column_name: str = field(
+        default="text",
+        metadata={"help": "The name of the dataset column containing the text data. Defaults to 'text'"},
+    )
+    max_duration_in_seconds: float = field(
+        default=20.0,
+        metadata={
+            "help": "Truncate training audio files that are longer than `max_duration_in_seconds` seconds to 'max_duration_in_seconds`"
+        },
+    )
+    min_duration_in_seconds: float = field(
+        default=0.0, metadata={"help": "Filter audio files that are shorter than `min_duration_in_seconds` seconds"}
+    )
+    max_eval_duration_in_seconds: float = field(
+        default=None,
+        metadata={
+            "help": "Truncate eval/test audio files that are longer than `max_duration_in_seconds` seconds to 'max_duration_in_seconds`"
+        },
+    )
+    max_target_length: Optional[int] = field(
+        default=128,
+        metadata={
+            "help": "The maximum total sequence length for target text after tokenization. Sequences longer "
+                    "than this will be truncated, sequences shorter will be padded."
+        },
+    )
+    min_target_length: Optional[int] = field(
+        default=2,
+        metadata={
+            "help": "The minimum total sequence length for target text after tokenization. Sequences shorter "
+                    "than this will be filtered."
+        },
+    )
+    preprocessing_only: bool = field(
+        default=False,
+        metadata={
+            "help": "Whether to only do data preprocessing and skip training. "
+                    "This is especially useful when data preprocessing errors out in distributed training due to timeout. "
+                    "In this case, one should run the preprocessing in a non-distributed setup with `preprocessing_only=True` "
+                    "so that the cached datasets can consequently be loaded in distributed training"
+        },
+    )
+    train_split_name: str = field(
+        default="train",
+        metadata={
+            "help": "The name of the training data set split to use (via the datasets library). Defaults to 'train'"
+        },
+    )
+    eval_split_name: str = field(
+        default="validation",
+        metadata={
+            "help": "The name of the evaluation data set split to use (via the datasets library). Defaults to 'validation'"
+        },
+    )
+    test_split_name: str = field(
+        default="test",
+        metadata={"help": "The name of the test data set split to use (via the datasets library). Defaults to 'test'"},
+    )
+    do_lower_case: bool = field(
+        default=True,
+        metadata={"help": "Whether the target text should be lower cased."},
+    )
+    wandb_project: str = field(
+        default="speech-recognition-rnnt",
+        metadata={"help": "The name of the wandb project."},
+    )
+
+
+def write_wandb_pred(pred_str, label_str, prefix="eval"):
+    # convert str data to a wandb compatible format
+    str_data = [[label_str[i], pred_str[i]] for i in range(len(pred_str))]
+    # we'll log all predictions for the last epoch
+    wandb.log(
+        {
+            f"{prefix}/predictions": wandb.Table(
+                columns=["label_str", "pred_str"], data=str_data
+            )
+        },
+    )
+
+
+def build_tokenizer(model_args, data_args, manifests):
+    """
+    Function to build a NeMo tokenizer from manifest file(s).
+    Copied from https://github.com/NVIDIA/NeMo/blob/66c7677cd4a68d78965d4905dd1febbf5385dff3/scripts/tokenizers/process_asr_text_tokenizer.py#L268
+    """
+    data_root = model_args.tokenizer_path
+    if isinstance(manifests, list):
+        joint_manifests = ",".join(manifests)
+    else:
+        joint_manifests = manifests
+    vocab_size = model_args.vocab_size
+    tokenizer = model_args.tokenizer_type
+    spe_type = model_args.spe_type
+    if not 0 <= model_args.cutoff_freq < 1:
+        raise ValueError(f"`cutoff_freq` must be between zero and one, got {model_args.cutoff_freq}")
+    spe_character_coverage = 1 - model_args.cutoff_freq
+
+    logger.info("Building tokenizer...")
+    if not os.path.exists(data_root):
+        os.makedirs(data_root)
+
+    text_corpus_path = nemo_build_document_from_manifests(data_root, joint_manifests)
+
+    tokenizer_path = nemo_process_data(
+        text_corpus_path,
+        data_root,
+        vocab_size,
+        tokenizer,
+        spe_type,
+        lower_case=data_args.do_lower_case,
+        spe_character_coverage=spe_character_coverage,
+        spe_sample_size=-1,
+        spe_train_extremely_large_corpus=False,
+        spe_max_sentencepiece_length=-1,
+        spe_bos=False,
+        spe_eos=False,
+        spe_pad=False,
+    )
+
+    print("Serialized tokenizer at location :", tokenizer_path)
+    logger.info('Done!')
+
+    # Tokenizer path
+    if tokenizer == 'spe':
+        tokenizer_dir = os.path.join(data_root, f"tokenizer_spe_{spe_type}_v{vocab_size}")
+        tokenizer_type_cfg = "bpe"
+    else:
+        tokenizer_dir = os.path.join(data_root, f"tokenizer_wpe_v{vocab_size}")
+        tokenizer_type_cfg = "wpe"
+
+    return tokenizer_dir, tokenizer_type_cfg
+
+
+def NeMoDataCollator(features: List[Dict[str, Union[List[int], torch.Tensor]]]) -> Dict[str, torch.Tensor]:
+    """
+    Data collator that will dynamically pad the inputs received.
+    Since NeMo models don't have a HF processor defined (feature extractor + tokenizer), we'll pad by hand...
+    The padding idx is arbitrary: we provide the model with the input lengths and label lengths, from which
+    all the relevant padding information is inferred. Thus, we'll use the default np.pad padding idx (0).
+    """
+    # split inputs and labels since they have to be of different lengths
+    # and need different padding methods
+    input_ids = [feature["input_ids"] for feature in features]
+    labels = [feature["labels"] for feature in features]
+
+    # first, pad the audio inputs to max_len
+    input_lengths = [feature["input_lengths"] for feature in features]
+    max_input_len = max(input_lengths)
+    input_ids = [np.pad(input_val, (0, max_input_len - input_len), 'constant') for input_val, input_len in
+                 zip(input_ids, input_lengths)]
+
+    # next, pad the target labels to max_len
+    label_lengths = [len(lab) for lab in labels]
+    max_label_len = max(label_lengths)
+    labels = [np.pad(lab, (0, max_label_len - lab_len), 'constant') for lab, lab_len in zip(labels, label_lengths)]
+
+    batch = {"input_lengths": input_lengths, "labels": labels, "label_lengths": label_lengths}
+
+    # return batch as a pt tensor (list -> np.array -> torch.tensor)
+    batch = {k: torch.tensor(np.array(v), requires_grad=False) for k, v in batch.items()}
+
+    # leave all ints as are, convert float64 to pt float
+    batch["input_ids"] = torch.tensor(np.array(input_ids, dtype=np.float32), requires_grad=False)
+
+    return batch
+
+
+def main():
+    # See all possible arguments in src/transformers/training_args.py
+    # or by passing the --help flag to this script.
+    # We now keep distinct sets of args, for a cleaner separation of concerns.
+
+    parser = HfArgumentParser((ModelArguments, DataTrainingArguments, Seq2SeqTrainingArguments))
+    if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
+        # If we pass only one argument to the script and it's the path to a json file,
+        # let's parse it to get our arguments.
+        model_args, data_args, training_args = parser.parse_json_file(json_file=os.path.abspath(sys.argv[1]))
+    else:
+        model_args, data_args, training_args = parser.parse_args_into_dataclasses()
+
+    # Set wandb project ID before instantiating the Trainer
+    os.environ["WANDB_PROJECT"] = data_args.wandb_project
+
+    # Detecting last checkpoint.
+    last_checkpoint = None
+    if os.path.isdir(training_args.output_dir) and training_args.do_train and not training_args.overwrite_output_dir:
+        last_checkpoint = get_last_checkpoint(training_args.output_dir)
+        if last_checkpoint is None and len(os.listdir(training_args.output_dir)) > 0:
+            raise ValueError(
+                f"Output directory ({training_args.output_dir}) already exists and is not empty. "
+                "Use --overwrite_output_dir to overcome."
+            )
+        elif last_checkpoint is not None:
+            logger.info(
+                f"Checkpoint detected, resuming training at {last_checkpoint}. To avoid this behavior, change "
+                "the `--output_dir` or add `--overwrite_output_dir` to train from scratch."
+            )
+
+    # Setup logging
+    logging.basicConfig(
+        format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
+        datefmt="%m/%d/%Y %H:%M:%S",
+        handlers=[logging.StreamHandler(sys.stdout)],
+    )
+    logger.setLevel(logging.INFO if is_main_process(training_args.local_rank) else logging.WARN)
+
+    # Log on each process the small summary:
+    logger.warning(
+        f"Process rank: {training_args.local_rank}, device: {training_args.device}, n_gpu: {training_args.n_gpu}"
+        f"distributed training: {bool(training_args.local_rank != -1)}, 16-bits training: {training_args.fp16}"
+    )
+    # Set the verbosity to info of the Transformers logger (on main process only):
+    if is_main_process(training_args.local_rank):
+        transformers.utils.logging.set_verbosity_info()
+    logger.info("Training/evaluation parameters %s", training_args)
+
+    # Set seed before initializing model.
+    set_seed(training_args.seed)
+
+    # load the model config (discarding optimiser and trainer attributes)
+    config = OmegaConf.load(model_args.config_path).model
+
+    # 4. Load dataset
+    raw_datasets = DatasetDict()
+
+    if training_args.do_train:
+        raw_datasets["train"] = load_dataset(
+            data_args.dataset_name,
+            data_args.dataset_config_name,
+            split=data_args.train_split_name,
+            cache_dir=data_args.dataset_cache_dir,
+            use_auth_token=True if model_args.use_auth_token else None,
+        )
+
+    if training_args.do_eval:
+        raw_datasets["eval"] = load_dataset(
+            data_args.dataset_name,
+            data_args.dataset_config_name,
+            split=data_args.eval_split_name,
+            cache_dir=data_args.dataset_cache_dir,
+            use_auth_token=True if model_args.use_auth_token else None,
+        )
+
+    if training_args.do_predict:
+        test_split = data_args.test_split_name.split("+")
+        for split in test_split:
+            raw_datasets[split] = load_dataset(
+                data_args.dataset_name,
+                data_args.dataset_config_name,
+                split=split,
+                cache_dir=data_args.dataset_cache_dir,
+                use_auth_token=True if model_args.use_auth_token else None,
+            )
+
+    if not training_args.do_train and not training_args.do_eval and not training_args.do_predict:
+        raise ValueError(
+            "Cannot not train, not do evaluation and not do prediction. At least one of "
+            "training, evaluation or prediction has to be done."
+        )
+
+    # if not training, there is no need to run multiple epochs
+    if not training_args.do_train:
+        training_args.num_train_epochs = 1
+
+    if data_args.audio_column_name not in next(iter(raw_datasets.values())).column_names:
+        raise ValueError(
+            f"--audio_column_name '{data_args.audio_column_name}' not found in dataset '{data_args.dataset_name}'. "
+            "Make sure to set `--audio_column_name` to the correct audio column - one of "
+            f"{', '.join(next(iter(raw_datasets.values())).column_names)}."
+        )
+
+    if data_args.text_column_name not in next(iter(raw_datasets.values())).column_names:
+        raise ValueError(
+            f"--text_column_name {data_args.text_column_name} not found in dataset '{data_args.dataset_name}'. "
+            "Make sure to set `--text_column_name` to the correct text column - one of "
+            f"{', '.join(next(iter(raw_datasets.values())).column_names)}."
+        )
+
+    # 6. Resample speech dataset ALWAYS
+    raw_datasets = raw_datasets.cast_column(
+            data_args.audio_column_name, datasets.features.Audio(sampling_rate=config.sample_rate)
+    )
+
+    # 7. Preprocessing the datasets.
+    # We need to read the audio files as arrays and tokenize the targets.
+    max_input_length = int(data_args.max_duration_in_seconds * config.sample_rate)
+    min_input_length = max(int(data_args.min_duration_in_seconds * config.sample_rate), 1)
+    max_eval_input_length = int(data_args.max_eval_duration_in_seconds * config.sample_rate) if data_args.max_eval_duration_in_seconds else None
+    audio_column_name = data_args.audio_column_name
+    num_workers = data_args.preprocessing_num_workers
+    text_column_name = data_args.text_column_name
+
+    if training_args.do_train and data_args.max_train_samples is not None:
+        raw_datasets["train"] = raw_datasets["train"].select(range(data_args.max_train_samples))
+
+    if training_args.do_eval and data_args.max_eval_samples is not None:
+        raw_datasets["eval"] = raw_datasets["eval"].select(range(data_args.max_eval_samples))
+
+    if training_args.do_predict and data_args.max_predict_samples is not None:
+        for split in test_split:
+            raw_datasets[split] = raw_datasets[split].select(range(data_args.max_predict_samples))
+
+    # Function to build a NeMo tokenizer manifest from a HF dataset
+    # TODO: with a bit of hacking around we can probably bypass this step entirely
+    def build_manifest(ds, manifest_path):
+        with open(manifest_path, 'w') as fout:
+            for sample in tqdm(ds[text_column_name]):
+                # Write the metadata to the manifest
+                metadata = {
+                    "text": sample
+                }
+                json.dump(metadata, fout)
+                fout.write('\n')
+
+    config.train_ds = config.validation_ds = config.test_ds = None
+
+    if not os.path.exists(model_args.manifest_path) and training_args.do_train:
+        os.makedirs(model_args.manifest_path)
+        manifest = os.path.join(model_args.manifest_path, "train.json")
+        logger.info(f"Building training manifest at {manifest}")
+        build_manifest(raw_datasets["train"], manifest)
+    else:
+        manifest = os.path.join(model_args.manifest_path, "train.json")
+        logger.info(f"Re-using training manifest at {manifest}")
+
+    tokenizer_dir, tokenizer_type_cfg = build_tokenizer(model_args, data_args, manifest)
+
+    # generalise the script later to load a pre-built tokenizer for eval only
+    config.tokenizer.dir = tokenizer_dir
+    config.tokenizer.type = tokenizer_type_cfg
+
+    # possibly fused-computation of prediction net + joint net + loss + WER calculation
+    config.joint.fuse_loss_wer = model_args.fuse_loss_wer
+    if model_args.fuse_loss_wer:
+        config.joint.fused_batch_size = model_args.fused_batch_size
+
+    if model_args.model_name_or_path is not None:
+        # load pre-trained model weights
+        model = RNNTBPEModel.from_pretrained(model_args.model_name_or_path, override_config_path=config,
+                                             map_location="cpu")
+        model.save_name = model_args.model_name_or_path
+
+        pretrained_decoder = model.decoder.state_dict()
+        pretrained_joint = model.joint.state_dict()
+        model.change_vocabulary(new_tokenizer_dir=tokenizer_dir, new_tokenizer_type=tokenizer_type_cfg)
+
+        # TODO: add checks for loading decoder/joint state dict
+        model.decoder.load_state_dict(pretrained_decoder)
+        model.joint.load_state_dict(pretrained_joint)
+
+    elif model_args.pretrained_model_name_or_path is not None:
+        model = RNNTBPEModel.restore_from(model_args.pretrained_model_name_or_path, override_config_path=config,
+                                          map_location="cpu")
+        model.save_name = model_args.config_path.split("/")[-1].split(".")[0]
+
+    else:
+        model = RNNTBPEModel(cfg=config)
+        model.save_name = model_args.config_path.split("/")[-1].split(".")[0]
+        model.change_vocabulary(new_tokenizer_dir=tokenizer_dir, new_tokenizer_type=tokenizer_type_cfg)
+
+    # now that we have our model and tokenizer defined, we can tokenize the text data
+    tokenizer = model.tokenizer.tokenizer.encode_as_ids
+
+    def prepare_dataset(batch):
+        # pre-process audio
+        sample = batch[audio_column_name]
+
+        # NeMo RNNT model performs the audio preprocessing in the `.forward()` call
+        # => we only need to supply it with the raw audio values
+        batch["input_ids"] = sample["array"]
+        batch["input_lengths"] = len(sample["array"])
+
+        batch["labels"] = tokenizer(batch[text_column_name])
+        return batch
+
+    vectorized_datasets = raw_datasets.map(
+        prepare_dataset,
+        remove_columns=next(iter(raw_datasets.values())).column_names,
+        num_proc=num_workers,
+        desc="preprocess train dataset",
+    )
+
+    # filter training data with inputs shorter than min_input_length or longer than max_input_length
+    def is_audio_in_length_range(length):
+        return min_input_length < length < max_input_length
+
+    vectorized_datasets = vectorized_datasets.filter(
+            is_audio_in_length_range,
+            num_proc=num_workers,
+            input_columns=["input_lengths"],
+        )
+
+    if max_eval_input_length is not None:
+        # filter training data with inputs longer than max_input_length
+        def is_eval_audio_in_length_range(length):
+            return min_input_length < length < max_eval_input_length
+
+        vectorized_datasets = vectorized_datasets.filter(
+            is_eval_audio_in_length_range,
+            num_proc=num_workers,
+            input_columns=["input_lengths"],
+        )
+
+    # for large datasets it is advised to run the preprocessing on a
+    # single machine first with `args.preprocessing_only` since there will mostly likely
+    # be a timeout when running the script in distributed mode.
+    # In a second step `args.preprocessing_only` can then be set to `False` to load the
+    # cached dataset
+    if data_args.preprocessing_only:
+        cache = {k: v.cache_files for k, v in vectorized_datasets.items()}
+        logger.info(f"Data preprocessing finished. Files cached at {cache}.")
+        return
+
+
+    def compute_metrics(pred):
+        # Tuple of WERs returned by the model during eval: (wer, wer_num, wer_denom)
+        wer_num = pred.predictions[1]
+        wer_denom = pred.predictions[2]
+        # compute WERs over concat batches
+        wer = sum(wer_num) / sum(wer_denom)
+        return {"wer": wer}
+
+
+    class NeMoTrainer(Trainer):
+        def _save(self, output_dir: Optional[str] = None, state_dict=None):
+            # If we are executing this function, we are the process zero, so we don't check for that.
+            output_dir = output_dir if output_dir is not None else self.args.output_dir
+            os.makedirs(output_dir, exist_ok=True)
+            logger.info(f"Saving model checkpoint to {output_dir}")
+            # Save a trained model and configuration using `save_pretrained()`.
+            # They can then be reloaded using `from_pretrained()`
+            self.model.save_to(save_path=os.path.join(output_dir, model.save_name + ".nemo"))
+            # Good practice: save your training arguments together with the trained model
+            torch.save(self.args, os.path.join(output_dir, "training_args.bin"))
+
+        def transcribe(self, test_dataset: Dataset) -> List[Any]:
+            self.model.eval()
+            test_dataloader = self.get_test_dataloader(test_dataset)
+            hypotheses = []
+            for test_batch in tqdm(test_dataloader, desc="Transcribing"):
+                inputs = self._prepare_inputs(test_batch)
+                best_hyp, all_hyp = self.model.transcribe(**inputs)
+                hypotheses += best_hyp
+                del test_batch
+            return hypotheses
+
+
+    # Initialize Trainer
+    trainer = NeMoTrainer(
+        model=model,
+        args=training_args,
+        compute_metrics=compute_metrics,
+        train_dataset=vectorized_datasets['train'] if training_args.do_train else None,
+        eval_dataset=vectorized_datasets['eval'] if training_args.do_eval else None,
+        data_collator=NeMoDataCollator,
+    )
+
+    # 8. Finally, we can start training
+
+    # Training
+    if training_args.do_train:
+
+        # use last checkpoint if exist
+        if last_checkpoint is not None:
+            checkpoint = last_checkpoint
+        elif model_args.model_name_or_path is not None and os.path.isdir(model_args.model_name_or_path):
+            checkpoint = model_args.model_name_or_path
+        else:
+            checkpoint = None
+
+        train_result = trainer.train(resume_from_checkpoint=checkpoint)
+        trainer.save_model()
+
+        metrics = train_result.metrics
+        max_train_samples = (
+            data_args.max_train_samples
+            if data_args.max_train_samples is not None
+            else len(vectorized_datasets["train"])
+        )
+        metrics["train_samples"] = min(max_train_samples, len(vectorized_datasets["train"]))
+
+        trainer.log_metrics("train", metrics)
+        trainer.save_metrics("train", metrics)
+        trainer.save_state()
+
+    # Change decoding strategy for final eval/predict
+    if training_args.do_eval or training_args.do_predict:
+        # set beam search decoding config
+        beam_decoding_config = copy.deepcopy(trainer.model.cfg.decoding)
+        beam_decoding_config.strategy = model_args.final_decoding_strategy
+        beam_decoding_config.beam.beam_size = model_args.final_num_beams
+
+        trainer.model.change_decoding_strategy(beam_decoding_config)
+
+    results = {}
+    if training_args.do_eval:
+        logger.info(f"*** Running Final Evaluation ({model_args.final_decoding_strategy}) ***")
+
+        predictions = trainer.transcribe(vectorized_datasets["eval"])
+        targets = model.tokenizer.ids_to_text(vectorized_datasets["eval"]["labels"])
+
+        cer_metric = load_metric("cer")
+        wer_metric = load_metric("wer")
+
+        cer = cer_metric.compute(predictions=predictions, references=targets)
+        wer = wer_metric.compute(predictions=predictions, references=targets)
+
+        metrics = {f"eval_cer": cer, f"eval_wer": wer}
+
+        max_eval_samples = (
+            data_args.max_eval_samples if data_args.max_eval_samples is not None else len(
+                vectorized_datasets["eval"])
+        )
+        metrics["eval_samples"] = min(max_eval_samples, len(vectorized_datasets["eval"]))
+
+        trainer.log_metrics("eval", metrics)
+        trainer.save_metrics("eval", metrics)
+
+        if "wandb" in training_args.report_to:
+            if not training_args.do_train:
+                wandb.init(name=training_args.run_name, project=data_args.wandb_project)
+            metrics = {os.path.join("eval", k[len("eval") + 1:]): v for k, v in metrics.items()}
+            # wandb.init(project=data_args.wandb_project, name=training_args.run_name)
+            wandb.log(metrics)
+            write_wandb_pred(predictions, targets, prefix="eval")
+
+    if training_args.do_predict:
+        logger.info(f"*** Running Final Prediction ({model_args.final_decoding_strategy}) ***")
+
+        for split in test_split:
+            predictions = trainer.transcribe(vectorized_datasets[split])
+            targets = model.tokenizer.ids_to_text(vectorized_datasets[split]["labels"])
+
+            cer_metric = load_metric("cer")
+            wer_metric = load_metric("wer")
+
+            cer = cer_metric.compute(predictions=predictions, references=targets)
+            wer = wer_metric.compute(predictions=predictions, references=targets)
+
+            metrics = {f"{split}_cer": cer, f"{split}_wer": wer}
+
+            max_predict_samples = (
+                data_args.max_predict_samples if data_args.max_predict_samples is not None else len(
+                    vectorized_datasets[split])
+            )
+            metrics[f"{split}_samples"] = min(max_predict_samples, len(vectorized_datasets[split]))
+
+            trainer.log_metrics(split, metrics)
+            trainer.save_metrics(split, metrics)
+
+            if "wandb" in training_args.report_to:
+                if not training_args.do_train or training_args.do_eval:
+                    wandb.init(name=training_args.run_name, project=data_args.wandb_project)
+                metrics = {os.path.join(split, k[len(split) + 1:]): v for k, v in metrics.items()}
+                wandb.log(metrics)
+                write_wandb_pred(predictions, targets, prefix=split)
+
+    # Write model card and (optionally) push to hub
+    config_name = data_args.dataset_config_name if data_args.dataset_config_name is not None else "na"
+    kwargs = {
+        "finetuned_from": model_args.model_name_or_path,
+        "tasks": "speech-recognition",
+        "tags": ["automatic-speech-recognition", data_args.dataset_name],
+        "dataset_args": (
+            f"Config: {config_name}, Training split: {data_args.train_split_name}, Eval split:"
+            f" {data_args.eval_split_name}"
+        ),
+        "dataset": f"{data_args.dataset_name.upper()} - {config_name.upper()}",
+    }
+    if "common_voice" in data_args.dataset_name:
+        kwargs["language"] = config_name
+
+    if training_args.push_to_hub:
+        trainer.push_to_hub(**kwargs)
+    #else:
+        #trainer.create_model_card(**kwargs)
+
+    return results
+
+
+if __name__ == "__main__":
+    main()

run_spgispeech.sh

@@ -0,0 +1,32 @@
+#!/usr/bin/env bash
+CUDA_VISIBLE_DEVICES=0 python run_speech_recognition_rnnt.py \
+        --config_path="conf/conformer_transducer_bpe_xlarge.yaml" \
+        --model_name_or_path="stt_en_conformer_transducer_xlarge" \
+        --dataset_name="esc-benchmark/esc-datasets" \
+        --tokenizer_path="tokenizer" \
+        --vocab_size="1024" \
+        --max_steps="100000" \
+        --dataset_config_name="spgispeech" \
+        --output_dir="./" \
+        --run_name="conformer-rnnt-spgispeech" \
+        --wandb_project="rnnt" \
+        --per_device_train_batch_size="8" \
+        --per_device_eval_batch_size="4" \
+        --logging_steps="50" \
+        --learning_rate="1e-4" \
+        --warmup_steps="500" \
+        --save_strategy="steps" \
+        --save_steps="20000" \
+        --evaluation_strategy="steps" \
+        --eval_steps="20000" \
+        --report_to="wandb" \
+        --preprocessing_num_workers="4" \
+        --fused_batch_size="4" \
+        --length_column_name="input_lengths" \
+        --fuse_loss_wer \
+        --group_by_length \
+        --overwrite_output_dir \
+        --do_train \
+        --do_eval \
+        --do_predict \
+        --use_auth_token

stt_en_conformer_transducer_xlarge.nemo

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6619304cc9b9924096b38299fc8ad959ff9ba079ac9cc980429a49d1cfc57d1b
+size 2577971200