-
Notifications
You must be signed in to change notification settings - Fork 1
/
train.py
172 lines (136 loc) · 7.39 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
import sys
from datetime import datetime
import argparse
from pathlib import Path
from tqdm import tqdm
import torch
import torch.nn as nn
import torch.optim as optim
from torch.utils.data import DataLoader
from torch.utils.tensorboard import SummaryWriter
from progan.models.utils.utils import load_progan
from models.frame_seed_generator import FrameSeedGenerator
from models.video_discriminator import VideoDiscriminator
from dataset.real_videos import RealVideos
from progan.visualization.visualizer import saveTensor
def train(fsg, progan, vdis, optimizer_D, optimizer_G, dataloader, start_epoch, epochs, name, log_writer):
n_frames = 8
time = torch.arange(n_frames).unsqueeze(1).cuda()
for epoch in range(start_epoch, epochs):
print(f'epoch: {epoch}')
dis_loss, gen_loss = 0, 0
total = 0
dis_out_fakes, dis_out_reals = 0, 0
acc_fakes, acc_reals = 0, 0
for iter, real_video in tqdm(enumerate(dataloader)):
real_video = real_video.cuda()
# --------------- update discriminator ---------------
optimizer_D.zero_grad()
# ------ real input ------
_, real_latent = progan.netD(real_video, getFeature=True) # (N, 512)
real_latent = real_latent.unsqueeze(0) # (1, N, 512)
dis_real = vdis(real_latent) # (1, 1)
label = torch.full([batch_size], 1., dtype=torch.float).cuda()
errD_real = criterion(dis_real.squeeze(0), label)
errD_real.backward()
D_x = dis_real.mean().item()
# ------ fake input ------
noise = torch.randn([1, 2047]).tile(n_frames, 1).cuda()
input = fsg(noise, time)
fake_video = progan.avgG(input)
_, fake_latent = progan.netD(fake_video.detach(), getFeature=True) # (N, 512)
fake_latent = fake_latent.unsqueeze(0) # (1, N, 512)
dis_fake = vdis(fake_latent) # (1, 1)
label.fill_(0.)
errD_fake = criterion(dis_fake.squeeze(0), label)
errD_fake.backward()
D_G_z1 = dis_fake.mean().item()
errD = errD_real + errD_fake
optimizer_D.step()
dis_out_fakes += D_G_z1
dis_out_reals += D_x
acc_fakes += 1 if round(D_G_z1) == 0 else 0
acc_reals += 1 if round(D_x) == 1 else 0
# --------------- update generator ---------------
optimizer_G.zero_grad()
_, fake_latent = progan.netD(fake_video, getFeature=True) # (N, 512)
dis_fake = vdis(fake_latent.unsqueeze(0)) # (1, 512, N) -> (1, 1)
label.fill_(1.)
errG = criterion(dis_fake.squeeze(0), label)
errG.backward()
D_G_z2 = dis_fake.mean().item()
optimizer_G.step()
dis_loss += errD.item()
gen_loss += errG.item()
total += 1
# --------------- log every 1000 iterations ---------------
if iter % 1000 == 0:
step = len(dataloader)*epoch+iter
log_writer.add_scalar('Discriminator loss', dis_loss/total, step)
log_writer.add_scalar('Generator loss', gen_loss/total, step)
log_writer.add_scalar('Discriminator output/fakes', dis_out_fakes/total, step)
log_writer.add_scalar('Discriminator output/reals', dis_out_reals/total, step)
log_writer.add_scalar('Accuracy/fakes', acc_fakes/total, step)
log_writer.add_scalar('Accuracy/reals', acc_reals/total, step)
dis_loss, gen_loss = 0, 0
total = 0
dis_out_fakes, dis_out_reals = 0, 0
acc_fakes, acc_reals = 0, 0
fake_video = fake_video.detach().cpu()
saveTensor(fake_video, (1024, 1024), f'fakes/{name}/video_{epoch}_{iter}.jpg')
# real_video = real_video.detach().cpu()
# saveTensor(real_video, (1024, 1024), f'reals/{name}/video_{epoch}_{iter}.jpg')
torch.save(fsg.state_dict(), f'checkpoints/{name}/frame_seed_generator_epoch_{epoch}.pt')
torch.save(vdis.state_dict(), f'checkpoints/{name}/video_discriminator_epoch_{epoch}.pt')
torch.save(optimizer_G.state_dict(), f'checkpoints/{name}/optimizer_G_epoch_{epoch}.pt')
torch.save(optimizer_D.state_dict(), f'checkpoints/{name}/optimizer_D_epoch_{epoch}.pt')
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument('-n', '--name', type=str, default='', help='Experiment name suffix')
parser.add_argument('-u', '--unfreeze_pgan_disc', default=False, action='store_true', help='Unfreeze ProGAN Discriminator')
parser.add_argument('-c', '--checkpoint_name', type=str, default=None, help='Checkpoint name to load model from (optional)')
parser.add_argument('-e', '--checkpoint_epoch', type=int, default=None, help='Checkpoint epoch to load model from (optional)')
parser.add_argument('-d', '--dataset_path', type=str, help='Path to dataset')
args = parser.parse_args()
now = datetime.now()
date = now.strftime("%Y.%m.%d_%H.%M.%S")
name = f'{date}_{args.name}'
log_writer = SummaryWriter(f'runs/{name}')
Path(f'fakes/{name}').mkdir(parents=True, exist_ok=True)
# Path(f'reals/{name}').mkdir(parents=True, exist_ok=True)
Path(f'checkpoints/{name}').mkdir(parents=True, exist_ok=True)
# --------------- load all model components ---------------
fsg = FrameSeedGenerator()
vdis = VideoDiscriminator()
if args.checkpoint_name and args.checkpoint_epoch:
fsg.load_state_dict(torch.load(f'checkpoints/{args.checkpoint_name}/frame_seed_generator_epoch_{args.checkpoint_epoch}.pt'))
vdis.load_state_dict(torch.load(f'checkpoints/{args.checkpoint_name}/video_discriminator_epoch_{args.checkpoint_epoch}.pt'))
vdis.cuda()
fsg.cuda()
progan = load_progan('jelito3d_batchsize8', 'output_networks/jelito3d_batchsize8', freeze_pgan_disc=not(args.unfreeze_pgan_disc))
# --------------- load optimizers ---------------
criterion = nn.BCELoss()
lr = 0.0002
beta1 = 0.5
optimizer_G = optim.Adam(fsg.parameters(), lr=lr, betas=(beta1, 0.999))
optimizer_D = optim.Adam(vdis.parameters(), lr=lr, betas=(beta1, 0.999))
if args.checkpoint_name and args.checkpoint_epoch:
optimizer_G.load_state_dict(torch.load(f'checkpoints/{args.checkpoint_name}/optimizer_G_epoch_{args.checkpoint_epoch}.pt'))
optimizer_D.load_state_dict(torch.load(f'checkpoints/{args.checkpoint_name}/optimizer_D_epoch_{args.checkpoint_epoch}.pt'))
if args.unfreeze_pgan_disc:
optimizer_D = optim.Adam(list(vdis.parameters()) + list(progan.netD.parameters()), lr=lr, betas=(beta1, 0.999))
# --------------- load data ---------------
real_videos = RealVideos(args.dataset_path)
dataloader = DataLoader(real_videos, batch_size=None, shuffle=True)
# --------------- train ---------------
start_epoch = args.checkpoint_epoch + 1 if args.checkpoint_epoch else 0
epochs = 100
batch_size = 1
fsg.train()
vdis.train()
if args.unfreeze_pgan_disc:
progan.netD.train()
else:
progan.netD.eval()
progan.avgG.eval()
train(fsg, progan, vdis, optimizer_D, optimizer_G, dataloader, start_epoch, epochs, name, log_writer)