-
Notifications
You must be signed in to change notification settings - Fork 2
/
train.py
277 lines (262 loc) · 12.5 KB
/
train.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
# https://huggingface.co/stabilityai/sd-vae-ft-mse-original/blob/main/README.md
from math import e
import os
from omegaconf import OmegaConf
import numpy as np
from datetime import datetime
from tqdm import tqdm
import torch
import torch.optim as optim
import torch.nn as nn
import torch.nn.functional as F
import random
import pytorch_lightning as pl
from torchvision import transforms
from torch.utils.data import Dataset
from pytorch_lightning import Trainer
from pytorch_lightning.loggers import WandbLogger
import wandb
from PIL import Image
from argparse import ArgumentParser
import lpips
from ldm.util import instantiate_from_config
from ldm.modules.ema import LitEma
from contextlib import contextmanager
torch.cuda.empty_cache()
class FinetuneFaceData(Dataset):
def __init__(self, data_dir:str,
img_list: list,
size:int=384,
):
self.data_dir = data_dir
self.img_list = img_list
self.size = size
self.transform = transforms.Compose(
[transforms.Resize((size, size)),
transforms.ToTensor(),
transforms.Normalize(mean=[0.5, 0.5, 0.5],
std=[0.5, 0.5, 0.5]),
])
def __len__(self):
return len(self.img_list)
def __getitem__(self, idx):
img_name = os.path.join(self.data_dir, self.img_list[idx])
image = Image.open(img_name)
return self.transform(image), self.img_list[idx]
class DataModule(pl.LightningDataModule):
def __init__(self, data_dir,
batch_size=64,
val_size=0.1,
size=384):
super().__init__()
self.data_dir = data_dir
self.batch_size = batch_size
self.val_size = val_size
self.size = size
self.setup('fit')
def setup(self, stage):
all_images = sorted([u for u in os.listdir(self.data_dir) if u.endswith(".png") or u.endswith(".jpg")])
random.shuffle(all_images)
train_size = int((1-self.val_size)*len(all_images))
train_images = all_images[:train_size]
val_images = all_images[train_size:]
self.train_ds = FinetuneFaceData(self.data_dir, train_images, self.size)
self.val_ds = FinetuneFaceData(self.data_dir, val_images, self.size)
print(f"Train size: {len(self.train_ds)}, Val size: {len(self.val_ds)}")
def train_dataloader(self):
return torch.utils.data.DataLoader(self.train_ds, batch_size=self.batch_size, shuffle=True, num_workers=4)
def val_dataloader(self):
return torch.utils.data.DataLoader(self.val_ds, batch_size=self.batch_size, shuffle=False, num_workers=4)
class FinetuneVAE(pl.LightningModule):
def __init__(self,
kl_weight=0.1,
lpips_loss_weight=0.1,
lr=1e-4,
momentum=0.9,
weight_decay=5e-4,
optim='sgd',
vae_config=None,
vae_weights=None,
device=torch.device('cuda'),
ema_decay=0.999,
precision=32,
log_dir=None):
super().__init__()
self.kl_weight = kl_weight
self.lpips_loss_weight = lpips_loss_weight
self.lpips_loss_fn = lpips.LPIPS(net='alex').to(device)
self.lpips_loss_fn.eval()
self.lr = lr
self.momentum = momentum
self.weight_decay = weight_decay
self.optim = optim
self.model = instantiate_from_config(vae_config)
self.model.load_state_dict(vae_weights, strict=True)
self.model.train()
self.precision = precision
self.log_dir = log_dir
self.log_one_batch = False
self.use_ema = ema_decay > 0
if self.use_ema :
self.ema_decay = ema_decay
assert 0. < ema_decay < 1.
self.model_ema = LitEma(self, decay=ema_decay)
print(f"Keeping EMAs of {len(list(self.model_ema.buffers()))}.")
def setup(self, stage=None):
if stage == 'fit' or stage is None:
# Assuming the DataModule is attached to the Trainer and accessible
self.train_ds = self.trainer.datamodule.train_ds
self.val_ds = self.trainer.datamodule.val_ds
print("Warning: The setup method is called")
@contextmanager
def ema_scope(self, context=None):
if self.use_ema:
self.model_ema.store(self.model.parameters())
self.model_ema.copy_to(self.model)
if context is not None:
print(f"{context}: Switched to EMA weights")
try:
yield None
finally:
if self.use_ema:
self.model_ema.restore(self.model.parameters())
if context is not None:
print(f"{context}: Restored training weights")
def forward(self, x):
return self.model(x)
def training_step(self, batch, batch_idx):
target, _ = batch
if self.precision == 16:
target = target.half()
posterior = self.model.encode(target)
z = posterior.sample()
pred = self.model.decode(z)
# kl_loss = posterior.kl()
# kl_loss = kl_loss.mean()
rec_loss = torch.abs(target.contiguous() - pred.contiguous())
if self.current_epoch < self.trainer.max_epochs // 3 * 2:
rec_loss = rec_loss.mean() * rec_loss.size(1)
else:
rec_loss = rec_loss.pow(2).mean() * rec_loss.size(1) #
lpips_loss = self.lpips_loss_fn(pred, target).mean()
loss = rec_loss + self.lpips_loss_weight * lpips_loss # + self.kl_weight * kl_loss
self.log('rec_loss', rec_loss, on_step=True, on_epoch=False, prog_bar=True, logger=False)
self.log('lpips_loss', lpips_loss, on_step=True, on_epoch=False, prog_bar=True, logger=False)
# self.log('kl_loss', kl_loss, on_step=True, on_epoch=False, prog_bar=True, logger=False)
return loss
def configure_optimizers(self):
if self.optim == 'sgd':
optimizer = optim.SGD(self.model.parameters(), lr=self.lr, momentum=self.momentum, weight_decay=self.weight_decay)
else:
raise NotImplementedError
return optimizer
def validation_step(self, batch, batch_idx):
target, name = batch
if self.precision == 16:
target = target.half()
posterior = self.model.encode(target)
z = posterior.mode()
pred = self.model.decode(z)
# kl_loss = posterior.kl()
# kl_loss = kl_loss.mean() # torch.sum(kl_loss) / kl_loss.shape[0]
rec_loss = torch.abs(target.contiguous() - pred.contiguous())
rec_loss = rec_loss.mean() # torch.sum(rec_loss) / (rec_loss.shape[0] * rec_loss.shape[2] * rec_loss.shape[3])
lpips_loss = self.lpips_loss_fn(pred, target).mean()
loss = rec_loss + self.lpips_loss_weight * lpips_loss # + self.kl_weight * kl_loss
# self.log('val_loss', loss, on_step=True, on_epoch=True, prog_bar=True, logger=True)
self.log_images(target, pred, name)
return {'val_loss': loss, "rec_loss": rec_loss, "lpips_loss": lpips_loss}
def log_images(self, input, output, names):
if self.log_one_batch:
return
for img1, img2, name in zip(input, output, names):
img1 = img1.cpu().detach().numpy().transpose(1, 2, 0)
img2 = img2.cpu().detach().numpy().transpose(1, 2, 0)
img1 = (img1 + 1) / 2
img2 = (img2 + 1) / 2
diff = abs(img1 - img2)
img = np.concatenate([img1, img2, diff], axis=1)
img = (img * 255).astype(np.uint8)
img = Image.fromarray(img)
os.makedirs(self.log_dir + "/" + str(self.current_epoch), exist_ok=True)
img.save(os.path.join(self.log_dir, str(self.current_epoch), name))
self.log_one_batch = True
def train_epoch_end(self, outputs):
if self.use_ema:
self.model_ema(self.model)
self.model_ema.copy_to(self.model)
if self.current_epoch == self.trainer.max_epochs // 3 * 2:
self.lpips_loss_weight = self.lpips_loss_weight * 0.1
def validation_epoch_end(self, validation_step_outputs):
self.log_one_batch = False
val_loss = torch.stack([x['val_loss'] for x in validation_step_outputs]).mean()
rec_loss = torch.stack([x['rec_loss'] for x in validation_step_outputs]).mean()
lpips_loss = torch.stack([x['lpips_loss'] for x in validation_step_outputs]).mean()
# kl_loss = torch.stack([x['kl_loss'] for x in validation_step_outputs]).mean()
self.log('val_loss', val_loss, on_step=False, on_epoch=True, prog_bar=False, logger=True)
self.log('val_rec_loss', rec_loss, on_step=False, on_epoch=True, prog_bar=False, logger=True)
self.log('val_lpips_loss', lpips_loss, on_step=False, on_epoch=True, prog_bar=False, logger=True)
# self.log('val_kl_loss', kl_loss, on_step=False, on_epoch=True, prog_bar=False, logger=True)
def get_vae_weights( input_path):
pretrained_weights = torch.load(input_path)
if 'state_dict' in pretrained_weights:
pretrained_weights = pretrained_weights['state_dict']
vae_weight = {}
for k in pretrained_weights.keys():
if "first_stage_model" in k:
vae_weight[k.replace("first_stage_model.", "")] = pretrained_weights[k]
return vae_weight
def argument_inputs():
parser = ArgumentParser()
parser.add_argument('--data_dir', type=str,
default='./dataset/',
help='The directory that contains the images, including original folder and the emotion folders.')
parser.add_argument('--ema_decay', type=float, default=0.99 ,help="Use use_ema")
parser.add_argument('--precision', type=int, default=16, choices=[16, 32])
parser.add_argument('--image_size', type=int, default=384)
parser.add_argument('--batch_size', type=int, default=32)
parser.add_argument('--val_size', type=float, default=0.1)
parser.add_argument('--lr', type=float, default=0.0001)
parser.add_argument('--kl_weight', type=float, default=1.)
parser.add_argument('--lpips_loss_weight', type=float, default=0.1)
parser.add_argument('--num_epochs', type=int, default=10)
parser.add_argument('--output_dir', type=str,
default='./vae_finetune',)
parser.add_argument('--note', type=str,
default='',)
args = parser.parse_args()
args.n_gpus = len(os.environ["CUDA_VISIBLE_DEVICES"].split(","))
args.devices = [i for i in range(args.n_gpus)]
args.strategy = "ddp" #"ddp"
return args
if __name__ == '__main__':
args = argument_inputs()
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
file_names = f"size_({args.image_size})_val({args.val_size})_ema({args.ema_decay})_bs({args.batch_size})_lr({args.lr})_epochs({args.num_epochs})_kl({args.kl_weight})_lpips({args.lpips_loss_weight})_{args.note}"
log_dir = f"{args.output_dir}/{file_names}"
os.makedirs(log_dir, exist_ok=True)
config = OmegaConf.load("./vae_config.yaml")
vae_config = config.model
input_path = "./sd_model/v1-5-pruned.ckpt"
vae_weight = get_vae_weights(input_path)
data_module = DataModule(args.data_dir,
batch_size=args.batch_size,
val_size=0.1,
size=args.image_size)
model = FinetuneVAE(vae_config=vae_config,
vae_weights=vae_weight,
kl_weight=args.kl_weight,
lpips_loss_weight=args.kl_weight,
lr=args.lr,
device=device,
log_dir=log_dir,
ema_decay=args.ema_decay)
trainer = Trainer(min_epochs=1,
max_epochs=args.num_epochs,
precision=args.precision,
strategy=args.strategy,
gpus=args.n_gpus,
num_sanity_val_steps=1 if args.val_size > 0 else 0,
default_root_dir=log_dir,)
trainer.fit(model, datamodule=data_module)
torch.save(model.model.state_dict(), f"{log_dir}/last_model.pth")