Spaces:
Runtime error
Runtime error
ShaoTengLiu
commited on
Commit
·
44fa1db
1
Parent(s):
ab4a868
update
Browse files- Video-P2P/run.py +0 -993
- Video-P2P/run_tuning.py +30 -5
- Video-P2P/run_videop2p.py +106 -69
- trainer.py +3 -1
Video-P2P/run.py
DELETED
@@ -1,993 +0,0 @@
|
|
1 |
-
import argparse
|
2 |
-
import datetime
|
3 |
-
import logging
|
4 |
-
import inspect
|
5 |
-
import math
|
6 |
-
import os
|
7 |
-
from typing import Optional, Union, Tuple, List, Callable, Dict
|
8 |
-
from omegaconf import OmegaConf
|
9 |
-
|
10 |
-
import torch
|
11 |
-
import torch.nn.functional as F
|
12 |
-
import torch.utils.checkpoint
|
13 |
-
|
14 |
-
import diffusers
|
15 |
-
import transformers
|
16 |
-
from accelerate import Accelerator
|
17 |
-
from accelerate.logging import get_logger
|
18 |
-
from accelerate.utils import set_seed
|
19 |
-
from diffusers import AutoencoderKL, DDPMScheduler, DDIMScheduler
|
20 |
-
from diffusers.optimization import get_scheduler
|
21 |
-
from diffusers.utils import check_min_version
|
22 |
-
from diffusers.utils.import_utils import is_xformers_available
|
23 |
-
from tqdm.auto import tqdm
|
24 |
-
from transformers import AutoTokenizer, CLIPTextModel, CLIPTokenizer
|
25 |
-
|
26 |
-
from tuneavideo.models.unet import UNet3DConditionModel
|
27 |
-
from tuneavideo.data.dataset import TuneAVideoDataset
|
28 |
-
from tuneavideo.pipelines.pipeline_tuneavideo import TuneAVideoPipeline
|
29 |
-
from tuneavideo.util import save_videos_grid, ddim_inversion
|
30 |
-
from einops import rearrange
|
31 |
-
|
32 |
-
import cv2
|
33 |
-
import abc
|
34 |
-
import ptp_utils
|
35 |
-
import seq_aligner
|
36 |
-
import shutil
|
37 |
-
from torch.optim.adam import Adam
|
38 |
-
from PIL import Image
|
39 |
-
import numpy as np
|
40 |
-
import decord
|
41 |
-
decord.bridge.set_bridge('torch')
|
42 |
-
|
43 |
-
# Will error if the minimal version of diffusers is not installed. Remove at your own risks.
|
44 |
-
check_min_version("0.10.0.dev0")
|
45 |
-
|
46 |
-
logger = get_logger(__name__, log_level="INFO")
|
47 |
-
|
48 |
-
|
49 |
-
def main(
|
50 |
-
pretrained_model_path: str,
|
51 |
-
output_dir: str,
|
52 |
-
train_data: Dict,
|
53 |
-
validation_data: Dict,
|
54 |
-
validation_steps: int = 100,
|
55 |
-
trainable_modules: Tuple[str] = (
|
56 |
-
"attn1.to_q",
|
57 |
-
"attn2.to_q",
|
58 |
-
"attn_temp",
|
59 |
-
),
|
60 |
-
train_batch_size: int = 1,
|
61 |
-
max_train_steps: int = 500,
|
62 |
-
learning_rate: float = 3e-5,
|
63 |
-
scale_lr: bool = False,
|
64 |
-
lr_scheduler: str = "constant",
|
65 |
-
lr_warmup_steps: int = 0,
|
66 |
-
adam_beta1: float = 0.9,
|
67 |
-
adam_beta2: float = 0.999,
|
68 |
-
adam_weight_decay: float = 1e-2,
|
69 |
-
adam_epsilon: float = 1e-08,
|
70 |
-
max_grad_norm: float = 1.0,
|
71 |
-
gradient_accumulation_steps: int = 1,
|
72 |
-
gradient_checkpointing: bool = True,
|
73 |
-
checkpointing_steps: int = 500,
|
74 |
-
resume_from_checkpoint: Optional[str] = None,
|
75 |
-
mixed_precision: Optional[str] = "fp16",
|
76 |
-
use_8bit_adam: bool = False,
|
77 |
-
enable_xformers_memory_efficient_attention: bool = True,
|
78 |
-
seed: Optional[int] = None,
|
79 |
-
# pretrained_model_path: str,
|
80 |
-
# image_path: str = None,
|
81 |
-
# prompt: str = None,
|
82 |
-
prompts: Tuple[str] = None,
|
83 |
-
eq_params: Dict = None,
|
84 |
-
save_name: str = None,
|
85 |
-
is_word_swap: bool = None,
|
86 |
-
blend_word: Tuple[str] = None,
|
87 |
-
cross_replace_steps: float = 0.2,
|
88 |
-
self_replace_steps: float = 0.5,
|
89 |
-
video_len: int = 8,
|
90 |
-
fast: bool = False,
|
91 |
-
mixed_precision_p2p: str = 'fp32',
|
92 |
-
):
|
93 |
-
*_, config = inspect.getargvalues(inspect.currentframe())
|
94 |
-
|
95 |
-
accelerator = Accelerator(
|
96 |
-
gradient_accumulation_steps=gradient_accumulation_steps,
|
97 |
-
mixed_precision=mixed_precision,
|
98 |
-
)
|
99 |
-
|
100 |
-
# Make one log on every process with the configuration for debugging.
|
101 |
-
logging.basicConfig(
|
102 |
-
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
|
103 |
-
datefmt="%m/%d/%Y %H:%M:%S",
|
104 |
-
level=logging.INFO,
|
105 |
-
)
|
106 |
-
logger.info(accelerator.state, main_process_only=False)
|
107 |
-
if accelerator.is_local_main_process:
|
108 |
-
transformers.utils.logging.set_verbosity_warning()
|
109 |
-
diffusers.utils.logging.set_verbosity_info()
|
110 |
-
else:
|
111 |
-
transformers.utils.logging.set_verbosity_error()
|
112 |
-
diffusers.utils.logging.set_verbosity_error()
|
113 |
-
|
114 |
-
# If passed along, set the training seed now.
|
115 |
-
if seed is not None:
|
116 |
-
set_seed(seed)
|
117 |
-
|
118 |
-
# Handle the output folder creation
|
119 |
-
if accelerator.is_main_process:
|
120 |
-
# now = datetime.datetime.now().strftime("%Y-%m-%dT%H-%M-%S")
|
121 |
-
# output_dir = os.path.join(output_dir, now)
|
122 |
-
os.makedirs(output_dir, exist_ok=True)
|
123 |
-
os.makedirs(f"{output_dir}/samples", exist_ok=True)
|
124 |
-
os.makedirs(f"{output_dir}/inv_latents", exist_ok=True)
|
125 |
-
OmegaConf.save(config, os.path.join(output_dir, 'config.yaml'))
|
126 |
-
|
127 |
-
# Load scheduler, tokenizer and models.
|
128 |
-
noise_scheduler = DDPMScheduler.from_pretrained(pretrained_model_path, subfolder="scheduler")
|
129 |
-
tokenizer = CLIPTokenizer.from_pretrained(pretrained_model_path, subfolder="tokenizer")
|
130 |
-
text_encoder = CLIPTextModel.from_pretrained(pretrained_model_path, subfolder="text_encoder")
|
131 |
-
vae = AutoencoderKL.from_pretrained(pretrained_model_path, subfolder="vae")
|
132 |
-
unet = UNet3DConditionModel.from_pretrained_2d(pretrained_model_path, subfolder="unet")
|
133 |
-
|
134 |
-
# Freeze vae and text_encoder
|
135 |
-
vae.requires_grad_(False)
|
136 |
-
text_encoder.requires_grad_(False)
|
137 |
-
|
138 |
-
unet.requires_grad_(False)
|
139 |
-
for name, module in unet.named_modules():
|
140 |
-
if name.endswith(tuple(trainable_modules)):
|
141 |
-
for params in module.parameters():
|
142 |
-
params.requires_grad = True
|
143 |
-
|
144 |
-
if enable_xformers_memory_efficient_attention:
|
145 |
-
if is_xformers_available():
|
146 |
-
unet.enable_xformers_memory_efficient_attention()
|
147 |
-
else:
|
148 |
-
raise ValueError("xformers is not available. Make sure it is installed correctly")
|
149 |
-
|
150 |
-
if gradient_checkpointing:
|
151 |
-
unet.enable_gradient_checkpointing()
|
152 |
-
|
153 |
-
if scale_lr:
|
154 |
-
learning_rate = (
|
155 |
-
learning_rate * gradient_accumulation_steps * train_batch_size * accelerator.num_processes
|
156 |
-
)
|
157 |
-
|
158 |
-
# Initialize the optimizer
|
159 |
-
if use_8bit_adam:
|
160 |
-
try:
|
161 |
-
import bitsandbytes as bnb
|
162 |
-
except ImportError:
|
163 |
-
raise ImportError(
|
164 |
-
"Please install bitsandbytes to use 8-bit Adam. You can do so by running `pip install bitsandbytes`"
|
165 |
-
)
|
166 |
-
|
167 |
-
optimizer_cls = bnb.optim.AdamW8bit
|
168 |
-
else:
|
169 |
-
optimizer_cls = torch.optim.AdamW
|
170 |
-
|
171 |
-
optimizer = optimizer_cls(
|
172 |
-
unet.parameters(),
|
173 |
-
lr=learning_rate,
|
174 |
-
betas=(adam_beta1, adam_beta2),
|
175 |
-
weight_decay=adam_weight_decay,
|
176 |
-
eps=adam_epsilon,
|
177 |
-
)
|
178 |
-
|
179 |
-
# Get the training dataset
|
180 |
-
train_dataset = TuneAVideoDataset(**train_data)
|
181 |
-
|
182 |
-
# Preprocessing the dataset
|
183 |
-
train_dataset.prompt_ids = tokenizer(
|
184 |
-
train_dataset.prompt, max_length=tokenizer.model_max_length, padding="max_length", truncation=True, return_tensors="pt"
|
185 |
-
).input_ids[0]
|
186 |
-
|
187 |
-
# DataLoaders creation:
|
188 |
-
train_dataloader = torch.utils.data.DataLoader(
|
189 |
-
train_dataset, batch_size=train_batch_size
|
190 |
-
)
|
191 |
-
|
192 |
-
# Get the validation pipeline
|
193 |
-
validation_pipeline = TuneAVideoPipeline(
|
194 |
-
vae=vae, text_encoder=text_encoder, tokenizer=tokenizer, unet=unet,
|
195 |
-
scheduler=DDIMScheduler.from_pretrained(pretrained_model_path, subfolder="scheduler")
|
196 |
-
)
|
197 |
-
validation_pipeline.enable_vae_slicing()
|
198 |
-
ddim_inv_scheduler = DDIMScheduler.from_pretrained(pretrained_model_path, subfolder='scheduler')
|
199 |
-
ddim_inv_scheduler.set_timesteps(validation_data.num_inv_steps)
|
200 |
-
|
201 |
-
# Scheduler
|
202 |
-
lr_scheduler = get_scheduler(
|
203 |
-
lr_scheduler,
|
204 |
-
optimizer=optimizer,
|
205 |
-
num_warmup_steps=lr_warmup_steps * gradient_accumulation_steps,
|
206 |
-
num_training_steps=max_train_steps * gradient_accumulation_steps,
|
207 |
-
)
|
208 |
-
|
209 |
-
# Prepare everything with our `accelerator`.
|
210 |
-
unet, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(
|
211 |
-
unet, optimizer, train_dataloader, lr_scheduler
|
212 |
-
)
|
213 |
-
|
214 |
-
# For mixed precision training we cast the text_encoder and vae weights to half-precision
|
215 |
-
# as these models are only used for inference, keeping weights in full precision is not required.
|
216 |
-
weight_dtype = torch.float32
|
217 |
-
if accelerator.mixed_precision == "fp16":
|
218 |
-
weight_dtype = torch.float16
|
219 |
-
elif accelerator.mixed_precision == "bf16":
|
220 |
-
weight_dtype = torch.bfloat16
|
221 |
-
|
222 |
-
# Move text_encode and vae to gpu and cast to weight_dtype
|
223 |
-
text_encoder.to(accelerator.device, dtype=weight_dtype)
|
224 |
-
vae.to(accelerator.device, dtype=weight_dtype)
|
225 |
-
|
226 |
-
# We need to recalculate our total training steps as the size of the training dataloader may have changed.
|
227 |
-
num_update_steps_per_epoch = math.ceil(len(train_dataloader) / gradient_accumulation_steps)
|
228 |
-
# Afterwards we recalculate our number of training epochs
|
229 |
-
num_train_epochs = math.ceil(max_train_steps / num_update_steps_per_epoch)
|
230 |
-
|
231 |
-
# We need to initialize the trackers we use, and also store our configuration.
|
232 |
-
# The trackers initializes automatically on the main process.
|
233 |
-
if accelerator.is_main_process:
|
234 |
-
accelerator.init_trackers("text2video-fine-tune")
|
235 |
-
|
236 |
-
# Train!
|
237 |
-
total_batch_size = train_batch_size * accelerator.num_processes * gradient_accumulation_steps
|
238 |
-
|
239 |
-
logger.info("***** Running training *****")
|
240 |
-
logger.info(f" Num examples = {len(train_dataset)}")
|
241 |
-
logger.info(f" Num Epochs = {num_train_epochs}")
|
242 |
-
logger.info(f" Instantaneous batch size per device = {train_batch_size}")
|
243 |
-
logger.info(f" Total train batch size (w. parallel, distributed & accumulation) = {total_batch_size}")
|
244 |
-
logger.info(f" Gradient Accumulation steps = {gradient_accumulation_steps}")
|
245 |
-
logger.info(f" Total optimization steps = {max_train_steps}")
|
246 |
-
global_step = 0
|
247 |
-
first_epoch = 0
|
248 |
-
|
249 |
-
# Potentially load in the weights and states from a previous save
|
250 |
-
if resume_from_checkpoint:
|
251 |
-
if resume_from_checkpoint != "latest":
|
252 |
-
path = os.path.basename(resume_from_checkpoint)
|
253 |
-
else:
|
254 |
-
# Get the most recent checkpoint
|
255 |
-
dirs = os.listdir(output_dir)
|
256 |
-
dirs = [d for d in dirs if d.startswith("checkpoint")]
|
257 |
-
dirs = sorted(dirs, key=lambda x: int(x.split("-")[1]))
|
258 |
-
path = dirs[-1]
|
259 |
-
accelerator.print(f"Resuming from checkpoint {path}")
|
260 |
-
accelerator.load_state(os.path.join(output_dir, path))
|
261 |
-
global_step = int(path.split("-")[1])
|
262 |
-
|
263 |
-
first_epoch = global_step // num_update_steps_per_epoch
|
264 |
-
resume_step = global_step % num_update_steps_per_epoch
|
265 |
-
|
266 |
-
# Only show the progress bar once on each machine.
|
267 |
-
progress_bar = tqdm(range(global_step, max_train_steps), disable=not accelerator.is_local_main_process)
|
268 |
-
progress_bar.set_description("Steps")
|
269 |
-
|
270 |
-
for epoch in range(first_epoch, num_train_epochs):
|
271 |
-
unet.train()
|
272 |
-
train_loss = 0.0
|
273 |
-
for step, batch in enumerate(train_dataloader):
|
274 |
-
# Skip steps until we reach the resumed step
|
275 |
-
if resume_from_checkpoint and epoch == first_epoch and step < resume_step:
|
276 |
-
if step % gradient_accumulation_steps == 0:
|
277 |
-
progress_bar.update(1)
|
278 |
-
continue
|
279 |
-
|
280 |
-
with accelerator.accumulate(unet):
|
281 |
-
# Convert videos to latent space
|
282 |
-
pixel_values = batch["pixel_values"].to(weight_dtype)
|
283 |
-
video_length = pixel_values.shape[1]
|
284 |
-
pixel_values = rearrange(pixel_values, "b f c h w -> (b f) c h w")
|
285 |
-
latents = vae.encode(pixel_values).latent_dist.sample()
|
286 |
-
latents = rearrange(latents, "(b f) c h w -> b c f h w", f=video_length)
|
287 |
-
latents = latents * 0.18215
|
288 |
-
|
289 |
-
# Sample noise that we'll add to the latents
|
290 |
-
noise = torch.randn_like(latents)
|
291 |
-
bsz = latents.shape[0]
|
292 |
-
# Sample a random timestep for each video
|
293 |
-
timesteps = torch.randint(0, noise_scheduler.num_train_timesteps, (bsz,), device=latents.device)
|
294 |
-
timesteps = timesteps.long()
|
295 |
-
|
296 |
-
# Add noise to the latents according to the noise magnitude at each timestep
|
297 |
-
# (this is the forward diffusion process)
|
298 |
-
noisy_latents = noise_scheduler.add_noise(latents, noise, timesteps)
|
299 |
-
|
300 |
-
# Get the text embedding for conditioning
|
301 |
-
encoder_hidden_states = text_encoder(batch["prompt_ids"])[0]
|
302 |
-
|
303 |
-
# Get the target for loss depending on the prediction type
|
304 |
-
if noise_scheduler.prediction_type == "epsilon":
|
305 |
-
target = noise
|
306 |
-
elif noise_scheduler.prediction_type == "v_prediction":
|
307 |
-
target = noise_scheduler.get_velocity(latents, noise, timesteps)
|
308 |
-
else:
|
309 |
-
raise ValueError(f"Unknown prediction type {noise_scheduler.prediction_type}")
|
310 |
-
|
311 |
-
# Predict the noise residual and compute loss
|
312 |
-
model_pred = unet(noisy_latents, timesteps, encoder_hidden_states).sample
|
313 |
-
loss = F.mse_loss(model_pred.float(), target.float(), reduction="mean")
|
314 |
-
|
315 |
-
# Gather the losses across all processes for logging (if we use distributed training).
|
316 |
-
avg_loss = accelerator.gather(loss.repeat(train_batch_size)).mean()
|
317 |
-
train_loss += avg_loss.item() / gradient_accumulation_steps
|
318 |
-
|
319 |
-
# Backpropagate
|
320 |
-
accelerator.backward(loss)
|
321 |
-
if accelerator.sync_gradients:
|
322 |
-
accelerator.clip_grad_norm_(unet.parameters(), max_grad_norm)
|
323 |
-
optimizer.step()
|
324 |
-
lr_scheduler.step()
|
325 |
-
optimizer.zero_grad()
|
326 |
-
|
327 |
-
# Checks if the accelerator has performed an optimization step behind the scenes
|
328 |
-
if accelerator.sync_gradients:
|
329 |
-
progress_bar.update(1)
|
330 |
-
global_step += 1
|
331 |
-
accelerator.log({"train_loss": train_loss}, step=global_step)
|
332 |
-
train_loss = 0.0
|
333 |
-
|
334 |
-
if global_step % checkpointing_steps == 0:
|
335 |
-
if accelerator.is_main_process:
|
336 |
-
save_path = os.path.join(output_dir, f"checkpoint-{global_step}")
|
337 |
-
accelerator.save_state(save_path)
|
338 |
-
logger.info(f"Saved state to {save_path}")
|
339 |
-
|
340 |
-
if global_step % validation_steps == 0:
|
341 |
-
if accelerator.is_main_process:
|
342 |
-
samples = []
|
343 |
-
generator = torch.Generator(device=latents.device)
|
344 |
-
generator.manual_seed(seed)
|
345 |
-
|
346 |
-
ddim_inv_latent = None
|
347 |
-
if validation_data.use_inv_latent:
|
348 |
-
inv_latents_path = os.path.join(output_dir, f"inv_latents/ddim_latent-{global_step}.pt")
|
349 |
-
ddim_inv_latent = ddim_inversion(
|
350 |
-
validation_pipeline, ddim_inv_scheduler, video_latent=latents,
|
351 |
-
num_inv_steps=validation_data.num_inv_steps, prompt="")[-1].to(weight_dtype)
|
352 |
-
torch.save(ddim_inv_latent, inv_latents_path)
|
353 |
-
|
354 |
-
for idx, prompt in enumerate(validation_data.prompts):
|
355 |
-
sample = validation_pipeline(prompt, generator=generator, latents=ddim_inv_latent,
|
356 |
-
**validation_data).videos
|
357 |
-
save_videos_grid(sample, f"{output_dir}/samples/sample-{global_step}/{prompt}.gif")
|
358 |
-
samples.append(sample)
|
359 |
-
samples = torch.concat(samples)
|
360 |
-
save_path = f"{output_dir}/samples/sample-{global_step}.gif"
|
361 |
-
save_videos_grid(samples, save_path)
|
362 |
-
logger.info(f"Saved samples to {save_path}")
|
363 |
-
|
364 |
-
logs = {"step_loss": loss.detach().item(), "lr": lr_scheduler.get_last_lr()[0]}
|
365 |
-
progress_bar.set_postfix(**logs)
|
366 |
-
|
367 |
-
if global_step >= max_train_steps:
|
368 |
-
break
|
369 |
-
|
370 |
-
# Create the pipeline using the trained modules and save it.
|
371 |
-
accelerator.wait_for_everyone()
|
372 |
-
if accelerator.is_main_process:
|
373 |
-
unet = accelerator.unwrap_model(unet)
|
374 |
-
pipeline = TuneAVideoPipeline.from_pretrained(
|
375 |
-
pretrained_model_path,
|
376 |
-
text_encoder=text_encoder,
|
377 |
-
vae=vae,
|
378 |
-
unet=unet,
|
379 |
-
)
|
380 |
-
pipeline.save_pretrained(output_dir)
|
381 |
-
|
382 |
-
accelerator.end_training()
|
383 |
-
|
384 |
-
torch.cuda.empty_cache()
|
385 |
-
|
386 |
-
# Video-P2P
|
387 |
-
scheduler = DDIMScheduler(beta_start=0.00085, beta_end=0.012, beta_schedule="scaled_linear", clip_sample=False, set_alpha_to_one=False)
|
388 |
-
MY_TOKEN = ''
|
389 |
-
LOW_RESOURCE = False
|
390 |
-
NUM_DDIM_STEPS = 50
|
391 |
-
GUIDANCE_SCALE = 7.5
|
392 |
-
MAX_NUM_WORDS = 77
|
393 |
-
device = torch.device('cuda:0') if torch.cuda.is_available() else torch.device('cpu')
|
394 |
-
|
395 |
-
# need to adjust sometimes
|
396 |
-
mask_th = (.3, .3)
|
397 |
-
|
398 |
-
|
399 |
-
pretrained_model_path = output_dir
|
400 |
-
image_path = train_data['video_path']
|
401 |
-
prompt = train_data['prompt']
|
402 |
-
# prompts = [prompt, ]
|
403 |
-
output_folder = os.path.join(pretrained_model_path, 'results')
|
404 |
-
if fast:
|
405 |
-
save_name_1 = os.path.join(output_folder, 'inversion_fast.gif')
|
406 |
-
save_name_2 = os.path.join(output_folder, '{}_fast.gif'.format(save_name))
|
407 |
-
else:
|
408 |
-
save_name_1 = os.path.join(output_folder, 'inversion.gif')
|
409 |
-
save_name_2 = os.path.join(output_folder, '{}.gif'.format(save_name))
|
410 |
-
if blend_word:
|
411 |
-
blend_word = (((blend_word[0],), (blend_word[1],)))
|
412 |
-
eq_params = dict(eq_params)
|
413 |
-
prompts = list(prompts)
|
414 |
-
cross_replace_steps = {'default_': cross_replace_steps,}
|
415 |
-
|
416 |
-
weight_dtype = torch.float32
|
417 |
-
if mixed_precision_p2p == "fp16":
|
418 |
-
weight_dtype = torch.float16
|
419 |
-
elif mixed_precision_p2p == "bf16":
|
420 |
-
weight_dtype = torch.bfloat16
|
421 |
-
|
422 |
-
if not os.path.exists(output_folder):
|
423 |
-
os.makedirs(output_folder)
|
424 |
-
|
425 |
-
# Load the tokenizer
|
426 |
-
tokenizer = CLIPTokenizer.from_pretrained(pretrained_model_path, subfolder="tokenizer")
|
427 |
-
# Load models and create wrapper for stable diffusion
|
428 |
-
text_encoder = CLIPTextModel.from_pretrained(
|
429 |
-
pretrained_model_path,
|
430 |
-
subfolder="text_encoder",
|
431 |
-
).to(device, dtype=weight_dtype)
|
432 |
-
vae = AutoencoderKL.from_pretrained(
|
433 |
-
pretrained_model_path,
|
434 |
-
subfolder="vae",
|
435 |
-
).to(device, dtype=weight_dtype)
|
436 |
-
unet = UNet3DConditionModel.from_pretrained(
|
437 |
-
pretrained_model_path, subfolder="unet"
|
438 |
-
).to(device)
|
439 |
-
ldm_stable = TuneAVideoPipeline(
|
440 |
-
vae=vae,
|
441 |
-
text_encoder=text_encoder,
|
442 |
-
tokenizer=tokenizer,
|
443 |
-
unet=unet,
|
444 |
-
scheduler=scheduler,
|
445 |
-
).to(device)
|
446 |
-
|
447 |
-
try:
|
448 |
-
ldm_stable.disable_xformers_memory_efficient_attention()
|
449 |
-
except AttributeError:
|
450 |
-
print("Attribute disable_xformers_memory_efficient_attention() is missing")
|
451 |
-
tokenizer = ldm_stable.tokenizer # Tokenizer of class: [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer)
|
452 |
-
# A tokenizer breaks a stream of text into tokens, usually by looking for whitespace (tabs, spaces, new lines).
|
453 |
-
|
454 |
-
class LocalBlend:
|
455 |
-
|
456 |
-
def get_mask(self, maps, alpha, use_pool):
|
457 |
-
k = 1
|
458 |
-
maps = (maps * alpha).sum(-1).mean(2)
|
459 |
-
if use_pool:
|
460 |
-
maps = F.max_pool2d(maps, (k * 2 + 1, k * 2 +1), (1, 1), padding=(k, k))
|
461 |
-
mask = F.interpolate(maps, size=(x_t.shape[3:]))
|
462 |
-
mask = mask / mask.max(2, keepdims=True)[0].max(3, keepdims=True)[0]
|
463 |
-
mask = mask.gt(self.th[1-int(use_pool)])
|
464 |
-
mask = mask[:1] + mask
|
465 |
-
return mask
|
466 |
-
|
467 |
-
def __call__(self, x_t, attention_store, step):
|
468 |
-
self.counter += 1
|
469 |
-
if self.counter > self.start_blend:
|
470 |
-
maps = attention_store["down_cross"][2:4] + attention_store["up_cross"][:3]
|
471 |
-
maps = [item.reshape(self.alpha_layers.shape[0], -1, 8, 16, 16, MAX_NUM_WORDS) for item in maps]
|
472 |
-
maps = torch.cat(maps, dim=2)
|
473 |
-
mask = self.get_mask(maps, self.alpha_layers, True)
|
474 |
-
if self.substruct_layers is not None:
|
475 |
-
maps_sub = ~self.get_mask(maps, self.substruct_layers, False)
|
476 |
-
mask = mask * maps_sub
|
477 |
-
mask = mask.float()
|
478 |
-
mask = mask.reshape(-1, 1, mask.shape[-3], mask.shape[-2], mask.shape[-1])
|
479 |
-
x_t = x_t[:1] + mask * (x_t - x_t[:1])
|
480 |
-
return x_t
|
481 |
-
|
482 |
-
def __init__(self, prompts: List[str], words: [List[List[str]]], substruct_words=None, start_blend=0.2, th=(.3, .3)):
|
483 |
-
alpha_layers = torch.zeros(len(prompts), 1, 1, 1, 1, MAX_NUM_WORDS)
|
484 |
-
for i, (prompt, words_) in enumerate(zip(prompts, words)):
|
485 |
-
if type(words_) is str:
|
486 |
-
words_ = [words_]
|
487 |
-
for word in words_:
|
488 |
-
ind = ptp_utils.get_word_inds(prompt, word, tokenizer)
|
489 |
-
alpha_layers[i, :, :, :, :, ind] = 1
|
490 |
-
|
491 |
-
if substruct_words is not None:
|
492 |
-
substruct_layers = torch.zeros(len(prompts), 1, 1, 1, 1, MAX_NUM_WORDS)
|
493 |
-
for i, (prompt, words_) in enumerate(zip(prompts, substruct_words)):
|
494 |
-
if type(words_) is str:
|
495 |
-
words_ = [words_]
|
496 |
-
for word in words_:
|
497 |
-
ind = ptp_utils.get_word_inds(prompt, word, tokenizer)
|
498 |
-
substruct_layers[i, :, :, :, :, ind] = 1
|
499 |
-
self.substruct_layers = substruct_layers.to(device)
|
500 |
-
else:
|
501 |
-
self.substruct_layers = None
|
502 |
-
self.alpha_layers = alpha_layers.to(device)
|
503 |
-
self.start_blend = int(start_blend * NUM_DDIM_STEPS)
|
504 |
-
self.counter = 0
|
505 |
-
self.th=th
|
506 |
-
|
507 |
-
|
508 |
-
class EmptyControl:
|
509 |
-
|
510 |
-
|
511 |
-
def step_callback(self, x_t):
|
512 |
-
return x_t
|
513 |
-
|
514 |
-
def between_steps(self):
|
515 |
-
return
|
516 |
-
|
517 |
-
def __call__(self, attn, is_cross: bool, place_in_unet: str):
|
518 |
-
return attn
|
519 |
-
|
520 |
-
|
521 |
-
class AttentionControl(abc.ABC):
|
522 |
-
|
523 |
-
def step_callback(self, x_t):
|
524 |
-
return x_t
|
525 |
-
|
526 |
-
def between_steps(self):
|
527 |
-
return
|
528 |
-
|
529 |
-
@property
|
530 |
-
def num_uncond_att_layers(self):
|
531 |
-
return self.num_att_layers if LOW_RESOURCE else 0
|
532 |
-
|
533 |
-
@abc.abstractmethod
|
534 |
-
def forward (self, attn, is_cross: bool, place_in_unet: str):
|
535 |
-
raise NotImplementedError
|
536 |
-
|
537 |
-
def __call__(self, attn, is_cross: bool, place_in_unet: str):
|
538 |
-
if self.cur_att_layer >= self.num_uncond_att_layers:
|
539 |
-
if LOW_RESOURCE:
|
540 |
-
attn = self.forward(attn, is_cross, place_in_unet)
|
541 |
-
else:
|
542 |
-
h = attn.shape[0]
|
543 |
-
attn[h // 2:] = self.forward(attn[h // 2:], is_cross, place_in_unet)
|
544 |
-
self.cur_att_layer += 1
|
545 |
-
if self.cur_att_layer == self.num_att_layers + self.num_uncond_att_layers:
|
546 |
-
self.cur_att_layer = 0
|
547 |
-
self.cur_step += 1
|
548 |
-
self.between_steps()
|
549 |
-
return attn
|
550 |
-
|
551 |
-
def reset(self):
|
552 |
-
self.cur_step = 0
|
553 |
-
self.cur_att_layer = 0
|
554 |
-
|
555 |
-
def __init__(self):
|
556 |
-
self.cur_step = 0
|
557 |
-
self.num_att_layers = -1
|
558 |
-
self.cur_att_layer = 0
|
559 |
-
|
560 |
-
class SpatialReplace(EmptyControl):
|
561 |
-
|
562 |
-
def step_callback(self, x_t):
|
563 |
-
if self.cur_step < self.stop_inject:
|
564 |
-
b = x_t.shape[0]
|
565 |
-
x_t = x_t[:1].expand(b, *x_t.shape[1:])
|
566 |
-
return x_t
|
567 |
-
|
568 |
-
def __init__(self, stop_inject: float):
|
569 |
-
super(SpatialReplace, self).__init__()
|
570 |
-
self.stop_inject = int((1 - stop_inject) * NUM_DDIM_STEPS)
|
571 |
-
|
572 |
-
|
573 |
-
class AttentionStore(AttentionControl):
|
574 |
-
|
575 |
-
@staticmethod
|
576 |
-
def get_empty_store():
|
577 |
-
return {"down_cross": [], "mid_cross": [], "up_cross": [],
|
578 |
-
"down_self": [], "mid_self": [], "up_self": []}
|
579 |
-
|
580 |
-
def forward(self, attn, is_cross: bool, place_in_unet: str):
|
581 |
-
key = f"{place_in_unet}_{'cross' if is_cross else 'self'}"
|
582 |
-
if attn.shape[1] <= 32 ** 2:
|
583 |
-
self.step_store[key].append(attn)
|
584 |
-
return attn
|
585 |
-
|
586 |
-
def between_steps(self):
|
587 |
-
if len(self.attention_store) == 0:
|
588 |
-
self.attention_store = self.step_store
|
589 |
-
else:
|
590 |
-
for key in self.attention_store:
|
591 |
-
for i in range(len(self.attention_store[key])):
|
592 |
-
self.attention_store[key][i] += self.step_store[key][i]
|
593 |
-
self.step_store = self.get_empty_store()
|
594 |
-
|
595 |
-
def get_average_attention(self):
|
596 |
-
average_attention = {key: [item / self.cur_step for item in self.attention_store[key]] for key in self.attention_store}
|
597 |
-
return average_attention
|
598 |
-
|
599 |
-
|
600 |
-
def reset(self):
|
601 |
-
super(AttentionStore, self).reset()
|
602 |
-
self.step_store = self.get_empty_store()
|
603 |
-
self.attention_store = {}
|
604 |
-
|
605 |
-
def __init__(self):
|
606 |
-
super(AttentionStore, self).__init__()
|
607 |
-
self.step_store = self.get_empty_store()
|
608 |
-
self.attention_store = {}
|
609 |
-
|
610 |
-
|
611 |
-
class AttentionControlEdit(AttentionStore, abc.ABC):
|
612 |
-
|
613 |
-
def step_callback(self, x_t):
|
614 |
-
if self.local_blend is not None:
|
615 |
-
x_t = self.local_blend(x_t, self.attention_store, self.cur_step)
|
616 |
-
return x_t
|
617 |
-
|
618 |
-
def replace_self_attention(self, attn_base, att_replace, place_in_unet):
|
619 |
-
if att_replace.shape[2] <= 32 ** 2:
|
620 |
-
attn_base = attn_base.unsqueeze(0).expand(att_replace.shape[0], *attn_base.shape)
|
621 |
-
return attn_base
|
622 |
-
else:
|
623 |
-
return att_replace
|
624 |
-
|
625 |
-
@abc.abstractmethod
|
626 |
-
def replace_cross_attention(self, attn_base, att_replace):
|
627 |
-
raise NotImplementedError
|
628 |
-
|
629 |
-
def forward(self, attn, is_cross: bool, place_in_unet: str):
|
630 |
-
super(AttentionControlEdit, self).forward(attn, is_cross, place_in_unet)
|
631 |
-
if is_cross or (self.num_self_replace[0] <= self.cur_step < self.num_self_replace[1]):
|
632 |
-
h = attn.shape[0] // (self.batch_size)
|
633 |
-
attn = attn.reshape(self.batch_size, h, *attn.shape[1:])
|
634 |
-
attn_base, attn_repalce = attn[0], attn[1:]
|
635 |
-
if is_cross:
|
636 |
-
alpha_words = self.cross_replace_alpha[self.cur_step]
|
637 |
-
attn_repalce_new = self.replace_cross_attention(attn_base, attn_repalce) * alpha_words + (1 - alpha_words) * attn_repalce
|
638 |
-
attn[1:] = attn_repalce_new
|
639 |
-
else:
|
640 |
-
attn[1:] = self.replace_self_attention(attn_base, attn_repalce, place_in_unet)
|
641 |
-
attn = attn.reshape(self.batch_size * h, *attn.shape[2:])
|
642 |
-
return attn
|
643 |
-
|
644 |
-
def __init__(self, prompts, num_steps: int,
|
645 |
-
cross_replace_steps: Union[float, Tuple[float, float], Dict[str, Tuple[float, float]]],
|
646 |
-
self_replace_steps: Union[float, Tuple[float, float]],
|
647 |
-
local_blend: Optional[LocalBlend]):
|
648 |
-
super(AttentionControlEdit, self).__init__()
|
649 |
-
self.batch_size = len(prompts)
|
650 |
-
self.cross_replace_alpha = ptp_utils.get_time_words_attention_alpha(prompts, num_steps, cross_replace_steps, tokenizer).to(device)
|
651 |
-
if type(self_replace_steps) is float:
|
652 |
-
self_replace_steps = 0, self_replace_steps
|
653 |
-
self.num_self_replace = int(num_steps * self_replace_steps[0]), int(num_steps * self_replace_steps[1])
|
654 |
-
self.local_blend = local_blend
|
655 |
-
|
656 |
-
class AttentionReplace(AttentionControlEdit):
|
657 |
-
|
658 |
-
def replace_cross_attention(self, attn_base, att_replace):
|
659 |
-
return torch.einsum('hpw,bwn->bhpn', attn_base, self.mapper)
|
660 |
-
|
661 |
-
def __init__(self, prompts, num_steps: int, cross_replace_steps: float, self_replace_steps: float,
|
662 |
-
local_blend: Optional[LocalBlend] = None):
|
663 |
-
super(AttentionReplace, self).__init__(prompts, num_steps, cross_replace_steps, self_replace_steps, local_blend)
|
664 |
-
self.mapper = seq_aligner.get_replacement_mapper(prompts, tokenizer).to(device)
|
665 |
-
|
666 |
-
|
667 |
-
class AttentionRefine(AttentionControlEdit):
|
668 |
-
|
669 |
-
def replace_cross_attention(self, attn_base, att_replace):
|
670 |
-
attn_base_replace = attn_base[:, :, self.mapper].permute(2, 0, 1, 3)
|
671 |
-
attn_replace = attn_base_replace * self.alphas + att_replace * (1 - self.alphas)
|
672 |
-
return attn_replace
|
673 |
-
|
674 |
-
def __init__(self, prompts, num_steps: int, cross_replace_steps: float, self_replace_steps: float,
|
675 |
-
local_blend: Optional[LocalBlend] = None):
|
676 |
-
super(AttentionRefine, self).__init__(prompts, num_steps, cross_replace_steps, self_replace_steps, local_blend)
|
677 |
-
self.mapper, alphas = seq_aligner.get_refinement_mapper(prompts, tokenizer)
|
678 |
-
self.mapper, alphas = self.mapper.to(device), alphas.to(device)
|
679 |
-
self.alphas = alphas.reshape(alphas.shape[0], 1, 1, alphas.shape[1])
|
680 |
-
|
681 |
-
|
682 |
-
class AttentionReweight(AttentionControlEdit):
|
683 |
-
|
684 |
-
def replace_cross_attention(self, attn_base, att_replace):
|
685 |
-
if self.prev_controller is not None:
|
686 |
-
attn_base = self.prev_controller.replace_cross_attention(attn_base, att_replace)
|
687 |
-
attn_replace = attn_base[None, :, :, :] * self.equalizer[:, None, None, :]
|
688 |
-
return attn_replace
|
689 |
-
|
690 |
-
def __init__(self, prompts, num_steps: int, cross_replace_steps: float, self_replace_steps: float, equalizer,
|
691 |
-
local_blend: Optional[LocalBlend] = None, controller: Optional[AttentionControlEdit] = None):
|
692 |
-
super(AttentionReweight, self).__init__(prompts, num_steps, cross_replace_steps, self_replace_steps, local_blend)
|
693 |
-
self.equalizer = equalizer.to(device)
|
694 |
-
self.prev_controller = controller
|
695 |
-
|
696 |
-
|
697 |
-
def get_equalizer(text: str, word_select: Union[int, Tuple[int, ...]], values: Union[List[float],
|
698 |
-
Tuple[float, ...]]):
|
699 |
-
if type(word_select) is int or type(word_select) is str:
|
700 |
-
word_select = (word_select,)
|
701 |
-
equalizer = torch.ones(1, 77)
|
702 |
-
|
703 |
-
for word, val in zip(word_select, values):
|
704 |
-
inds = ptp_utils.get_word_inds(text, word, tokenizer)
|
705 |
-
equalizer[:, inds] = val
|
706 |
-
return equalizer
|
707 |
-
|
708 |
-
def aggregate_attention(attention_store: AttentionStore, res: int, from_where: List[str], is_cross: bool, select: int):
|
709 |
-
out = []
|
710 |
-
attention_maps = attention_store.get_average_attention()
|
711 |
-
num_pixels = res ** 2
|
712 |
-
for location in from_where:
|
713 |
-
for item in attention_maps[f"{location}_{'cross' if is_cross else 'self'}"]:
|
714 |
-
if item.shape[1] == num_pixels:
|
715 |
-
cross_maps = item.reshape(8, 8, res, res, item.shape[-1])
|
716 |
-
out.append(cross_maps)
|
717 |
-
out = torch.cat(out, dim=1)
|
718 |
-
out = out.sum(1) / out.shape[1]
|
719 |
-
return out.cpu()
|
720 |
-
|
721 |
-
|
722 |
-
def make_controller(prompts: List[str], is_replace_controller: bool, cross_replace_steps: Dict[str, float], self_replace_steps: float, blend_words=None, equilizer_params=None, mask_th=(.3,.3)) -> AttentionControlEdit:
|
723 |
-
if blend_words is None:
|
724 |
-
lb = None
|
725 |
-
else:
|
726 |
-
lb = LocalBlend(prompts, blend_word, th=mask_th)
|
727 |
-
if is_replace_controller:
|
728 |
-
controller = AttentionReplace(prompts, NUM_DDIM_STEPS, cross_replace_steps=cross_replace_steps, self_replace_steps=self_replace_steps, local_blend=lb)
|
729 |
-
else:
|
730 |
-
controller = AttentionRefine(prompts, NUM_DDIM_STEPS, cross_replace_steps=cross_replace_steps, self_replace_steps=self_replace_steps, local_blend=lb)
|
731 |
-
if equilizer_params is not None:
|
732 |
-
eq = get_equalizer(prompts[1], equilizer_params["words"], equilizer_params["values"])
|
733 |
-
controller = AttentionReweight(prompts, NUM_DDIM_STEPS, cross_replace_steps=cross_replace_steps,
|
734 |
-
self_replace_steps=self_replace_steps, equalizer=eq, local_blend=lb, controller=controller)
|
735 |
-
return controller
|
736 |
-
|
737 |
-
|
738 |
-
def load_512_seq(image_path, left=0, right=0, top=0, bottom=0, n_sample_frame=video_len, sampling_rate=1):
|
739 |
-
vr = decord.VideoReader(image_path, width=512, height=512)
|
740 |
-
sample_index = list(range(0, len(vr), sampling_rate))[:n_sample_frame]
|
741 |
-
video = vr.get_batch(sample_index)
|
742 |
-
return video.numpy()
|
743 |
-
|
744 |
-
|
745 |
-
class NullInversion:
|
746 |
-
|
747 |
-
def prev_step(self, model_output: Union[torch.FloatTensor, np.ndarray], timestep: int, sample: Union[torch.FloatTensor, np.ndarray]):
|
748 |
-
prev_timestep = timestep - self.scheduler.config.num_train_timesteps // self.scheduler.num_inference_steps
|
749 |
-
alpha_prod_t = self.scheduler.alphas_cumprod[timestep]
|
750 |
-
alpha_prod_t_prev = self.scheduler.alphas_cumprod[prev_timestep] if prev_timestep >= 0 else self.scheduler.final_alpha_cumprod
|
751 |
-
beta_prod_t = 1 - alpha_prod_t
|
752 |
-
pred_original_sample = (sample - beta_prod_t ** 0.5 * model_output) / alpha_prod_t ** 0.5
|
753 |
-
pred_sample_direction = (1 - alpha_prod_t_prev) ** 0.5 * model_output
|
754 |
-
prev_sample = alpha_prod_t_prev ** 0.5 * pred_original_sample + pred_sample_direction
|
755 |
-
return prev_sample
|
756 |
-
|
757 |
-
def next_step(self, model_output: Union[torch.FloatTensor, np.ndarray], timestep: int, sample: Union[torch.FloatTensor, np.ndarray]):
|
758 |
-
timestep, next_timestep = min(timestep - self.scheduler.config.num_train_timesteps // self.scheduler.num_inference_steps, 999), timestep
|
759 |
-
alpha_prod_t = self.scheduler.alphas_cumprod[timestep] if timestep >= 0 else self.scheduler.final_alpha_cumprod
|
760 |
-
alpha_prod_t_next = self.scheduler.alphas_cumprod[next_timestep]
|
761 |
-
beta_prod_t = 1 - alpha_prod_t
|
762 |
-
next_original_sample = (sample - beta_prod_t ** 0.5 * model_output) / alpha_prod_t ** 0.5
|
763 |
-
next_sample_direction = (1 - alpha_prod_t_next) ** 0.5 * model_output
|
764 |
-
next_sample = alpha_prod_t_next ** 0.5 * next_original_sample + next_sample_direction
|
765 |
-
return next_sample
|
766 |
-
|
767 |
-
def get_noise_pred_single(self, latents, t, context):
|
768 |
-
noise_pred = self.model.unet(latents, t, encoder_hidden_states=context)["sample"]
|
769 |
-
return noise_pred
|
770 |
-
|
771 |
-
def get_noise_pred(self, latents, t, is_forward=True, context=None):
|
772 |
-
latents_input = torch.cat([latents] * 2)
|
773 |
-
if context is None:
|
774 |
-
context = self.context
|
775 |
-
guidance_scale = 1 if is_forward else GUIDANCE_SCALE
|
776 |
-
noise_pred = self.model.unet(latents_input, t, encoder_hidden_states=context)["sample"]
|
777 |
-
noise_pred_uncond, noise_prediction_text = noise_pred.chunk(2)
|
778 |
-
noise_pred = noise_pred_uncond + guidance_scale * (noise_prediction_text - noise_pred_uncond)
|
779 |
-
if is_forward:
|
780 |
-
latents = self.next_step(noise_pred, t, latents)
|
781 |
-
else:
|
782 |
-
latents = self.prev_step(noise_pred, t, latents)
|
783 |
-
return latents
|
784 |
-
|
785 |
-
@torch.no_grad()
|
786 |
-
def latent2image(self, latents, return_type='np'):
|
787 |
-
latents = 1 / 0.18215 * latents.detach()
|
788 |
-
image = self.model.vae.decode(latents)['sample']
|
789 |
-
if return_type == 'np':
|
790 |
-
image = (image / 2 + 0.5).clamp(0, 1)
|
791 |
-
image = image.cpu().permute(0, 2, 3, 1).numpy()[0]
|
792 |
-
image = (image * 255).astype(np.uint8)
|
793 |
-
return image
|
794 |
-
|
795 |
-
@torch.no_grad()
|
796 |
-
def latent2image_video(self, latents, return_type='np'):
|
797 |
-
latents = 1 / 0.18215 * latents.detach()
|
798 |
-
latents = latents[0].permute(1, 0, 2, 3)
|
799 |
-
image = self.model.vae.decode(latents)['sample']
|
800 |
-
if return_type == 'np':
|
801 |
-
image = (image / 2 + 0.5).clamp(0, 1)
|
802 |
-
image = image.cpu().permute(0, 2, 3, 1).numpy()
|
803 |
-
image = (image * 255).astype(np.uint8)
|
804 |
-
return image
|
805 |
-
|
806 |
-
@torch.no_grad()
|
807 |
-
def image2latent(self, image):
|
808 |
-
with torch.no_grad():
|
809 |
-
if type(image) is Image:
|
810 |
-
image = np.array(image)
|
811 |
-
if type(image) is torch.Tensor and image.dim() == 4:
|
812 |
-
latents = image
|
813 |
-
else:
|
814 |
-
image = torch.from_numpy(image).float() / 127.5 - 1
|
815 |
-
image = image.permute(2, 0, 1).unsqueeze(0).to(device, dtype=weight_dtype)
|
816 |
-
latents = self.model.vae.encode(image)['latent_dist'].mean
|
817 |
-
latents = latents * 0.18215
|
818 |
-
return latents
|
819 |
-
|
820 |
-
@torch.no_grad()
|
821 |
-
def image2latent_video(self, image):
|
822 |
-
with torch.no_grad():
|
823 |
-
image = torch.from_numpy(image).float() / 127.5 - 1
|
824 |
-
image = image.permute(0, 3, 1, 2).to(device).to(device, dtype=weight_dtype)
|
825 |
-
latents = self.model.vae.encode(image)['latent_dist'].mean
|
826 |
-
latents = rearrange(latents, "(b f) c h w -> b c f h w", b=1)
|
827 |
-
latents = latents * 0.18215
|
828 |
-
return latents
|
829 |
-
|
830 |
-
@torch.no_grad()
|
831 |
-
def init_prompt(self, prompt: str):
|
832 |
-
uncond_input = self.model.tokenizer(
|
833 |
-
[""], padding="max_length", max_length=self.model.tokenizer.model_max_length,
|
834 |
-
return_tensors="pt"
|
835 |
-
)
|
836 |
-
uncond_embeddings = self.model.text_encoder(uncond_input.input_ids.to(self.model.device))[0]
|
837 |
-
text_input = self.model.tokenizer(
|
838 |
-
[prompt],
|
839 |
-
padding="max_length",
|
840 |
-
max_length=self.model.tokenizer.model_max_length,
|
841 |
-
truncation=True,
|
842 |
-
return_tensors="pt",
|
843 |
-
)
|
844 |
-
text_embeddings = self.model.text_encoder(text_input.input_ids.to(self.model.device))[0]
|
845 |
-
self.context = torch.cat([uncond_embeddings, text_embeddings])
|
846 |
-
self.prompt = prompt
|
847 |
-
|
848 |
-
@torch.no_grad()
|
849 |
-
def ddim_loop(self, latent):
|
850 |
-
uncond_embeddings, cond_embeddings = self.context.chunk(2)
|
851 |
-
all_latent = [latent]
|
852 |
-
latent = latent.clone().detach()
|
853 |
-
for i in range(NUM_DDIM_STEPS):
|
854 |
-
t = self.model.scheduler.timesteps[len(self.model.scheduler.timesteps) - i - 1]
|
855 |
-
noise_pred = self.get_noise_pred_single(latent, t, cond_embeddings)
|
856 |
-
latent = self.next_step(noise_pred, t, latent)
|
857 |
-
all_latent.append(latent)
|
858 |
-
return all_latent
|
859 |
-
|
860 |
-
@property
|
861 |
-
def scheduler(self):
|
862 |
-
return self.model.scheduler
|
863 |
-
|
864 |
-
@torch.no_grad()
|
865 |
-
def ddim_inversion(self, image):
|
866 |
-
latent = self.image2latent_video(image)
|
867 |
-
image_rec = self.latent2image_video(latent)
|
868 |
-
ddim_latents = self.ddim_loop(latent)
|
869 |
-
return image_rec, ddim_latents
|
870 |
-
|
871 |
-
def null_optimization(self, latents, num_inner_steps, epsilon):
|
872 |
-
uncond_embeddings, cond_embeddings = self.context.chunk(2)
|
873 |
-
uncond_embeddings_list = []
|
874 |
-
latent_cur = latents[-1]
|
875 |
-
# bar = tqdm(total=num_inner_steps * NUM_DDIM_STEPS)
|
876 |
-
for i in range(NUM_DDIM_STEPS):
|
877 |
-
uncond_embeddings = uncond_embeddings.clone().detach()
|
878 |
-
uncond_embeddings.requires_grad = True
|
879 |
-
optimizer = Adam([uncond_embeddings], lr=1e-2 * (1. - i / 100.))
|
880 |
-
latent_prev = latents[len(latents) - i - 2]
|
881 |
-
t = self.model.scheduler.timesteps[i]
|
882 |
-
with torch.no_grad():
|
883 |
-
noise_pred_cond = self.get_noise_pred_single(latent_cur, t, cond_embeddings)
|
884 |
-
for j in range(num_inner_steps):
|
885 |
-
noise_pred_uncond = self.get_noise_pred_single(latent_cur, t, uncond_embeddings)
|
886 |
-
noise_pred = noise_pred_uncond + GUIDANCE_SCALE * (noise_pred_cond - noise_pred_uncond)
|
887 |
-
latents_prev_rec = self.prev_step(noise_pred, t, latent_cur)
|
888 |
-
loss = F.mse_loss(latents_prev_rec, latent_prev)
|
889 |
-
optimizer.zero_grad()
|
890 |
-
loss.backward()
|
891 |
-
optimizer.step()
|
892 |
-
loss_item = loss.item()
|
893 |
-
# bar.update()
|
894 |
-
if loss_item < epsilon + i * 2e-5:
|
895 |
-
break
|
896 |
-
# for j in range(j + 1, num_inner_steps):
|
897 |
-
# bar.update()
|
898 |
-
uncond_embeddings_list.append(uncond_embeddings[:1].detach())
|
899 |
-
with torch.no_grad():
|
900 |
-
context = torch.cat([uncond_embeddings, cond_embeddings])
|
901 |
-
latent_cur = self.get_noise_pred(latent_cur, t, False, context)
|
902 |
-
# bar.close()
|
903 |
-
return uncond_embeddings_list
|
904 |
-
|
905 |
-
def invert(self, image_path: str, prompt: str, offsets=(0,0,0,0), num_inner_steps=10, early_stop_epsilon=1e-5, verbose=False):
|
906 |
-
self.init_prompt(prompt)
|
907 |
-
ptp_utils.register_attention_control(self.model, None)
|
908 |
-
image_gt = load_512_seq(image_path, *offsets)
|
909 |
-
if verbose:
|
910 |
-
print("DDIM inversion...")
|
911 |
-
image_rec, ddim_latents = self.ddim_inversion(image_gt)
|
912 |
-
if verbose:
|
913 |
-
print("Null-text optimization...")
|
914 |
-
uncond_embeddings = self.null_optimization(ddim_latents, num_inner_steps, early_stop_epsilon)
|
915 |
-
return (image_gt, image_rec), ddim_latents[-1], uncond_embeddings
|
916 |
-
|
917 |
-
def invert_(self, image_path: str, prompt: str, offsets=(0,0,0,0), num_inner_steps=10, early_stop_epsilon=1e-5, verbose=False):
|
918 |
-
self.init_prompt(prompt)
|
919 |
-
ptp_utils.register_attention_control(self.model, None)
|
920 |
-
image_gt = load_512_seq(image_path, *offsets)
|
921 |
-
if verbose:
|
922 |
-
print("DDIM inversion...")
|
923 |
-
image_rec, ddim_latents = self.ddim_inversion(image_gt)
|
924 |
-
if verbose:
|
925 |
-
print("Null-text optimization...")
|
926 |
-
return (image_gt, image_rec), ddim_latents[-1], None
|
927 |
-
|
928 |
-
def __init__(self, model):
|
929 |
-
scheduler = DDIMScheduler(beta_start=0.00085, beta_end=0.012, beta_schedule="scaled_linear", clip_sample=False,
|
930 |
-
set_alpha_to_one=False)
|
931 |
-
self.model = model
|
932 |
-
self.tokenizer = self.model.tokenizer
|
933 |
-
self.model.scheduler.set_timesteps(NUM_DDIM_STEPS)
|
934 |
-
self.prompt = None
|
935 |
-
self.context = None
|
936 |
-
|
937 |
-
null_inversion = NullInversion(ldm_stable)
|
938 |
-
|
939 |
-
###############
|
940 |
-
# Custom APIs:
|
941 |
-
|
942 |
-
ldm_stable.enable_xformers_memory_efficient_attention()
|
943 |
-
|
944 |
-
if fast:
|
945 |
-
(image_gt, image_enc), x_t, uncond_embeddings = null_inversion.invert_(image_path, prompt, offsets=(0,0,0,0), verbose=True)
|
946 |
-
else:
|
947 |
-
(image_gt, image_enc), x_t, uncond_embeddings = null_inversion.invert(image_path, prompt, offsets=(0,0,0,0), verbose=True)
|
948 |
-
|
949 |
-
##### load uncond #####
|
950 |
-
# uncond_embeddings_load = np.load(uncond_embeddings_path)
|
951 |
-
# uncond_embeddings = []
|
952 |
-
# for i in range(uncond_embeddings_load.shape[0]):
|
953 |
-
# uncond_embeddings.append(torch.from_numpy(uncond_embeddings_load[i]).to(device))
|
954 |
-
#######################
|
955 |
-
|
956 |
-
##### save uncond #####
|
957 |
-
# uncond_embeddings = torch.cat(uncond_embeddings)
|
958 |
-
# uncond_embeddings = uncond_embeddings.cpu().numpy()
|
959 |
-
#######################
|
960 |
-
|
961 |
-
print("Start Video-P2P!")
|
962 |
-
controller = make_controller(prompts, is_word_swap, cross_replace_steps, self_replace_steps, blend_word, eq_params, mask_th=mask_th)
|
963 |
-
ptp_utils.register_attention_control(ldm_stable, controller)
|
964 |
-
generator = torch.Generator(device=device)
|
965 |
-
with torch.no_grad():
|
966 |
-
sequence = ldm_stable(
|
967 |
-
prompts,
|
968 |
-
generator=generator,
|
969 |
-
latents=x_t,
|
970 |
-
uncond_embeddings_pre=uncond_embeddings,
|
971 |
-
controller = controller,
|
972 |
-
video_length=video_len,
|
973 |
-
fast=fast,
|
974 |
-
).videos
|
975 |
-
sequence1 = rearrange(sequence[0], "c t h w -> t h w c")
|
976 |
-
sequence2 = rearrange(sequence[1], "c t h w -> t h w c")
|
977 |
-
inversion = []
|
978 |
-
videop2p = []
|
979 |
-
for i in range(sequence1.shape[0]):
|
980 |
-
inversion.append( Image.fromarray((sequence1[i] * 255).numpy().astype(np.uint8)) )
|
981 |
-
videop2p.append( Image.fromarray((sequence2[i] * 255).numpy().astype(np.uint8)) )
|
982 |
-
|
983 |
-
# inversion[0].save(save_name_1, save_all=True, append_images=inversion[1:], optimize=False, loop=0, duration=250)
|
984 |
-
videop2p[0].save(save_name_2, save_all=True, append_images=videop2p[1:], optimize=False, loop=0, duration=250)
|
985 |
-
|
986 |
-
|
987 |
-
if __name__ == "__main__":
|
988 |
-
parser = argparse.ArgumentParser()
|
989 |
-
parser.add_argument("--config", type=str, default="./configs/tuneavideo.yaml")
|
990 |
-
parser.add_argument("--fast", action='store_true')
|
991 |
-
args = parser.parse_args()
|
992 |
-
|
993 |
-
main(**OmegaConf.load(args.config), fast=args.fast)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
Video-P2P/run_tuning.py
CHANGED
@@ -1,12 +1,10 @@
|
|
1 |
-
# From https://github.com/showlab/Tune-A-Video/blob/main/train_tuneavideo.py
|
2 |
-
|
3 |
import argparse
|
4 |
import datetime
|
5 |
import logging
|
6 |
import inspect
|
7 |
import math
|
8 |
import os
|
9 |
-
from typing import
|
10 |
from omegaconf import OmegaConf
|
11 |
|
12 |
import torch
|
@@ -23,7 +21,7 @@ from diffusers.optimization import get_scheduler
|
|
23 |
from diffusers.utils import check_min_version
|
24 |
from diffusers.utils.import_utils import is_xformers_available
|
25 |
from tqdm.auto import tqdm
|
26 |
-
from transformers import CLIPTextModel, CLIPTokenizer
|
27 |
|
28 |
from tuneavideo.models.unet import UNet3DConditionModel
|
29 |
from tuneavideo.data.dataset import TuneAVideoDataset
|
@@ -31,6 +29,16 @@ from tuneavideo.pipelines.pipeline_tuneavideo import TuneAVideoPipeline
|
|
31 |
from tuneavideo.util import save_videos_grid, ddim_inversion
|
32 |
from einops import rearrange
|
33 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
34 |
|
35 |
# Will error if the minimal version of diffusers is not installed. Remove at your own risks.
|
36 |
check_min_version("0.10.0.dev0")
|
@@ -68,6 +76,19 @@ def main(
|
|
68 |
use_8bit_adam: bool = False,
|
69 |
enable_xformers_memory_efficient_attention: bool = True,
|
70 |
seed: Optional[int] = None,
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
71 |
):
|
72 |
*_, config = inspect.getargvalues(inspect.currentframe())
|
73 |
|
@@ -96,6 +117,8 @@ def main(
|
|
96 |
|
97 |
# Handle the output folder creation
|
98 |
if accelerator.is_main_process:
|
|
|
|
|
99 |
os.makedirs(output_dir, exist_ok=True)
|
100 |
os.makedirs(f"{output_dir}/samples", exist_ok=True)
|
101 |
os.makedirs(f"{output_dir}/inv_latents", exist_ok=True)
|
@@ -358,10 +381,12 @@ def main(
|
|
358 |
|
359 |
accelerator.end_training()
|
360 |
|
|
|
361 |
|
362 |
if __name__ == "__main__":
|
363 |
parser = argparse.ArgumentParser()
|
364 |
parser.add_argument("--config", type=str, default="./configs/tuneavideo.yaml")
|
|
|
365 |
args = parser.parse_args()
|
366 |
|
367 |
-
main(**OmegaConf.load(args.config))
|
|
|
|
|
|
|
1 |
import argparse
|
2 |
import datetime
|
3 |
import logging
|
4 |
import inspect
|
5 |
import math
|
6 |
import os
|
7 |
+
from typing import Optional, Union, Tuple, List, Callable, Dict
|
8 |
from omegaconf import OmegaConf
|
9 |
|
10 |
import torch
|
|
|
21 |
from diffusers.utils import check_min_version
|
22 |
from diffusers.utils.import_utils import is_xformers_available
|
23 |
from tqdm.auto import tqdm
|
24 |
+
from transformers import AutoTokenizer, CLIPTextModel, CLIPTokenizer
|
25 |
|
26 |
from tuneavideo.models.unet import UNet3DConditionModel
|
27 |
from tuneavideo.data.dataset import TuneAVideoDataset
|
|
|
29 |
from tuneavideo.util import save_videos_grid, ddim_inversion
|
30 |
from einops import rearrange
|
31 |
|
32 |
+
import cv2
|
33 |
+
import abc
|
34 |
+
import ptp_utils
|
35 |
+
import seq_aligner
|
36 |
+
import shutil
|
37 |
+
from torch.optim.adam import Adam
|
38 |
+
from PIL import Image
|
39 |
+
import numpy as np
|
40 |
+
import decord
|
41 |
+
decord.bridge.set_bridge('torch')
|
42 |
|
43 |
# Will error if the minimal version of diffusers is not installed. Remove at your own risks.
|
44 |
check_min_version("0.10.0.dev0")
|
|
|
76 |
use_8bit_adam: bool = False,
|
77 |
enable_xformers_memory_efficient_attention: bool = True,
|
78 |
seed: Optional[int] = None,
|
79 |
+
# pretrained_model_path: str,
|
80 |
+
# image_path: str = None,
|
81 |
+
# prompt: str = None,
|
82 |
+
prompts: Tuple[str] = None,
|
83 |
+
eq_params: Dict = None,
|
84 |
+
save_name: str = None,
|
85 |
+
is_word_swap: bool = None,
|
86 |
+
blend_word: Tuple[str] = None,
|
87 |
+
cross_replace_steps: float = 0.2,
|
88 |
+
self_replace_steps: float = 0.5,
|
89 |
+
video_len: int = 8,
|
90 |
+
fast: bool = False,
|
91 |
+
mixed_precision_p2p: str = 'fp32',
|
92 |
):
|
93 |
*_, config = inspect.getargvalues(inspect.currentframe())
|
94 |
|
|
|
117 |
|
118 |
# Handle the output folder creation
|
119 |
if accelerator.is_main_process:
|
120 |
+
# now = datetime.datetime.now().strftime("%Y-%m-%dT%H-%M-%S")
|
121 |
+
# output_dir = os.path.join(output_dir, now)
|
122 |
os.makedirs(output_dir, exist_ok=True)
|
123 |
os.makedirs(f"{output_dir}/samples", exist_ok=True)
|
124 |
os.makedirs(f"{output_dir}/inv_latents", exist_ok=True)
|
|
|
381 |
|
382 |
accelerator.end_training()
|
383 |
|
384 |
+
torch.cuda.empty_cache()
|
385 |
|
386 |
if __name__ == "__main__":
|
387 |
parser = argparse.ArgumentParser()
|
388 |
parser.add_argument("--config", type=str, default="./configs/tuneavideo.yaml")
|
389 |
+
parser.add_argument("--fast", action='store_true')
|
390 |
args = parser.parse_args()
|
391 |
|
392 |
+
main(**OmegaConf.load(args.config), fast=args.fast)
|
Video-P2P/run_videop2p.py
CHANGED
@@ -1,54 +1,113 @@
|
|
1 |
-
|
2 |
-
|
|
|
|
|
|
|
3 |
import os
|
4 |
from typing import Optional, Union, Tuple, List, Callable, Dict
|
5 |
-
from
|
|
|
6 |
import torch
|
7 |
-
|
8 |
-
import torch.
|
9 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
10 |
import abc
|
11 |
import ptp_utils
|
12 |
import seq_aligner
|
13 |
import shutil
|
14 |
from torch.optim.adam import Adam
|
15 |
from PIL import Image
|
16 |
-
|
17 |
-
|
|
|
18 |
|
19 |
-
|
20 |
-
|
21 |
|
22 |
-
|
23 |
-
import argparse
|
24 |
-
from omegaconf import OmegaConf
|
25 |
|
26 |
-
scheduler = DDIMScheduler(beta_start=0.00085, beta_end=0.012, beta_schedule="scaled_linear", clip_sample=False, set_alpha_to_one=False)
|
27 |
-
MY_TOKEN = ''
|
28 |
-
LOW_RESOURCE = False
|
29 |
-
NUM_DDIM_STEPS = 50
|
30 |
-
GUIDANCE_SCALE = 7.5
|
31 |
-
MAX_NUM_WORDS = 77
|
32 |
-
device = torch.device('cuda:0') if torch.cuda.is_available() else torch.device('cpu')
|
33 |
-
|
34 |
-
# need to adjust sometimes
|
35 |
-
mask_th = (.3, .3)
|
36 |
|
37 |
def main(
|
38 |
pretrained_model_path: str,
|
39 |
-
|
40 |
-
|
41 |
-
|
42 |
-
|
43 |
-
|
44 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
45 |
blend_word: Tuple[str] = None,
|
46 |
cross_replace_steps: float = 0.2,
|
47 |
self_replace_steps: float = 0.5,
|
48 |
video_len: int = 8,
|
49 |
fast: bool = False,
|
50 |
-
|
51 |
):
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
52 |
output_folder = os.path.join(pretrained_model_path, 'results')
|
53 |
if fast:
|
54 |
save_name_1 = os.path.join(output_folder, 'inversion_fast.gif')
|
@@ -63,9 +122,9 @@ def main(
|
|
63 |
cross_replace_steps = {'default_': cross_replace_steps,}
|
64 |
|
65 |
weight_dtype = torch.float32
|
66 |
-
if
|
67 |
weight_dtype = torch.float16
|
68 |
-
elif
|
69 |
weight_dtype = torch.bfloat16
|
70 |
|
71 |
if not os.path.exists(output_folder):
|
@@ -106,8 +165,8 @@ def main(
|
|
106 |
k = 1
|
107 |
maps = (maps * alpha).sum(-1).mean(2)
|
108 |
if use_pool:
|
109 |
-
maps =
|
110 |
-
mask =
|
111 |
mask = mask / mask.max(2, keepdims=True)[0].max(3, keepdims=True)[0]
|
112 |
mask = mask.gt(self.th[1-int(use_pool)])
|
113 |
mask = mask[:1] + mask
|
@@ -385,33 +444,10 @@ def main(
|
|
385 |
|
386 |
|
387 |
def load_512_seq(image_path, left=0, right=0, top=0, bottom=0, n_sample_frame=video_len, sampling_rate=1):
|
388 |
-
|
389 |
-
|
390 |
-
|
391 |
-
|
392 |
-
sequence_length = (n_sample_frame - 1) * sampling_rate + 1
|
393 |
-
if n_images < sequence_length:
|
394 |
-
raise ValueError
|
395 |
-
frames = []
|
396 |
-
for index in range(n_sample_frame):
|
397 |
-
p = os.path.join(image_path, images[index])
|
398 |
-
image = np.array(Image.open(p).convert("RGB"))
|
399 |
-
h, w, c = image.shape
|
400 |
-
left = min(left, w-1)
|
401 |
-
right = min(right, w - left - 1)
|
402 |
-
top = min(top, h - left - 1)
|
403 |
-
bottom = min(bottom, h - top - 1)
|
404 |
-
image = image[top:h-bottom, left:w-right]
|
405 |
-
h, w, c = image.shape
|
406 |
-
if h < w:
|
407 |
-
offset = (w - h) // 2
|
408 |
-
image = image[:, offset:offset + h]
|
409 |
-
elif w < h:
|
410 |
-
offset = (h - w) // 2
|
411 |
-
image = image[offset:offset + w]
|
412 |
-
image = np.array(Image.fromarray(image).resize((512, 512)))
|
413 |
-
frames.append(image)
|
414 |
-
return np.stack(frames)
|
415 |
|
416 |
|
417 |
class NullInversion:
|
@@ -544,7 +580,7 @@ def main(
|
|
544 |
uncond_embeddings, cond_embeddings = self.context.chunk(2)
|
545 |
uncond_embeddings_list = []
|
546 |
latent_cur = latents[-1]
|
547 |
-
bar = tqdm(total=num_inner_steps * NUM_DDIM_STEPS)
|
548 |
for i in range(NUM_DDIM_STEPS):
|
549 |
uncond_embeddings = uncond_embeddings.clone().detach()
|
550 |
uncond_embeddings.requires_grad = True
|
@@ -557,21 +593,21 @@ def main(
|
|
557 |
noise_pred_uncond = self.get_noise_pred_single(latent_cur, t, uncond_embeddings)
|
558 |
noise_pred = noise_pred_uncond + GUIDANCE_SCALE * (noise_pred_cond - noise_pred_uncond)
|
559 |
latents_prev_rec = self.prev_step(noise_pred, t, latent_cur)
|
560 |
-
loss =
|
561 |
optimizer.zero_grad()
|
562 |
loss.backward()
|
563 |
optimizer.step()
|
564 |
loss_item = loss.item()
|
565 |
-
bar.update()
|
566 |
if loss_item < epsilon + i * 2e-5:
|
567 |
break
|
568 |
-
for j in range(j + 1, num_inner_steps):
|
569 |
-
|
570 |
uncond_embeddings_list.append(uncond_embeddings[:1].detach())
|
571 |
with torch.no_grad():
|
572 |
context = torch.cat([uncond_embeddings, cond_embeddings])
|
573 |
latent_cur = self.get_noise_pred(latent_cur, t, False, context)
|
574 |
-
bar.close()
|
575 |
return uncond_embeddings_list
|
576 |
|
577 |
def invert(self, image_path: str, prompt: str, offsets=(0,0,0,0), num_inner_steps=10, early_stop_epsilon=1e-5, verbose=False):
|
@@ -652,12 +688,13 @@ def main(
|
|
652 |
inversion.append( Image.fromarray((sequence1[i] * 255).numpy().astype(np.uint8)) )
|
653 |
videop2p.append( Image.fromarray((sequence2[i] * 255).numpy().astype(np.uint8)) )
|
654 |
|
655 |
-
inversion[0].save(save_name_1, save_all=True, append_images=inversion[1:], optimize=False, loop=0, duration=250)
|
656 |
videop2p[0].save(save_name_2, save_all=True, append_images=videop2p[1:], optimize=False, loop=0, duration=250)
|
657 |
|
|
|
658 |
if __name__ == "__main__":
|
659 |
parser = argparse.ArgumentParser()
|
660 |
-
parser.add_argument("--config", type=str, default="./configs/
|
661 |
parser.add_argument("--fast", action='store_true')
|
662 |
args = parser.parse_args()
|
663 |
|
|
|
1 |
+
import argparse
|
2 |
+
import datetime
|
3 |
+
import logging
|
4 |
+
import inspect
|
5 |
+
import math
|
6 |
import os
|
7 |
from typing import Optional, Union, Tuple, List, Callable, Dict
|
8 |
+
from omegaconf import OmegaConf
|
9 |
+
|
10 |
import torch
|
11 |
+
import torch.nn.functional as F
|
12 |
+
import torch.utils.checkpoint
|
13 |
+
|
14 |
+
import diffusers
|
15 |
+
import transformers
|
16 |
+
from accelerate import Accelerator
|
17 |
+
from accelerate.logging import get_logger
|
18 |
+
from accelerate.utils import set_seed
|
19 |
+
from diffusers import AutoencoderKL, DDPMScheduler, DDIMScheduler
|
20 |
+
from diffusers.optimization import get_scheduler
|
21 |
+
from diffusers.utils import check_min_version
|
22 |
+
from diffusers.utils.import_utils import is_xformers_available
|
23 |
+
from tqdm.auto import tqdm
|
24 |
+
from transformers import AutoTokenizer, CLIPTextModel, CLIPTokenizer
|
25 |
+
|
26 |
+
from tuneavideo.models.unet import UNet3DConditionModel
|
27 |
+
from tuneavideo.data.dataset import TuneAVideoDataset
|
28 |
+
from tuneavideo.pipelines.pipeline_tuneavideo import TuneAVideoPipeline
|
29 |
+
from tuneavideo.util import save_videos_grid, ddim_inversion
|
30 |
+
from einops import rearrange
|
31 |
+
|
32 |
+
import cv2
|
33 |
import abc
|
34 |
import ptp_utils
|
35 |
import seq_aligner
|
36 |
import shutil
|
37 |
from torch.optim.adam import Adam
|
38 |
from PIL import Image
|
39 |
+
import numpy as np
|
40 |
+
import decord
|
41 |
+
decord.bridge.set_bridge('torch')
|
42 |
|
43 |
+
# Will error if the minimal version of diffusers is not installed. Remove at your own risks.
|
44 |
+
check_min_version("0.10.0.dev0")
|
45 |
|
46 |
+
logger = get_logger(__name__, log_level="INFO")
|
|
|
|
|
47 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
48 |
|
49 |
def main(
|
50 |
pretrained_model_path: str,
|
51 |
+
output_dir: str,
|
52 |
+
train_data: Dict,
|
53 |
+
validation_data: Dict,
|
54 |
+
validation_steps: int = 100,
|
55 |
+
trainable_modules: Tuple[str] = (
|
56 |
+
"attn1.to_q",
|
57 |
+
"attn2.to_q",
|
58 |
+
"attn_temp",
|
59 |
+
),
|
60 |
+
train_batch_size: int = 1,
|
61 |
+
max_train_steps: int = 500,
|
62 |
+
learning_rate: float = 3e-5,
|
63 |
+
scale_lr: bool = False,
|
64 |
+
lr_scheduler: str = "constant",
|
65 |
+
lr_warmup_steps: int = 0,
|
66 |
+
adam_beta1: float = 0.9,
|
67 |
+
adam_beta2: float = 0.999,
|
68 |
+
adam_weight_decay: float = 1e-2,
|
69 |
+
adam_epsilon: float = 1e-08,
|
70 |
+
max_grad_norm: float = 1.0,
|
71 |
+
gradient_accumulation_steps: int = 1,
|
72 |
+
gradient_checkpointing: bool = True,
|
73 |
+
checkpointing_steps: int = 500,
|
74 |
+
resume_from_checkpoint: Optional[str] = None,
|
75 |
+
mixed_precision: Optional[str] = "fp16",
|
76 |
+
use_8bit_adam: bool = False,
|
77 |
+
enable_xformers_memory_efficient_attention: bool = True,
|
78 |
+
seed: Optional[int] = None,
|
79 |
+
# pretrained_model_path: str,
|
80 |
+
# image_path: str = None,
|
81 |
+
# prompt: str = None,
|
82 |
+
prompts: Tuple[str] = None,
|
83 |
+
eq_params: Dict = None,
|
84 |
+
save_name: str = None,
|
85 |
+
is_word_swap: bool = None,
|
86 |
blend_word: Tuple[str] = None,
|
87 |
cross_replace_steps: float = 0.2,
|
88 |
self_replace_steps: float = 0.5,
|
89 |
video_len: int = 8,
|
90 |
fast: bool = False,
|
91 |
+
mixed_precision_p2p: str = 'fp32',
|
92 |
):
|
93 |
+
|
94 |
+
# Video-P2P
|
95 |
+
scheduler = DDIMScheduler(beta_start=0.00085, beta_end=0.012, beta_schedule="scaled_linear", clip_sample=False, set_alpha_to_one=False)
|
96 |
+
MY_TOKEN = ''
|
97 |
+
LOW_RESOURCE = False
|
98 |
+
NUM_DDIM_STEPS = 50
|
99 |
+
GUIDANCE_SCALE = 7.5
|
100 |
+
MAX_NUM_WORDS = 77
|
101 |
+
device = torch.device('cuda:0') if torch.cuda.is_available() else torch.device('cpu')
|
102 |
+
|
103 |
+
# need to adjust sometimes
|
104 |
+
mask_th = (.3, .3)
|
105 |
+
|
106 |
+
|
107 |
+
pretrained_model_path = output_dir
|
108 |
+
image_path = train_data['video_path']
|
109 |
+
prompt = train_data['prompt']
|
110 |
+
# prompts = [prompt, ]
|
111 |
output_folder = os.path.join(pretrained_model_path, 'results')
|
112 |
if fast:
|
113 |
save_name_1 = os.path.join(output_folder, 'inversion_fast.gif')
|
|
|
122 |
cross_replace_steps = {'default_': cross_replace_steps,}
|
123 |
|
124 |
weight_dtype = torch.float32
|
125 |
+
if mixed_precision_p2p == "fp16":
|
126 |
weight_dtype = torch.float16
|
127 |
+
elif mixed_precision_p2p == "bf16":
|
128 |
weight_dtype = torch.bfloat16
|
129 |
|
130 |
if not os.path.exists(output_folder):
|
|
|
165 |
k = 1
|
166 |
maps = (maps * alpha).sum(-1).mean(2)
|
167 |
if use_pool:
|
168 |
+
maps = F.max_pool2d(maps, (k * 2 + 1, k * 2 +1), (1, 1), padding=(k, k))
|
169 |
+
mask = F.interpolate(maps, size=(x_t.shape[3:]))
|
170 |
mask = mask / mask.max(2, keepdims=True)[0].max(3, keepdims=True)[0]
|
171 |
mask = mask.gt(self.th[1-int(use_pool)])
|
172 |
mask = mask[:1] + mask
|
|
|
444 |
|
445 |
|
446 |
def load_512_seq(image_path, left=0, right=0, top=0, bottom=0, n_sample_frame=video_len, sampling_rate=1):
|
447 |
+
vr = decord.VideoReader(image_path, width=512, height=512)
|
448 |
+
sample_index = list(range(0, len(vr), sampling_rate))[:n_sample_frame]
|
449 |
+
video = vr.get_batch(sample_index)
|
450 |
+
return video.numpy()
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
451 |
|
452 |
|
453 |
class NullInversion:
|
|
|
580 |
uncond_embeddings, cond_embeddings = self.context.chunk(2)
|
581 |
uncond_embeddings_list = []
|
582 |
latent_cur = latents[-1]
|
583 |
+
# bar = tqdm(total=num_inner_steps * NUM_DDIM_STEPS)
|
584 |
for i in range(NUM_DDIM_STEPS):
|
585 |
uncond_embeddings = uncond_embeddings.clone().detach()
|
586 |
uncond_embeddings.requires_grad = True
|
|
|
593 |
noise_pred_uncond = self.get_noise_pred_single(latent_cur, t, uncond_embeddings)
|
594 |
noise_pred = noise_pred_uncond + GUIDANCE_SCALE * (noise_pred_cond - noise_pred_uncond)
|
595 |
latents_prev_rec = self.prev_step(noise_pred, t, latent_cur)
|
596 |
+
loss = F.mse_loss(latents_prev_rec, latent_prev)
|
597 |
optimizer.zero_grad()
|
598 |
loss.backward()
|
599 |
optimizer.step()
|
600 |
loss_item = loss.item()
|
601 |
+
# bar.update()
|
602 |
if loss_item < epsilon + i * 2e-5:
|
603 |
break
|
604 |
+
# for j in range(j + 1, num_inner_steps):
|
605 |
+
# bar.update()
|
606 |
uncond_embeddings_list.append(uncond_embeddings[:1].detach())
|
607 |
with torch.no_grad():
|
608 |
context = torch.cat([uncond_embeddings, cond_embeddings])
|
609 |
latent_cur = self.get_noise_pred(latent_cur, t, False, context)
|
610 |
+
# bar.close()
|
611 |
return uncond_embeddings_list
|
612 |
|
613 |
def invert(self, image_path: str, prompt: str, offsets=(0,0,0,0), num_inner_steps=10, early_stop_epsilon=1e-5, verbose=False):
|
|
|
688 |
inversion.append( Image.fromarray((sequence1[i] * 255).numpy().astype(np.uint8)) )
|
689 |
videop2p.append( Image.fromarray((sequence2[i] * 255).numpy().astype(np.uint8)) )
|
690 |
|
691 |
+
# inversion[0].save(save_name_1, save_all=True, append_images=inversion[1:], optimize=False, loop=0, duration=250)
|
692 |
videop2p[0].save(save_name_2, save_all=True, append_images=videop2p[1:], optimize=False, loop=0, duration=250)
|
693 |
|
694 |
+
|
695 |
if __name__ == "__main__":
|
696 |
parser = argparse.ArgumentParser()
|
697 |
+
parser.add_argument("--config", type=str, default="./configs/tuneavideo.yaml")
|
698 |
parser.add_argument("--fast", action='store_true')
|
699 |
args = parser.parse_args()
|
700 |
|
trainer.py
CHANGED
@@ -145,7 +145,9 @@ class Trainer:
|
|
145 |
with open(config_path, 'w') as f:
|
146 |
OmegaConf.save(config, f)
|
147 |
|
148 |
-
command = f'accelerate launch Video-P2P/
|
|
|
|
|
149 |
subprocess.run(shlex.split(command))
|
150 |
save_model_card(save_dir=output_dir,
|
151 |
base_model=base_model,
|
|
|
145 |
with open(config_path, 'w') as f:
|
146 |
OmegaConf.save(config, f)
|
147 |
|
148 |
+
command = f'accelerate launch Video-P2P/run_tuning.py --config {config_path}'
|
149 |
+
subprocess.run(shlex.split(command))
|
150 |
+
command = f'python Video-P2P/run_tuning.py --config {config_path} --fast'
|
151 |
subprocess.run(shlex.split(command))
|
152 |
save_model_card(save_dir=output_dir,
|
153 |
base_model=base_model,
|