-
Notifications
You must be signed in to change notification settings - Fork 1
/
LatentTrainer.py
239 lines (188 loc) · 9.59 KB
/
LatentTrainer.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
import os
import json
from tensorboardX import SummaryWriter
import torch
import torch.nn as nn
from Models.utils import ShiftDistribution, make_noise, make_interpolation_chart, fig_to_image
from Models.utils import MeanTracker, DeformatorType
class Params(object):
def __init__(self, **kwargs):
self.shift_scale = 6.0
self.min_shift = 0.5
self.shift_distribution = ShiftDistribution.UNIFORM
self.deformator_lr = 0.0001
self.shift_predictor_lr = 0.0001
self.n_steps = int(3e+3)
self.batch_size = 128
self.directions_count = None
self.max_latent_dim = None
self.label_weight = 1.0
self.shift_weight = 0.25
self.steps_per_log = 10
self.steps_per_save = 10000
self.steps_per_img_log = 1000
self.steps_per_backup = 1000
self.truncation = None
for key, val in kwargs.items():
if val is not None:
self.__dict__[key] = val
class Trainer(object):
def __init__(self, params=Params(), out_dir='', verbose=False, device='cpu'):
if verbose:
print('Trainer inited with:\n{}'.format(str(params.__dict__)))
self.p = params
self.log_dir = os.path.join(out_dir, 'logs')
os.makedirs(self.log_dir, exist_ok=True)
self.cross_entropy = nn.CrossEntropyLoss()
tb_dir = os.path.join(out_dir, 'tensorboard')
self.models_dir = os.path.join(out_dir, 'models')
self.images_dir = os.path.join(self.log_dir, 'images')
os.makedirs(tb_dir, exist_ok=True)
os.makedirs(self.models_dir, exist_ok=True)
os.makedirs(self.images_dir, exist_ok=True)
self.checkpoint = os.path.join(out_dir, 'checkpoint.pt')
self.writer = SummaryWriter(tb_dir)
self.out_json = os.path.join(self.log_dir, 'stat.json')
self.fixed_test_noise = None
self.device = device
def make_shifts(self, latent_dim):
target_indices = torch.randint(
0, self.p.directions_count, [self.p.batch_size], device=self.device)
if self.p.shift_distribution == ShiftDistribution.NORMAL:
shifts = torch.randn(target_indices.shape, device=self.device)
elif self.p.shift_distribution == ShiftDistribution.UNIFORM:
shifts = 2.0 * torch.rand(target_indices.shape, device=self.device) - 1.0
shifts = self.p.shift_scale * shifts
shifts[(shifts < self.p.min_shift) & (shifts > 0)] = self.p.min_shift
shifts[(shifts > -self.p.min_shift) & (shifts < 0)] = -self.p.min_shift
try:
latent_dim[0]
latent_dim = list(latent_dim)
except Exception:
latent_dim = [latent_dim]
z_shift = torch.zeros([self.p.batch_size] + latent_dim, device=self.device)
for i, (index, val) in enumerate(zip(target_indices, shifts)):
z_shift[i][index] += val
return target_indices, shifts, z_shift
def log_train(self, step, should_print=True, stats=()):
if should_print:
out_text = '{}% [step {}]'.format(int(100 * step / self.p.n_steps), step)
for named_value in stats:
out_text += (' | {}: {:.2f}'.format(*named_value))
print(out_text)
for named_value in stats:
self.writer.add_scalar(named_value[0], named_value[1], step)
with open(self.out_json, 'w') as out:
stat_dict = {named_value[0]: named_value[1] for named_value in stats}
json.dump(stat_dict, out)
def log_interpolation(self, G, deformator, step):
noise = make_noise(1, G.dim_z, self.p.truncation).to(self.device)
if self.fixed_test_noise is None:
self.fixed_test_noise = noise.clone()
for z, prefix in zip([noise, self.fixed_test_noise], ['rand', 'fixed']):
fig = make_interpolation_chart(
G, deformator, z=z, shifts_r=3 * self.p.shift_scale, shifts_count=3, dims_count=15,
dpi=500, device=self.device)
self.writer.add_figure('{}_deformed_interpolation'.format(prefix), fig, step)
fig_to_image(fig).convert("RGB").save(
os.path.join(self.images_dir, '{}_{}.jpg'.format(prefix, step)))
def start_from_checkpoint(self, deformator, shift_predictor):
step = 0
if os.path.isfile(self.checkpoint):
state_dict = torch.load(self.checkpoint)
step = state_dict['step']
deformator.load_state_dict(state_dict['deformator'])
shift_predictor.load_state_dict(state_dict['shift_predictor'])
print('starting from step {}'.format(step))
return step
def save_checkpoint(self, deformator, shift_predictor, step):
state_dict = {
'step': step,
'deformator': deformator.state_dict(),
'shift_predictor': shift_predictor.state_dict(),
}
torch.save(state_dict, self.checkpoint)
def save_models(self, deformator, shift_predictor, step):
torch.save(deformator.state_dict(),
os.path.join(self.models_dir, 'deformator_{}.pt'.format(step)))
torch.save(shift_predictor.state_dict(),
os.path.join(self.models_dir, 'shift_predictor_{}.pt'.format(step)))
def log_accuracy(self, G, deformator, shift_predictor, step):
deformator.eval()
shift_predictor.eval()
accuracy = validate_classifier(G, deformator, shift_predictor, trainer=self)
self.writer.add_scalar('accuracy', accuracy.item(), step)
deformator.train()
shift_predictor.train()
return accuracy
def log(self, G, deformator, shift_predictor, step, avgs):
if step % self.p.steps_per_log == 0:
self.log_train(step, True, [avg.flush() for avg in avgs])
if step % self.p.steps_per_img_log == 0:
self.log_interpolation(G, deformator, step)
if step % self.p.steps_per_backup == 0 and step > 0:
self.save_checkpoint(deformator, shift_predictor, step)
accuracy = self.log_accuracy(G, deformator, shift_predictor, step)
print('Step {} accuracy: {:.3}'.format(step, accuracy.item()))
if step % self.p.steps_per_save == 0 and step > 0:
self.save_models(deformator, shift_predictor, step)
def log_final(self, G, deformator, shift_predictor, step, avgs):
self.log_train(step, True, [avg.flush() for avg in avgs])
self.log_interpolation(G, deformator, step)
self.save_checkpoint(deformator, shift_predictor, step)
accuracy = self.log_accuracy(G, deformator, shift_predictor, step)
print('Step {} accuracy: {:.3}'.format(step, accuracy.item()))
self.save_models(deformator, shift_predictor, step)
def train(self, G, deformator, shift_predictor, multi_gpu=False):
G.to(self.device).eval()
deformator.to(self.device).train()
shift_predictor.to(self.device).train()
deformator_opt = torch.optim.Adam(deformator.parameters(), lr=self.p.deformator_lr) \
if deformator.type not in [DeformatorType.ID, DeformatorType.RANDOM] else None
shift_predictor_opt = torch.optim.Adam(
shift_predictor.parameters(), lr=self.p.shift_predictor_lr)
avgs = MeanTracker('percent'), MeanTracker('loss'), MeanTracker('direction_loss'),\
MeanTracker('shift_loss')
avg_correct_percent, avg_loss, avg_label_loss, avg_shift_loss = avgs
recovered_step = self.start_from_checkpoint(deformator, shift_predictor)
for step in range(recovered_step, self.p.n_steps, 1):
G.zero_grad()
deformator.zero_grad()
shift_predictor.zero_grad()
z = make_noise(self.p.batch_size, G.dim_z, self.p.truncation).to(self.device)
target_indices, shifts, basis_shift = self.make_shifts(deformator.input_dim)
# Deformation
shift = deformator(basis_shift)
imgs = G(z)
imgs_shifted = G.gen_shifted(z, shift)
logits, shift_prediction = shift_predictor(imgs, imgs_shifted)
logit_loss = self.p.label_weight * self.cross_entropy(logits, target_indices)
shift_loss = self.p.shift_weight * torch.mean(torch.abs(shift_prediction - shifts))
# total loss
loss = logit_loss + shift_loss
loss.backward()
if deformator_opt is not None:
deformator_opt.step()
shift_predictor_opt.step()
# update statistics trackers
avg_correct_percent.add(torch.mean(
(torch.argmax(logits, dim=1) == target_indices).to(torch.float32)).detach())
avg_loss.add(loss.item())
avg_label_loss.add(logit_loss.item())
avg_shift_loss.add(shift_loss)
self.log(G, deformator, shift_predictor, step, avgs)
self.log_final(G, deformator, shift_predictor, step, avgs)
@torch.no_grad()
def validate_classifier(G, deformator, shift_predictor, params_dict=None, trainer=None, device='cpu'):
n_steps = 100
if trainer is None:
trainer = Trainer(params=Params(**params_dict), verbose=False)
percents = torch.empty([n_steps])
for step in range(n_steps):
z = make_noise(trainer.p.batch_size, G.dim_z, trainer.p.truncation).to(trainer.device)
target_indices, shifts, basis_shift = trainer.make_shifts(deformator.input_dim)
imgs = G(z)
imgs_shifted = G.gen_shifted(z, deformator(basis_shift))
logits, _ = shift_predictor(imgs, imgs_shifted)
percents[step] = (torch.argmax(logits, dim=1) == target_indices).to(torch.float32).mean()
return percents.mean()