-
Notifications
You must be signed in to change notification settings - Fork 11
/
train.py
305 lines (258 loc) · 10.9 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
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
import os
import glob
import time
import argparse
import numpy as np
from sklearn import metrics
import torch
import torch.nn.functional as F
import torch.optim as optim
from torch.utils.data import DataLoader
from gcn import BatchGCN
from gat import BatchGAT
from data_loader import ChunkSampler
from data_loader import InteractionDataSet
# Training settings
parser = argparse.ArgumentParser()
parser.add_argument(
'--model', type=str, default='gcn', help="Base model maybe gcn or gat.")
parser.add_argument(
'--task', type=str, default='gender', help="Task maybe gender or age.")
parser.add_argument(
'--no-cuda', action='store_true', default=False, help='Disable CUDA training.')
parser.add_argument(
'--seed', type=int, default=42, help='Random seed.')
parser.add_argument(
'--epochs', type=int, default=1000, help='Number of epochs to train.')
parser.add_argument(
'--lr', type=float, default=1e-3, help='Initial learning rate.')
parser.add_argument(
'--weight-decay', type=float, default=5e-4, help='Weight decay (L2 loss on parameters).')
parser.add_argument(
'--dropout', type=float, default=0.2, help='Dropout rate (1 - keep probability).')
parser.add_argument(
'--hidden-units', type=str, default="16,8", help="Hidden units in each hidden layer, splitted with comma.")
parser.add_argument(
'--feature-size', type=str, default="200,200", help="feature size of item and user, splitted with comma.")
parser.add_argument(
'--heads', type=str, default="1,1,1", help="Heads in each layer, splitted with comma.")
parser.add_argument(
'--batch', type=int, default=64, help="Batch size.")
parser.add_argument(
'--patience', type=int, default=10, help="Patience.")
parser.add_argument(
'--instance-normalization', action='store_true', default=False, help="Enable instance normalization.")
parser.add_argument(
'--shuffle', action='store_true', default=False, help="Shuffle dataset.")
parser.add_argument(
'--data-dir', type=str, default='data', help="Data file directory.")
parser.add_argument(
'--pkl-dir', type=str, default='00', help="Model file directory.")
parser.add_argument(
'--train-ratio', type=float, default=75, help="Training ratio (0, 100).")
parser.add_argument(
'--valid-ratio', type=float, default=12.5, help="Validation ratio (0, 100).")
parser.add_argument(
'--class-weight-balanced', action='store_true', default=False, help="Adjust weights inversely proportional to class frequencies.")
parser.add_argument(
'--use-user-feature', action='store_true', default=False, help="Whether to use users' structural features(include user node).")
parser.add_argument(
'--use-item-feature', action='store_true', default=False, help="Whether to use items' structural features(include user and item nodes).")
parser.add_argument(
'--use-word-feature', action='store_true', default=False, help="Whether to use words' structural features(include user, item and word nodes).")
args = parser.parse_args()
args.cuda = not args.no_cuda and torch.cuda.is_available()
np.random.seed(args.seed)
torch.manual_seed(args.seed)
if args.cuda:
torch.cuda.manual_seed(args.seed)
interaction_dataset = InteractionDataSet(
args.data_dir, args.seed, args.shuffle, args.model, args.task)
N = len(interaction_dataset)
n_classes = interaction_dataset.get_num_class()
class_weight = interaction_dataset.get_class_weight() \
if args.class_weight_balanced else torch.ones(n_classes)
print("class_weight: " + str(list(class_weight.numpy())))
feature_dim = interaction_dataset.get_feature_dimension()
item_dim, user_dim = [int(x) for x in args.feature_size.strip().split(",")]
n_units = [int(x) for x in args.hidden_units.strip().split(",")] + [n_classes]
print("feature dimension={}".format(feature_dim))
print("number of classes={}".format(n_classes))
train_start, valid_start, test_start = 0, \
int(N * args.train_ratio / 100), \
int(N * (args.train_ratio + args.valid_ratio) / 100)
train_loader = DataLoader(
interaction_dataset, batch_size=args.batch,
sampler=ChunkSampler(valid_start - train_start, 0))
valid_loader = DataLoader(
interaction_dataset, batch_size=args.batch,
sampler=ChunkSampler(test_start - valid_start, valid_start))
test_loader = DataLoader(
interaction_dataset, batch_size=args.batch,
sampler=ChunkSampler(N - test_start, test_start))
interaction_item = interaction_dataset.get_interaction_item()
interaction_word = interaction_dataset.get_interaction_word()
if args.cuda:
interaction_item = interaction_item.cuda()
interaction_word = interaction_word.cuda()
# Model and optimizer
if args.model == "gcn":
model = BatchGCN(
word_feature=interaction_dataset.get_word_features(),
interaction_item=interaction_item,
interaction_word=interaction_word,
use_user_feature=args.use_user_feature,
use_item_feature=args.use_item_feature,
use_word_feature=args.use_word_feature,
n_units=n_units,
item_dim=item_dim, user_dim=user_dim,
dropout=args.dropout,
instance_normalization=args.instance_normalization)
elif args.model == "gat":
n_heads = [int(x) for x in args.heads.strip().split(",")]
model = BatchGAT(
word_feature=interaction_dataset.get_word_features(),
interaction_item=interaction_item,
interaction_word=interaction_word,
use_user_feature=args.use_user_feature,
use_item_feature=args.use_item_feature,
use_word_feature=args.use_word_feature,
n_units=n_units, n_heads=n_heads,
item_dim=item_dim, user_dim=user_dim,
dropout=args.dropout,
instance_normalization=args.instance_normalization)
else:
raise NotImplementedError
if args.cuda:
model.cuda()
class_weight = class_weight.cuda()
params = [{'params': filter(lambda p: p.requires_grad, model.parameters())}]
optimizer = optim.Adagrad(params, lr=args.lr, weight_decay=args.weight_decay)
def classification_metrics(y_true, y_pred):
acc = float(metrics.accuracy_score(y_true, y_pred))
# micro_f1 = float(metrics.f1_score(y_true, y_pred, average="micro"))
macro_f1 = float(metrics.f1_score(y_true, y_pred, average="macro"))
# return acc, micro_f1, macro_f1
return acc, macro_f1
def train(epoch, train_loader, valid_loader):
t = time.time()
model.train()
loss_train = 0.
total_train = 0.
y_true_train, y_pred_train, y_score_train = [], [], []
for _, batch in enumerate(train_loader):
graph, labels, vertices = batch
bs = graph.size(0)
if args.cuda:
graph = graph.cuda()
labels = labels.cuda()
vertices = vertices.cuda()
optimizer.zero_grad()
output, _ = model(vertices, graph)
if args.model == "gcn" or args.model == "gat":
output = output[:, -1, :]
loss_train_batch = F.nll_loss(output, labels, class_weight)
loss_train += bs * loss_train_batch.item()
y_true_train += labels.data.tolist()
y_pred_train += output.max(1)[1].data.tolist()
y_score_train += output[:, 1].data.tolist()
total_train += bs
loss_train_batch.backward()
optimizer.step()
model.eval()
loss_val = 0.
total_val = 0.
y_true_val, y_pred_val, y_score_val = [], [], []
for _, batch in enumerate(valid_loader):
graph, labels, vertices = batch
bs = graph.size(0)
if args.cuda:
graph = graph.cuda()
labels = labels.cuda()
vertices = vertices.cuda()
output, _ = model(vertices, graph)
output = output[:, -1, :]
loss_val_batch = F.nll_loss(output, labels, class_weight)
loss_val += bs * loss_val_batch.item()
y_true_val += labels.data.tolist()
y_pred_val += output.max(1)[1].data.tolist()
y_score_val += output[:, 1].data.tolist()
total_val += bs
# acc_train, macro_f1_train = classification_metrics(y_true_train, y_pred_train)
acc_val, macro_f1_val = classification_metrics(y_true_val, y_pred_val)
print('Epoch: {:04d}'.format(epoch),
'loss_train: {:.4f}'.format(loss_train / total_train),
# 'acc_train: {:.4f}'.format(acc_train),
# 'macro_f1_train: {:.4f}'.format(macro_f1_train),
'||',
'loss_val: {:.4f}'.format(loss_val / total_val),
'||',
'acc_val: {:.4f}'.format(acc_val),
'macro_f1_val: {:.4f}'.format(macro_f1_val),
'time: {:.4f}s'.format(time.time() - t))
return loss_val / total_val
def compute_test(test_loader):
model.eval()
loss_test = 0.
total_test = 0.
y_true, y_pred, y_score = [], [], []
emb_test = []
for _, batch in enumerate(test_loader):
graph, labels, vertices = batch
bs = graph.size(0)
if args.cuda:
graph = graph.cuda()
labels = labels.cuda()
vertices = vertices.cuda()
output, test_emb = model(vertices, graph)
emb_test += test_emb.data.tolist()
output = output[:, -1, :]
loss_test_batch = F.nll_loss(output, labels, class_weight)
loss_test += bs * loss_test_batch.item()
y_true += labels.data.tolist()
y_pred += output.max(1)[1].data.tolist()
y_score += output[:, 1].data.tolist()
total_test += bs
acc_test, macro_f1_test = classification_metrics(y_true, y_pred)
# for t-SNE
np.save(os.path.join(args.pkl_dir, 'emb_test.npy'), np.array(emb_test))
np.save(os.path.join(args.pkl_dir, 'label_test.npy'), np.array(y_true))
print("Test set results:",
"loss= {:.4f}".format(loss_test / total_test),
"accuracy= {:.4f}".format(acc_test),
'macro_f1=: {:.4f}'.format(macro_f1_test))
# print("confusion matrix: ", metrics.confusion_matrix(y_true, y_pred))
# Train model
t_total = time.time()
loss_values = []
bad_counter = 0
best = args.epochs + 1
best_epoch = 0
for epoch in range(args.epochs):
loss_values.append(train(epoch, train_loader, valid_loader))
torch.save(model.state_dict(), os.path.join(args.pkl_dir, '{}.pkl'.format(epoch)))
if loss_values[-1] < best:
best = loss_values[-1]
best_epoch = epoch
bad_counter = 0
else:
bad_counter += 1
if bad_counter == args.patience:
break
files = glob.glob(os.path.join(args.pkl_dir, '*.pkl'))
for file in files:
epoch_nb = int(file.split('/')[-1].split('.')[0])
if epoch_nb < best_epoch:
os.remove(file)
files = glob.glob(os.path.join(args.pkl_dir, '*.pkl'))
for file in files:
epoch_nb = int(file.split('/')[-1].split('.')[0])
if epoch_nb > best_epoch:
os.remove(file)
print("Optimization Finished!")
print("Total time elapsed: {:.4f}s".format(time.time() - t_total))
# Restore best model
print('Loading {}th epoch'.format(best_epoch))
model.load_state_dict(torch.load(os.path.join(args.pkl_dir, '{}.pkl'.format(best_epoch))))
# Testing
compute_test(test_loader)