def build_policy(self):
dilate_input = self._expand_model_graph(self.input, train=True)
dilate_input = tf.reduce_mean(dilate_input, axis=1)
logits = tf.matmul(dilate_input, self.softmax_w, transpose_b=True)
logits = tf.nn.bias_add(logits, self.output_bias)
logits = tf.reshape(logits, [-1, self.action_num, 2])
logits = tf.sigmoid(logits)
self.logits_check = tf.nn.softmax(logits)
action = gumbel_softmax_(logits, hard=True, temperature=self.temp)
self.action_predict = action[:, :, 0]
def build_policy(self, input, train):
tf.get_variable_scope().reuse_variables()
dilate_input = self._hidden_forward(input, train=train)
dilate_input = tf.reduce_mean(dilate_input, axis=1)
logits = tf.matmul(dilate_input, self.softmax_w, transpose_b=True)
logits = tf.nn.bias_add(logits, self.output_bias)
logits = tf.reshape(logits, [-1, self.action_num, 2])
logits = tf.sigmoid(logits)
logits_check = tf.nn.softmax(logits)
action = gumbel_softmax_(logits, hard=True, temperature=self.temp)
action_predict = action[:, :, 0]
return logits_check, action_predict
def build_policy(self):
context_seq = self.input[:, 0:-1] # [B, S]
dilate_input = tf.nn.embedding_lookup(
self.item_embedding, context_seq, name="context_embedding"
) # [B, S, C]
dilate_input = self.gru(dilate_input) # [B, C]
logits = tf.matmul(dilate_input, self.softmax_w, transpose_b=True)
logits = tf.nn.bias_add(logits, self.output_bias)
logits = tf.reshape(logits, [-1, self.action_num, 2])
logits = tf.sigmoid(logits)
self.logits_check = tf.nn.softmax(logits)
action = gumbel_softmax_(logits, hard=True, temperature=self.temp)
self.action_predict = action[:, :, 0]
def build_policy(self, context, train):
hidden = self._expand_model_graph(context, train=train) # # [B, L, C]
hidden = tf.reduce_mean(hidden, axis=1) # [B, C]
logits = tf.matmul(hidden, self.softmax_w,
transpose_b=True) # [B, C] * [C, A*2]=[B,#block*2]
logits = tf.nn.bias_add(logits, self.output_bias) # [B,#block*2]
logits = tf.reshape(logits, [-1, self.action_num, 2]) # [B, #block, 2]
if self.hard_policy:
logits = tf.sigmoid(logits)
action = gumbel_softmax_(logits, temperature=self.temp,
hard=True) # [B, #block, 2]
else: # soft policy
action = tf.sigmoid(logits) # [B, #block, 2]
logits_check = tf.nn.softmax(logits)
action_predict = action[:, :, 0] # [B, #block]
return logits_check, action_predict
def build_policy(self, context):
hidden = tf.nn.embedding_lookup(self.item_embedding,
context,
name="context_embedding") # [B, L, C]
hidden = self.gru(hidden) # [B, C]
logits = tf.matmul(hidden, self.softmax_w,
transpose_b=True) # [B, C] * [C, A*2]=[B,#block*2]
logits = tf.nn.bias_add(logits, self.output_bias) # [B,#block*2]
logits = tf.reshape(logits, [-1, self.action_num, 2]) # [B, #block, 2]
if self.hard_policy:
logits = tf.sigmoid(logits)
action = gumbel_softmax_(logits, temperature=self.temp,
hard=True) # [B, #block, 2]
else: # soft policy
action = tf.sigmoid(logits) # [B, #block, 2]
logits_check = tf.nn.softmax(logits)
action_predict = action[:, :, 0] # [B, #block]
return logits_check, action_predict