-
Notifications
You must be signed in to change notification settings - Fork 48
/
main.py
375 lines (302 loc) · 18.1 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
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
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
from datetime import datetime
import math
import os
import random
import sys
from time import time
from tqdm import tqdm
import pickle
import numpy as np
import scipy.sparse as sp
from scipy.sparse import csr_matrix
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
import torch.sparse as sparse
from torch import autograd
import random
import copy
os.environ['CUDA_VISIBLE_DEVICES'] = '0'
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
from utility.parser import parse_args
from Models import MM_Model, Decoder
from utility.batch_test import *
from utility.logging import Logger
from utility.norm import build_sim, build_knn_normalized_graph
import setproctitle
args = parse_args()
class Trainer(object):
def __init__(self, data_config):
self.task_name = "%s_%s_%s" % (datetime.now().strftime('%Y-%m-%d %H:%M:%S'), args.dataset, args.cf_model,)
self.logger = Logger(filename=self.task_name, is_debug=args.debug)
self.logger.logging("PID: %d" % os.getpid())
self.logger.logging(str(args))
self.mess_dropout = eval(args.mess_dropout)
self.lr = args.lr
self.emb_dim = args.embed_size
self.batch_size = args.batch_size
self.weight_size = eval(args.weight_size)
self.n_layers = len(self.weight_size)
self.regs = eval(args.regs)
self.decay = self.regs[0]
self.image_feats = np.load(args.data_path + '{}/image_feat.npy'.format(args.dataset))
self.text_feats = np.load(args.data_path + '{}/text_feat.npy'.format(args.dataset))
self.image_feat_dim = self.image_feats.shape[-1]
self.text_feat_dim = self.text_feats.shape[-1]
self.ui_graph = self.ui_graph_raw = pickle.load(open(args.data_path + args.dataset + '/train_mat','rb'))
# get user embedding
augmented_user_init_embedding = pickle.load(open(args.data_path + args.dataset + '/augmented_user_init_embedding','rb'))
augmented_user_init_embedding_list = []
for i in range(len(augmented_user_init_embedding)):
augmented_user_init_embedding_list.append(augmented_user_init_embedding[i])
augmented_user_init_embedding_final = np.array(augmented_user_init_embedding_list)
pickle.dump(augmented_user_init_embedding_final, open(args.data_path + args.dataset + '/augmented_user_init_embedding_final','wb'))
self.user_init_embedding = pickle.load(open(args.data_path + args.dataset + '/augmented_user_init_embedding_final','rb'))
# get separate embedding matrix
if args.dataset=='preprocessed_raw_MovieLens':
augmented_total_embed_dict = {'title':[] , 'genre':[], 'director':[], 'country':[], 'language':[]}
elif args.dataset=='netflix_valid_item':
augmented_total_embed_dict = {'year':[] , 'title':[], 'director':[], 'country':[], 'language':[]}
augmented_atttribute_embedding_dict = pickle.load(open(args.data_path + args.dataset + '/augmented_atttribute_embedding_dict','rb'))
for value in augmented_atttribute_embedding_dict.keys():
for i in range(len(augmented_atttribute_embedding_dict[value])):
augmented_total_embed_dict[value].append(augmented_atttribute_embedding_dict[value][i])
augmented_total_embed_dict[value] = np.array(augmented_total_embed_dict[value])
pickle.dump(augmented_total_embed_dict, open(args.data_path + args.dataset + '/augmented_total_embed_dict','wb'))
self.item_attribute_embedding = pickle.load(open(args.data_path + args.dataset + '/augmented_total_embed_dict','rb'))
self.image_ui_index = {'x':[], 'y':[]}
self.text_ui_index = {'x':[], 'y':[]}
self.n_users = self.ui_graph.shape[0]
self.n_items = self.ui_graph.shape[1]
self.iu_graph = self.ui_graph.T
self.ui_graph = self.csr_norm(self.ui_graph, mean_flag=True)
self.iu_graph = self.csr_norm(self.iu_graph, mean_flag=True)
self.ui_graph = self.matrix_to_tensor(self.ui_graph)
self.iu_graph = self.matrix_to_tensor(self.iu_graph)
self.image_ui_graph = self.text_ui_graph = self.ui_graph
self.image_iu_graph = self.text_iu_graph = self.iu_graph
self.model_mm = MM_Model(self.n_users, self.n_items, self.emb_dim, self.weight_size, self.mess_dropout, self.image_feats, self.text_feats, self.user_init_embedding, self.item_attribute_embedding)
self.model_mm = self.model_mm.cuda()
self.decoder = Decoder(self.user_init_embedding.shape[1]).cuda()
self.optimizer = optim.AdamW(
[
{'params':self.model_mm.parameters()},
]
, lr=self.lr)
self.de_optimizer = optim.AdamW(
[
{'params':self.decoder.parameters()},
]
, lr=args.de_lr)
def csr_norm(self, csr_mat, mean_flag=False):
rowsum = np.array(csr_mat.sum(1))
rowsum = np.power(rowsum+1e-8, -0.5).flatten()
rowsum[np.isinf(rowsum)] = 0.
rowsum_diag = sp.diags(rowsum)
colsum = np.array(csr_mat.sum(0))
colsum = np.power(colsum+1e-8, -0.5).flatten()
colsum[np.isinf(colsum)] = 0.
colsum_diag = sp.diags(colsum)
if mean_flag == False:
return rowsum_diag*csr_mat*colsum_diag
else:
return rowsum_diag*csr_mat
def matrix_to_tensor(self, cur_matrix):
if type(cur_matrix) != sp.coo_matrix:
cur_matrix = cur_matrix.tocoo() #
indices = torch.from_numpy(np.vstack((cur_matrix.row, cur_matrix.col)).astype(np.int64)) #
values = torch.from_numpy(cur_matrix.data) #
shape = torch.Size(cur_matrix.shape)
return torch.sparse.FloatTensor(indices, values, shape).to(torch.float32).cuda() #
def innerProduct(self, u_pos, i_pos, u_neg, j_neg):
pred_i = torch.sum(torch.mul(u_pos,i_pos), dim=-1)
pred_j = torch.sum(torch.mul(u_neg,j_neg), dim=-1)
return pred_i, pred_j
def weights_init(self, m):
if isinstance(m, nn.Linear):
nn.init.kaiming_normal_(m.weight)
m.bias.data.fill_(0)
def sim(self, z1, z2):
z1 = F.normalize(z1)
z2 = F.normalize(z2)
return torch.mm(z1, z2.t())
def feat_reg_loss_calculation(self, g_item_image, g_item_text, g_user_image, g_user_text):
feat_reg = 1./2*(g_item_image**2).sum() + 1./2*(g_item_text**2).sum() \
+ 1./2*(g_user_image**2).sum() + 1./2*(g_user_text**2).sum()
feat_reg = feat_reg / self.n_items
feat_emb_loss = args.feat_reg_decay * feat_reg
return feat_emb_loss
def prune_loss(self, pred, drop_rate):
ind_sorted = np.argsort(pred.cpu().data).cuda()
loss_sorted = pred[ind_sorted]
remember_rate = 1 - drop_rate
num_remember = int(remember_rate * len(loss_sorted))
ind_update = ind_sorted[:num_remember]
loss_update = pred[ind_update]
return loss_update.mean()
def mse_criterion(self, x, y, alpha=3):
x = F.normalize(x, p=2, dim=-1)
y = F.normalize(y, p=2, dim=-1)
tmp_loss = (1 - (x * y).sum(dim=-1)).pow_(alpha)
tmp_loss = tmp_loss.mean()
loss = F.mse_loss(x, y)
return loss
def sce_criterion(self, x, y, alpha=1):
x = F.normalize(x, p=2, dim=-1)
y = F.normalize(y, p=2, dim=-1)
loss = (1-(x*y).sum(dim=-1)).pow_(alpha)
loss = loss.mean()
return loss
def test(self, users_to_test, is_val):
self.model_mm.eval()
with torch.no_grad():
ua_embeddings, ia_embeddings, *rest = self.model_mm(self.ui_graph, self.iu_graph, self.image_ui_graph, self.image_iu_graph, self.text_ui_graph, self.text_iu_graph)
result = test_torch(ua_embeddings, ia_embeddings, users_to_test, is_val)
return result
def train(self):
now_time = datetime.now()
run_time = datetime.strftime(now_time,'%Y_%m_%d__%H_%M_%S')
training_time_list = []
stopping_step = 0
n_batch = data_generator.n_train // args.batch_size + 1
best_recall = 0
for epoch in range(args.epoch):
t1 = time()
loss, mf_loss, emb_loss, reg_loss = 0., 0., 0., 0.
contrastive_loss = 0.
n_batch = data_generator.n_train // args.batch_size + 1
sample_time = 0.
build_item_graph = True
self.gene_u, self.gene_real, self.gene_fake = None, None, {}
self.topk_p_dict, self.topk_id_dict = {}, {}
for idx in tqdm(range(n_batch)):
self.model_mm.train()
sample_t1 = time()
users, pos_items, neg_items = data_generator.sample()
# augment samples
augmented_sample_dict = pickle.load(open(args.data_path + args.dataset + '/augmented_sample_dict','rb'))
users_aug = random.sample(users, int(len(users)*args.aug_sample_rate))
pos_items_aug = [augmented_sample_dict[user][0] for user in users_aug if (augmented_sample_dict[user][0]<self.n_items and augmented_sample_dict[user][1]<self.n_items)]
neg_items_aug = [augmented_sample_dict[user][1] for user in users_aug if (augmented_sample_dict[user][0]<self.n_items and augmented_sample_dict[user][1]<self.n_items)]
users_aug = [user for user in users_aug if (augmented_sample_dict[user][0]<self.n_items and augmented_sample_dict[user][1]<self.n_items)]
self.new_batch_size = len(users_aug)
users += users_aug
pos_items += pos_items_aug
neg_items += neg_items_aug
sample_time += time() - sample_t1
user_presentation_h, item_presentation_h, image_i_feat, text_i_feat, image_u_feat, text_u_feat \
, user_prof_feat_pre, item_prof_feat_pre, user_prof_feat, item_prof_feat, user_att_feats, item_att_feats, i_mask_nodes, u_mask_nodes \
= self.model_mm(self.ui_graph, self.iu_graph, self.image_ui_graph, self.image_iu_graph, self.text_ui_graph, self.text_iu_graph)
u_bpr_emb = user_presentation_h[users]
i_bpr_pos_emb = item_presentation_h[pos_items]
i_bpr_neg_emb = item_presentation_h[neg_items]
batch_mf_loss, batch_emb_loss, batch_reg_loss = self.bpr_loss(u_bpr_emb, i_bpr_pos_emb, i_bpr_neg_emb)
# modal feat
image_u_bpr_emb = image_u_feat[users]
image_i_bpr_pos_emb = image_i_feat[pos_items]
image_i_bpr_neg_emb = image_i_feat[neg_items]
image_batch_mf_loss, image_batch_emb_loss, image_batch_reg_loss = self.bpr_loss(image_u_bpr_emb, image_i_bpr_pos_emb, image_i_bpr_neg_emb)
text_u_bpr_emb = text_u_feat[users]
text_i_bpr_pos_emb = text_i_feat[pos_items]
text_i_bpr_neg_emb = text_i_feat[neg_items]
text_batch_mf_loss, text_batch_emb_loss, text_batch_reg_loss = self.bpr_loss(text_u_bpr_emb, text_i_bpr_pos_emb, text_i_bpr_neg_emb)
mm_mf_loss = image_batch_mf_loss + text_batch_mf_loss
batch_mf_loss_aug = 0
for index,value in enumerate(item_att_feats): #
u_g_embeddings_aug = user_prof_feat[users]
pos_i_g_embeddings_aug = item_att_feats[value][pos_items]
neg_i_g_embeddings_aug = item_att_feats[value][neg_items]
tmp_batch_mf_loss_aug, batch_emb_loss_aug, batch_reg_loss_aug = self.bpr_loss(u_g_embeddings_aug, pos_i_g_embeddings_aug, neg_i_g_embeddings_aug)
batch_mf_loss_aug += tmp_batch_mf_loss_aug
feat_emb_loss = self.feat_reg_loss_calculation(image_i_feat, text_i_feat, image_u_feat, text_u_feat)
att_re_loss = 0
if args.mask:
input_i = {}
for index,value in enumerate(item_att_feats.keys()):
input_i[value] = item_att_feats[value][i_mask_nodes]
decoded_u, decoded_i = self.decoder(torch.tensor(user_prof_feat[u_mask_nodes]), input_i)
if args.feat_loss_type=='mse':
att_re_loss += self.mse_criterion(decoded_u, torch.tensor(self.user_init_embedding[u_mask_nodes]).cuda(), alpha=args.alpha_l)
for index,value in enumerate(item_att_feats.keys()):
att_re_loss += self.mse_criterion(decoded_i[index], torch.tensor(self.item_attribute_embedding[value][i_mask_nodes]).cuda(), alpha=args.alpha_l)
elif args.feat_loss_type=='sce':
att_re_loss += self.sce_criterion(decoded_u, torch.tensor(self.user_init_embedding[u_mask_nodes]).cuda(), alpha=args.alpha_l)
for index,value in enumerate(item_att_feats.keys()):
att_re_loss += self.sce_criterion(decoded_i[index], torch.tensor(self.item_attribute_embedding[value][i_mask_nodes]).cuda(), alpha=args.alpha_l)
batch_loss = batch_mf_loss + batch_emb_loss + batch_reg_loss + feat_emb_loss + args.aug_mf_rate*batch_mf_loss_aug + args.mm_mf_rate*mm_mf_loss + args.att_re_rate*att_re_loss
nn.utils.clip_grad_norm_(self.model_mm.parameters(), max_norm=1.0) #+ ssl_loss2 #+ batch_contrastive_loss
self.optimizer.zero_grad()
batch_loss.backward(retain_graph=False)
self.optimizer.step()
loss += float(batch_loss)
mf_loss += float(batch_mf_loss)
emb_loss += float(batch_emb_loss)
reg_loss += float(batch_reg_loss)
del user_presentation_h, item_presentation_h, u_bpr_emb, i_bpr_neg_emb, i_bpr_pos_emb
if math.isnan(loss) == True:
self.logger.logging('ERROR: loss is nan.')
sys.exit()
if (epoch + 1) % args.verbose != 0:
perf_str = 'Epoch %d [%.1fs]: train==[%.5f=%.5f + %.5f + %.5f + %.5f]' % (
epoch, time() - t1, loss, mf_loss, emb_loss, reg_loss, contrastive_loss)
training_time_list.append(time() - t1)
self.logger.logging(perf_str)
t2 = time()
users_to_test = list(data_generator.test_set.keys())
users_to_val = list(data_generator.val_set.keys())
ret = self.test(users_to_test, is_val=False) #^-^
training_time_list.append(t2 - t1)
t3 = time()
if args.verbose > 0:
perf_str = 'Epoch %d [%.1fs + %.1fs]: train==[%.5f=%.5f + %.5f + %.5f], recall=[%.5f, %.5f, %.5f, %.5f], ' \
'precision=[%.5f, %.5f, %.5f, %.5f], hit=[%.5f, %.5f, %.5f, %.5f], ndcg=[%.5f, %.5f, %.5f, %.5f]' % \
(epoch, t2 - t1, t3 - t2, loss, mf_loss, emb_loss, reg_loss, ret['recall'][0], ret['recall'][1], ret['recall'][2],
ret['recall'][-1],
ret['precision'][0], ret['precision'][1], ret['precision'][2], ret['precision'][-1], ret['hit_ratio'][0], ret['hit_ratio'][1], ret['hit_ratio'][2], ret['hit_ratio'][-1],
ret['ndcg'][0], ret['ndcg'][1], ret['ndcg'][2], ret['ndcg'][-1])
self.logger.logging(perf_str)
if ret['recall'][1] > best_recall:
best_recall = ret['recall'][1]
test_ret = self.test(users_to_test, is_val=False)
self.logger.logging("Test_Recall@%d: %.5f, precision=[%.5f], ndcg=[%.5f]" % (eval(args.Ks)[1], test_ret['recall'][1], test_ret['precision'][1], test_ret['ndcg'][1]))
stopping_step = 0
elif stopping_step < args.early_stopping_patience:
stopping_step += 1
self.logger.logging('#####Early stopping steps: %d #####' % stopping_step)
else:
self.logger.logging('#####Early stop! #####')
break
self.logger.logging(str(test_ret))
return best_recall, run_time
def bpr_loss(self, users, pos_items, neg_items):
pos_scores = torch.sum(torch.mul(users, pos_items), dim=1)
neg_scores = torch.sum(torch.mul(users, neg_items), dim=1)
regularizer = 1./(2*(users**2).sum()+1e-8) + 1./(2*(pos_items**2).sum()+1e-8) + 1./(2*(neg_items**2).sum()+1e-8)
regularizer = regularizer / self.batch_size
maxi = F.logsigmoid(pos_scores - neg_scores+1e-8)
mf_loss = - self.prune_loss(maxi, args.prune_loss_drop_rate)
emb_loss = self.decay * regularizer
reg_loss = 0.0
return mf_loss, emb_loss, reg_loss
def sparse_mx_to_torch_sparse_tensor(self, sparse_mx):
"""Convert a scipy sparse matrix to a torch sparse tensor."""
sparse_mx = sparse_mx.tocoo().astype(np.float32)
indices = torch.from_numpy(
np.vstack((sparse_mx.row, sparse_mx.col)).astype(np.int64))
values = torch.from_numpy(sparse_mx.data)
shape = torch.Size(sparse_mx.shape)
return torch.sparse.FloatTensor(indices, values, shape)
def set_seed(seed):
np.random.seed(seed)
random.seed(seed)
torch.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
if __name__ == '__main__':
os.environ["CUDA_VISIBLE_DEVICES"] = str(args.gpu_id)
set_seed(args.seed)
config = dict()
config['n_users'] = data_generator.n_users
config['n_items'] = data_generator.n_items
trainer = Trainer(data_config=config)
trainer.train()