def _forward_pass(self, ):
# 用户的向量表示
with tf.name_scope('user_express'):
# 用户隐向量
self.Usr_Emb = tf.nn.embedding_lookup(
self.Wu_Emb, tf.cast(self.user_indices,
tf.int32)) # [batch_size, dim_k]
self.Usr_Feat = tf.nn.embedding_lookup(
self.user_emb_feat, tf.cast(self.user_indices,
tf.int32)) # [batch_size, 128]
self.Usr_Feat_Emb = tf.matmul(
self.Usr_Feat, self.W_usr_feat_emb) # [batch_size, dim_k]
self.Usr_Feat_Emb = tf.nn.relu(self.Usr_Feat_Emb)
# self.Usr_Feat_Emb = tf.layers.dropout(self.Usr_Feat_Emb, self.dropout_emb)
# self.Usr_Feat_Emb = self._batch_norm_layer(self.Usr_Feat_Emb, self.train_phase, 'user_emb_bn')
self.Usr_Expr_a = self.Usr_Emb + self.Usr_Feat_Emb # [batch_size, dim_k]
# 环境的向量表示
with tf.name_scope('context_express'):
self.att_oh_ctx = []
self.Ctx_oh_emb = []
self.att_u_ctx = tf.matmul(
self.Usr_Emb, self.Wu_ctx_oh_Att) # [batch_size, att_dim_k]
for i in range(len(self.ctx_oh_dims)):
Ctx_emb_temp = tf.nn.embedding_lookup(
self.W_ctx_oh[i], tf.cast(self.ctx_oh[:, i],
tf.int32)) # [batch_size, dim_k]
att_oh_temp = tf.matmul(
Ctx_emb_temp,
self.Woh_ctx_Att[i]) # [batch_size, att_dim_k]
att_temp = tf.nn.relu(self.att_u_ctx + att_oh_temp +
self.b_oh_Att) # [batch_size, att_dim_k]
att_temp = tf.matmul(
att_temp, self.w_oh_Att) + self.c_oh_Att # [batch_size, 1]
self.att_oh_ctx.append(att_temp)
self.Ctx_oh_emb.append(Ctx_emb_temp)
self.att_oh_ctx = tf.nn.softmax(tf.concat(
self.att_oh_ctx, axis=1)) # [batch_size, ctx_dim]
for i in range(0, len(self.ctx_oh_dims)):
self.Ctx_oh_emb[i] = self.Ctx_oh_emb[i] * self.att_u_ctx[:,
i:i +
1]
self.Ctx_Emb = tf.add_n(self.Ctx_oh_emb) # [batch_size, dim_k]
# 物品的向量表示
with tf.name_scope('item_express'):
self.I_Wds_a = tf.SparseTensor(
indices=tf.cast(self.item_words_indices_a, dtype=np.int64),
values=self.item_words_values_a,
dense_shape=[
tf.cast(self.batch_size, dtype=np.int64), self.num_words
])
self.att_u_a = tf.matmul(self.Usr_Emb,
self.Wu_oh_Att) # [batch_size, att_dim_k]
self.att_ctx = tf.matmul(self.Ctx_Emb, self.Wctx_Att)
self.att_oh = []
self.I_Wds_Emb_a = tf.sparse_tensor_dense_matmul(
self.I_Wds_a, self.Wwords_Emb) # [batch_size, dim_k]
self.I_Wds_Emb_a = tf.nn.relu(self.I_Wds_Emb_a)
self.att_I_Wds = tf.matmul(self.I_Wds_Emb_a,
self.WW_Att) # 词的attention
self.att_I_Wds = tf.nn.relu(self.att_u_a + self.att_ctx +
self.att_I_Wds + self.b_oh_Att)
self.att_I_Wds = tf.matmul(self.att_I_Wds,
self.w_oh_Att) + self.c_oh_Att
self.att_oh.append(self.att_I_Wds)
vae_encoder = VAE(input_dim=2048,
hidden_encoder_dim=1024,
latent_dim=96,
lam=0.001,
kld_loss=0.001)
self.I_visual_Emb, self.vae_loss = vae_encoder.get_vae_embbeding(
self.visual_emb_feat)
# TODO
self.I_visual_Emb = self._batch_norm_layer(self.I_visual_Emb,
self.train_phase,
'vis_bn')
# self.I_visual_Emb = tf.layers.dropout(self.I_visual_Emb, self.dropout_emb)
self.att_I_visual = tf.matmul(self.I_visual_Emb,
self.W_visual_Att) # 图像的attention
self.att_I_visual = tf.nn.relu(self.att_u_a + self.att_ctx +
self.att_I_visual + self.b_oh_Att)
self.att_I_visual = tf.matmul(self.att_I_visual,
self.w_oh_Att) + self.c_oh_Att
self.att_oh.append(self.att_I_visual)
self.I_LDA_Emb = tf.matmul(self.words_lda,
self.W_LDA_emb) # [batch_size, dim_k]
self.att_I_LDA = tf.matmul(self.I_LDA_Emb,
self.W_LDA_Att) # LDA的attention
self.att_I_LDA = tf.nn.relu(self.att_u_a + self.att_ctx +
self.att_I_LDA + self.b_oh_Att)
self.att_I_LDA = tf.matmul(self.att_I_LDA,
self.w_oh_Att) + self.c_oh_Att
self.att_oh.append(self.att_I_LDA)
self.I_One_hot_a = []
for i in range(len(self.one_hots_dims)):
I_Emb_temp_a = tf.nn.embedding_lookup(
self.W_one_hots[i],
tf.cast(self.one_hots_a[:, i],
tf.int32)) # [batch_size, dim_k]
att_oh_temp = tf.matmul(
I_Emb_temp_a, self.Woh_Att[i]) # [batch_size, att_dim_k]
att_temp = tf.nn.relu(self.att_u_a + self.att_ctx +
att_oh_temp +
self.b_oh_Att) # [batch_size, att_dim_k]
att_temp = tf.matmul(
att_temp, self.w_oh_Att) + self.c_oh_Att # [batch_size, 1]
self.att_oh.append(att_temp)
self.I_One_hot_a.append(I_Emb_temp_a)
self.att_oh = tf.nn.softmax(tf.concat(
self.att_oh, axis=1)) # [batch_size, oh_dim] 第一列是词attention
self.I_Wds_Emb_a = self.I_Wds_Emb_a * self.att_oh[:, 0:1]
self.I_visual_Emb = self.I_visual_Emb * self.att_oh[:, 1:2]
self.I_LDA_Emb = self.I_LDA_Emb * self.att_oh[:, 2:3]
for i in range(3, len(self.one_hots_dims) + 3):
self.I_One_hot_a[
i - 3] = self.I_One_hot_a[i - 3] * self.att_oh[:, i:i + 1]
self.Item_Expr_a = tf.add_n(
self.I_One_hot_a +
[self.I_visual_Emb, self.I_Wds_Emb_a, self.I_LDA_Emb])
with tf.name_scope('deep'):
if self.use_deep:
self.Usr_emb_deep = tf.nn.embedding_lookup(
self.Wu_deep_emb, tf.cast(self.user_indices,
tf.int32)) # [batch_size, dim_k]
self.I_Wds_emb_deep = tf.sparse_tensor_dense_matmul(
self.I_Wds_a, self.Wwords_deep_emb) # [batch_size, dim_k]
self.I_one_hot_deep = []
for i in range(len(self.one_hots_dims)):
I_Emb_temp_a = tf.nn.embedding_lookup(
self.W_deep_one_hots[i],
tf.cast(self.one_hots_a[:, i],
tf.int32)) # [batch_size, dim_k]
self.I_one_hot_deep.append(I_Emb_temp_a)
# self.deep_input = tf.concat([self.num_features, self.Usr_Feat, self.face_num, self.visual_emb_feat],
# axis=1) # [batch_size, input_dim]
self.deep_input = tf.concat(
[self.num_features, self.visual_emb_feat] +
self.I_one_hot_deep,
axis=1)
# 输入加入batch_norm
# self.deep_input = self._batch_norm_layer(self.deep_input, self.train_phase, 'input_bn')
for i, deep_dim in enumerate(self.deep_dims):
self.deep_input = tf.layers.dense(
inputs=self.deep_input,
kernel_initializer=tf.glorot_uniform_initializer(),
units=deep_dim,
activation=tf.nn.relu,
)
if i == 0:
self.deep_input = self._batch_norm_layer(
self.deep_input, self.train_phase,
'deep_bn_{}'.format(i))
# 加入dropout
self.deep_input = tf.layers.dropout(inputs=self.deep_input,
rate=self.dropout_deep)
self.deep_output = tf.layers.dense(self.deep_input,
1,
activation=None)
self.deep_output = tf.reshape(self.deep_output,
[-1]) # [batch_size]
with tf.name_scope('output'):
self.cf_out = tf.layers.dropout(self.Item_Expr_a * self.Usr_Expr_a,
rate=self.dropout_emb)
self.ctx_usr_out = tf.layers.dropout(self.Ctx_Emb *
self.Usr_Expr_a,
rate=self.dropout_emb)
self.ctx_item_out = tf.layers.dropout(self.Ctx_Emb *
self.Item_Expr_a,
rate=self.dropout_emb)
# self.cf_out, self.ctx_usr_out, self.ctx_item_out = self._attentioned_layer(
# [self.cf_out, self.ctx_usr_out, self.ctx_item_out])
# self.in_prd_out = tf.add_n([self.cf_out, self.ctx_usr_out, self.ctx_item_out])
# if self.use_deep:
# self.concated = tf.concat([self.in_prd_out, self.deep_input], axis=1)
# else:
# self.concated = tf.concat([self.in_prd_out], axis=1)
if self.use_deep:
self.concated = tf.concat([
self.cf_out, self.ctx_usr_out, self.ctx_item_out,
self.deep_input
],
axis=1)
else:
self.concated = tf.concat(
[self.cf_out, self.ctx_usr_out, self.ctx_item_out], axis=1)
self.hidden = self.concated
for i, dim in enumerate(self.dim_hidden_out):
self.hidden = tf.layers.dense(
inputs=self.hidden,
kernel_initializer=tf.glorot_uniform_initializer(),
units=dim,
activation=tf.nn.relu)
self.hidden = tf.layers.dropout(inputs=self.hidden,
rate=self.dropout_deep)
self.bu = tf.nn.embedding_lookup(
self.bias_u, tf.cast(self.user_indices, dtype=tf.int32))
self.y_ui_a = tf.layers.dense(self.hidden, 1,
activation=None) + self.bu
self.y_ui_a = tf.reshape(tf.nn.sigmoid(self.y_ui_a), [-1])
def _forward_pass(self, ):
# 用户的向量表示
with tf.name_scope('user_express'):
# 用户隐向量
self.Usr_Emb = tf.nn.embedding_lookup(self.Wu_Emb,
tf.cast(self.user_indices, tf.int32)) # [batch_size, dim_k]
self.Usr_Feat = tf.nn.embedding_lookup(self.user_emb_feat,
tf.cast(self.user_indices, tf.int32)) # [batch_size, dim_cf_emb]
self.bias_usr_feat_emb = tf.get_variable(shape=[self.dim_k], initializer=tf.zeros_initializer(),
dtype=tf.float32,
name='bias_usr_feat_emb')
self.Usr_Feat_Emb = tf.matmul(self.Usr_Feat,
self.W_usr_feat_emb) + self.bias_usr_feat_emb # [batch_size, dim_k]
self.Usr_Feat_Emb = tf.nn.relu(self.Usr_Feat_Emb)
# self.Usr_Feat_Emb = tf.layers.dropout(self.Usr_Feat_Emb, self.dropout_emb)
# self.Usr_Feat_Emb = self._batch_norm_layer(self.Usr_Feat_Emb, self.train_phase, 'user_emb_bn')
self.Usr_Expr_a = self.Usr_Emb + self.Usr_Feat_Emb # [batch_size, dim_k]
# 环境的向量表示
with tf.name_scope('context_express'):
self.bias_ctx_emb = tf.get_variable(shape=[self.dim_k], initializer=tf.zeros_initializer(),
dtype=tf.float32,
name='bias_ctx_emb')
self.Ctx_Emb = tf.matmul(self.num_features, self.W_Ctx) + self.bias_ctx_emb # [batch_size, dim_k]
self.Ctx_Emb = self._batch_norm_layer(self.Ctx_Emb, self.train_phase, 'ctx_bn')
self.Ctx_Emb = tf.nn.relu(self.Ctx_Emb)
# 物品的向量表示
with tf.name_scope('item_express'):
self.att_u_a = tf.matmul(self.Usr_Expr_a, self.Wu_oh_Att) # [batch_size, dim_att]
self.att_ctx = tf.matmul(self.Ctx_Emb, self.Wctx_Att)
self.att_oh = []
vae_encoder = VAE(input_dim=2048, hidden_encoder_dim=1024, latent_dim=self.dim_k, lam=0.001, kld_loss=0.001)
self.I_visual_Emb, self.vae_loss = vae_encoder.get_vae_embbeding(self.visual_emb_feat)
# TODO
self.I_visual_Emb = self._batch_norm_layer(self.I_visual_Emb, self.train_phase, 'vis_bn')
# self.I_visual_Emb = tf.layers.dropout(self.I_visual_Emb, self.dropout_emb)
self.att_I_visual = tf.matmul(self.I_visual_Emb, self.W_visual_Att) # 图像的attention
self.att_I_visual = tf.nn.relu(self.att_u_a + self.att_ctx + self.att_I_visual + self.b_oh_Att)
self.att_I_visual = tf.matmul(self.att_I_visual, self.w_oh_Att) + self.c_oh_Att
self.att_oh.append(self.att_I_visual)
self.bias_lda_emb = tf.get_variable(shape=[self.dim_k], initializer=tf.zeros_initializer(),
dtype=tf.float32,
name='bias_lda_emb')
self.I_LDA_Emb = tf.matmul(self.words_lda, self.W_LDA_emb) + self.bias_lda_emb # [batch_size, dim_k]
self.I_LDA_Emb = tf.nn.relu(self.I_LDA_Emb)
self.att_I_LDA = tf.matmul(self.I_LDA_Emb, self.W_LDA_Att) # LDA的attention
self.att_I_LDA = tf.nn.relu(self.att_u_a + self.att_ctx + self.att_I_LDA + self.b_oh_Att)
self.att_I_LDA = tf.matmul(self.att_I_LDA, self.w_oh_Att) + self.c_oh_Att
self.att_oh.append(self.att_I_LDA)
self.I_One_hot_a = []
for i in range(len(self.one_hots_dims)):
I_Emb_temp_a = tf.nn.embedding_lookup(self.W_one_hots[i],
tf.cast(self.one_hots_a[:, i],
tf.int32)) # [batch_size, dim_k]
att_oh_temp = tf.matmul(I_Emb_temp_a, self.Woh_Att[i]) # [batch_size, att_dim_k]
att_temp = tf.nn.relu(
self.att_u_a + self.att_ctx + att_oh_temp + self.b_oh_Att) # [batch_size, att_dim_k]
att_temp = tf.matmul(att_temp, self.w_oh_Att) + self.c_oh_Att # [batch_size, 1]
self.att_oh.append(att_temp)
self.I_One_hot_a.append(I_Emb_temp_a)
self.att_oh = tf.nn.softmax(tf.concat(self.att_oh, axis=1)) # [batch_size, oh_dim] 第一列是词attention
self.I_visual_Emb = self.I_visual_Emb * self.att_oh[:, 0:1]
self.I_LDA_Emb = self.I_LDA_Emb * self.att_oh[:, 1:2]
for i in range(2, len(self.one_hots_dims) + 2):
self.I_One_hot_a[i - 2] = self.I_One_hot_a[i - 2] * self.att_oh[:, i:i + 1]
self.Item_Expr_a = tf.add_n(self.I_One_hot_a + [self.I_visual_Emb, self.I_LDA_Emb])
with tf.name_scope('deep'):
if self.use_deep:
self.Usr_emb_deep = tf.nn.embedding_lookup(self.Wu_deep_emb,
tf.cast(self.user_indices, tf.int32)) # [batch_size, dim_k]
self.I_one_hot_deep = []
for i in range(len(self.one_hots_dims)):
I_Emb_temp_a = tf.nn.embedding_lookup(self.W_deep_one_hots[i], tf.cast(self.one_hots_a[:, i],
tf.int32)) # [batch_size, dim_k]
self.I_one_hot_deep.append(I_Emb_temp_a)
# self.deep_input = tf.concat([self.num_features, self.Usr_Feat, self.face_num, self.visual_emb_feat],
# axis=1) # [batch_size, input_dim]
self.deep_input = tf.concat(
[self.num_features, self.visual_emb_feat] + self.I_one_hot_deep, axis=1)
# 输入加入batch_norm
# self.deep_input = self._batch_norm_layer(self.deep_input, self.train_phase, 'input_bn')
for i, deep_dim in enumerate(self.deep_dims):
self.deep_input = tf.layers.dense(inputs=self.deep_input,
kernel_initializer=tf.glorot_uniform_initializer(),
units=deep_dim, activation=tf.nn.relu, )
if i == 0:
self.deep_input = self._batch_norm_layer(self.deep_input, self.train_phase,
'deep_bn_{}'.format(i))
# 加入dropout
self.deep_input = tf.layers.dropout(inputs=self.deep_input, rate=self.dropout_deep)
self.deep_output = tf.layers.dense(self.deep_input, 1, activation=None)
self.deep_output = tf.reshape(self.deep_output, [-1]) # [batch_size]
with tf.name_scope('output'):
self.ctx_item_out = tf.layers.dropout(self.Ctx_Emb * self.Item_Expr_a, rate=self.dropout_emb)
self.cf_outs = []
for usr_expr in [self.Usr_Emb, self.Usr_Feat_Emb]:
cf_out = tf.layers.dropout(self.Item_Expr_a * usr_expr, rate=self.dropout_emb)
ctx_usr_out = tf.layers.dropout(self.Ctx_Emb * usr_expr, rate=self.dropout_emb)
self.cf_outs.extend([cf_out, ctx_usr_out])
if self.use_deep:
self.concated = tf.concat(self.cf_outs + [self.ctx_item_out, self.deep_input], axis=1)
else:
self.concated = tf.concat(self.cf_outs + [self.ctx_item_out], axis=1)
self.hidden = self.concated
for i, dim in enumerate(self.dim_hidden_out):
self.hidden = tf.layers.dense(inputs=self.hidden, kernel_initializer=tf.glorot_uniform_initializer(),
units=dim, activation=tf.nn.relu)
self.hidden = tf.layers.dropout(inputs=self.hidden, rate=self.dropout_deep)
self.bu = tf.nn.embedding_lookup(self.bias_u, tf.cast(self.user_indices, dtype=tf.int32))
self.y_ui_a = tf.layers.dense(self.hidden, 1, activation=None) + self.bu
self.y_ui_a = tf.reshape(tf.nn.sigmoid(self.y_ui_a), [-1])
def _forward_pass(self, ):
# 用户的向量表示
with tf.name_scope('user_express'):
# 用户隐向量
self.Usr_Emb = tf.nn.embedding_lookup(
self.Wu_Emb, tf.cast(self.user_indices,
tf.int32)) # [batch_size, dim_k]
self.Usr_Feat = tf.nn.embedding_lookup(
self.user_emb_feat,
tf.cast(self.user_indices,
tf.int32)) # [batch_size, dim_cf_emb]
self.bias_usr_feat_emb = tf.get_variable(
shape=[self.dim_k],
initializer=tf.zeros_initializer(),
dtype=tf.float32,
name='bias_usr_feat_emb')
self.Usr_Feat_Emb = tf.matmul(
self.Usr_Feat, self.W_usr_feat_emb
) + self.bias_usr_feat_emb # [batch_size, dim_k]
self.Usr_Feat_Emb = tf.nn.relu(self.Usr_Feat_Emb)
# self.Usr_Feat_Emb = tf.layers.dropout(self.Usr_Feat_Emb, self.dropout_emb)
# self.Usr_Feat_Emb = self._batch_norm_layer(self.Usr_Feat_Emb, self.train_phase, 'user_emb_bn')
self.Usr_Expr_a = self.Usr_Emb + self.Usr_Feat_Emb # [batch_size, dim_k]
# 环境的向量表示
with tf.name_scope('context_express'):
self.bias_ctx_emb = tf.get_variable(
shape=[self.dim_k],
initializer=tf.zeros_initializer(),
dtype=tf.float32,
name='bias_ctx_emb')
self.Ctx_Emb = tf.matmul(
self.num_features,
self.W_Ctx) + self.bias_ctx_emb # [batch_size, dim_k]
self.Ctx_Emb = self._batch_norm_layer(self.Ctx_Emb,
self.train_phase, 'ctx_bn')
self.Ctx_Emb = tf.nn.relu(self.Ctx_Emb)
# 物品的向量表示
with tf.name_scope('item_express'):
self.I_Wds_a = tf.SparseTensor(
indices=tf.cast(self.item_words_indices_a, dtype=np.int64),
values=self.item_words_values_a,
dense_shape=[
tf.cast(self.batch_size, dtype=np.int64), self.num_words
])
self.att_u_a = tf.matmul(self.Usr_Expr_a,
self.Wu_oh_Att) # [batch_size, dim_att]
self.att_ctx = tf.matmul(self.Ctx_Emb, self.Wctx_Att)
self.att_oh = []
self.bias_wds_emb = tf.get_variable(
shape=[self.dim_k],
initializer=tf.zeros_initializer(),
dtype=tf.float32,
name='bias_wds_emb')
self.I_Wds_Emb_a = tf.sparse_tensor_dense_matmul(
self.I_Wds_a,
self.Wwords_Emb) + self.bias_wds_emb # [batch_size, dim_k]
self.I_Wds_Emb_a = tf.nn.relu(self.I_Wds_Emb_a)
self.att_I_Wds = tf.matmul(self.I_Wds_Emb_a,
self.WW_Att) # 词的attention
self.att_I_Wds = tf.nn.relu(self.att_u_a + self.att_ctx +
self.att_I_Wds + self.b_oh_Att)
self.att_I_Wds = tf.matmul(self.att_I_Wds,
self.w_oh_Att) + self.c_oh_Att
self.att_oh.append(self.att_I_Wds)
vae_encoder = VAE(input_dim=2048,
hidden_encoder_dim=1024,
latent_dim=96,
lam=0.001,
kld_loss=0.001)
self.I_visual_Emb, self.vae_loss = vae_encoder.get_vae_embbeding(
self.visual_emb_feat)
# TODO
self.I_visual_Emb = self._batch_norm_layer(self.I_visual_Emb,
self.train_phase,
'vis_bn')
self.bias_lda_emb = tf.get_variable(
shape=[self.dim_k],
initializer=tf.zeros_initializer(),
dtype=tf.float32,
name='bias_lda_emb')
self.I_LDA_Emb = tf.matmul(
self.words_lda,
self.W_LDA_emb) + self.bias_lda_emb # [batch_size, dim_k]
self.I_LDA_Emb = tf.nn.relu(self.I_LDA_Emb)
self.att_I_LDA = tf.matmul(self.I_LDA_Emb,
self.W_LDA_Att) # LDA的attention
self.att_I_LDA = tf.nn.relu(self.att_u_a + self.att_ctx +
self.att_I_LDA + self.b_oh_Att)
self.att_I_LDA = tf.matmul(self.att_I_LDA,
self.w_oh_Att) + self.c_oh_Att
self.att_oh.append(self.att_I_LDA)
self.I_One_hot_a = []
for i in range(len(self.one_hots_dims)):
I_Emb_temp_a = tf.nn.embedding_lookup(
self.W_one_hots[i],
tf.cast(self.one_hots_a[:, i],
tf.int32)) # [batch_size, dim_k]
att_oh_temp = tf.matmul(
I_Emb_temp_a, self.Woh_Att[i]) # [batch_size, att_dim_k]
att_temp = tf.nn.relu(self.att_u_a + self.att_ctx +
att_oh_temp +
self.b_oh_Att) # [batch_size, att_dim_k]
att_temp = tf.matmul(
att_temp, self.w_oh_Att) + self.c_oh_Att # [batch_size, 1]
self.att_oh.append(att_temp)
self.I_One_hot_a.append(I_Emb_temp_a)
self.att_oh = tf.nn.softmax(tf.concat(
self.att_oh, axis=1)) # [batch_size, oh_dim] 第一列是词attention
self.I_Wds_Emb_a = self.I_Wds_Emb_a * self.att_oh[:, 0:1]
self.I_LDA_Emb = self.I_LDA_Emb * self.att_oh[:, 1:2]
for i in range(2, len(self.one_hots_dims) + 2):
self.I_One_hot_a[
i - 2] = self.I_One_hot_a[i - 2] * self.att_oh[:, i:i + 1]
with tf.name_scope('cross'):
self.cross_inputs = [
self.Usr_Emb, self.Usr_Feat_Emb, self.num_features,
self.I_visual_Emb, self.I_Wds_Emb_a, self.I_LDA_Emb
] + self.I_One_hot_a
self.cross_x0 = tf.concat(self.cross_inputs,
axis=1) # [batch_size, self.dim_cross]
self.cross_x0 = self._batch_norm_layer(self.cross_x0,
self.train_phase,
'cross_bn')
self.dim_cross = self.cross_x0.get_shape().as_list()[1]
# self.cross_x0 = tf.reshape(self.cross_x0, [-1, self.dim_cross, 1])
self.cross_xl = self.cross_x0
for l in range(self.num_cross_layer):
self.cross_xl = self._cross_layer(self.cross_xl,
'cross_{}'.format(l))
self.cross_xl = tf.layers.dropout(inputs=self.cross_xl,
rate=self.dropout_emb)
self.cross_output = self.cross_xl
# self.cross_output = tf.reshape(self.cross_xl, [-1, self.dim_cross])
with tf.name_scope('deep'):
if self.use_deep:
self.I_one_hot_deep = []
for i in range(len(self.one_hots_dims)):
I_Emb_temp_a = tf.nn.embedding_lookup(
self.W_deep_one_hots[i],
tf.cast(self.one_hots_a[:, i],
tf.int32)) # [batch_size, dim_k]
self.I_one_hot_deep.append(I_Emb_temp_a)
#
self.deep_input = tf.concat(
[self.num_features, self.visual_emb_feat] +
self.I_one_hot_deep,
axis=1)
# 输入加入batch_norm
for i, deep_dim in enumerate(self.deep_dims):
if i == 0:
self.deep_input = tf.layers.dense(
inputs=self.deep_input,
kernel_initializer=tf.glorot_uniform_initializer(),
units=deep_dim,
activation=tf.nn.relu,
)
self.deep_input = self._batch_norm_layer(
self.deep_input, self.train_phase,
'deep_bn_{}'.format(i))
else:
self.deep_input = self._highway_layer(
self.deep_input, deep_dim,
'deep_highway_{}'.format(i))
# 加入dropout
self.deep_input = tf.layers.dropout(inputs=self.deep_input,
rate=self.dropout_deep)
self.deep_output = tf.layers.dense(self.deep_input,
1,
activation=None)
self.deep_output = tf.reshape(self.deep_output,
[-1]) # [batch_size]
with tf.name_scope('output'):
if self.use_deep:
self.concated = tf.concat([self.cross_output, self.deep_input],
axis=1)
else:
self.concated = self.cross_output
self.hidden = self.concated
for i, dim in enumerate(self.dim_hidden_out):
if i == 0:
self.hidden = tf.layers.dense(
inputs=self.hidden,
kernel_initializer=tf.glorot_uniform_initializer(),
units=dim,
activation=tf.nn.relu)
else:
self.hidden = self._highway_layer(
self.hidden, dim, 'out_highway_{}'.format(i))
self.hidden = tf.layers.dropout(inputs=self.hidden,
rate=self.dropout_deep)
self.bu = tf.nn.embedding_lookup(
self.bias_u, tf.cast(self.user_indices, dtype=tf.int32))
self.y_ui_a = tf.layers.dense(self.hidden, 1,
activation=None) + self.bu
self.y_ui_a = tf.reshape(tf.nn.sigmoid(self.y_ui_a), [-1])