-
Notifications
You must be signed in to change notification settings - Fork 0
/
train.py
200 lines (184 loc) · 8.83 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
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
from __future__ import division
from __future__ import print_function
import os
import random
import argparse
import numpy as np
import time
import torch
import math
from sklearn.utils import shuffle
from utils import load_data,get_metrics
from models import NodeP
parser = argparse.ArgumentParser()
parser.add_argument('--seed', type=int, default=36, help='Random seed.')
parser.add_argument('--epochs', type=int, default=400, help='Number of epochs to train.')
parser.add_argument('--hidden', type=int, default=150, help='Number of hidden units.')
parser.add_argument('--b_sz', type=int, default=70, help='Number of batch_size.')
parser.add_argument('--class_num', type=int, default=4, help='Number of labels')
args = parser.parse_args()
torch.cuda.set_device(1)
test_single=True
rng = np.random.RandomState(seed=args.seed)
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
features, labels,label_emb,neighbor = load_data()
model = NodeP(batch_size=args.b_sz,
lstm_hid_dim= args.hidden,
n_classes=args.class_num,
label_embed=label_emb,
)
model=model.cuda()
features=features.cuda()
def train(model,idx_train, idx_val,crit,opt,epochs,b_sz,neighbor,features):
# if not os.path.isdir('./sala_log'):
# os.makedirs('./sala_log')
# trace_file = './sala_log/trace.txt'
for epoch in range(epochs):
print("Running EPOCH",epoch+1)
train_loss = []
micro_f1_ep = []
micro_precision_ep = []
micro_recall_ep = []
train_nodes = shuffle(idx_train)
batches = math.floor(len(train_nodes) / b_sz)
for index in range(batches):
opt.zero_grad()
trn_batch = train_nodes[index * b_sz:(index + 1) * b_sz]
labels_batch = torch.FloatTensor(labels[trn_batch]).cuda()
trn_node_emb=features[trn_batch]
trn_batch_neighbor = neighbor[trn_batch].numpy()
#y_pred,weight_all,weight_one= attention_model(x)
trn_neighbor_emb=features[trn_batch_neighbor]
y_pred= model(trn_node_emb,trn_neighbor_emb)
loss = crit(y_pred, labels_batch.float())
loss.backward()
opt.step()
labels_cpu = labels_batch.data.cpu().float()
pred_cpu = y_pred.data.cpu()
pred_2 = np.int64(pred_cpu.numpy() > 0.5)
hamming_loss, micro_f1, micro_precision, micro_recall = get_metrics(labels_cpu.numpy(), pred_2)
train_loss.append(float(loss))
micro_f1_ep.append(micro_f1)
micro_precision_ep.append(micro_precision)
micro_recall_ep.append(micro_recall)
avg_loss = np.mean(train_loss)
avg_micro_f1 = np.mean(micro_f1_ep)
avg_micro_precision = np.mean(micro_precision_ep)
avg_micro_recall = np.mean(micro_recall_ep)
print("epoch %2d train end : avg_loss = %.4f" % (epoch+1, avg_loss))
print(' micro_f1: %.4f | micro_precision: %.4f | micro_recall: %.4f'
% ( avg_micro_f1, avg_micro_precision, avg_micro_recall))
print("start validation!!!!!")
test_loss = []
test_hamming = []
test_micro_prec = []
test_micro_recall = []
test_micro_f1 = []
val_nodes = shuffle(idx_val)
batches = math.floor(len(val_nodes) / b_sz)
for index in range(batches):
t = time.time()
test_batch = val_nodes[index * b_sz:(index + 1) * b_sz]
test_labels_batch = torch.FloatTensor(labels[test_batch]).cuda()
tst_node_embs = features[test_batch]
tst_batch_neighbor = neighbor[test_batch].numpy()
tst_neighbor_emb = features[tst_batch_neighbor]
val_y= model(tst_node_embs, tst_neighbor_emb)
loss = crit(val_y, test_labels_batch.float())/b_sz
labels_cpu = test_labels_batch.data.cpu().float()
pred_cpu = val_y.data.cpu()
pred_2 = np.int64(pred_cpu.numpy() > 0.5)
hamming_loss, micro_f1, micro_precision, micro_recall = get_metrics(labels_cpu.numpy(), pred_2)
test_loss.append(float(loss))
test_hamming.append(hamming_loss)
test_micro_prec.append(micro_precision)
test_micro_recall.append(micro_recall)
test_micro_f1.append(micro_f1)
avg_test_loss = np.mean(test_loss)
avg_test_micro_prec = np.mean(test_micro_prec)
avg_test_micro_recall = np.mean(test_micro_recall)
avg_test_micro_f1 = np.mean(test_micro_f1)
print("epoch %2d test end : avg_loss = %.4f" % (epoch+1, avg_test_loss))
print('micro_f1: %.4f | micro_precision: %.4f | micro_recall: %.4f'
% ( avg_test_micro_f1, avg_test_micro_prec, avg_test_micro_recall))
print('time: {:.4f}s'.format(time.time() - t))
# if epoch % 5 == 0:
# p = './sala_log/best_%d.pth' % epoch
# name = model.save(path=p)
# print("save done", name)
def iterative_sampling(Y, labeled_idx, fold, rng):
ratio_per_fold = 1 / fold
folds = [[] for i in range(fold)]
number_of_examples_per_fold = np.array([(1 / fold) * np.shape(Y[labeled_idx, :])[0] for i in range(fold)])
blacklist_samples = np.array([])
number_of_examples_per_label = np.sum(Y[labeled_idx, :], 0)
blacklist_labels = np.where(number_of_examples_per_label < fold)[0]
print(blacklist_labels)
desired_examples_per_label = number_of_examples_per_label * ratio_per_fold
subset_label_desire = np.array([desired_examples_per_label for i in range(fold)])
total_index = np.sum(labeled_idx)
max_label_occurance = np.max(number_of_examples_per_label) + 1
sel_labels = np.setdiff1d(range(Y.shape[1]), blacklist_labels)
while total_index > 0:
try:
min_label_index = np.where(number_of_examples_per_label == np.min(number_of_examples_per_label))[0]
for index in labeled_idx:
if (Y[index, min_label_index[0]] == 1 and index != -1) and (min_label_index[0] not in blacklist_labels):
m = np.where(
subset_label_desire[:, min_label_index[0]] == subset_label_desire[:, min_label_index[0]].max())[0]
if len(m) == 1:
folds[m[0]].append(index)
subset_label_desire[m[0], Y[index, :].astype(np.bool)] -= 1
labeled_idx[np.where(labeled_idx == index)] = -1
number_of_examples_per_fold[m[0]] -= 1
total_index = total_index - index
else:
m2 = np.where(number_of_examples_per_fold[m] == np.max(number_of_examples_per_fold[m]))[0]
if len(m2) > 1:
m = m[rng.choice(m2, 1)[0]]
folds[m].append(index)
subset_label_desire[m, Y[index, :].astype(np.bool)] -= 1
labeled_idx[np.where(labeled_idx == index)] = -1
number_of_examples_per_fold[m] -= 1
total_index = total_index - index
else:
m = m[m2[0]]
folds[m].append(index)
subset_label_desire[m, Y[index, :].astype(np.bool)] -= 1
labeled_idx[np.where(labeled_idx == index)] = -1
number_of_examples_per_fold[m] -= 1
total_index = total_index - index
elif (Y[index, min_label_index[0]] == 1 and index != -1):
if (min_label_index[0] in blacklist_labels) and np.any(Y[index, sel_labels]) == False:
np.append(blacklist_samples, index)
labeled_idx[np.where(labeled_idx == index)] = -1
total_index = total_index - index
number_of_examples_per_label[min_label_index[0]] = max_label_occurance
except:
traceback.print_exc(file=sys.stdout)
exit()
Y = Y[:, sel_labels]
return folds, Y, blacklist_samples
crit = torch.nn.BCELoss())
opt = torch.optim.Adam(model.parameters(), lr=0.001, betas=(0.9, 0.99))
labeled_idx = np.where(labels.sum(-1) > 0)[0]
num_runs = 5
cv_splits, Y, blacklist_samples = iterative_sampling(labels, labeled_idx, num_runs, rng)
print("Done loading training data..")
for i in range(num_runs):
training_samples = []
testing_samples = []
for j in range(len(cv_splits)):
if test_single:
if j != i:
training_samples += cv_splits[j]
else:
testing_samples = cv_splits[j]
else:
if j == i:
training_samples = cv_splits[j]
else:
testing_samples += cv_splits[j]
train(model,training_samples,testing_samples,crit,opt,args.epochs,args.b_sz,neighbor,features)