-
Notifications
You must be signed in to change notification settings - Fork 10
/
model.py
246 lines (192 loc) · 10.1 KB
/
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
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
import tensorflow as tf
from reader import batch_review_normalize, PAD_INDEX, START_INDEX
from utils import load_glove, get_shape
class Model:
def __init__(self, total_users, total_items, global_rating,
num_factors, img_dims, vocab_size,
word_dim, lstm_dim, max_length, dropout_rate):
self.total_users = total_users
self.total_items = total_items
self.global_rating = global_rating
self.dropout_rate = dropout_rate
self.F = num_factors
self.L = img_dims[0]
self.D = img_dims[1]
self.V = vocab_size
self.W = word_dim
self.C = lstm_dim
self.T = max_length
self.weight_initializer = tf.contrib.layers.xavier_initializer()
self.const_initializer = tf.zeros_initializer()
self.users = tf.placeholder(tf.int32, shape=[None])
self.items = tf.placeholder(tf.int32, shape=[None])
self.ratings = tf.placeholder(tf.float32, shape=[None])
self.images = tf.placeholder(tf.float32, shape=[None, self.L, self.D])
self.reviews = tf.placeholder(tf.int32, shape=[None, None])
self.is_training = tf.placeholder(tf.bool)
self._init_embeddings()
self.user_emb = tf.nn.embedding_lookup(self.user_matrix, self.users)
self.item_emb = tf.nn.embedding_lookup(self.item_matrix, self.items)
self.sentiment_features = self._get_features(self.user_emb, self.item_emb)
self.sentiment_features = self._batch_norm(self.sentiment_features, name='review/sentiment')
self.visual_features = self._batch_norm(self.images, name='review/visual')
self.visual_projection = self._visual_projection(self.visual_features)
self._build_rating_predictor()
self._build_review_generator()
self._build_review_sampler(max_decode_length=self.T)
def _init_embeddings(self):
self.user_matrix = tf.get_variable(
name='user_matrix',
shape=[self.total_users, self.F],
initializer=self.weight_initializer,
dtype=tf.float32
)
self.item_matrix = tf.get_variable(
name='item_matrix',
shape=[self.total_items, self.F],
initializer=self.weight_initializer,
dtype=tf.float32
)
self.word_matrix = tf.get_variable(
name='word_matrix',
shape=[self.V, self.W],
initializer=tf.constant_initializer(load_glove(self.V, self.W)),
dtype=tf.float32
)
def _get_features(self, user_emb, item_emb, num_layers=1):
with tf.variable_scope('features', reuse=tf.AUTO_REUSE):
features = tf.concat([user_emb, item_emb], axis=1)
for layer in range(num_layers):
w = tf.get_variable('w{}'.format(layer), [2 * self.F, 2 * self.F], initializer=self.weight_initializer)
b = tf.get_variable('b{}'.format(layer), [2 * self.F], initializer=self.const_initializer)
features = tf.matmul(features, w) + b
features = tf.nn.tanh(features, 'h{}'.format(layer))
return features
def _build_rating_predictor(self):
features = self._get_features(self.user_emb, self.item_emb)
with tf.variable_scope('rating'):
rating_labels = tf.reshape(self.ratings, [-1, 1])
rating_preds = self.global_rating + tf.layers.dense(features, units=1, name='prediction')
self.rating_loss = tf.losses.mean_squared_error(rating_labels, rating_preds)
self.rating_preds = tf.clip_by_value(rating_preds, clip_value_min=1.0, clip_value_max=5.0)
self.mae, mae_update = tf.metrics.mean_absolute_error(rating_labels, self.rating_preds, name='metrics/MAE')
self.rmse, rmse_update = tf.metrics.root_mean_squared_error(rating_labels, self.rating_preds, name='metrics/RMSE')
metric_vars = tf.get_collection(tf.GraphKeys.LOCAL_VARIABLES, scope="rating/metrics")
self.init_metrics = tf.variables_initializer(var_list=metric_vars)
self.update_metrics = tf.group([mae_update, rmse_update])
def _batch_norm(self, x, name=None):
return tf.contrib.layers.batch_norm(inputs=x,
decay=0.95,
center=True,
scale=True,
is_training=self.is_training,
updates_collections=None,
scope=(name + '_batch_norm'))
def _visual_projection(self, features):
with tf.variable_scope('review/visual_projection'):
w = tf.get_variable('w', [self.D, self.D], initializer=self.weight_initializer)
features_flat = tf.reshape(features, [-1, self.D])
features_proj = tf.matmul(features_flat, w)
features_proj = tf.reshape(features_proj, [-1, self.L, self.D])
return features_proj
def _attention_layer(self, h, features, features_proj):
with tf.variable_scope('attention'):
L = get_shape(features)[1]
w = tf.get_variable('w', [self.C, self.D], initializer=self.weight_initializer)
b = tf.get_variable('b', [self.D], initializer=self.const_initializer)
w_att = tf.get_variable('w_att', [self.D, 1], initializer=self.weight_initializer)
b_att = tf.get_variable('b_att', [1], initializer=self.const_initializer)
h_att = tf.nn.tanh(features_proj + tf.expand_dims(tf.matmul(h, w), 1) + b)
out_att = tf.reshape(tf.matmul(tf.reshape(h_att, [-1, self.D]), w_att) + b_att, [-1, L])
alpha = tf.nn.softmax(out_att)
context = tf.reduce_sum(features * tf.expand_dims(alpha, 2), 1, name='context')
return context, alpha
def _fusion_gate(self, x, h, s_features, v_features):
with tf.variable_scope('fusion_gate'):
w_x = tf.get_variable('w_x', [self.W, 1], initializer=self.weight_initializer)
w_h = tf.get_variable('w_h', [self.C, 1], initializer=self.weight_initializer)
b = tf.get_variable('b', [1], initializer=self.const_initializer)
beta = tf.nn.sigmoid(tf.matmul(x, w_x) + tf.matmul(h, w_h) + b) # (N, 1)
weighted_features = tf.multiply(beta, s_features) + tf.multiply((1. - beta), v_features)
return weighted_features, beta
def _init_lstm(self):
with tf.variable_scope('init_lstm'):
user_item_emb = tf.concat([self.user_emb, self.item_emb], axis=1)
w_h_ui = tf.get_variable('w_h_ui', [self.D, self.C], initializer=self.weight_initializer)
b_h = tf.get_variable('b_h', [self.C], initializer=self.const_initializer)
h = tf.matmul(user_item_emb, w_h_ui) + b_h
w_c_ui = tf.get_variable('w_c_ui', [self.D, self.C], initializer=self.weight_initializer)
b_c = tf.get_variable('b_c', [self.C], initializer=self.const_initializer)
c = tf.matmul(user_item_emb, w_c_ui) + b_c
h = tf.nn.tanh(h)
c = tf.nn.tanh(c)
return c, h
def _decode_lstm(self, x, h, context):
with tf.variable_scope('decode_lstm'):
w_h = tf.get_variable('w_h', [self.C, self.W], initializer=self.weight_initializer)
b_h = tf.get_variable('b_h', [self.W], initializer=self.const_initializer)
w_out = tf.get_variable('w_out', [self.W, self.V], initializer=self.weight_initializer)
b_out = tf.get_variable('b_out', [self.V], initializer=self.const_initializer)
h = tf.layers.dropout(h, self.dropout_rate, training=self.is_training)
h_logits = tf.matmul(h, w_h) + b_h
w_ctx2out = tf.get_variable('w_ctx2out', [get_shape(context)[1], self.W], initializer=self.weight_initializer)
h_logits += tf.matmul(context, w_ctx2out)
h_logits += x
h_logits = tf.nn.tanh(h_logits)
h_logits = tf.layers.dropout(h_logits, self.dropout_rate, training=self.is_training)
out_logits = tf.matmul(h_logits, w_out) + b_out
return out_logits
def _build_review_generator(self):
with tf.variable_scope('review', reuse=tf.AUTO_REUSE):
reviews_inputs = self.reviews[:, :self.T - 1]
reviews_emb = tf.nn.embedding_lookup(self.word_matrix, reviews_inputs)
reviews_labels = self.reviews[:, 1:]
mask = tf.to_float(tf.not_equal(reviews_labels, PAD_INDEX))
loss = 0.0
self.cell = tf.nn.rnn_cell.BasicLSTMCell(num_units=self.C, name='LSTM_Cell')
c, h = self._init_lstm()
for t in range(self.T - 1):
x = reviews_emb[:, t, :]
visual_context, alpha = self._attention_layer(h, self.visual_features, self.visual_projection)
context, beta = self._fusion_gate(x, h, self.sentiment_features, visual_context)
cell_input = tf.concat([x, context], axis=1)
_, (c, h) = self.cell(inputs=cell_input, state=[c, h])
logits = self._decode_lstm(x, h, context)
loss += tf.reduce_sum(
tf.nn.sparse_softmax_cross_entropy_with_logits(labels=reviews_labels[:, t], logits=logits) * mask[:, t])
self.review_loss = loss / tf.reduce_sum(mask)
def _build_review_sampler(self, max_decode_length):
with tf.variable_scope('review', reuse=tf.AUTO_REUSE):
sampled_word_list = []
beta_list = []
alpha_list = []
c, h = self._init_lstm()
batch_size = tf.shape(self.users)[0]
sampled_word = tf.fill([batch_size], START_INDEX)
for t in range(max_decode_length):
x = tf.nn.embedding_lookup(self.word_matrix, sampled_word)
visual_context, alpha = self._attention_layer(h, self.visual_features, self.visual_projection)
alpha_list.append(alpha)
context, beta = self._fusion_gate(x, h, self.sentiment_features, visual_context)
beta_list.append(beta)
cell_input = tf.concat([x, context], axis=1)
_, (c, h) = self.cell(inputs=cell_input, state=[c, h])
logits = self._decode_lstm(x, h, context)
sampled_word = tf.argmax(logits, 1)
sampled_word_list.append(sampled_word)
self.sampled_reviews = tf.transpose(tf.stack(sampled_word_list), (1, 0)) # (N, max_len)
self.alphas = tf.transpose(tf.stack(alpha_list), (1, 0, 2)) # (N, T, L)
self.betas = tf.transpose(tf.squeeze(tf.stack(beta_list), axis=2), (1, 0)) # (N, T)
def feed_dict(self, users, items, ratings=None, images=None, reviews=None, is_training=False):
fd = {
self.users: users,
self.items: items,
self.is_training: is_training
}
if ratings is not None:
fd[self.ratings] = ratings
if images is not None:
fd[self.images] = images
if reviews is not None:
fd[self.reviews] = batch_review_normalize(reviews, self.T)
return fd