def feed_previous_decode(feed_previous_bool):
weights = tf.matmul(self.dec_weights,
self.type_embedding,
transpose_b=True)
biases = tf.matmul(tf.expand_dims(self.dec_biases, 0),
self.type_embedding,
transpose_b=True)
biases = array_ops.reshape(biases, [self.num_types])
dec_embed, loop_fn = seq2seq.get_decoder_embedding(
self.decoder_input,
self.num_types,
self.num_type_dim,
output_projection=(weights, biases),
feed_previous=feed_previous_bool)
#dec_embed is of dimension self.max_steps_label * batch_size * self.num_type_dim
concatenated_input = self.dec_concat_ip(
dec_embed, sentence_states, sentence_mention_states)
#concat_dec_inputs is a list of self.max_steps_label tensors of dimension batch_size * (self.num_type_dim + self.hidden_size)
dec_output, _ = self.decode(
concatenated_input, loop_fn, init_state, sentence_states,
sentence_outputs, sentence_mention_states,
sentence_mention_outputs, scope, attn_len, attn_size,
attn_t_given_context_per_level,
attn_t_given_mention_per_level, initial_state_attention)
return dec_output
def feed_previous_decode(feed_previous_bool):
dec_embed, loop_fn = seq2seq.get_decoder_embedding(decoder_inputs, self.decoder_words,
self.text_embedding_size,
output_projection=(weights, biases),
feed_previous=feed_previous_bool)
# dec_embed is of dimension max_len * batch_size * self.text_embedding_size
# utterance_output is of dimension batch_size * cell_size
concatenated_input = self.get_dec_concat_ip(dec_embed, utterance_output)
dec_output, _ = self.decode(concatenated_input, loop_fn, self.dec_cells_text, init_state,
utterance_output, scope, attention_states)
# dec_output is a max_len sized list of tensors of dimension batch_size * cell_size
return dec_output
def feed_previous_decode(feed_previous_bool):
dec_embed_word, loop_fn = seq2seq.get_decoder_embedding(
decoder_inputs,
self.decoder_words,
self.text_embedding_size,
output_projection=(weight_word, bias_word),
feed_previous=feed_previous_bool)
concatenated_input_word = self.get_dec_concat_ip(
dec_embed_word, utterance_outputs)
dec_output_word, _ = self.decode(concatenated_input_word,
loop_fn, self.dec_cells_text,
init_state, utterance_outputs,
scope)
return dec_output_word
def feed_prev_decode(feed_previous_bool):
'''Makes two seperate graphs based on feed_previous input given at training and test time.
Args:
feed_previous_bool: Boolean tensor which is True at time of validation and testing and False at time of training.
Return: dec_output which is a list having tensor of size batch_size*self.cell_size.'''
dec_embed, loop_fn = seq2seq.get_decoder_embedding(
decoder_inputs,
self.decoder_words,
self.embedding_size,
output_projection=(weights, biases),
feed_previous=feed_previous_bool
) #look for get_decoder_embedding in seq2seq.py
concatenated_ip = self.get_dec_concat_ip(
dec_embed, utterance_output)
dec_output, _ = self.decode(concatenated_ip, loop_fn, dec_cell,
init_state, utterance_output,
scope)
return dec_output