-
Notifications
You must be signed in to change notification settings - Fork 7
/
Copy pathautoencode.py
137 lines (104 loc) · 4.91 KB
/
autoencode.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
import torch
from torch.utils.data import DataLoader
from torch.utils.tensorboard import SummaryWriter
from tqdm import tqdm
from colorama import Fore, Style
from torch.optim import Adam
from torch.optim.lr_scheduler import ReduceLROnPlateau
import torch.nn as nn
import statistics as stats
from keypoints.models import vgg, knn, autoencoder
from utils import get_lr, UniImageViewer, make_grid
from keypoints.ds import datasets as ds
from apex import amp
from config import config
scale = 4
view_in = UniImageViewer('in', screen_resolution=(128 * 2 * scale, 128 * scale))
view_z = UniImageViewer('z', screen_resolution=(128//2 * 5 * scale, 128//2 * 4 * scale))
def log(phase):
writer.add_scalar(f'{phase}_loss', loss.item(), global_step)
if i % args.display_freq == 0:
recon = torch.cat((x[0], x_[0]), dim=2)
latent = make_grid(z[0].unsqueeze(1), 4, 4)
if args.display:
view_in.render(recon)
view_z.render(latent)
writer.add_image(f'{phase}_recon', recon, global_step)
writer.add_image(f'{phase}_latent', latent.squeeze(0), global_step)
if __name__ == '__main__':
args = config()
torch.cuda.set_device(args.device)
""" variables """
best_loss = 100.0
run_dir = f'data/models/{args.tag}/autoencode/{args.model_type}/run_{args.run_id}'
writer = SummaryWriter(log_dir=run_dir)
global_step = 0
""" data """
datapack = ds.datasets[args.dataset]
train, test = datapack.make(args.dataset_train_len, args.dataset_test_len, data_root=args.data_root)
train_l = DataLoader(train, batch_size=args.batch_size, shuffle=True, drop_last=True, pin_memory=True)
test_l = DataLoader(test, batch_size=args.batch_size, shuffle=True, drop_last=True, pin_memory=True)
def add_co_ords_channels(x):
""" adds 2 channels that carry co-ordinate information """
b, h, w = x.size(0), x.size(2), x.size(3)
hm = torch.linspace(0, 1, h, dtype=x.dtype, device=x.device).reshape(1, 1, h, 1).repeat(b, 1, 1, w)
wm = torch.linspace(0, 1, w, dtype=x.dtype, device=x.device).reshape(1, 1, 1, w).repeat(b, 1, h, 1)
return torch.cat((x, hm, wm), dim=1)
""" model """
nonlinearity, kwargs = nn.LeakyReLU, {"inplace": True}
encoder_core = vgg.make_layers(vgg.vgg_cfg[args.model_type], nonlinearity=nonlinearity, nonlinearity_kwargs=kwargs)
encoder = knn.Unit(args.model_in_channels, args.model_z_channels, encoder_core)
decoder_core = vgg.make_layers(vgg.decoder_cfg[args.model_type], nonlinearity=nonlinearity, nonlinearity_kwargs=kwargs)
decoder = knn.Unit(args.model_z_channels, args.model_in_channels, decoder_core)
auto_encoder = autoencoder.AutoEncoder(encoder, decoder, init_weights=args.load is None).to(args.device)
if args.load is not None:
auto_encoder.load(args.load)
""" optimizer """
optim = Adam(auto_encoder.parameters(), lr=args.lr)
scheduler = ReduceLROnPlateau(optim, mode='min')
""" apex mixed precision """
if args.device != 'cpu':
model, optimizer = amp.initialize(auto_encoder, optim, opt_level=args.opt_level)
""" loss function """
criterion = nn.MSELoss()
for epoch in range(1, args.epochs + 1):
""" training """
batch = tqdm(train_l, total=len(train) // args.batch_size)
for i, (x, _) in enumerate(batch):
x = x.to(args.device)
#x = add_co_ords_channels(x)
optim.zero_grad()
z, x_ = auto_encoder(x)
loss = criterion(x_, x)
if not args.demo:
if args.device != 'cpu':
with amp.scale_loss(loss, optim) as scaled_loss:
scaled_loss.backward()
else:
loss.backward()
optim.step()
batch.set_description(f'Epoch: {epoch} LR: {get_lr(optim)} Train Loss: {loss.item()}')
log('train')
if i % args.checkpoint_freq == 0 and args.demo == 0:
auto_encoder.save(run_dir + '/checkpoint')
global_step += 1
""" test """
with torch.no_grad():
ll = []
batch = tqdm(test_l, total=len(test) // args.batch_size)
for i, (x, _) in enumerate(batch):
x = x.to(args.device)
z, x_ = auto_encoder(x)
loss = criterion(x_, x)
batch.set_description(f'Epoch: {epoch} Test Loss: {loss.item()}')
ll.append(loss.item())
log('test')
global_step += 1
""" check improvement """
ave_loss = stats.mean(ll)
scheduler.step(ave_loss)
best_loss = ave_loss if ave_loss <= best_loss else best_loss
print(f'{Fore.CYAN}ave loss: {ave_loss} {Fore.LIGHTBLUE_EX}best loss: {best_loss} {Style.RESET_ALL}')
""" save if model improved """
if ave_loss <= best_loss and not args.demo:
auto_encoder.save(run_dir + '/best')