-
Notifications
You must be signed in to change notification settings - Fork 1
/
train_sampler.py
304 lines (254 loc) · 10.8 KB
/
train_sampler.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
from numpy.core.fromnumeric import mean
import torch
import numpy as np
import copy
import time
import os
from tqdm import tqdm
from models import VQAutoEncoder, Generator
from hparams import get_sampler_hparams
from utils.data_utils import get_data_loaders, cycle
from utils.sampler_utils import generate_latent_ids, get_latent_loaders, retrieve_autoencoder_components_state_dicts,\
get_samples, get_sampler
from utils.train_utils import EMA, optim_warmup
from utils.log_utils import log, log_stats, set_up_visdom, config_log, start_training_log, \
save_stats, load_stats, save_model, load_model, save_images, \
display_images
# torch.backends.cudnn.benchmark = True
def main(H, vis):
latents_fp_suffix = '_flipped' if H.horizontal_flip else ''
latents_filepath = f'latents/{H.dataset}_{H.latent_shape[-1]}_train_latents{latents_fp_suffix}' # we don't have this one
train_with_validation_dataset = False
if H.steps_per_eval:
train_with_validation_dataset = True
if not os.path.exists(latents_filepath): # True
print("!!!latent no exist!!!!")
ae_state_dict = retrieve_autoencoder_components_state_dicts(
H, ['encoder', 'quantize', 'generator']
) # corresponding learned weights loaded
ae = VQAutoEncoder(H)
ae.load_state_dict(ae_state_dict, strict=False)
# val_loader will be assigned to None if not training with validation dataest
train_loader, val_loader = get_data_loaders(
H.dataset,
H.img_size,
H.batch_size,
drop_last=False,
shuffle=False,
get_flipped=H.horizontal_flip,
get_val_dataloader=train_with_validation_dataset
)
log("Transferring autoencoder to GPU to generate latents...")
ae = ae.cuda() # put ae on GPU for generating
generate_latent_ids(H, ae, train_loader, val_loader) # latent_id saved at ./latents/
log("Deleting autoencoder to conserve GPU memory...")
ae = ae.cpu()
ae = None # release ae
train_latent_loader, val_latent_loader = get_latent_loaders(H, get_validation_loader=train_with_validation_dataset)
quanitzer_and_generator_state_dict = retrieve_autoencoder_components_state_dicts(
H,
['quantize', 'generator'],
remove_component_from_key=True
)
embedding_weight = quanitzer_and_generator_state_dict.pop(
'embedding.weight') # codebook
if H.deepspeed:
embedding_weight = embedding_weight.half()
embedding_weight = embedding_weight.cuda()
generator = Generator(H) # decoder
generator.load_state_dict(quanitzer_and_generator_state_dict, strict=False)
generator = generator.cuda()
### absorbing diffusion sampler ###
sampler = get_sampler(H, embedding_weight).cuda()
optim = torch.optim.Adam(sampler.parameters(), lr=H.lr)
if H.ema:
ema = EMA(H.ema_beta)
ema_sampler = copy.deepcopy(sampler)
# initialise before loading so as not to overwrite loaded stats
losses = np.array([])
val_losses = np.array([])
elbo = np.array([])
val_elbos = np.array([])
mean_losses = np.array([])
start_step = 0
log_start_step = 0
if H.load_step > 0:
start_step = H.load_step + 1
sampler = load_model(sampler, H.sampler, H.load_step, H.load_dir).cuda()
if H.ema:
# if EMA has not been generated previously, recopy newly loaded model
try:
ema_sampler = load_model(
ema_sampler, f'{H.sampler}_ema', H.load_step, H.load_dir)
except Exception:
ema_sampler = copy.deepcopy(sampler)
if H.load_optim:
optim = load_model(
optim, f'{H.sampler}_optim', H.load_step, H.load_dir)
# only used when changing learning rates and reloading from checkpoint
for param_group in optim.param_groups:
param_group['lr'] = H.lr
try:
train_stats = load_stats(H, H.load_step)
except Exception:
train_stats = None
if train_stats is not None:
losses, mean_losses, val_losses, elbo, H.steps_per_log
losses = train_stats["losses"],
mean_losses = train_stats["mean_losses"],
val_losses = train_stats["val_losses"],
val_elbos = train_stats["val_elbos"]
elbo = train_stats["elbo"],
H.steps_per_log = train_stats["steps_per_log"]
log_start_step = 0
losses = losses[0]
mean_losses = mean_losses[0]
val_losses = val_losses[0]
val_elbos = val_elbos[0]
elbo = elbo[0]
# initialise plots
vis.line(
mean_losses,
list(range(log_start_step, start_step, H.steps_per_log)),
win='loss',
opts=dict(title='Loss')
)
vis.line(
elbo,
list(range(log_start_step, start_step, H.steps_per_log)),
win='ELBO',
opts=dict(title='ELBO')
)
vis.line(
val_losses,
list(range(H.steps_per_eval, start_step, H.steps_per_eval)),
win='Val_loss',
opts=dict(title='Validation Loss')
)
else:
log('No stats file found for loaded model, displaying stats from load step only.')
log_start_step = start_step
scaler = torch.cuda.amp.GradScaler()
train_iterator = cycle(train_latent_loader)
# val_iterator = cycle(val_latent_loader)
log(f"Sampler params total: {sum(p.numel() for p in sampler.parameters())}")
for step in range(start_step, H.train_steps):
step_start_time = time.time()
# lr warmup
if H.warmup_iters:
if step <= H.warmup_iters:
optim_warmup(H, step, optim)
x = next(train_iterator) # x : [b, 256] -> 256 is latent size in vector-quantized code space
x = x.cuda()
if H.amp: # True
optim.zero_grad()
with torch.cuda.amp.autocast():
stats = sampler.train_iter(x) # loss calculation
scaler.scale(stats['loss']).backward()
scaler.step(optim)
scaler.update()
else:
stats = sampler.train_iter(x)
if torch.isnan(stats['loss']).any():
log(f'Skipping step {step} with NaN loss')
continue
optim.zero_grad()
stats['loss'].backward()
optim.step()
losses = np.append(losses, stats['loss'].item())
if step % H.steps_per_log == 0:
step_time_taken = time.time() - step_start_time
stats['step_time'] = step_time_taken
mean_loss = np.mean(losses)
stats['mean_loss'] = mean_loss
mean_losses = np.append(mean_losses, mean_loss)
losses = np.array([])
vis.line(
np.array([mean_loss]),
np.array([step]),
win='loss',
update=('append' if step > 0 else 'replace'),
opts=dict(title='Loss')
)
log_stats(step, stats)
if H.sampler == 'absorbing':
elbo = np.append(elbo, stats['vb_loss'].item())
vis.bar(
sampler.loss_history,
list(range(sampler.loss_history.size(0))),
win='loss_bar',
opts=dict(title='loss_bar')
)
vis.line(
np.array([stats['vb_loss'].item()]),
np.array([step]),
win='ELBO',
update=('append' if step > 0 else 'replace'),
opts=dict(title='ELBO')
)
if H.ema and step % H.steps_per_update_ema == 0 and step > 0:
ema.update_model_average(ema_sampler, sampler)
images = None
if step % H.steps_per_display_output == 0 and step > 0:
images = get_samples(H, generator, ema_sampler if H.ema else sampler)
display_images(vis, images, H, win_name=f'{H.sampler}_samples')
if step % H.steps_per_save_output == 0 and step > 0:
if images is None:
images = get_samples(H, generator, ema_sampler if H.ema else sampler)
save_images(images, 'samples', step, H.log_dir, H.save_individually)
if H.steps_per_eval and step % H.steps_per_eval == 0 and step > 0:
# calculate validation loss
valid_loss, valid_elbo, num_samples = 0.0, 0.0, 0
eval_repeats = 5
log("Evaluating")
for _ in tqdm(range(eval_repeats)):
for x in val_latent_loader:
with torch.no_grad():
stats = sampler.train_iter(x.cuda())
valid_loss += stats['loss'].item()
if H.sampler == 'absorbing':
valid_elbo += stats['vb_loss'].item()
num_samples += x.size(0)
valid_loss = valid_loss / num_samples
if H.sampler == 'absorbing':
valid_elbo = valid_elbo / num_samples
val_losses = np.append(val_losses, valid_loss)
val_elbos = np.append(val_elbos, valid_elbo)
vis.line(
np.array([valid_loss]),
np.array([step]),
win='Val_loss',
update=('append' if step > 0 else 'replace'),
opts=dict(title='Validation Loss')
)
if H.sampler == 'absorbing':
vis.line(
np.array([valid_elbo]),
np.array([step]),
win='Val_elbo',
update=('append' if step > 0 else 'replace'),
opts=dict(title='Validation ELBO')
)
if step % H.steps_per_checkpoint == 0 and step > H.load_step:
save_model(sampler, H.sampler, step, H.log_dir)
save_model(optim, f'{H.sampler}_optim', step, H.log_dir)
if H.ema:
save_model(ema_sampler, f'{H.sampler}_ema', step, H.log_dir)
train_stats = {
'losses': losses,
'mean_losses': mean_losses,
'val_losses': val_losses,
'elbo': elbo,
'val_elbos': val_elbos,
'steps_per_log': H.steps_per_log,
'steps_per_eval': H.steps_per_eval,
}
save_stats(H, train_stats, step)
if __name__ == '__main__':
H = get_sampler_hparams()
vis = set_up_visdom(H)
config_log(H.log_dir)
log('---------------------------------')
log(f'Setting up training for {H.sampler}')
start_training_log(H)
main(H, vis)