-
Notifications
You must be signed in to change notification settings - Fork 6
/
Copy pathtrain_vae.py
163 lines (136 loc) · 6.34 KB
/
train_vae.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
import argparse
import numpy as np
import os
import matplotlib.pyplot as plt
import torch
import torchvision.utils as vutils
import torch.optim as optim
from torchvision import transforms
from torch.utils.data import DataLoader
from dataset.Dataset import ImageDataset
from logger import Logger
from vae import VAE
from tqdm import tqdm
from PIL import Image
class LinearSchedule:
def __init__(self, start, end=None, steps=None):
if end is None:
end = start
steps = 1
self.inc = (end - start) / float(steps)
self.current = start
self.end = end
if end > start:
self.bound = min
else:
self.bound = max
def __call__(self, steps=1):
val = self.current
self.current = self.bound(self.current + self.inc * steps, self.end)
return val
def save_plots(data, title, fpath):
plt.figure(figsize=(10,5))
plt.title(title)
plt.plot(data)
plt.xlabel("iterations")
plt.ylabel("Loss")
plt.savefig(fpath)
def parse_arguments():
parser = argparse.ArgumentParser()
parser.add_argument('--dir', type=str, default="dataset/reach_target-vision-v0/wrist_rgb", help='path to image folders')
parser.add_argument('--seed', type=int, default=0, help='seed number for reproducibility')
parser.add_argument('--num_workers', type=int, default=8, help='number of workers for dataloaders')
parser.add_argument('--batch_size', type=int, default=16, help='batch size')
parser.add_argument('--epochs', type=int, default=50, help='Number of epochs')
parser.add_argument('--beta', type=float, default=1, help='Weighing value for KLD in B-VAE')
parser.add_argument('--lr', type=float, default=1e-4, help='Learning Rate')
parser.add_argument('--ngpu', type=int, default=1, help='number of gpus to use')
parser.add_argument('--save_freq', type=int, default=2000, help='save weights every <x> iterations')
parser.add_argument('--log_freq', type=int, default=50, help='log losses every <x> iterations')
parser.add_argument('--save_dir', type=str, default="output", help='path to save weights and logs')
parser.add_argument('--load', type=str, default=None, help='path to load weights')
return parser.parse_args()
def main():
device = "cuda" if torch.cuda.is_available() else "cpu"
args = parse_arguments()
torch.manual_seed(args.seed)
np.random.seed(args.seed)
os.makedirs(os.path.join(args.save_dir, "reconstruction"), exist_ok=True)
os.makedirs(os.path.join(args.save_dir, "weights"), exist_ok=True)
transform=transforms.Compose([
transforms.ColorJitter(brightness=0.2, contrast=0.2, saturation=0.2, hue=0),
transforms.RandomRotation(10),
transforms.RandomHorizontalFlip(p=0.5),
transforms.ToTensor(),
transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))
])
val_transform = transforms.Compose([
transforms.ToTensor(),
transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))
])
dataset = ImageDataset(args.dir, transform=transform)
dataloader = DataLoader(dataset, batch_size=args.batch_size, shuffle=True, num_workers=args.num_workers)
# Get 8 sample images to test the decoder on to visualise training progress
sample_dataset = ImageDataset(args.dir, transform=val_transform)
sample_dl = DataLoader(sample_dataset, batch_size=8, shuffle=True)
vae = VAE(beta=args.beta).to(device)
if args.load is not None:
vae.load_weights(args.load)
if args.ngpu>1:
vae.dataparallel(args.ngpu)
optimizer = optim.Adam(vae.parameters(), lr=args.lr)
logger = Logger(output_dir=args.save_dir)
frames = []
p_schedule = LinearSchedule(start=0.2, end=0.1, steps=args.epochs)
for epoch in range(1, args.epochs+1):
p = p_schedule()
logger.running_log_dict = {
'elbo': 0,
'recon_loss': 0,
'kl': 0
}
for i, image in enumerate(tqdm(dataloader), 1):
vae.train()
image = image.to(device)
optimizer.zero_grad()
elbo_loss, log_dict = vae.get_elbo_loss(image, p)
elbo_loss.backward()
optimizer.step()
for k, v in log_dict.items():
logger.running_log_dict[k] += v
logger.store(**log_dict)
if i%args.log_freq==0:
print()
print("#"* 50)
print(f"p: {p}")
print(f"Epoch: [{epoch}/{args.epochs}]\t Iter:[{i}/{len(dataloader)}]")
for k, v in logger.running_log_dict.items():
print(f"{k}: {v/args.log_freq}")
logger.running_log_dict = {
'elbo': 0,
'recon_loss': 0,
'kl': 0
}
print("#"*50)
# Check how the model is doing by saving decoder's output on original input
if (i % args.save_freq == 0) or ((epoch == args.epochs-1) and (i == len(dataloader)-1)):
vae.eval()
sample_images = next(iter(sample_dl)).to(device)
with torch.no_grad():
im = vae(sample_images).cpu()
im = torch.cat([sample_images.cpu(), im], dim=0)
im = vutils.make_grid(im, padding=2).permute(1, 2, 0).numpy() * 0.5 + 0.5
im = Image.fromarray((im*255).astype(np.uint8))
im.save(os.path.join(args.save_dir, "reconstruction", f"epoch{epoch}_iter{i}.png"))
frames.append(im)
# vae.save_weights(os.path.join(args.save_dir, "weights", f"{epoch}_{i}.pth"))
logger.dump()
latest_weights_fname = '_'.join(args.dir.split('/')[1:]) + '.pth'
vae.save_weights(os.path.join(args.save_dir, latest_weights_fname))
frames[0].save(os.path.join(args.save_dir, 'animation.gif'), format='GIF', append_images=frames[1:], save_all=True, duration=500, loop=0)
elbo, kl, recon_loss = logger.load_results(['elbo', 'kl', 'recon_loss'])
save_plots(elbo, title="ELBO loss curve", fpath=os.path.join(args.save_dir, "elbo_curve.png"))
save_plots(kl, title="KL loss curve", fpath=os.path.join(args.save_dir, "kl_curve.png"))
save_plots(recon_loss, title="reconstruction loss curve", fpath=os.path.join(args.save_dir, "recon_loss_curve.png"))
if __name__ == '__main__':
main()