-
Notifications
You must be signed in to change notification settings - Fork 0
/
irgan.py
274 lines (242 loc) · 14 KB
/
irgan.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
import pandas as pd
import os
import datetime
import numpy as np
import pickle
from tools import log_time_delta
import time
from multiprocessing import Pool
from multiprocessing import cpu_count
from scipy.sparse import csr_matrix,csr_matrix
import math
from config import Singleton
import sklearn
import tensorflow as tf
from Discrimiator import Dis
from Generator import Gen
from dataHelper import FLAGS,helper
#from oldMF import DIS
import time
import random
os.environ['CUDA_VISIBLE_DEVICES'] = '0'
g1 = tf.Graph()
g2 = tf.Graph()
sess1 = tf.InteractiveSession(graph=g1)
sess2 = tf.InteractiveSession(graph=g2)
paras=None
with g1.as_default():
gen = Gen(itm_cnt = helper.i_cnt,
usr_cnt = helper.u_cnt,
dim_hidden = FLAGS.rnn_embedding_dim,
n_time_step = FLAGS.item_windows_size,
learning_rate = FLAGS.learning_rate,
grad_clip = 0.2,
emb_dim = FLAGS.mf_embedding_dim,
lamda = FLAGS.lamda,
initdelta = 0.05,
MF_paras=paras,
model_type=FLAGS.model_type,
update_rule = 'sgd',
use_sparse_tensor=FLAGS.sparse_tensor
)
gen.build_pretrain()
init1=tf.global_variables_initializer()
saver1 = tf.train.Saver(max_to_keep=50)
sess1.run(init1)
# checkpoint_filepath= "model/joint-25-0.25933.ckpt"
# saver1.restore(sess1,checkpoint_filepath)
with g2.as_default():
dis = Dis(itm_cnt = helper.i_cnt,
usr_cnt = helper.u_cnt,
dim_hidden = FLAGS.rnn_embedding_dim,
n_time_step = FLAGS.item_windows_size,
learning_rate = FLAGS.learning_rate,
grad_clip = 0.2,
emb_dim = FLAGS.mf_embedding_dim,
lamda = FLAGS.lamda,
initdelta = 0.05,
MF_paras=paras,
model_type=FLAGS.model_type,
update_rule = 'sgd',
use_sparse_tensor=FLAGS.sparse_tensor
)
dis.build_pretrain()
init2=tf.global_variables_initializer()
saver2 = tf.train.Saver(max_to_keep=50)
sess2.run(init2)
checkpoint_filepath= "model/Dis/joint-25-0.26067-0.28800.ckpt"
saver2.restore(sess2,checkpoint_filepath)
print(helper.evaluateMultiProcess(sess2, dis))
print(helper.evaluateMultiProcess(sess1, gen))
setting =False
df = helper.test
def sigmoid(x):
return 1 / (1 + math.exp(-x))
def main(checkpoint_dir="model/"):
for e in range(20):
for g_epoch in range(100):
rewardes,pg_losses=[],[]
for user in df["uid"].unique():
# generate pesudo labels for the given user
all_rating = gen.predictionItems(sess1,user) # todo delete the pos ones
exp_rating = np.exp(np.array(all_rating) * 50)
prob = exp_rating / np.sum(exp_rating)
# sorted(prob,reverse=True)[:10]
negative_items_sampled = np.random.choice(np.arange(helper.i_cnt), size=32, p=prob)
# negative_items_sampled = np.argsort(prob)[::-1][:32]
negative_samples = []
unlabeled_rewards=[]
u_seq_neg=[]
i_seq_neg = []
u_neg=[]
i_neg=[]
if setting==True:
for item in negative_items_sampled:
u_seqs,i_seqs = helper.getSeqInTime(user,item,0)
negative_samples.append((u_seqs,i_seqs,user,item ))
u_seq_neg,i_seq_neg = [[ s[j].toarray() for s in negative_samples ] for j in range(2)]
u_neg,i_neg = [[ s[j] for s in negative_samples ] for j in range(2,4)]
unlabeled_rewards = np.array(dis.prediction(sess2,u_seq_neg,i_seq_neg,u_neg,i_neg,sparse=False))
unlabeled_rewards = list((unlabeled_rewards-np.mean(unlabeled_rewards))/np.std(unlabeled_rewards))
#unlabeled_rewards = [2* (sigmoid(v)-0.5) for v in unlabeled_rewards]
positive_samples=[]
labeled_rewards=[]
u_seq_pos=[]
i_seq_pos = []
u_pos=[]
i_pos=[]
if setting==False:
pos_items_time_dict = helper.user_item_pos_rating_time_dict.get(user,{})
if len(pos_items_time_dict)==0:
continue
for ind in np.random.randint(len(pos_items_time_dict), size=32):
positive_item,t = list(pos_items_time_dict.items())[ind]
u_seqs,i_seqs = helper.getSeqInTime(user,positive_item,t)
positive_samples.append((u_seqs,i_seqs,user,positive_item))
u_seq_pos,i_seq_pos = [[ s[j].toarray() for s in positive_samples ] for j in range(2)]
u_pos,i_pos = [[ s[j] for s in positive_samples ] for j in range(2,4)]
labeled_rewards = dis.prediction(sess2,u_seq_pos,i_seq_pos,u_pos,i_pos,sparse=False)
# labeled_rewards = [2* (sigmoid(v)-0.5) + 0.1 for v in labeled_rewards]
labeled_rewards=list((labeled_rewards-np.mean(labeled_rewards))/np.std(labeled_rewards))
# pg_loss = gen.unsupervised_train_step(sess, u_seq_neg,i_seq_neg,u_neg,i_neg, unlabeled_rewards)
pg_loss = gen.unsupervised_train_step(sess1, u_seq_neg + u_seq_pos,
i_seq_neg + i_seq_pos,
u_neg + u_pos,i_neg + i_pos, unlabeled_rewards + labeled_rewards)
pg_losses.append(pg_loss)
rewardes.extend(unlabeled_rewards)
# with open("test_lr.txt", "a") as myfile:
# myfile.write("pg loss : %.5f reward : %.5f "%(np.mean(np.array(pg_losses)),np.sum(np.array(rewardes)))+"\n")
if g_epoch % 3 == 0:
print("pg loss : %.5f reward : %.5f "%(np.mean(np.array(pg_losses)),np.sum(np.array(rewardes))))
g = helper.evaluateMultiProcess(sess1, gen)
print (g)
# for d_epoch in range(2):
# rnn_losses,mf_losses,joint_losses=[],[],[]
# positive_samples = []
# negative_samples = []
# for user in df["uid"].unique():
# pos_items_time_dict = helper.user_item_pos_rating_time_dict.get(user,{}) # null
# if len(pos_items_time_dict)==0:
# continue
#
# # generate pesudo labels for the given user
# all_rating = gen.predictionItems(sess,user)
# exp_rating = np.exp(np.array(all_rating) * helper.conf.temperature)
# prob = exp_rating / np.sum(exp_rating)
#
## negative_items_argmax = np.argsort(prob)[::-1][:16]
# negative_items_sampled = np.random.choice(np.arange(helper.i_cnt), size=32, p=prob)
# for item in negative_items_sampled:
# # the pesudo labels are regarded as high-quality negative labels but at the beginning, the pesudo labels are very low-quality ones.
# u_seqs,i_seqs = helper.getSeqInTime(user,item,0)
# negative_samples.append((u_seqs,i_seqs,user,item,0 ))
#
# # sample positive examples in a random manner
# positive_item,t = list(pos_items_time_dict.items())[np.random.randint(len(pos_items_time_dict), size=1)[0]]
# u_seqs,i_seqs = helper.getSeqInTime(user,positive_item,t)
# positive_samples.append((u_seqs,i_seqs,user,positive_item,1))
# samples = negative_samples + positive_samples
#
# random.shuffle(samples)
#
# u_seq,i_seq = [[ s[j].toarray() for s in samples ] for j in range(2)]
# u,i = [[ s[j] for s in samples ] for j in range(2,4)]
# ratings = [ s[4] for s in samples ]
# _,loss_mf,loss_rnn,joint_loss,rnn,mf = dis.pretrain_step(sess,ratings, u, i,u_seq,i_seq)
# rnn_losses.append(loss_rnn)
# mf_losses.append(loss_mf)
# joint_losses.append(joint_loss)
# with open("test1.txt", "a") as myfile:
# myfile.write("rnn loss : %.5f mf loss : %.5f : joint loss %.5f"%(np.mean(np.array(loss_rnn)),np.mean(np.array(loss_mf)),np.mean(np.array(joint_loss)))+"\n")
# print("rnn loss : %.5f mf loss : %.5f : joint loss %.5f"%(np.mean(np.array(loss_rnn)),np.mean(np.array(loss_mf)),np.mean(np.array(joint_loss))))
#
# d = helper.evaluateMultiProcess(sess, dis)
# g = helper.evaluateMultiProcess(sess, gen)
# with open("test1.txt", "a") as myfile:
# myfile.write("\n".join(str(elem) for elem in d))
# myfile.write("\n")
# myfile.write("\n".join(str(elem) for elem in g))
# print(d)
# print(g)
for d_epoch in range(3):
rnn_losses,mf_losses,joint_losses=[],[],[]
user_item_neg_rating_time_dict = lambda group:{item:t for i,(item,t) in group[group.rating<=2][["itemid","user_granularity"]].iterrows()}
user_item_neg_rating_time_dict = helper.train.groupby("uid").apply(user_item_neg_rating_time_dict).to_dict()
for user in df["uid"].unique():
all_rating = gen.predictionItems(sess1,user) # todo delete the pos ones
exp_rating = np.exp(np.array(all_rating) * 50)
prob = exp_rating / np.sum(exp_rating)
pesudo_positive_items = np.argsort(prob)[::-1][:32]
pesudo_positive_samples = []
for item in pesudo_positive_items:
u_seqs,i_seqs = helper.getSeqInTime(user,item,0)
pesudo_positive_samples.append((u_seqs,i_seqs,user,item,1))
negative_samples = []
neg_items_time_dict = user_item_neg_rating_time_dict.get(user,{})
if len(neg_items_time_dict)==0:
continue
for ind in np.random.randint(len(neg_items_time_dict), size=32):
negative_item,t = list(neg_items_time_dict.items())[ind]
u_seqs,i_seqs = helper.getSeqInTime(user,negative_item,t)
negative_samples.append((u_seqs,i_seqs,user,negative_item,0))
samples = pesudo_positive_samples + negative_samples
u_seq,i_seq = [[ s[j].toarray() for s in samples ] for j in range(2)]
u,i = [[ s[j] for s in samples ] for j in range(2,4)]
ratings = [ s[4] for s in samples ]
_,loss_mf,loss_rnn,joint_loss,rnn,mf = dis.pretrain_step(sess1,ratings, u, i,u_seq,i_seq)
rnn_losses.append(loss_rnn)
mf_losses.append(loss_mf)
joint_losses.append(joint_loss)
if d_epoch % 3 == 0:
print("rnn loss : %.5f mf loss : %.5f : joint loss %.5f"%(np.mean(np.array(loss_rnn)),np.mean(np.array(loss_mf)),np.mean(np.array(joint_loss))))
d = helper.evaluateMultiProcess(sess2, dis)
# with open("test_lr.txt", "a") as myfile:
# myfile.write("\n".join(str(elem) for elem in d))
# myfile.write("\n")
# myfile.write("\n".join(str(elem) for elem in g))
print(d)
print(g)
# for i, (u_seqs,i_seqs,ratings,userids,itemids) in enumerate(helper.getBatchFromDNS(dns=True,sess=sess,model=gen,fresh=False)):
# _,loss_mf,loss_rnn,joint_loss,rnn,mf = dis.pretrain_step(sess,ratings, userids, itemids,u_seqs,i_seqs)
# rnn_losses.append(loss_rnn)
# mf_losses.append(loss_mf)
# joint_losses.append(joint_loss)
# print("rnn loss : %.5f mf loss : %.5f : joint loss %.5f"%(np.mean(np.array(loss_rnn)),np.mean(np.array(loss_mf)),np.mean(np.array(joint_loss))))
#
# if helper.conf.lastone:
# else:
# u_seqss,i_seqss= helper.getSeqOverAlltime(user,item)
# predicted = gen.prediction(sess,u_seqss,i_seqss, [user]*len(u_seqss),[item]*len(u_seqss),sparse=True)
# index=np.argmax(predicted)
# samples.append((u_seqss[index],i_seqss[index],user,item ))
# samples=[]
# for item in sampled_items:
# u_seqs,i_seqs = helper.getSeqInTime(user,item,0)
# labeled_row = df.loc[(df.uid==user) & (df.itemid==item)]
# samples.append((u_seqs,i_seqs,user,item, (1 if len(labeled_row)>0 else 0),
# int(labeled_row.rating if len(labeled_row)>0 else 0)))
# rewards = dis.getRewards(sess,gen, samples, sparse=True)
# labeled_rewards = np.zeros(len(samples))
# return 2 * (self.sigmoid(unlabeled_rewards) - 0.5)
if __name__== "__main__":
main()