/
peris_model.py
187 lines (156 loc) · 10.2 KB
/
peris_model.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
import tensorflow as tf
import time
from utils import Eval
import numpy as np
from sampler import *
from mylayers import RCEncoding, EuclideanDistillation
samplers = [SamplerModel, PopularSamplerModel, ClusterSamplerModel, ClusterPopularSamplerModel, ExactSamplerModel]
class EvaluateCallback(tf.keras.callbacks.Callback):
def __init__(self, round_):
self.round = round_
super(EvaluateCallback, self).__init__()
def on_epoch_begin(self, epoch, logs=None):
self.starttime = time.time()
def on_epoch_end(self, epoch, logs=None):
elapsed = time.time() - self.starttime
print('Epoch={} - {}s - loss={:.4f}'.format(self.round + epoch + 1, int(elapsed), logs['loss']))
def compute_loss(pred, prob, weighted):
if weighted:
importance = tf.nn.softmax(tf.negative(pred) - tf.log(prob))
else:
importance = tf.nn.softmax(tf.ones_like(pred))
weight_loss = tf.multiply(importance, tf.negative(tf.log_sigmoid(pred)))
loss = tf.reduce_sum(weight_loss, -1, keepdims=True)
return loss
def identity_loss(y_true, y_pred):
return tf.reduce_mean(y_pred - 0 * y_true)
class PerisModel:
def __init__(self, config):
user_id = tf.keras.Input(shape=(1,), name='user_id')
pos_id = tf.keras.Input(shape=(1,), name='pos_id')
neg_id = tf.keras.Input(shape=(config.neg_num,), name='neg_id')
neg_prob = tf.keras.Input(shape=(config.neg_num,), name='neg_prob', dtype='float32')
item_embed_layer = tf.keras.layers.Embedding(config.num_item, config.d, name='item_embedding',
embeddings_initializer=tf.keras.initializers.glorot_normal(),
activity_regularizer=tf.keras.regularizers.l2(config.coef / config.batch_size))
user_embed = tf.keras.layers.Embedding(config.num_user, config.d, name='user_embedding',
embeddings_initializer=tf.keras.initializers.glorot_normal(),
activity_regularizer=tf.keras.regularizers.l2(config.coef / config.batch_size))(user_id)
pos_item_embed = item_embed_layer(pos_id)
neg_item_embed = item_embed_layer(neg_id)
pos_score = tf.keras.layers.dot([user_embed, pos_item_embed], axes=-1)
neg_score = tf.keras.layers.dot([user_embed, neg_item_embed], axes=-1)
ruij = tf.keras.layers.Flatten()(tf.keras.layers.subtract([pos_score, neg_score]))
loss = tf.keras.layers.Lambda(lambda x: compute_loss(*x, config.weighted))([ruij, neg_prob])
self.model = tf.keras.Model(inputs=[user_id, pos_id, neg_id, neg_prob], outputs=loss)
self.model.compile(loss=identity_loss, optimizer=tf.keras.optimizers.Adam(lr=config.learning_rate))
self.config = config
def get_uv(self):
user_embed = self.model.get_layer('user_embedding')
item_embed = self.model.get_layer('item_embedding')
u = user_embed.get_weights()[0]
v = item_embed.get_weights()[0]
return u, v
def fit(self, train):
steps_per_epoch = int((train.nnz + self.config.batch_size - 1) / self.config.batch_size)
opt_para = [{}, {'mode': self.config.mode}, {'num_clusters': self.config.num_clusters},
{'num_clusters': self.config.num_clusters, 'mode': self.config.mode}, {}]
if self.config.sampler in {0, 1}:
sampler = samplers[self.config.sampler](train, **opt_para[self.config.sampler])\
.negative_sampler(neg=self.config.neg_num)
dataset = IO.construct_dataset(sampler, self.config.neg_num).shuffle(50000)\
.batch(self.config.batch_size).repeat(self.config.epochs)
self.model.fit(dataset, epochs=self.config.epochs, steps_per_epoch=steps_per_epoch, verbose=0,
callbacks=[EvaluateCallback(0)])
elif self.config.sampler in {2, 3, 4}:
sampler = samplers[self.config.sampler].__bases__[0](train, **opt_para[self.config.sampler % 2])\
.negative_sampler(neg=self.config.neg_num)
for i in range(int(self.config.epochs / self.config.epochs_)):
dataset = IO.construct_dataset(sampler, self.config.neg_num).shuffle(50000)\
.batch(self.config.batch_size).repeat(self.config.epochs_)
self.model.fit(dataset, epochs=self.config.epochs_, steps_per_epoch=steps_per_epoch, verbose=0,
callbacks=[EvaluateCallback(i * self.config.epochs_)])
u, v = self.get_uv()
sampler = samplers[self.config.sampler](train, {'U': u, 'V': v}, **opt_para[self.config.sampler])\
.negative_sampler(self.config.neg_num)
def evaluate(self, train, test):
m, n = train.shape
u, v = self.get_uv()
users = np.random.choice(m, min(m, 50000), False)
m = Eval.evaluate_item(train[users, :], test[users, :], u[users, :], v, topk=-1)
return m
class PerisJointModel:
def __init__(self, config):
user_id = tf.keras.Input(shape=(1,), name='user_id')
pos_id = tf.keras.Input(shape=(1,), name='pos_id')
neg_id = tf.keras.Input(shape=(config.neg_num,), name='neg_id')
neg_prob = tf.keras.Input(shape=(config.neg_num,), name='neg_prob', dtype='float32')
item_embed_layer = tf.keras.layers.Embedding(config.num_item, config.d, name='item_embedding',
embeddings_initializer=tf.keras.initializers.glorot_normal(),
activity_regularizer=tf.keras.regularizers.l2(config.coef / config.batch_size))
user_embed = tf.keras.layers.Embedding(config.num_user, config.d, name='user_embedding',
embeddings_initializer=tf.keras.initializers.glorot_normal(),
activity_regularizer=tf.keras.regularizers.l2(config.coef / config.batch_size))(user_id)
pos_item_embed = item_embed_layer(pos_id)
neg_item_embed = item_embed_layer(neg_id)
pos_score = tf.keras.layers.dot([user_embed, pos_item_embed], axes=-1)
neg_score = tf.keras.layers.dot([user_embed, neg_item_embed], axes=-1)
ruij = tf.keras.layers.Flatten()(tf.keras.layers.subtract([pos_score, neg_score]))
loss = tf.keras.layers.Lambda(lambda x: compute_loss(*x, config.weighted))([ruij, neg_prob])
num_clusters = config.num_clusters
reg = tf.keras.layers.ActivityRegularization(l2=config.coef2 / config.batch_size)
dist = EuclideanDistillation(coef=config.coef_kd)
def transform(x):
return reg(dist(x))
stop_grad = tf.keras.layers.Lambda(lambda x: tf.stop_gradient(x))
num_codewords = [num_clusters]
item_rce_layer = RCEncoding(num_codewords, att_mode='bilinear', rnn_mode='none', name='rcencoding')
pos_item_embed_stop = stop_grad(pos_item_embed)
neg_item_embed_stop = stop_grad(neg_item_embed)
pos_item_embed_, pos_item_cluster_idx = item_rce_layer(pos_item_embed_stop)
neg_item_embed_, _ = item_rce_layer(neg_item_embed_stop)
user_embed_ = tf.keras.layers.Dense(config.d, use_bias=False, name='user_dense',
activity_regularizer=tf.keras.regularizers.l2(config.coef2 / config.batch_size))(stop_grad(user_embed))
pos_score_ = tf.keras.layers.dot([user_embed_, transform([pos_item_embed_stop, pos_item_embed_])], axes=-1)
neg_score_ = tf.keras.layers.dot([user_embed_, transform([neg_item_embed_stop, neg_item_embed_])], axes=-1)
ruij_ = tf.keras.layers.Flatten()(tf.keras.layers.subtract([pos_score_, neg_score_]))
loss_ = tf.keras.layers.Lambda(lambda x: compute_loss(*x, config.weighted))([ruij_, neg_prob])
loss2 = tf.keras.layers.Lambda(lambda x: x[0] + x[1])([loss, loss_])
self.model = tf.keras.Model(inputs=[user_id, pos_id, neg_id, neg_prob], outputs=loss2)
self.model.compile(loss=identity_loss, optimizer=tf.keras.optimizers.Adam())
self.config = config
def get_cluster(self, m, n):
cluster = self.model.get_layer('rcencoding')
model_pred_item_cluster = tf.keras.Model(inputs=self.model.input[1], outputs=cluster.output[1])
user_embed_layer = self.model.get_layer('user_dense')
model_pred_user = tf.keras.Model(inputs=self.model.input[0], outputs=user_embed_layer.output)
item_code = np.squeeze(model_pred_item_cluster.predict(np.arange(n)))
item_center = cluster.get_weights()[0]
U = np.squeeze(model_pred_user.predict(np.arange(m)))
return U, item_code, item_center
def get_uv(self):
user_embed = self.model.get_layer('user_embedding')
item_embed = self.model.get_layer('item_embedding')
U = user_embed.get_weights()[0]
V = item_embed.get_weights()[0]
return U, V
def fit(self, train):
steps_per_epoch = int((train.nnz + self.config.batch_size - 1) / self.config.batch_size)
opt_para = {} if self.config.sampler == 2 else {'mode': self.config.mode}
if self.config.sampler in {2, 3}:
sampler = samplers[self.config.sampler].__bases__[0](train, **opt_para)\
.negative_sampler(neg=self.config.neg_num)
for i in range(int(self.config.epochs / self.config.epochs_)):
dataset = IO.construct_dataset(sampler, self.config.neg_num).shuffle(50000)\
.batch(self.config.batch_size).repeat(self.config.epochs_)
self.model.fit(dataset, epochs=self.config.epochs_, steps_per_epoch=steps_per_epoch, verbose=0,
callbacks=[EvaluateCallback(i * self.config.epochs_)])
u, code, center = self.get_cluster(self.config.num_user, self.config.num_item)
sampler = ClusterPopularSamplerModel(train, {'U': u, 'code': code, 'center': center}, **opt_para)\
.negative_sampler(self.config.neg_num)
def evaluate(self, train, test):
m, n = train.shape
u, v = self.get_uv()
users = np.random.choice(m, min(m, 50000), False)
m = Eval.evaluate_item(train[users, :], test[users, :], u[users, :], v, topk=-1)
return m