def _build_rnn_inputs(self, word_embedder, time_major):
inputs = word_embedder.embed(self.inputs, zero_pad=True)
self.word_embeddings = inputs
if not time_major:
inputs = transpose_first_two_dims(
inputs) # (seq_len, batch_size, input_size)
return inputs
def _build_output(self, output_dict):
'''
Take RNN outputs and produce logits over the vocab and the attentions.
'''
logits = super(CopyGraphDecoder, self)._build_output(output_dict) # (batch_size, seq_len, num_symbols)
attn_scores = transpose_first_two_dims(output_dict['attn_scores']) # (batch_size, seq_len, num_nodes)
return tf.concat(2, [logits, attn_scores])
def _build_output(self, output_dict):
'''
Take RNN outputs and produce logits over the vocab.
'''
outputs = output_dict['outputs']
outputs = transpose_first_two_dims(outputs) # (batch_size, seq_len, output_size)
logits = batch_linear(outputs, self.num_symbols, True)
#logits = BasicDecoder.penalize_repetition(logits)
return logits
def _build_rnn_inputs(self, word_embedder, time_major):
inputs = super(GraphDecoder, self)._build_rnn_inputs(word_embedder, time_major)
checklists = tf.cumsum(tf.one_hot(self.entities, self.num_nodes, on_value=1, off_value=0), axis=1) + self.init_checklists
# cumsum can cause >1 indicator
checklists = tf.cast(tf.greater(checklists, 0), tf.float32)
self.output_dict['checklists'] = checklists
checklists = transpose_first_two_dims(checklists) # (seq_len, batch_size, num_nodes)
return inputs, checklists
def _build_rnn_inputs(self, word_embedder, time_major):
'''
Concatenate word embedding with entity/node embedding.
'''
word_embeddings = word_embedder.embed(self.inputs, zero_pad=True)
self.word_embeddings = word_embeddings
if self.node_embed_in_rnn_inputs:
entity_embeddings = self._get_node_embedding(self.context[0], self.entities)
inputs = tf.concat(2, [word_embeddings, entity_embeddings])
else:
inputs = word_embeddings
if not time_major:
inputs = transpose_first_two_dims(inputs) # (seq_len, batch_size, input_size)
return inputs
def _get_final_state(self, states):
'''
Return the final non-pad state from tf.scan outputs.
'''
with tf.name_scope(type(self).__name__+'/get_final_state'):
flat_states = nest.flatten(states)
flat_last_states = []
for state in flat_states:
state = transpose_first_two_dims(state) # (batch_size, time_seq, state_size)
# NOTE: when state has dim=4, it's the context which does not change in a seq; just take the last one.
if len(state.get_shape()) == 4:
last_state = state[:, -1, :, :]
else:
last_state = tf.squeeze(batch_embedding_lookup(state, tf.reshape(self.last_inds, [-1, 1])), [1])
flat_last_states.append(last_state)
last_states = nest.pack_sequence_as(states, flat_last_states)
return last_states
