-
Notifications
You must be signed in to change notification settings - Fork 9
/
Copy pathtrain.py
171 lines (140 loc) · 6.47 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
import os.path as path
import time
import torch
import torch.nn as nn
import numpy as np
from torch.autograd import Variable
from constants import *
from data_loaders import *
from stats import *
from tools import *
from model import CapsNet
from options import create_options
from tqdm import tqdm
print(torch.__version__)
def get_alpha(epoch):
# WARNING: Does not support alpha value saving when continuning training from a saved model
if opts.anneal_alpha == "none":
alpha = opts.alpha
if opts.anneal_alpha == "1":
alpha = opts.alpha * float(np.tanh(epoch/DEFAULT_ANNEAL_TEMPERATURE - np.pi) + 1) / 2
if opts.anneal_alpha == "2":
alpha = opts.alpha * float(np.tanh(epoch/(2 * DEFAULT_ANNEAL_TEMPERATURE)))
return alpha
def onehot(tensor, num_classes=10):
return torch.eye(num_classes).cuda().index_select(dim=0, index=tensor) # One-hot encode
def transform_data(data,target,use_gpu, num_classes=10):
data, target = Variable(data), Variable(target)
if use_gpu:
data, target = data.cuda(), target.cuda()
target = onehot(target, num_classes=num_classes)
return data, target
class GPUParallell(nn.DataParallel):
def __init__(self, capsnet, device_ids):
super(Test, self).__init__(capsnet, device_ids=device_ids)
self.capsnet = capsnet
self.num_classes = capsnet.num_classes
def loss(self, images,labels, capsule_output, reconstruction):
return self.capsnet.loss(images, labels, capsule_output, reconstruction)
def forward(self, x, target=None):
return self.capsnet(x, target)
def get_network(opts):
if opts.dataset == "mnist":
capsnet = CapsNet(reconstruction_type=opts.decoder,
routing_iterations = opts.routing_iterations,
batchnorm=opts.batch_norm,
loss=opts.loss_type,
leaky_routing=opts.leaky_routing)
if opts.dataset == "small_norb":
if opts.decoder == "Conv":
opts.decoder = "Conv32"
capsnet = CapsNet(reconstruction_type=opts.decoder,
imsize=32,
num_classes=5,
routing_iterations = opts.routing_iterations,
primary_caps_gridsize=8,
num_primary_capsules=32,
batchnorm=opts.batch_norm,
loss = opts.loss_type,
leaky_routing=opts.leaky_routing)
if opts.dataset == "cifar10":
if opts.decoder == "Conv":
opts.decoder = "Conv32"
capsnet = CapsNet(reconstruction_type=opts.decoder,
imsize=32,
routing_iterations = opts.routing_iterations,
primary_caps_gridsize=8,
img_channels=3,
batchnorm=opts.batch_norm,
num_primary_capsules=32,
loss=opts.loss_type,
leaky_routing=opts.leaky_routing)
if opts.use_gpu:
capsnet.cuda()
if opts.gpu_ids:
capsnet = GPUParallell(capsnet, opts.gpu_ids)
print("Training on GPU IDS:", opts.gpu_ids)
return capsnet
def load_model(opts, capsnet):
model_path = path.join(SAVE_DIR, opts.filepath)
if path.isfile(model_path):
print("Saved model found")
capsnet.load_state_dict(torch.load(model_path))
else:
print("Saved model not found; Model initialized.")
initialize_weights(capsnet)
def get_dataset(opts):
if opts.dataset == 'mnist':
return load_mnist(opts.batch_size)
if opts.dataset == 'small_norb':
return load_small_norb(opts.batch_size)
if opts.dataset == 'cifar10':
return load_cifar10(opts.batch_size)
raise ValueError("Dataset not supported:" + opts.dataset)
def main(opts):
capsnet = get_network(opts)
optimizer = torch.optim.Adam(capsnet.parameters(), lr=opts.learning_rate)
""" Load saved model"""
load_model(opts, capsnet)
train_loader, valid_loader, test_loader = get_dataset(opts)
stats = Statistics(LOG_DIR, opts.model)
for epoch in range(opts.epochs):
capsnet.train()
# Annealing alpha
alpha = get_alpha(epoch)
for batch, (data, target) in tqdm(list(enumerate(train_loader)), ascii=True, desc="Epoch{:3d}".format(epoch)):
optimizer.zero_grad()
data, target = transform_data(data, target, opts.use_gpu, num_classes=capsnet.num_classes)
capsule_output, reconstructions, _ = capsnet(data, target)
predictions = torch.norm(capsule_output.squeeze(), dim=2)
data = denormalize(data)
loss, rec_loss, marg_loss = capsnet.loss(data, target, capsule_output, reconstructions, alpha)
loss.backward()
optimizer.step()
stats.track_train(loss.data.detach().item(), rec_loss.detach().item(), marg_loss.detach().item(), target.detach(), predictions.detach())
"""Evaluate on test set"""
capsnet.eval()
for batch_id, (data, target) in tqdm(list(enumerate(test_loader)), ascii=True, desc="Test {:3d}".format(epoch)):
data, target = transform_data(data, target, opts.use_gpu, num_classes=capsnet.num_classes)
capsule_output, reconstructions, predictions = capsnet(data)
data = denormalize(data)
loss, rec_loss, marg_loss = capsnet.loss(data, target, capsule_output, reconstructions, alpha)
stats.track_test(loss.data.detach().item(),rec_loss.detach().item(), marg_loss.detach().item(), target.detach(), predictions.detach())
stats.save_stats(epoch)
# Save reconstruction image from testing set
if opts.save_images:
data, target = iter(test_loader).next()
data, _ = transform_data(data, target, opts.use_gpu)
_, reconstructions, _ = capsnet(data)
filename = "reconstruction_epoch_{}.png".format(epoch)
if opts.dataset == 'cifar10':
save_images_cifar10(IMAGES_SAVE_DIR, filename, data, reconstructions)
else:
save_images(IMAGES_SAVE_DIR, filename, data, reconstructions, imsize=capsnet.imsize)
# Save model
model_path = get_path(SAVE_DIR, "model{}.pt".format(epoch))
torch.save(capsnet.state_dict(), model_path)
capsnet.train()
if __name__ == '__main__':
opts = create_options()
main(opts)