tokenflow / preprocess_utils.py
Linoy Tsaban
add ddpm inversion (#4)
b34b4e8 verified
from transformers import CLIPTextModel, CLIPTokenizer, logging
from diffusers import AutoencoderKL, UNet2DConditionModel, DDIMScheduler
from diffusers.utils.torch_utils import randn_tensor
# suppress partial model loading warning
logging.set_verbosity_error()
import os
from tqdm import tqdm, trange
import torch
import torch.nn as nn
import argparse
from torchvision.io import write_video
from pathlib import Path
from utils import *
import torchvision.transforms as T
import cv2
import numpy as np
def get_timesteps(scheduler, num_inference_steps, strength, device):
# get the original timestep using init_timestep
init_timestep = min(int(num_inference_steps * strength), num_inference_steps)
t_start = max(num_inference_steps - init_timestep, 0)
timesteps = scheduler.timesteps[t_start:]
return timesteps, num_inference_steps - t_start
class Preprocess(nn.Module):
def __init__(self, device, opt, vae, tokenizer, text_encoder, unet,scheduler, hf_key=None):
super().__init__()
self.device = device
self.sd_version = opt["sd_version"]
self.use_depth = False
self.config = opt
print(f'[INFO] loading stable diffusion...')
if hf_key is not None:
print(f'[INFO] using hugging face custom model key: {hf_key}')
model_key = hf_key
elif self.sd_version == '2.1':
model_key = "stabilityai/stable-diffusion-2-1-base"
elif self.sd_version == '2.0':
model_key = "stabilityai/stable-diffusion-2-base"
elif self.sd_version == '1.5' or self.sd_version == 'ControlNet':
model_key = "runwayml/stable-diffusion-v1-5"
elif self.sd_version == 'depth':
model_key = "stabilityai/stable-diffusion-2-depth"
else:
raise ValueError(f'Stable-diffusion version {self.sd_version} not supported.')
self.model_key = model_key
# Create model
# self.vae = AutoencoderKL.from_pretrained(model_key, subfolder="vae", revision="fp16",
# torch_dtype=torch.float16).to(self.device)
# self.tokenizer = CLIPTokenizer.from_pretrained(model_key, subfolder="tokenizer")
# self.text_encoder = CLIPTextModel.from_pretrained(model_key, subfolder="text_encoder", revision="fp16",
# torch_dtype=torch.float16).to(self.device)
# self.unet = UNet2DConditionModel.from_pretrained(model_key, subfolder="unet", revision="fp16",
# torch_dtype=torch.float16).to(self.device)
self.vae = vae
self.tokenizer = tokenizer
self.text_encoder = text_encoder
self.unet = unet
self.scheduler=scheduler
self.total_inverted_latents = {}
self.noise_total = None # will contain all zs if inversion == 'ddpm', var name chosen to match the save path of zs used in pr https://github.com/omerbt/TokenFlow/pull/24/files#
self.paths, self.frames, self.latents = self.get_data(self.config["data_path"], self.config["n_frames"])
print("self.frames", self.frames.shape)
print("self.latents", self.latents.shape)
if self.sd_version == 'ControlNet':
from diffusers import ControlNetModel, StableDiffusionControlNetPipeline
controlnet = ControlNetModel.from_pretrained("lllyasviel/sd-controlnet-canny", torch_dtype=torch.float16).to(self.device)
control_pipe = StableDiffusionControlNetPipeline.from_pretrained(
"runwayml/stable-diffusion-v1-5", controlnet=controlnet, torch_dtype=torch.float16
).to(self.device)
self.unet = control_pipe.unet
self.controlnet = control_pipe.controlnet
self.canny_cond = self.get_canny_cond()
elif self.sd_version == 'depth':
self.depth_maps = self.prepare_depth_maps()
self.scheduler = scheduler
self.unet.enable_xformers_memory_efficient_attention()
print(f'[INFO] loaded stable diffusion!')
@torch.no_grad()
def prepare_depth_maps(self, model_type='DPT_Large', device='cuda'):
depth_maps = []
midas = torch.hub.load("intel-isl/MiDaS", model_type)
midas.to(device)
midas.eval()
midas_transforms = torch.hub.load("intel-isl/MiDaS", "transforms")
if model_type == "DPT_Large" or model_type == "DPT_Hybrid":
transform = midas_transforms.dpt_transform
else:
transform = midas_transforms.small_transform
for i in range(len(self.paths)):
img = cv2.imread(self.paths[i])
img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
latent_h = img.shape[0] // 8
latent_w = img.shape[1] // 8
input_batch = transform(img).to(device)
prediction = midas(input_batch)
depth_map = torch.nn.functional.interpolate(
prediction.unsqueeze(1),
size=(latent_h, latent_w),
mode="bicubic",
align_corners=False,
)
depth_min = torch.amin(depth_map, dim=[1, 2, 3], keepdim=True)
depth_max = torch.amax(depth_map, dim=[1, 2, 3], keepdim=True)
depth_map = 2.0 * (depth_map - depth_min) / (depth_max - depth_min) - 1.0
depth_maps.append(depth_map)
return torch.cat(depth_maps).to(self.device).to(torch.float16)
@torch.no_grad()
def get_canny_cond(self):
canny_cond = []
for image in self.frames.cpu().permute(0, 2, 3, 1):
image = np.uint8(np.array(255 * image))
low_threshold = 100
high_threshold = 200
image = cv2.Canny(image, low_threshold, high_threshold)
image = image[:, :, None]
image = np.concatenate([image, image, image], axis=2)
image = torch.from_numpy((image.astype(np.float32) / 255.0))
canny_cond.append(image)
canny_cond = torch.stack(canny_cond).permute(0, 3, 1, 2).to(self.device).to(torch.float16)
return canny_cond
def controlnet_pred(self, latent_model_input, t, text_embed_input, controlnet_cond):
down_block_res_samples, mid_block_res_sample = self.controlnet(
latent_model_input,
t,
encoder_hidden_states=text_embed_input,
controlnet_cond=controlnet_cond,
conditioning_scale=1,
return_dict=False,
)
# apply the denoising network
noise_pred = self.unet(
latent_model_input,
t,
encoder_hidden_states=text_embed_input,
cross_attention_kwargs={},
down_block_additional_residuals=down_block_res_samples,
mid_block_additional_residual=mid_block_res_sample,
return_dict=False,
)[0]
return noise_pred
@torch.no_grad()
def encode_text(self, prompts, device=None):
if device is None:
device = self.device
text_inputs = self.tokenizer(
prompts,
padding="max_length",
max_length=self.tokenizer.model_max_length,
return_tensors="pt",
)
text_input_ids = text_inputs.input_ids
if text_input_ids.shape[-1] > self.tokenizer.model_max_length:
removed_text = self.tokenizer.batch_decode(text_input_ids[:, self.tokenizer.model_max_length:])
print(
"The following part of your input was truncated because CLIP can only handle sequences up to"
f" {self.tokenizer.model_max_length} tokens: {removed_text}"
)
text_input_ids = text_input_ids[:, : self.tokenizer.model_max_length]
text_embeddings = self.text_encoder(text_input_ids.to(device))[0]
return text_embeddings
@torch.no_grad()
def get_text_embeds(self, prompt, negative_prompt, device="cuda"):
text_embeddings = self.encode_text(prompt, device=device)
uncond_embeddings = self.encode_text(negative_prompt, device=device)
text_embeddings = torch.cat([uncond_embeddings, text_embeddings])
return text_embeddings
@torch.no_grad()
def decode_latents(self, latents):
decoded = []
batch_size = 8
for b in range(0, latents.shape[0], batch_size):
latents_batch = 1 / 0.18215 * latents[b:b + batch_size]
imgs = self.vae.decode(latents_batch).sample
imgs = (imgs / 2 + 0.5).clamp(0, 1)
decoded.append(imgs)
return torch.cat(decoded)
@torch.no_grad()
def encode_imgs(self, imgs, batch_size=10, deterministic=True):
imgs = 2 * imgs - 1
latents = []
for i in range(0, len(imgs), batch_size):
posterior = self.vae.encode(imgs[i:i + batch_size]).latent_dist
latent = posterior.mean if deterministic else posterior.sample()
latents.append(latent * self.vae.config.scaling_factor)
latents = torch.cat(latents)
return latents
def get_data(self, frames_path, n_frames):
# load frames
if not self.config["frames"]:
paths = [f"{frames_path}/%05d.png" % i for i in range(n_frames)]
print(paths)
if not os.path.exists(paths[0]):
paths = [f"{frames_path}/%05d.jpg" % i for i in range(n_frames)]
self.paths = paths
frames = [Image.open(path).convert('RGB') for path in paths]
if frames[0].size[0] == frames[0].size[1]:
frames = [frame.resize((512, 512), resample=Image.Resampling.LANCZOS) for frame in frames]
else:
frames = self.config["frames"][:n_frames]
frames = torch.stack([T.ToTensor()(frame) for frame in frames]).to(torch.float16).to(self.device)
# encode to latents
latents = self.encode_imgs(frames, deterministic=True).to(torch.float16).to(self.device)
print("frames", frames.shape)
print("latents", latents.shape)
if not self.config["frames"]:
return paths, frames, latents
else:
return None, frames, latents
@torch.no_grad()
def ddim_inversion(self, cond, latent_frames, save_path, batch_size, save_latents=True, timesteps_to_save=None):
timesteps = reversed(self.scheduler.timesteps)
timesteps_to_save = timesteps_to_save if timesteps_to_save is not None else timesteps
return_inverted_latents = self.config["frames"] is not None
for i, t in enumerate(tqdm(timesteps)):
for b in range(0, latent_frames.shape[0], int(batch_size)):
x_batch = latent_frames[b:b + batch_size]
model_input = x_batch
cond_batch = cond.repeat(x_batch.shape[0], 1, 1)
if self.sd_version == 'depth':
depth_maps = torch.cat([self.depth_maps[b: b + batch_size]])
model_input = torch.cat([x_batch, depth_maps],dim=1)
alpha_prod_t = self.scheduler.alphas_cumprod[t]
alpha_prod_t_prev = (
self.scheduler.alphas_cumprod[timesteps[i - 1]]
if i > 0 else self.scheduler.final_alpha_cumprod
)
mu = alpha_prod_t ** 0.5
mu_prev = alpha_prod_t_prev ** 0.5
sigma = (1 - alpha_prod_t) ** 0.5
sigma_prev = (1 - alpha_prod_t_prev) ** 0.5
eps = self.unet(model_input, t, encoder_hidden_states=cond_batch).sample if self.sd_version != 'ControlNet' \
else self.controlnet_pred(x_batch, t, cond_batch, torch.cat([self.canny_cond[b: b + batch_size]]))
pred_x0 = (x_batch - sigma_prev * eps) / mu_prev
latent_frames[b:b + batch_size] = mu * pred_x0 + sigma * eps
if return_inverted_latents and t in timesteps_to_save:
self.total_inverted_latents[f'noisy_latents_{t}'] = latent_frames.clone()
if save_latents and t in timesteps_to_save:
torch.save(latent_frames, os.path.join(save_path, 'latents', f'noisy_latents_{t}.pt'))
if save_latents:
torch.save(latent_frames, os.path.join(save_path, 'latents', f'noisy_latents_{t}.pt'))
if return_inverted_latents:
self.total_inverted_latents[f'noisy_latents_{t}'] = latent_frames.clone()
return latent_frames
@torch.no_grad()
def ddpm_inversion(self, cond,
latent_frames,
batch_size,
num_inversion_steps,
save_path=None,
save_latents=True,
eta: float = 1.0,
skip_steps=20):
timesteps = self.scheduler.timesteps
return_inverted_latents = self.config["frames"] is not None
variance_noise_shape = (
num_inversion_steps,
*latent_frames.shape)
x0 = latent_frames
t_to_idx = {int(v): k for k, v in enumerate(timesteps)}
xts = torch.zeros(size=variance_noise_shape, device=self.device, dtype=cond.dtype)
for t in reversed(timesteps):
idx = t_to_idx[int(t)]
for b in range(0, x0.shape[0], batch_size):
x_batch = x0[b:b + batch_size]
noise = randn_tensor(shape=x_batch.shape, device=self.device, dtype=x0.dtype)
xts[idx, b:b + batch_size] = self.scheduler.add_noise(x_batch, noise, t)
xts = torch.cat([xts, x0.unsqueeze(0)], dim=0)
zs = torch.zeros(size=variance_noise_shape, device=self.device, dtype=cond.dtype)
for t in tqdm(timesteps):
idx = t_to_idx[int(t)]
# 1. predict noise residual
for b in range(0, x0.shape[0], batch_size):
xt = xts[idx, b:b + batch_size]
cond_batch = cond.repeat(xt.shape[0], 1, 1)
noise_pred = self.unet(xt, timestep=t, encoder_hidden_states=cond_batch).sample
xtm1 = xts[idx + 1, b:b + batch_size]
z, xtm1_corrected = compute_noise(self.scheduler, xtm1, xt, t, noise_pred, eta)
zs[idx, b:b + batch_size] = z
# correction to avoid error accumulation
xts[idx + 1, b:b + batch_size] = xtm1_corrected
if save_latents:
torch.save(xts[idx], os.path.join(save_path, 'latents', f'noisy_latents_{t}.pt'))
if return_inverted_latents:
self.total_inverted_latents[f'noisy_latents_{t}'] = xts[idx].clone()
if save_path:
torch.save(xts[idx], os.path.join(save_path, 'latents', f'noisy_latents_{t}.pt'))
torch.save(zs, os.path.join(save_path, 'latents', f'noise_total.pt'))
if return_inverted_latents:
self.total_inverted_latents[f'noisy_latents_{t}'] = xts[idx].clone()
self.noise_total = zs.clone()
return xts[skip_steps].expand(latent_frames.shape[0], -1, -1, -1), zs
def prepare_extra_step_kwargs(self, eta):
# prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
# eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
# eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
# and should be between [0, 1]
accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
extra_step_kwargs = {}
if accepts_eta:
extra_step_kwargs["eta"] = eta
# check if the scheduler accepts generator
accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
return extra_step_kwargs
@torch.no_grad()
def ddpm_sample(self, init_latents, cond, batch_size, num_inversion_steps, skip_steps, eta, zs_all,
guidance_scale=0):
use_ddpm = True
do_classifier_free_guidance = guidance_scale > 1.0
total_latents = init_latents
self.scheduler.set_timesteps(num_inversion_steps, device=device)
timesteps = self.scheduler.timesteps
zs_total = zs_all[skip_steps:]
if use_ddpm:
t_to_idx = {int(v): k for k, v in enumerate(timesteps[-zs_total.shape[0]:])}
timesteps = timesteps[-zs_total.shape[0]:]
num_warmup_steps = len(timesteps) - num_inversion_steps * self.scheduler.order
extra_step_kwargs = self.prepare_extra_step_kwargs(eta)
for i, t in enumerate(tqdm(timesteps)):
for b in range(0, total_latents.shape[0], batch_size):
latents = total_latents[b:b + batch_size]
if do_classifier_free_guidance:
latent_model_input = torch.cat([latents] * 2)
else:
latent_model_input = latents
cond_batch = cond.repeat(latents.shape[0], 1, 1)
latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
noise_pred = self.unet(
latent_model_input,
t,
encoder_hidden_states=cond_batch,
return_dict=False,
)[0]
if do_classifier_free_guidance:
noise_pred_out = noise_pred.chunk(2) # [b,4, 64, 64]
noise_pred_uncond, noise_pred_text = noise_pred_out[0], noise_pred_out[1]
# default text guidance
noise_guidance = guidance_scale * (noise_pred_text - noise_pred_uncond)
noise_pred = noise_pred_uncond + noise_guidance
idx = t_to_idx[int(t)]
zs = zs_total[idx, b:b + batch_size]
latents = self.scheduler.step(noise_pred, t, latents, variance_noise=zs,
**extra_step_kwargs).prev_sample
total_latents[b:b + batch_size] = latents
return total_latents
@torch.no_grad()
def ddim_sample(self, x, cond, batch_size):
timesteps = self.scheduler.timesteps
for i, t in enumerate(tqdm(timesteps)):
for b in range(0, x.shape[0], batch_size):
x_batch = x[b:b + batch_size]
model_input = x_batch
cond_batch = cond.repeat(x_batch.shape[0], 1, 1)
if self.sd_version == 'depth':
depth_maps = torch.cat([self.depth_maps[b: b + batch_size]])
model_input = torch.cat([x_batch, depth_maps],dim=1)
alpha_prod_t = self.scheduler.alphas_cumprod[t]
alpha_prod_t_prev = (
self.scheduler.alphas_cumprod[timesteps[i + 1]]
if i < len(timesteps) - 1
else self.scheduler.final_alpha_cumprod
)
mu = alpha_prod_t ** 0.5
sigma = (1 - alpha_prod_t) ** 0.5
mu_prev = alpha_prod_t_prev ** 0.5
sigma_prev = (1 - alpha_prod_t_prev) ** 0.5
eps = self.unet(model_input, t, encoder_hidden_states=cond_batch).sample if self.sd_version != 'ControlNet' \
else self.controlnet_pred(x_batch, t, cond_batch, torch.cat([self.canny_cond[b: b + batch_size]]))
pred_x0 = (x_batch - sigma * eps) / mu
x[b:b + batch_size] = mu_prev * pred_x0 + sigma_prev * eps
return x
@torch.no_grad()
def extract_latents(self,
num_steps,
save_path,
batch_size,
timesteps_to_save,
inversion_prompt='',
skip_steps=20,
inversion_type='ddim',
eta=1.0,
reconstruction=False):
self.scheduler.set_timesteps(num_steps)
cond = self.get_text_embeds(inversion_prompt, "")[1].unsqueeze(0)
latent_frames = self.latents
if inversion_type == 'ddim':
inverted_x= self.ddim_inversion(cond,
latent_frames,
save_path,
batch_size=batch_size,
save_latents=True if save_path else False,
timesteps_to_save=timesteps_to_save)
if reconstruction:
latent_reconstruction = self.ddim_sample(inverted_x, cond, batch_size=batch_size)
rgb_reconstruction = self.decode_latents(latent_reconstruction)
return (self.frames, self.latents, self.total_inverted_latents), rgb_reconstruction
else:
return (self.frames, self.latents, self.total_inverted_latents), None
elif inversion_type == 'ddpm':
inverted_x, zs = self.ddpm_inversion(cond,
latent_frames,
save_path= save_path,
batch_size=batch_size,
save_latents=True if save_path else False,
num_inversion_steps=num_steps,
eta=eta,
skip_steps=skip_steps)
cond = self.encode_text(inversion_prompt)
if reconstruction:
latent_reconstruction = self.ddpm_sample(init_latents=inverted_x,
cond=cond, batch_size=batch_size,
num_inversion_steps=num_steps, skip_steps=skip_steps,
eta=eta, zs_all=zs)
rgb_reconstruction = self.decode_latents(latent_reconstruction)
return (self.frames, self.latents, self.total_inverted_latents, self.noise_total), rgb_reconstruction
else:
return (self.frames, self.latents, self.total_inverted_latents, self.noise_total), None
else:
raise NotImplementedError()
def compute_noise(scheduler, prev_latents, latents, timestep, noise_pred, eta):
# 1. get previous step value (=t-1)
prev_timestep = timestep - scheduler.config.num_train_timesteps // scheduler.num_inference_steps
# 2. compute alphas, betas
alpha_prod_t = scheduler.alphas_cumprod[timestep]
alpha_prod_t_prev = (
scheduler.alphas_cumprod[prev_timestep] if prev_timestep >= 0 else scheduler.final_alpha_cumprod
)
beta_prod_t = 1 - alpha_prod_t
# 3. compute predicted original sample from predicted noise also called
# "predicted x_0" of formula (12) from https://arxiv.org/pdf/2010.02502.pdf
pred_original_sample = (latents - beta_prod_t ** (0.5) * noise_pred) / alpha_prod_t ** (0.5)
# 4. Clip "predicted x_0"
if scheduler.config.clip_sample:
pred_original_sample = torch.clamp(pred_original_sample, -1, 1)
# 5. compute variance: "sigma_t(η)" -> see formula (16)
# σ_t = sqrt((1 − α_t−1)/(1 − α_t)) * sqrt(1 − α_t/α_t−1)
variance = scheduler._get_variance(timestep, prev_timestep)
std_dev_t = eta * variance ** (0.5)
# 6. compute "direction pointing to x_t" of formula (12) from https://arxiv.org/pdf/2010.02502.pdf
pred_sample_direction = (1 - alpha_prod_t_prev - std_dev_t ** 2) ** (0.5) * noise_pred
# modifed so that updated xtm1 is returned as well (to avoid error accumulation)
mu_xt = alpha_prod_t_prev ** (0.5) * pred_original_sample + pred_sample_direction
noise = (prev_latents - mu_xt) / (variance ** (0.5) * eta)
return noise, mu_xt + (eta * variance ** 0.5) * noise
def prep(opt):
# timesteps to save
if opt["sd_version"] == '2.1':
model_key = "stabilityai/stable-diffusion-2-1-base"
elif opt["sd_version"] == '2.0':
model_key = "stabilityai/stable-diffusion-2-base"
elif opt["sd_version"] == '1.5' or opt["sd_version"] == 'ControlNet':
model_key = "runwayml/stable-diffusion-v1-5"
elif opt["sd_version"] == 'depth':
model_key = "stabilityai/stable-diffusion-2-depth"
toy_scheduler = DDIMScheduler.from_pretrained(model_key, subfolder="scheduler")
toy_scheduler.set_timesteps(opt["save_steps"])
timesteps_to_save, num_inference_steps = get_timesteps(toy_scheduler, num_inference_steps=opt["save_steps"],
strength=1.0,
device=device)
seed_everything(opt["seed"])
if not opt["frames"]: # original non demo setting
save_path = os.path.join(opt["save_dir"],
f'inversion_{opt[inversion]}',
f'sd_{opt["sd_version"]}',
Path(opt["data_path"]).stem,
f'steps_{opt["steps"]}',
f'nframes_{opt["n_frames"]}')
os.makedirs(os.path.join(save_path, f'latents'), exist_ok=True)
if opt[inversion] == 'ddpm':
os.makedirs(os.path.join(save_path, f'latents'), exist_ok=True)
add_dict_to_yaml_file(os.path.join(opt["save_dir"], 'inversion_prompts.yaml'), Path(opt["data_path"]).stem, opt["inversion_prompt"])
# save inversion prompt in a txt file
with open(os.path.join(save_path, 'inversion_prompt.txt'), 'w') as f:
f.write(opt["inversion_prompt"])
else:
save_path = None
model = Preprocess(device,
config,
vae=vae,
text_encoder=text_encoder,
scheduler=scheduler,
tokenizer=tokenizer,
unet=unet)
frames_and_latents, rgb_reconstruction = model.extract_latents(
num_steps=model.config["steps"],
save_path=save_path,
batch_size=model.config["batch_size"],
timesteps_to_save=timesteps_to_save,
inversion_prompt=model.config["inversion_prompt"],
inversion_type=model.config[inversion],
skip_steps=model.config[skip_steps],
reconstruction=model.config[reconstruct]
)
if model.config[inversion] == 'ddpm':
frames, latents, total_inverted_latents, zs = frames_and_latents
return frames, latents, total_inverted_latents, zs, rgb_reconstruction
else:
frames, latents, total_inverted_latents = frames_and_latents
return frames, latents, total_inverted_latents, rgb_reconstruction