-
Notifications
You must be signed in to change notification settings - Fork 6
/
trainer.py
132 lines (100 loc) · 4.19 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
import logging
import torch
import tqdm
import loss
logger = logging.getLogger(__name__)
# ToDo
# Implement and test different learning rates for generator and discriminator and
# do multiple discriminator steps per generator step. See [1] and [2]
#
# [1] Heusel et. al., "GANs Trained by a Two Time-Scale Update Rule Converge to a Local Nash Equilibrium", 2018
# [2] Zhang et. al., "Self-Attention Generative Adversarial Networks", 2019
class Trainer:
def __init__(
self, g_net, d_net, src=["dem", "seg"], dest="rgb", feat_loss=None
):
self.rank = torch.distributed.get_rank()
self.src = src
self.dest = dest
self.d_net = d_net
self.g_net = g_net
# parameters taken from SPADE https://github.com/NVlabs/SPADE/issues/50#issuecomment-494217696
self.g_optim = torch.optim.Adam(
self.g_net.parameters(), lr=0.0001, betas=(0, 0.9)
)
self.d_optim = torch.optim.Adam(
self.d_net.parameters(), lr=0.0004, betas=(0, 0.9)
)
self.g_loss = loss.HingeGenerator()
self.g_feat_lambda = feat_loss
if feat_loss is not None:
self.g_feat_loss = torch.nn.functional.l1_loss
self.d_net.module.return_intermed = True
self.d_loss = loss.HingeDiscriminator()
def sample2gen_input(self, sample):
return {src: sample[src] for src in self.src}
def g_one_step(self, sample):
self.g_optim.zero_grad()
g_input = self.sample2gen_input(sample)
dest_fake = self.g_net(g_input)
d_output_fake = self.d_net(g_input, dest_fake)
loss_val = sum(self.g_loss(o) for o in d_output_fake.final)
if self.g_feat_lambda is not None:
dest_real = sample[self.dest]
d_output_real = self.d_net(g_input, dest_real)
if not d_output_real.features:
logger.error("Trying to compute feature loss on empty list")
raise RuntimeError("Trying to compute feature loss on empty list")
feat_loss = sum(
self.g_feat_loss(fake, real)
for real, fake in zip(
d_output_real.features, d_output_fake.features
)
)
loss_val += self.g_feat_lambda * feat_loss
loss_val.backward()
self.g_optim.step()
return loss_val
def d_one_step(self, sample):
self.d_optim.zero_grad()
g_input = self.sample2gen_input(sample)
# call detach to not compute gradients for generator
dest_fake = self.g_net(g_input).detach()
dest_real = sample[self.dest]
disc_real = self.d_net(g_input, dest_real).final
disc_fake = self.d_net(g_input, dest_fake).final
loss_val = sum(
self.d_loss(*disc_out) for disc_out in zip(disc_real, disc_fake)
)
loss_val.backward()
self.d_optim.step()
return loss_val
def train(self, dataloader, n_epochs):
pbar = tqdm.tqdm(total=n_epochs)
for n_epoch in range(1, n_epochs + 1):
running_g_loss = torch.tensor(0.0, requires_grad=False)
running_d_loss = torch.tensor(0.0, requires_grad=False)
for idx, sample in enumerate(dataloader):
g_loss = self.g_one_step(sample)
torch.distributed.all_reduce(g_loss)
running_g_loss += g_loss.item()
d_loss = self.d_one_step(sample)
torch.distributed.all_reduce(d_loss)
running_d_loss += d_loss.item()
if self.rank == 0:
logger.debug(
"batch idx {:3d}, g_loss:{:7.3f}, d_loss:{:7.3f}".format(
idx, g_loss.item(), d_loss.item()
)
)
running_g_loss /= len(dataloader)
running_d_loss /= len(dataloader)
info_str = "epoch {:3d}, g_loss:{:7.3f}, d_loss:{:7.3f}".format(
n_epoch, running_g_loss, running_d_loss
)
pbar.update(1)
pbar.set_description(info_str)
if self.rank == 0:
pbar.write(info_str)
logger.info(info_str)
return None