def build_train_graph(self, policy_action):
# tf.get_variable_scope().reuse_variables()
label_seq, dilate_input = self._expand_model_graph(self.input_train,
policy_action,
train=True)
if self.using_negative_sampling:
logits_2d = tf.reshape(dilate_input, [-1, self.dilated_channels])
label_flat = tf.reshape(label_seq, [-1, 1]) # [B*(SessLen-1), 1]
num_sampled = int(self.negative_sampling_ratio * self.item_size)
loss = tf.nn.sampled_softmax_loss(
self.softmax_w,
self.softmax_b,
labels=label_flat,
inputs=logits_2d,
num_sampled=num_sampled,
num_classes=self.item_size,
) # [B*(SessLen-1), 1]
else:
logits = conv1d(tf.nn.relu(dilate_input),
output_channels=self.item_size,
name="logits") # [B,SessLen-1,ItemSize]
logits_2d = tf.reshape(
logits, [-1, self.item_size]) # [B*(SessLen-1),ItemSize]
label_flat = tf.reshape(label_seq, [-1]) # [B*(SessLen-1), 1]
loss = tf.nn.sparse_softmax_cross_entropy_with_logits(
labels=label_flat, logits=logits_2d) # [B*(SessLen-1), 1]
self.loss = tf.reduce_mean(loss)
def build_test_graph(self, source_hidden):
tf.get_variable_scope().reuse_variables()
logits = conv1d(tf.nn.relu(source_hidden[:, -1:, :]),
self.target_item_size,
name="logits")
logits_2d = tf.reshape(logits, [-1, self.target_item_size])
self.test_probs = tf.nn.softmax(logits_2d)
def build_train_graph(self, source_hidden):
# source_hidden: [B, L, C]
logits = conv1d(tf.nn.relu(source_hidden[:, -1:, :]),
output_channels=self.target_item_size,
name="logits")
logits_2d = tf.reshape(logits, [-1, self.target_item_size])
label_flat = tf.reshape(self.target, [-1])
loss = tf.nn.sparse_softmax_cross_entropy_with_logits(
labels=label_flat, logits=logits_2d)
self.train_loss = tf.reduce_mean(loss)
def _get_test_result(self, dilate_input, label_seq):
if self.using_negative_sampling:
logits_2d = tf.reshape(dilate_input[:, -1:, :],
[-1, self.dilated_channels])
logits_2d = tf.matmul(logits_2d, tf.transpose(self.softmax_w))
logits_2d = tf.nn.bias_add(logits_2d,
self.softmax_b) # [B, ItemSize]
else:
logits = conv1d(tf.nn.relu(dilate_input[:, -1:, :]),
self.item_size,
name="logits")
logits_2d = tf.reshape(logits, [-1, self.item_size])
label_flat = tf.reshape(label_seq[:, -1], [-1]) # [B,]
loss = tf.nn.sparse_softmax_cross_entropy_with_logits(
labels=label_flat, logits=logits_2d) # [B,]
loss = tf.reduce_mean(loss)
probs = tf.nn.softmax(logits_2d) # [B, ItemSize]
return loss, probs
def build_train_graph(self):
label_seq, dilate_input = self._expand_model_graph(self.input_train, train=True)
# label_seq: [B, SessLen - 1]
# dilate_input: [B, SessLen - 1, DilatedChannels]
if self.using_negative_sampling:
logits_2d = tf.reshape(
dilate_input, [-1, self.channels]
) # [B*(SessLen-1), DilatedChannels]
label_flat = tf.reshape(label_seq, [-1, 1]) # [B*(SessLen-1), 1]
num_sampled = int(self.negative_sampling_ratio * self.item_size)
loss = tf.nn.sampled_softmax_loss(
self.softmax_w,
self.softmax_b,
labels=label_flat,
inputs=logits_2d,
num_sampled=num_sampled,
num_classes=self.item_size,
) # [B*(SessLen-1), 1]
else:
logits = conv1d(
tf.nn.relu(dilate_input), output_channels=self.item_size, name="logits"
) # [B,SessLen-1,ItemSize]
logits_2d = tf.reshape(
logits, [-1, self.item_size]
) # [B*(SessLen-1),ItemSize]
label_flat = tf.reshape(label_seq, [-1]) # [B*(SessLen-1), 1]
loss = tf.nn.sparse_softmax_cross_entropy_with_logits(
labels=label_flat, logits=logits_2d
) # [B*(SessLen-1), 1]
loss_train = tf.reduce_mean(loss) # output of training steps, [1]
return loss_train