-
Notifications
You must be signed in to change notification settings - Fork 17
/
run_training_lstm.py
117 lines (93 loc) · 4.29 KB
/
run_training_lstm.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
import torch
import torchvision
import torch.optim as optim
import torch.nn as nn
import torch.nn.functional as F
import numpy as np
from net.fmri_lstm import fMRI_LSTM
import random
from scipy import stats
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
print(device)
###### **model parameters**
W = 100 # window size
TS = 64 # number of voters per test subject
###### **training parameters**
LR = 0.001 # learning rate
batch_size = 64
###### setup model & data
net = fMRI_LSTM(256, 22, 1, batch_size=batch_size)
net.to(device)
criterion = nn.BCELoss() # CrossEntropyLoss()
optimizer = optim.Adam(net.parameters(), lr=LR, weight_decay=0.001)
# state_dict = torch.load('checkpoint.pth')
# net.load_state_dict(state_dict)
train_data = np.load('data/train_data_1200_1.npy')
train_label = np.load('data/train_label_1200_1.npy')
test_data = np.load('data/test_data_1200_1.npy')
test_label = np.load('data/test_label_1200_1.npy')
print(train_data.shape)
###### start training model
training_loss = 0.0
for epoch in range(100001): # number of mini-batches
# select a random sub-set of subjects
net.hidden = net.init_hidden(batch_size)
net.train()
idx_batch = np.random.permutation(int(train_data.shape[0]))
idx_batch = idx_batch[:int(batch_size)]
# construct a mini-batch by sampling a window W for each subject
train_data_batch = np.zeros((batch_size, 1, W, 22, 1))
train_label_batch = train_label[idx_batch]
for i in range(batch_size):
r1 = random.randint(0, train_data.shape[2] - W)
train_data_batch[i] = train_data[idx_batch[i], :, r1:r1 + W, :, :]
train_data_batch_dev = torch.from_numpy(train_data_batch).float().to(device)
train_label_batch_dev = torch.from_numpy(train_label_batch).float().to(device)
train_data_batch_dev = train_data_batch_dev.squeeze()
# forward + backward + optimize
optimizer.zero_grad()
#net.hidden = net.init_hidden(batch_size)
outputs = net(train_data_batch_dev)
loss = criterion(outputs, train_label_batch_dev)
loss.backward()
optimizer.step()
# print training statistics
training_loss += loss.item()
if epoch % 1000 == 0: # print every T mini-batches
# print(outputs)
outputs = outputs.data.cpu().numpy() > 0.5
train_acc = sum(outputs[:, 0] == train_label_batch) / train_label_batch.shape[0]
print('[%d] training loss: %.3f training batch acc %f' % (epoch + 1, training_loss/1000, train_acc))
training_loss = 0.0
# validate on test subjects by voting
if epoch % 1000 == 0: # print every K mini-batches
net.eval()
idx_batch = np.random.permutation(int(test_data.shape[0]))
idx_batch = idx_batch[:int(batch_size)]
test_label_batch = test_label[idx_batch]
prediction = np.zeros((test_data.shape[0],))
voter = np.zeros((test_data.shape[0],))
for v in range(TS):
idx = np.random.permutation(int(test_data.shape[0]))
# testing also performed batch by batch (otherwise it produces error)
for k in range(int(test_data.shape[0] / batch_size)):
idx_batch = idx[int(batch_size * k):int(batch_size * (k + 1))]
# construct random sub-sequences from a batch of test subjects
test_data_batch = np.zeros((batch_size, 1, W, 22, 1))
for i in range(64):
r1 = random.randint(0, test_data.shape[2] - W)
test_data_batch[i] = test_data[idx_batch[i], :, r1:r1 + W, :, :]
test_data_batch_dev = torch.from_numpy(test_data_batch).float().to(device)
test_data_batch_dev = test_data_batch_dev.squeeze()
#net.eval()
outputs = net(test_data_batch_dev)
outputs = outputs.data.cpu().numpy()
#net.train()
prediction[idx_batch] = prediction[idx_batch] + outputs[:, 0];
voter[idx_batch] = voter[idx_batch] + 1;
# average voting
prediction = prediction / voter;
# print(prediction)
print(sum((prediction > 0.5) == test_label) / test_label.shape[0])
torch.save(net.state_dict(), 'checkpoint.pth')
net.train()