-
Notifications
You must be signed in to change notification settings - Fork 4
/
trainer.py
216 lines (183 loc) · 9.27 KB
/
trainer.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
import torch
import torch.nn as nn
from torch.utils.data import DataLoader
from tensorboardX import SummaryWriter
from tqdm import tqdm
import torchvision.utils as vutils
import itertools
from utils.models import get_current_LR
from utils.models import isnan
import numpy as np
class Trainer():
def __init__(self, config, init_iter, netD, netG,
dataset, z_dim, loss, sampler_fn, optimizerD,
optimizerG, logdir, device, netG_avg, netG_params):
self.config = config
self.init_iter = init_iter
self.netD = self.ckpt_D = netD
self.netG = self.ckpt_G = netG
self.z_dim = z_dim
self.netG_avg = netG_avg
self.dataset = dataset
self.loss = loss
self.sampler_fn = sampler_fn
self.optimizerD = optimizerD
self.optimizerG = optimizerG
self.logdir = logdir
self.device = device
self.batch_size = config.batch_size
self.num_workers = config.num_workers
self.ngpu = config.ngpu
self.netG_params = netG_params
self.local_rank = self.config.local_rank
if self.ngpu > 1:
if config.fp16:
from apex.parallel import DistributedDataParallel as DDP
from apex import amp
self.netD = DDP(netD, delay_allreduce=True)
self.netG = DDP(netG, delay_allreduce=True)
else:
self.netD = nn.parallel.DistributedDataParallel(netD,
device_ids=[self.local_rank], output_device=self.local_rank)
self.netG = nn.parallel.DistributedDataParallel(netG,
device_ids=[self.local_rank], output_device=self.local_rank)
if self.local_rank == 0:
self.summary_writer = SummaryWriter(self.logdir)
if self.ngpu > 1:
self.sampler = torch.utils.data.distributed.DistributedSampler(self.dataset,
num_replicas=self.config.ngpu, rank=self.local_rank)
else:
self.sampler = None
self.data_loader = DataLoader(self.dataset, batch_size=self.batch_size,
num_workers=self.num_workers, pin_memory=True,
sampler=self.sampler, drop_last=True,
shuffle=self.sampler is None)
self.log_step = config.log_step
self.n_epochs = config.n_epochs
self.d_steps = config.d_steps
self.save_sample_step = config.save_sample_step
self.save_checkpoint_ep = config.save_checkpoint_ep
self.n_samples = config.n_samples
def tqdm_fn(self, x, **kwargs):
if self.local_rank == 0:
return tqdm(x, **kwargs)
else:
return x
def run(self):
fixed_z = self.sampler_fn(self.n_samples,
self.z_dim).to(self.device)
fixed_y = []
for i in range(self.n_samples):
emb = self.dataset.embeddings[i][0]
fixed_y.append(torch.tensor(emb).unsqueeze(0))
fixed_y = torch.cat(fixed_y, dim=0).to(self.device)
fixed_mask = None
if self.config.fp16:
fixed_z, fixed_y = fixed_z.half(), fixed_y.half()
it_count = 0
for ep in self.tqdm_fn(range(self.init_iter, self.n_epochs),
initial=self.init_iter,
total=self.n_epochs):
if (ep % self.save_checkpoint_ep == 0 or ep == (self.n_epochs-1)) and self.local_rank == 0:
tqdm.write('saving checkpoints...')
self.save_ckpt(ep, self.optimizerG, self.ckpt_G,
self.logdir / 'netG_epoch_{}.pth'.format(ep))
self.save_ckpt(ep, self.optimizerD, self.ckpt_D,
self.logdir / 'netD_epoch_{}.pth'.format(ep))
if self.netG_avg is not None:
self.save_ckpt(ep, None, self.netG_avg,
self.logdir / 'netG_avg_epoch_{}.pth'.format(ep),
netG_params=self.netG_params)
d_steps = 0
for it, (x, sent_emb) in enumerate(self.tqdm_fn(self.data_loader)):
for p in self.netD.parameters():
p.requires_grad = True
x, sent_emb = x.to(self.device), sent_emb.to(self.device)
self.netD.zero_grad()
# ---- update Discriminator
with torch.no_grad():
z = self.sampler_fn(x.size(0), self.z_dim)
fake = self.netG(z, sent_emb)
d_logits_fake = self.netD(fake, sent_emb)
d_loss_fake, retain_graph = self.loss.loss_d_fake(d_logits_fake, None)
d_loss_fake.backward(retain_graph=retain_graph)
d_logits_real = self.netD(x, sent_emb)
d_loss_real, retain_graph = self.loss.loss_d_real(d_logits_real, None)
d_loss_real.backward(retain_graph=False)
d_loss = d_loss_real.item() + d_loss_fake.item()
self.optimizerD.step()
d_steps +=1
if d_steps < self.d_steps:
update_g = False
else:
update_g = True
d_steps = 0
if update_g:
for p in self.netD.parameters():
p.requires_grad = False
# # --- update Generator
self.netG.zero_grad()
z = self.sampler_fn(sent_emb.size(0), self.z_dim)
fake = self.netG(z, sent_emb)
d_logits_fake = self.netD(fake, sent_emb)
g_loss, retain_graph = self.loss.loss_g(d_logits_fake, None)
g_loss.backward(retain_graph=retain_graph)
self.optimizerG.step()
if self.netG_avg:
self.update_EMA(self.netG_avg, self.netG, self.device)
if it_count % self.log_step == 0 and self.local_rank == 0 and it > self.d_steps:
tqdm.write('[%d] Loss_D: %.4f Loss_G: %.4f Loss_D_real: ' \
'%.4f Loss_D_fake %.4f' % (it_count,
d_loss, g_loss.item(), d_loss_real.item(),
d_loss_fake.item()))
self.summary_writer.add_scalar('D/D_LR',
get_current_LR(self.optimizerD),
it_count)
self.summary_writer.add_scalar('D/D_loss',
d_loss, it_count)
self.summary_writer.add_scalar('D/D_loss_fake',
d_loss_fake.item(), it_count)
self.summary_writer.add_scalar('D/D_loss_real',
d_loss_real.item(), it_count)
self.summary_writer.add_scalar('G/G_LR',
get_current_LR(self.optimizerG),
it_count)
self.summary_writer.add_scalar('G/G_loss',
g_loss.item(), it_count)
if it_count % self.save_sample_step == 0:
if self.local_rank == 0:
tqdm.write("Saving samples...")
self.netG.eval()
with torch.no_grad():
fake = self.netG(fixed_z, fixed_y)
fake = fake.mul(0.5).add(0.5).clamp_(0, 1).cpu()
if self.local_rank == 0:
self.summary_writer.add_image('G/Samples',
vutils.make_grid(fake, nrow=4), it_count)
self.netG.train()
if self.netG_avg is not None:
with torch.no_grad():
fake = self.netG_avg(fixed_z, fixed_y)
fake = fake.mul(0.5).add(0.5).clamp_(0, 1).cpu()
if self.local_rank == 0:
self.summary_writer.add_image('G/Avg_Samples',
vutils.make_grid(fake, nrow=4), it_count)
it_count += 1
def save_ckpt(self, it, optim, net, fname, **kwargs):
opt_state_dict = optim.state_dict() if optim is not None else None
ckpt = {
'it': it,
'optimizer': opt_state_dict,
'net': net.state_dict(),
}
ckpt.update(kwargs)
torch.save(ckpt, fname)
def update_EMA(self, avg_net, net, device, mu=0.999):
for (avg_k, avg_v), (k, v) in zip(avg_net.state_dict().items(),
net.state_dict().items()):
if ('weight' in avg_k or 'bias' in avg_k) and \
not ('weight_v' in avg_k or 'weight_u' in avg_k):
avg = (1.0 - mu) * v.to(device).data + mu * avg_v.data
avg_v.data.copy_(avg)
else:
avg_v.data.copy_(v.to(device))