-
Notifications
You must be signed in to change notification settings - Fork 1
/
unet_reward_lcm.py
383 lines (327 loc) · 19.7 KB
/
unet_reward_lcm.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
# Copyright 2022 Luping Liu
#
# 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.
import os
import sys
import math
import json
import inspect
import random
import itertools
import argparse
import platform
import numpy as np
import wandb.util
from tqdm.auto import tqdm
from PIL import Image
import matplotlib.pyplot as plt
import torch
from torch.utils import checkpoint
import torch.nn.functional as F
import torch.utils.checkpoint
from torch.utils.data import Dataset
from torchvision import transforms
from accelerate import Accelerator
from diffusers.optimization import get_scheduler
from diffusers import AutoencoderKL, DDPMScheduler, UNet2DConditionModel
from transformers import AutoTokenizer, T5EncoderModel, CLIPTextModel
from modules.lora import inject_trainable_lora_extended, save_lora_weight
from modules.lora import monkeypatch_or_replace_lora_extended, collapse_lora, monkeypatch_remove_lora
from modules.adapters import TextAdapter
from longclip import longclip
from tools import load_dataset, collate_fn, caption2embed, sample_images
from tools import scalings_for_boundary_conditions, get_predicted_original, DDIMSolver
from loss import pick_score, hpsv2, dense_score, dense_score_plus
# Arguments
def parse_args():
parser = argparse.ArgumentParser(description="LongAlign Training")
parser.add_argument("--pretrained_decoder", type=str, default="stabilityai/stable-diffusion-2-1")
parser.add_argument("--mixed_precision", type=str, default="no")
parser.add_argument("--token_length", type=int, default=77)
parser.add_argument("--resolution", type=int, default=512)
parser.add_argument("--lora_rank", type=int, default=32)
parser.add_argument("--loss_type", type=str, default="l2")
parser.add_argument("--batch_size", type=int, default=1)
parser.add_argument("--accumulation_steps", type=int, default=1)
parser.add_argument("--learning_rate", type=float, default=1e-5)
parser.add_argument("--gradient_checkpointing", action="store_true")
parser.add_argument("--max_train_epochs", type=int, default=1)
parser.add_argument("--validation_steps", type=int, default=500)
parser.add_argument("--save_steps", type=int, default=1250)
parser.add_argument("--local_rank", type=int, default=-1)
parser.add_argument("--output_dir", type=str, default="")
parser.add_argument("--ckpt_dir", type=str, default="")
parser.add_argument("--resume", action="store_true")
parser.add_argument("--debug", action="store_true")
args = parser.parse_args()
env_local_rank = int(os.environ.get("LOCAL_RANK", -1))
if env_local_rank != -1 and env_local_rank != args.local_rank:
args.local_rank = env_local_rank
return args
def main(args):
accelerator = Accelerator(
mixed_precision=args.mixed_precision,
gradient_accumulation_steps=args.accumulation_steps,
log_with='wandb' if not args.debug else None,
)
if accelerator.is_main_process and not args.debug:
os.makedirs(args.output_dir, exist_ok=True)
weight_dtype = torch.float32
if args.mixed_precision == "fp16":
weight_dtype = torch.float16
elif args.mixed_precision == "bf16":
weight_dtype = torch.bfloat16
# Blocks to inject LoRA
# note: result without CrossAttention is poor
VIS_REPLACE_MODULES = {"ResnetBlock2D", "CrossAttention", "Attention", "GEGLU"}
# Modules of T2I diffusion models
vae = AutoencoderKL.from_pretrained(args.pretrained_decoder, subfolder="vae", torch_dtype=weight_dtype)
vis = UNet2DConditionModel.from_pretrained(args.pretrained_decoder, subfolder="unet", torch_dtype=weight_dtype)
# teacher_vis = UNet2DConditionModel.from_pretrained(args.pretrained_decoder, subfolder="unet",
# torch_dtype=weight_dtype)
tokenizer_clip = AutoTokenizer.from_pretrained(args.pretrained_decoder, subfolder="tokenizer",
torch_dtype=weight_dtype, use_fast=False)
text_encoder_clip = CLIPTextModel.from_pretrained(args.pretrained_decoder, subfolder="text_encoder",
torch_dtype=weight_dtype)
tokenizer_t5 = AutoTokenizer.from_pretrained("google-t5/t5-large", torch_dtype=weight_dtype, model_max_length=512)
text_encoder_t5 = T5EncoderModel.from_pretrained("google-t5/t5-large", torch_dtype=weight_dtype)
# download from https://huggingface.co/shihaozhao/LaVi-Bridge/tree/main/t5_unet/adapter --> ./model/LaVi-Bridge
adapter = TextAdapter.from_pretrained('./model/LaVi-Bridge')
noise_scheduler = DDPMScheduler.from_pretrained(args.pretrained_decoder, subfolder="scheduler",
torch_dtype=weight_dtype)
alpha_schedule = torch.sqrt(noise_scheduler.alphas_cumprod)
sigma_schedule = torch.sqrt(1 - noise_scheduler.alphas_cumprod)
args.num_ddim_timesteps = 50
solver = DDIMSolver(noise_scheduler.alphas_cumprod.numpy(),
timesteps=noise_scheduler.config.num_train_timesteps,
ddim_timesteps=args.num_ddim_timesteps)
if args.ckpt_dir != "":
args.ckpt_dir = args.ckpt_dir.split(", ")
for ckpt_dir in args.ckpt_dir:
if os.path.exists(os.path.join(ckpt_dir, f"adapter")):
adapter = TextAdapter.from_pretrained(os.path.join(ckpt_dir, f"adapter"))
# else:
# print(f"adapter ckpt not found in {args.ckpt_dir}")
# LoRA
monkeypatch_or_replace_lora_extended(
vis,
torch.load(os.path.join(ckpt_dir, f"lora_vis.pt"), map_location="cpu"),
r=32,
target_replace_module=VIS_REPLACE_MODULES,
)
# merge LoRA
collapse_lora(vis, VIS_REPLACE_MODULES)
monkeypatch_remove_lora(vis)
# LoRA injection
vis_lora_params, _ = inject_trainable_lora_extended(
vis,
r=args.lora_rank,
target_replace_module=VIS_REPLACE_MODULES,
)
if args.gradient_checkpointing:
vis.enable_gradient_checkpointing()
# Dataset and dataloader
train_dataset, args.validation_prompts = load_dataset(args)
train_dataloader = torch.utils.data.DataLoader(train_dataset, batch_size=args.batch_size, prefetch_factor=3,
num_workers=6 if not args.debug else 6, collate_fn=collate_fn,
shuffle=True, drop_last=True, pin_memory=True)
args.max_train_steps = len(train_dataset) * args.max_train_epochs // \
(args.batch_size * accelerator.num_processes * accelerator.gradient_accumulation_steps)
# Optimizer and scheduler
optimizer_class = torch.optim.AdamW
params_to_optimize = ([
{"params": itertools.chain(*vis_lora_params)},
# {"params": adapter.parameters()},
])
# todo check 1e-6 and 1e-8 optim_eps
optimizer = optimizer_class(params_to_optimize, lr=args.learning_rate, betas=(0.9, 0.999),
weight_decay=1e-2, eps=1e-8)
lr_scheduler = get_scheduler("constant_with_warmup", optimizer=optimizer,
num_warmup_steps=2000 * accelerator.num_processes,
num_training_steps=args.max_train_steps * accelerator.num_processes)
vae.to(accelerator.device, weight_dtype)
# teacher_vis.to(accelerator.device, weight_dtype)
text_encoder_clip.to(accelerator.device, weight_dtype)
# text_encoder_t5.to(accelerator.device, weight_dtype)
# text_encoder_longclip.to(accelerator.device, weight_dtype)
alpha_schedule = alpha_schedule.to(accelerator.device, weight_dtype)
sigma_schedule = sigma_schedule.to(accelerator.device, weight_dtype)
get_predicted_original_sample = lambda model_output, timesteps, sample: get_predicted_original(
model_output, timesteps, sample, noise_scheduler.config.prediction_type, alpha_schedule, sigma_schedule)
solver = solver.to(accelerator.device, weight_dtype)
vae.eval()
# teacher_vis.eval()
text_encoder_clip.eval()
# text_encoder_t5.eval()
# text_encoder_longclip.eval()
vis, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(
vis, optimizer, train_dataloader, lr_scheduler, )
caption2embed_simple = lambda captions: caption2embed(captions, [tokenizer_clip, tokenizer_t5], [text_encoder_clip, text_encoder_t5],
args, accelerator.device, weight_dtype)
if not os.path.exists(f"{args.output_dir}/latest_status"):
args.resume = False
if args.resume:
accelerator.load_state(f"{args.output_dir}/latest_status")
resume_json = json.load(open(f"{args.output_dir}/latest_status/resume.json", 'r'))
global_step = resume_json['global_step']
last_save = global_step // accelerator.gradient_accumulation_steps
num_train_epochs = args.max_train_epochs - global_step // len(train_dataset)
wandb_id = resume_json['wandb_id']
else:
global_step = last_save = 0
num_train_epochs = args.max_train_epochs
wandb_id = wandb.util.generate_id()
resume_json = {'global_step': global_step, 'wandb_id': wandb_id}
# Log
if accelerator.is_main_process:
if not args.debug:
tracker_config = dict(vars(args))
accelerator.init_trackers(project_name='dense-final-sea', config=tracker_config,
init_kwargs={'wandb': {'name': args.output_dir.split('/')[-1],
'resume': args.resume, 'id': wandb_id}, })
# adapter.eval()
vis.eval()
sample_images(vae, adapter, caption2embed_simple, vis, args, accelerator, weight_dtype, global_step=0)
print(f"Comuting node = {platform.node()}")
print(f"Num examples = {len(train_dataset)}")
print(f"Total batch size = {args.batch_size * accelerator.num_processes * accelerator.gradient_accumulation_steps}")
print(f"Num Epochs = {num_train_epochs}")
print(f"Total optimization steps = {args.max_train_steps}")
accelerator.wait_for_everyone()
progress_bar = tqdm(range(args.max_train_steps), disable=not accelerator.is_local_main_process, dynamic_ncols=True)
progress_bar.set_description("Steps")
if args.resume:
progress_bar.update(global_step // accelerator.gradient_accumulation_steps)
ortho_weight = 1. - float(os.environ.get('REWEIGHT', 0.3))
# Training
for _ in range(num_train_epochs):
# adapter.train()
vis.train()
for _, batch in enumerate(train_dataloader):
with accelerator.accumulate(vis):
with torch.no_grad():
# Latent preparation
latents = []
batch['pixel_values'] = batch['pixel_values'].to(weight_dtype)
for i in range(0, batch['pixel_values'].shape[0], 8):
latents.append(vae.encode(batch['pixel_values'][i: i + 8]).latent_dist.sample())
latents = torch.cat(latents, dim=0)
# latents = vae.encode(batch['pixel_values'].to(weight_dtype)).latent_dist.sample()
latents = latents * vae.config.scaling_factor
bsz = latents.shape[0]
if global_step % 2 == 1:
start_timesteps = torch.randint(noise_scheduler.config.num_train_timesteps // 2,
noise_scheduler.config.num_train_timesteps,
(latents.shape[0],), device=latents.device).long()
else:
start_timesteps = torch.randint(8, noise_scheduler.config.num_train_timesteps // 2,
(latents.shape[0],), device=latents.device).long()
timesteps = start_timesteps - solver.step_ratio
timesteps = torch.where(timesteps < 0, torch.zeros_like(timesteps), timesteps)
noise = torch.randn_like(latents).to(weight_dtype)
noisy_model_input = noise_scheduler.add_noise(latents, noise, start_timesteps)
# noisy_model_input_pre = noise_scheduler.add_noise(latents, noise, timesteps)
# 5. Sample a random guidance scale w from U[w_min, w_max]
# Note that for LCM-LoRA distillation it is not necessary to use a guidance scale embedding
# args.w_min, args.w_max = 4, 10
args.w_min = args.w_max = 7.5
w = (args.w_max - args.w_min) * torch.rand((bsz,)) + args.w_min
w = w.reshape(bsz, 1, 1, 1)
w = w.to(device=latents.device, dtype=latents.dtype)
# 6. Prepare prompt embeds and unet_added_conditions
# captions = [c if random.random() > 0.1 else '' for c in batch['caption']]
captions = batch['caption']
# encoder_hidden_states_pre = text_encoder(batch['input_ids'])[0]
encoder_hidden_states_pre = caption2embed_simple([batch['caption_split'], batch['caption_index']])
encoder_hidden_states_clip = encoder_hidden_states_pre['encoder_hidden_states_clip_concat']
encoder_hidden_states_t5 = encoder_hidden_states_pre["encoder_hidden_states_t5"]
encoder_hidden_states_t5 = adapter(encoder_hidden_states_t5).sample
encoder_hidden_states_ct5 = torch.cat([encoder_hidden_states_clip, encoder_hidden_states_t5], dim=1)
x_prev = noise
for t in [999, 749, 499, 249]:
# noise_pred = vis(x_prev.detach(), t * torch.ones_like(timesteps), encoder_hidden_states_ct5).sample
noise_pred = checkpoint.checkpoint(vis, x_prev.detach(), t * torch.ones_like(timesteps), # .detach()
encoder_hidden_states_ct5, use_reentrant=False).sample
pred_x_0 = get_predicted_original_sample(noise_pred, t * torch.ones_like(timesteps), x_prev)
x_prev = solver.ddim_step(pred_x_0, torch.randn_like(noise_pred), (t - 250) * torch.ones_like(timesteps),
is_prev=True).to(weight_dtype)
pred_x_0 = vae.decode(pred_x_0.to(weight_dtype) / vae.config.scaling_factor, return_dict=False)[0]
pred_x_0 = (pred_x_0 * 0.5 + 0.5).clamp(0, 1)
score = dense_score(pred_x_0, batch['caption_split'], batch['caption_index'], do_ortho=True, accelerator=accelerator)
# score = dense_score_plus(pred_x_0, batch['caption_split'], batch['caption_index'], do_ortho=True, accelerator=accelerator)
# loss = 0.5 - (torch.diagonal(score).mean() - score.mean() * 0.3)
loss = (1 - score.mean()) * 0.5 - 0.09 * ortho_weight
# loss = ((1 - dense_score(pred_x_0, batch['caption_split'], batch['caption_index'])) * 0.5).clamp(0.365).mean()
# # Optimization
# if args.loss_type == "huber":
# args.huber_c = 0.001
# loss = torch.sqrt((noise_pred.float() - noise.float()) ** 2 + args.huber_c ** 2).mean() - args.huber_c
# # note: cannot directly combine with mse_loss of noise
# # loss_1 = F.mse_loss(noise_pred.float(), noise.float(), reduction="mean")
# else:
# loss = F.mse_loss(noise_pred.float(), noise.float(), reduction="mean")
# # if args.pretrained_decoder == "stabilityai/stable-diffusion-2-1":
# # loss = loss * 0.2
accelerator.backward(loss)
if accelerator.sync_gradients:
params_to_clip = (itertools.chain(vis.parameters()))
accelerator.clip_grad_norm_(params_to_clip, 0.3)
optimizer.step()
lr_scheduler.step()
global_step += 1
optimizer.zero_grad()
# Saving
if accelerator.sync_gradients: # global_step % accelerator.gradient_accumulation_steps == 0:
progress_bar.update(1)
global_step_ = global_step // accelerator.gradient_accumulation_steps
# if 10500 <= global_step_ <= 12500 and global_step_ % 500 == 0:
# print(accelerator.sync_gradients)
if accelerator.is_main_process: # accelerator.sync_gradients and
if global_step_ % 5 == 0:
accelerator.log({"train_loss": loss.detach().item(), "lr": lr_scheduler.get_last_lr()[0]},
step=global_step_)
if global_step_ % args.validation_steps == 0:
# adapter.eval()
vis.eval()
sample_images(vae, adapter, caption2embed_simple, vis, args, accelerator, weight_dtype,
global_step_)
# adapter.train()
vis.train()
if global_step_ % 500 == 0:
# training status
accelerator.save_state(f"{args.output_dir}/latest_status")
resume_json['global_step'] = global_step
json.dump(resume_json, open(f"{args.output_dir}/latest_status/resume.json", 'w'))
if global_step_ - last_save >= args.save_steps or global_step_ >= args.max_train_steps:
accepts_keep_fp32_wrapper = "keep_fp32_wrapper" in set(
inspect.signature(accelerator.unwrap_model).parameters.keys())
extra_args = ({"keep_fp32_wrapper": True} if accepts_keep_fp32_wrapper else {})
save_lora_weight(
accelerator.unwrap_model(vis, **extra_args),
f"{args.output_dir}/s{global_step_}_lora_vis.pt",
target_replace_module=VIS_REPLACE_MODULES,
)
# accelerator.unwrap_model(adapter, **extra_args).save_pretrained(
# f"{args.output_dir}/s{global_step_}_adapter")
last_save = global_step_
logs = {"loss": loss.detach().item(), "lr": lr_scheduler.get_last_lr()[0]}
progress_bar.set_postfix(**logs)
# accelerator.log(logs, step=global_step)
if global_step_ >= args.max_train_steps:
accelerator.wait_for_everyone()
accelerator.end_training()
accelerator.wait_for_everyone()
accelerator.end_training()
if __name__ == "__main__":
args = parse_args()
main(args)