forked from phanideepgampa/stacked-capsule-networks
-
Notifications
You must be signed in to change notification settings - Fork 4
/
Copy pathmain.py
169 lines (142 loc) · 6.03 KB
/
main.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
import os
import argparse
import logging
import random
import pickle
from collections import namedtuple
import time
import traceback
import torch
import torch.nn as nn
import torch.optim as optim
import torch.nn.functional as F
from torch.autograd import Variable
import torchvision.datasets
import torchvision.transforms
import numpy as np
from tqdm import tqdm
import model
np.set_printoptions(precision=4, suppress=True)
def seed_everything(seed=1234):
random.seed(seed)
os.environ['PYTHONHASHSEED'] = str(seed)
np.random.seed(seed)
torch.manual_seed(seed)
torch.cuda.manual_seed(seed)
torch.backends.cudnn.deterministic = True
def evaluate(model,train,test,K,device):
model.eval()
with torch.no_grad():
prev_max = -1e+6*torch.ones(K).to(device)
prev_labels = -1*torch.LongTensor(K).fill_(1).to(device)
for batch_idx, (x, target) in tqdm(enumerate(train),desc="train"):
x = Variable(x).to(device)
target = Variable(target).to(device)
out = model(x,device,mode='test')
a_k = out[2].squeeze(-1) #(B,K)
max_act,max_ex = a_k.max(0).values.view(-1),a_k.max(0).indices.view(-1) #(K)
if (max_act>prev_max).sum()!=0:
prev_labels[max_act>prev_max]= target[max_ex]
prev_max[max_act>prev_max]= max_act[max_act>prev_max]
count = 0
total_count = 0
with torch.no_grad():
for batch_idx, (x, target) in tqdm(enumerate(test),desc='test'):
x = Variable(x).to(device)
target = Variable(target).to(device)
out = model(x,device,mode='test')
a_k = out[2].squeeze(-1) #(B,K)
max_act,max_ex = a_k.max(-1).values.view(-1),a_k.max(-1).indices.view(-1)
pred = prev_labels[max_ex]
count+=(pred == target.data).sum()
total_count += x.data.size()[0]
accuracy = count/total_count
return accuracy
def train(args,train,test,device):
scae = model.SCAE()
if torch.cuda.device_count() > 1:
scae = nn.DataParallel(scae)
scae.to(device)
total_loss = model.SCAE_LOSS()
model_name = "model/scae.model"
log_name = "log/SCAE"
prev_best_accuracy = 0.
K = args.K
C = args.C
B = args.batch_size
if args.load_ext:
model_name = args.model_file
print("loading existing model:->%s" % model_name)
scae = torch.load(model_name, map_location=lambda storage, loc: storage)
scae.to(device)
log_name = 'log/'+model_name.split('/')[-1]
prev_best_accuracy = evaluate(scae,train,test,K=K,device=device)
logging.basicConfig(filename='%s.log' % log_name,
level=logging.DEBUG, format='%(asctime)s %(levelname)-10s %(message)s')
optimizer = optim.RMSprop(scae.parameters(), lr=args.lr, momentum=0.9,eps=(1/(10*args.batch_size)**2) )
k_c = torch.tensor(float(K/C)).to(device)
b_c = torch.tensor(float(B/C)).to(device)
for epoch in range(args.epochs):
ave_loss = 0
for batch_idx, (x, target) in tqdm(enumerate(train)):
optimizer.zero_grad()
scae.train()
x = Variable(x).to(device)
out = scae(x,device,mode='train')
loss = total_loss(out,b_c=b_c,k_c=k_c)
ave_loss = ave_loss * 0.9 + loss.mean().data * 0.1
loss.mean().backward()
#torch.nn.utils.clip_grad_norm_(scae.parameters(),5)
#print(loss)
optimizer.step()
if (batch_idx+1) % 50 == 0 or (batch_idx+1) == len(train):
logging.info('==>>> epoch: {}, batch index: {}, train loss: {:.6f}'.format(
epoch, batch_idx+1, ave_loss))
if (epoch+1)%50 == 0:
scae.eval()
accuracy = evaluate(scae,train=train,test=test,K=K,device=device)
if accuracy>prev_best_accuracy:
prev_best_accuracy = accuracy
torch.save(scae, model_name)
logging.debug("saving model"+str(model_name)+" "+"with test_accuracy:"+ str(accuracy))
logging.debug('epoch ' + str(epoch) + 'test-accuracy: '
+ str(accuracy))
return
def main():
seed_everything()
parser = argparse.ArgumentParser()
parser.add_argument('--mode', type=str, default='train')
parser.add_argument('--model_file', type=str, default='model/scae.model')
parser.add_argument('--batch_size', type=int, default=128)
parser.add_argument('--epochs', type=int, default=600)
parser.add_argument('--lr', type=float, default=1e-5)
parser.add_argument('--K', type=int, default=24)
parser.add_argument('--C', type=int, default=10)
parser.add_argument('--load_ext', action = "store_true")
args = parser.parse_args()
root = './data'
if not os.path.exists(root):
os.mkdir(root)
trans = torchvision.transforms.Compose([torchvision.transforms.ToTensor(), torchvision.transforms.Normalize((0.1307,), (0.3081,))])
train_set = torchvision.datasets.MNIST(root=root, train=True, transform=trans, download=True)
test_set = torchvision.datasets.MNIST(root=root, train=False, transform=trans, download=True)
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
train_loader = torch.utils.data.DataLoader(
dataset=train_set,
batch_size=args.batch_size,
shuffle=True)
test_loader = torch.utils.data.DataLoader(
dataset=test_set,
batch_size=args.batch_size,
shuffle=False)
if args.mode == 'train':
train(args,train=train_loader,test=test_loader,device=device)
else:
model_name = args.model_file
print("loading existing model:->%s" % model_name)
scae = torch.load(model_name, map_location=lambda storage, loc: storage)
scae.to(device)
accuracy = evaluate(scae,train=train_loader,test=test_loader,K=args.K,device=device)
print("accuracy: %0.4f"%accuracy)
if __name__ == "__main__":
main()