-
Notifications
You must be signed in to change notification settings - Fork 3
/
Copy pathtrain.py
321 lines (272 loc) · 12 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
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
import numpy as np
from torch.utils.tensorboard import SummaryWriter
import os
from tqdm import tqdm
from matplotlib import pyplot as plt
from utils.file_util import *
from utils.dist_util import *
import argparse
import time
class Trainer:
def __init__(
self,
synthesizer,
device,
optim,
guidance_loss,
regularization_losses,
work_dir,
iterations,
log_image_interval,
save_ckpt_interval,
search_cfg,
max_images,
vis_grad=False,
vis_x0=False,
):
self.device = device
self.work_dir = work_dir
self.make_work_dir()
if get_rank() == 0:
self.writer = SummaryWriter(self.work_dir)
else:
self.writer = None
self.iterations = iterations
self.iter = 0
pbar = range(int(iterations) + 1)
if get_rank() == 0:
self.pbar = tqdm(pbar, initial=0, dynamic_ncols=True, smoothing=0.01)
else:
self.pbar = pbar
self.optim = optim
self.synthesizer = synthesizer
self.guidance_loss = instantiate_from_config(guidance_loss)
if self.guidance_loss.t_policy == 'predetermined':
self.guidance_loss.t_policy = 'predetermined_{}'.format(str(self.search(**search_cfg)))
self.synthesizer.reset_data()
self.regularization_losses = []
for cfg in regularization_losses:
self.regularization_losses.append(instantiate_from_config(cfg))
self.max_images = max_images
self.log_image_interval = log_image_interval
self.save_ckpt_interval = save_ckpt_interval
self.vis_grad = vis_grad
self.vis_x0 = vis_x0
def train(self):
for idx in self.pbar:
self.iter = idx
if self.iter > self.iterations:
print("Done!")
break
batch = self.synthesizer.synthesize_image()
self.optim.zero_grad()
loss_dict = dict()
# guidance loss backward
gl = self.guidance_loss(batch, self.iter, backward=True)
loss_dict['guidance_loss'] = gl
grad = self.synthesizer.get_image_gradient().clone()
# regularization loss backward
for regularization_loss in self.regularization_losses:
if regularization_loss.space == 'w+' or regularization_loss.space =='style':
rl = regularization_loss(self.synthesizer.before_latent, self.synthesizer.current_latent)
elif regularization_loss.space == 'image':
rl = regularization_loss(self.synthesizer.before_image, self.synthesizer.before_image)
rl.backward()
loss_dict[type(regularization_loss).__name__] = rl
self.optim.step()
# after train
self.log_metric(loss_dict)
if (self.iter) % self.log_image_interval == 0:
self.synthesizer.log_images(str(self.iter).zfill(6), self.sample_dir)
if self.vis_grad:
self.synthesizer.log_grads(str(self.iter).zfill(6), self.sample_dir, grad)
if (self.iter) % self.save_ckpt_interval == 0 or self.iter == self.iterations:
if self.synthesizer.saving_input_latents:
self.synthesizer.save_input_latents(self.work_dir)
self.save_ckpt()
def search(self,
metrics_type,
select_by,
min_step,
max_step,
search_step,
vis_step_interval,
num_sample=100):
f, (ax1, ax2) = plt.subplots(1, 2, sharey=True)
all_metrics_dict = dict()
for image_idx in range(num_sample):
batch = self.synthesizer.synthesize_image()
batch = torch.unsqueeze(batch[0], dim=0)
metrics_list = np.zeros(int((max_step-min_step)/search_step+1))
all_t = []
grad_image_seq = None
# compute grad from t = 900 to t = 100
search_dir = os.path.join(self.search_dir, 'img{}'.format(image_idx))
os.makedirs(search_dir, exist_ok=True)
for idx, t in enumerate(range(min_step, max_step+1, search_step)):
self.t = t
all_t.append(t)
self.guidance_loss.t_policy = 'fixed_{}'.format(t)
gl = self.guidance_loss(batch, self.iter, backward=True, return_x0=False)
grad = self.synthesizer.get_image_gradient().clone()
self.synthesizer.current_image.grad.data.zero_()
metrics = self.compute_t_metrics(metrics_type, grad)
metrics_list[idx] += metrics
with open('{}/{}'.format(search_dir, 'metrics.txt'), 'a') as f:
f.write("metrics-{}_t{}: {}\n".format(metrics_type, t, metrics))
print("metrics-{}_t{}: {}".format(metrics_type, t, metrics))
if t % vis_step_interval == 0:
grad_resized = F.interpolate(grad, (256, 256), mode='bilinear', align_corners=False)
if grad_image_seq is None:
grad_image_seq = grad_resized
else:
grad_image_seq = torch.cat([grad_image_seq, grad_resized])
torchvision.utils.save_image(grad, f"{search_dir}/grad_{str(t).zfill(6)}.jpg",
normalize=True,
range=(0, 1))
torchvision.utils.save_image(grad_image_seq, f"{search_dir}/grad_seq.jpg", nrow=grad_image_seq.shape[0])
torchvision.utils.save_image(self.synthesizer.before_image, f"{search_dir}/img_ori.jpg", normalize=True, range=(-1, 1))
all_metrics_dict[image_idx] = metrics_list
# plot
ax1.set_title(metrics_type)
ax1.plot(all_t, metrics_list)
ax2.plot(all_t, metrics_list)
plt.savefig('{}/{}'.format(search_dir, 't_fig.png'))
ax2.clear()
f2, ax = plt.subplots(1, 1)
sum_metrics = [sum([all_metrics_dict[idx][i] for idx in all_metrics_dict])/len(all_metrics_dict) for i in range(len(all_t))]
print(all_t)
ax1.set_title(metrics_type)
ax.plot(all_t, sum_metrics)
plt.savefig('{}/{}'.format(self.search_dir, 'data_avg_metrics.png'))
all_t = np.array(all_t, dtype=int)
sum_metrics = np.array(sum_metrics, dtype=float)
if select_by == 'min':
ret = all_t[sum_metrics.argmin()]
elif select_by == 'max':
ret = all_t[sum_metrics.argmax()]
elif select_by.startswith('below_'):
threshold = float(select_by.split('_')[-1])
ret = all_t[np.where(sum_metrics < threshold)].tolist()
elif select_by.startswith('above_'):
threshold = float(select_by.split('_')[-1])
ret = all_t[np.where(sum_metrics > threshold)].tolist()
else:
raise NotImplementedError
print(ret)
with open('{}/t_result.txt'.format(self.search_dir), 'w') as f:
f.write(str(ret))
return ret
def inference(self, ckpt, data):
self.synthesizer.load(ckpt)
dataset = instantiate_from_config(data)
sampler = data_sampler(
dataset, shuffle=False, distributed=False
)
dataloader = torch.utils.data.dataloader.DataLoader(
dataset,
batch_size=data['bs_per_gpu'],
sampler=sampler,
drop_last=False,
num_workers=data['num_workers'],
)
self.inference_dir = os.path.join(self.work_dir, 'inference/')
os.makedirs(os.path.dirname(os.path.join(self.inference_dir, 'original/')), exist_ok=True)
os.makedirs(os.path.dirname(os.path.join(self.inference_dir, 'edited/')), exist_ok=True)
for idx, batch in enumerate(dataloader):
print('{}/{}'.format(idx, len(dataloader)))
batch = batch.cuda()
t = time.time()
edited_image = self.synthesizer.synthesize_image(batch)
print('cost: {}'.format(time.time()- t))
torchvision.utils.save_image(self.synthesizer.before_image,
'{}/{}/{}.jpg'.format(self.inference_dir, 'original', idx), normalize=True,
range=(-1, 1))
torchvision.utils.save_image(edited_image,
'{}/{}/{}.jpg'.format(self.inference_dir, 'edited', idx), normalize=True,
range=(-1, 1))
self.synthesizer.save_input_latents(self.work_dir)
def compute_t_metrics(self, metrics_type, img_grad):
def entropy(p):
return -torch.where(p > 0, p * p.log(), p.new([0.0])).sum().item()
if metrics_type == 'mean':
return torch.mean(img_grad).item()
if metrics_type == 'entropy':
b, c, h, w = img_grad.shape
img_grad = torch.mean(img_grad, dim=1)
img_grad = img_grad.view(b, 1, -1)
img_grad = torch.nn.functional.softmax(img_grad, dim=-1)
img_grad = img_grad.view(b, 1, h, w)
return entropy(img_grad)
raise NotImplementedError
def log_metric(self, dict):
if get_rank() == 0:
# self.pbar.set_description(
# (
# ' '.join([f"{k}: {v.mean().item():.4f}" for k, v in dict.items()])
# )
# )
for k, v in dict.items():
if isinstance(v, float):
self.writer.add_scalar(f'train/{k}', v, self.iter)
else:
self.writer.add_scalar(f'train/{k}', (v).mean(), self.iter)
def save_ckpt(self):
if get_rank() == 0:
saved_state_dict, name = self.synthesizer.get_saved_state_dict()
torch.save(
{
name: saved_state_dict,
"optimizer": self.optim.state_dict(),
},
f"{self.checkpoint_dir}{name}_{str(self.iter).zfill(6)}.pt"
)
def make_work_dir(self):
self.sample_dir = os.path.join(self.work_dir, 'sample/')
self.search_dir = os.path.join(self.work_dir, 'search/')
self.checkpoint_dir = os.path.join(self.work_dir, 'checkpoint/')
if get_rank() == 0:
os.makedirs(os.path.dirname(self.sample_dir), exist_ok=True)
os.makedirs(os.path.dirname(self.search_dir), exist_ok=True)
os.makedirs(os.path.dirname(self.checkpoint_dir), exist_ok=True)
def main():
device = "cuda"
# parse necessary information
parser = argparse.ArgumentParser()
parser.add_argument("--config", type=str, default='')
parser.add_argument("--local_rank", type=int, default=0)
parser.add_argument("--work_dir", type=str, default='')
args = parser.parse_args()
# read config
f = open(args.config, 'r', encoding='utf-8')
d = yaml.safe_load(f)
# dump config
os.makedirs(os.path.dirname(args.work_dir), exist_ok=True)
config_path = os.path.join(args.work_dir, 'config_dump.yml')
save_dict_to_yaml(d, config_path)
# set seed
if 'seed' in d:
torch.manual_seed(d['seed'])
else:
torch.manual_seed(1010)
# prepare synthesizer
synthesizer = instantiate_from_config(d['synthesizer'])
optimized_target, _ = synthesizer.get_optimized_target()
optimizer = optim.AdamW(
optimized_target, lr=d['optimizer']['params']['lr'],
weight_decay=d['optimizer']['params']['weight_decay']
)
# start training
trainer = Trainer(
synthesizer=synthesizer,
optim=optimizer,
device=device,
work_dir=args.work_dir,
guidance_loss=d['guidance_loss'],
regularization_losses=d['regularization_losses'],
**d['train'],
search_cfg=d['search']
)
trainer.train()
if __name__ == "__main__":
main()