-
Notifications
You must be signed in to change notification settings - Fork 7
/
Copy pathtrainer.py
169 lines (138 loc) · 6.17 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
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
########################################
#### Licensed under the MIT license ####
########################################
import torch
import torch.nn as nn
import torch.optim as optim
import os
from numpy import prod
from datetime import datetime
from model import CapsuleNetwork
from loss import CapsuleLoss
from time import time
from ranger21 import Ranger21
from conflicting_bundles import bundle_entropy
from momentumnet import transform_to_momentumnet
import matplotlib.pyplot as plt
from mem_profile import get_gpu_memory_map
SAVE_MODEL_PATH = 'checkpoints/'
if not os.path.exists(SAVE_MODEL_PATH):
os.mkdir(SAVE_MODEL_PATH)
class CapsNetTrainer:
"""
Wrapper object for handling training and evaluation
"""
def __init__(self, loaders, args, device):
self.device = device
self.multi_gpu = args.multi_gpu
self.batch_size = args.batch_size
self.cb_batch_size = args.cb_batch_size
self.conflicts = args.conflicts
self.modelname = args.modelname
self.loaders = loaders
img_shape = self.loaders['train'].dataset[0][0].numpy().shape
self.net = CapsuleNetwork(args, img_shape=img_shape, channels=256, primary_dim=8, num_classes=args.num_classes,
out_dim=16, device=self.device).to(self.device)
# print(self.net)
if args.momentum:
self.net = transform_to_momentumnet(
self.net,
["blocks." + str(i) +
".functions" for i in range(args.num_res_blocks)],
gamma=args.gamma,
use_backprop=False,
is_residual=True,
keep_first_layer=False,
)
if self.multi_gpu:
self.net = nn.DataParallel(self.net)
self.criterion = CapsuleLoss(loss_lambda=0.5, recon_loss_scale=5e-4)
if args.optimizer == 'ranger21':
print('using ranger21 optimizer...')
self.optimizer = Ranger21(
self.net.parameters(), lr=args.learning_rate, num_epochs=args.epochs, num_batches_per_epoch=args.batch_size)
else:
print('using adam optimizer...')
self.optimizer = optim.Adam(
self.net.parameters(), lr=args.learning_rate)
self.scheduler = optim.lr_scheduler.ExponentialLR(
self.optimizer, gamma=args.lr_decay)
print(8*'#', 'PyTorch Model built'.upper(), 8*'#')
print('Num params:', sum([prod(p.size())
for p in self.net.parameters()]))
def __repr__(self):
return repr(self.net)
def run(self, epochs, classes):
print(8*'#', 'Run started'.upper(), 8*'#')
print('MODE;EPOCH;LOSS:ACCURACY;TIME')
eye = torch.eye(len(classes)).to(self.device)
max_memory_usage = 0
for epoch in range(1, epochs+1):
for phase in ['train', 'test']:
if phase == 'train':
self.net.train()
else:
self.net.eval()
t0 = time()
running_loss = 0.0
correct = 0
total = 0
for i, (images, labels) in enumerate(self.loaders[phase]):
t1 = time()
images, labels = images.to(
self.device), labels.to(self.device)
# One-hot encode labels
labels = eye[labels]
self.optimizer.zero_grad()
outputs, reconstructions, layers = self.net(images)
loss = self.criterion(
outputs, labels, images, reconstructions)
if phase == 'train':
loss.backward()
self.optimizer.step()
running_loss += loss.item()
_, predicted = torch.max(outputs, 1)
_, labels = torch.max(labels, 1)
total += labels.size(0)
correct += (predicted == labels).sum()
accuracy = float(correct) / float(total)
# measure conflicting bundles
if i == 0 and self.conflicts:
cb_batch_size = min(
self.cb_batch_size, self.batch_size)
# For each layer
for i, l in enumerate(layers):
a_batch = l[:cb_batch_size]
nb, be = bundle_entropy(
a_batch, labels, num_classes=len(classes))
print("Layer %d | Bundle entropy %.3f" % (i, be))
print("Layer %d | Number of bundles %d" % (i, nb))
print(
f'{phase.upper()};{epoch};{running_loss/(i+1)};{accuracy};{round(time()-t0, 3)}s')
# check memory usage
current_memory_usage = get_gpu_memory_map()[0]
if current_memory_usage > max_memory_usage:
max_memory_usage = current_memory_usage
if hasattr(self, 'scheduler'):
self.scheduler.step()
now = str(datetime.now()).replace(" ", "-")
error_rate = round((1-accuracy)*100, 2)
torch.save(self.net.state_dict(), os.path.join(
SAVE_MODEL_PATH, f'{error_rate}_{now}.pth.tar'))
class_correct = list(0. for _ in classes)
class_total = list(0. for _ in classes)
for images, labels in self.loaders['test']:
images, labels = images.to(self.device), labels.to(self.device)
outputs, reconstructions, _ = self.net(images)
_, predicted = torch.max(outputs, 1)
c = (predicted == labels).squeeze()
for i in range(labels.size(0)):
label = labels[i]
class_correct[label] += c[i].item()
class_total[label] += 1
print(8*'#', 'Run ended'.upper(), 8*'#')
print('Accuracy of ', self.modelname)
for i in range(len(classes)):
print('Accuracy of %5s : %2d %%' % (
classes[i], 100 * class_correct[i] / class_total[i]))
print('max. memory usage: ', max_memory_usage)