Python seq2seq.GreedyEmbeddingHelper方法代码示例

# 需要导入模块: from tensorflow.contrib import seq2seq [as 别名]
# 或者: from tensorflow.contrib.seq2seq import GreedyEmbeddingHelper [as 别名]
def next_inputs(self, time, outputs, state, stop_id, name=None):
        """next_inputs_fn for GreedyEmbeddingHelper."""
        del time  # unused by next_inputs_fn
        finished = math_ops.equal(stop_id, 1)  # 1 is stop signal
        all_finished = math_ops.reduce_all(finished)
        next_inputs = control_flow_ops.cond(
            all_finished,
            # If we're finished, the next_inputs value doesn't matter
            lambda: self._start_inputs,
            lambda: outputs)
        return (finished, next_inputs, state) 

# 需要导入模块: from tensorflow.contrib import seq2seq [as 别名]
# 或者: from tensorflow.contrib.seq2seq import GreedyEmbeddingHelper [as 别名]
def _build_decoder_test_greedy(self):
        r"""
        Builds the greedy test decoder, which feeds the most likely decoded symbol as input for the
        next timestep
        """
        self._helper_greedy = seq2seq.GreedyEmbeddingHelper(
            embedding=self._embedding_matrix,
            start_tokens=tf.tile([self._GO_ID], [self._batch_size]),
            end_token=self._EOS_ID)

        if self._hparams.enable_attention is True:
            cells, initial_state = add_attention(
                cells=self._decoder_cells,
                attention_types=self._hparams.attention_type[1],
                num_units=self._hparams.decoder_units_per_layer[-1],
                memory=self._encoder_memory,
                memory_len=self._encoder_features_len,
                beam_search=False,
                batch_size=self._batch_size,
                initial_state=self._decoder_initial_state,
                mode=self._mode,
                dtype=self._hparams.dtype,
                fusion_type='linear_fusion',
                write_attention_alignment=self._hparams.write_attention_alignment)
        else:
            cells = self._decoder_cells
            initial_state = self._decoder_initial_state

        self._decoder_inference = seq2seq.BasicDecoder(
            cell=cells,
            helper=self._helper_greedy,
            initial_state=initial_state,
            output_layer=self._dense_layer)

        outputs, states, lengths = seq2seq.dynamic_decode(
            self._decoder_inference,
            impute_finished=True,
            swap_memory=False,
            maximum_iterations=self._hparams.max_label_length)

        self.inference_outputs = outputs.rnn_output
        self.inference_predicted_ids = outputs.sample_id

        if self._hparams.write_attention_alignment is True:
            self.attention_summary = self._create_attention_alignments_summary(states) 

# 需要导入模块: from tensorflow.contrib import seq2seq [as 别名]
# 或者: from tensorflow.contrib.seq2seq import GreedyEmbeddingHelper [as 别名]
def _build_decoder_greedy(self):

        batch_size, _ = tf.unstack(tf.shape(self._labels))
        self._helper_greedy = seq2seq.GreedyEmbeddingHelper(
            embedding=self._embedding_matrix,
            start_tokens=tf.tile([self._GO_ID], [batch_size]),
            end_token=self._EOS_ID)

        if self._hparams.enable_attention is True:
            attention_mechanisms, layer_sizes = self._create_attention_mechanisms()

            attention_cells = seq2seq.AttentionWrapper(
                cell=self._decoder_cells,
                attention_mechanism=attention_mechanisms,
                attention_layer_size=layer_sizes,
                initial_cell_state=self._decoder_initial_state,
                alignment_history=self._hparams.write_attention_alignment,
                output_attention=self._output_attention
            )
            attn_zero = attention_cells.zero_state(
                dtype=self._hparams.dtype, batch_size=batch_size
            )
            initial_state = attn_zero.clone(
                cell_state=self._decoder_initial_state
            )
            cells = attention_cells
        else:
            cells = self._decoder_cells
            initial_state = self._decoder_initial_state

        self._decoder_inference = seq2seq.BasicDecoder(
            cell=cells,
            helper=self._helper_greedy,
            initial_state=initial_state,
            output_layer=self._dense_layer)

        outputs, states, lengths = seq2seq.dynamic_decode(
            self._decoder_inference,
            impute_finished=True,
            swap_memory=False,
            maximum_iterations=self._hparams.max_label_length)

        # self._result = outputs, states, lengths
        self.inference_outputs = outputs.rnn_output
        self.inference_predicted_ids = outputs.sample_id

        if self._hparams.write_attention_alignment is True:
            self.attention_summary = self._create_attention_alignments_summary(states) 

# 需要导入模块: from tensorflow.contrib import seq2seq [as 别名]
# 或者: from tensorflow.contrib.seq2seq import GreedyEmbeddingHelper [as 别名]
def _build_model(self):
        with tf.variable_scope("embeddings"):
            self.source_embs = tf.get_variable(name="source_embs", shape=[self.cfg.source_vocab_size, self.cfg.emb_dim],
                                               dtype=tf.float32, trainable=True)
            self.target_embs = tf.get_variable(name="embeddings", shape=[self.cfg.vocab_size, self.cfg.emb_dim],
                                               dtype=tf.float32, trainable=True)
            source_emb = tf.nn.embedding_lookup(self.source_embs, self.enc_source)
            target_emb = tf.nn.embedding_lookup(self.target_embs, self.dec_target_in)
            print("source embedding shape: {}".format(source_emb.get_shape().as_list()))
            print("target input embedding shape: {}".format(target_emb.get_shape().as_list()))

        with tf.variable_scope("encoder"):
            if self.cfg.use_bi_rnn:
                with tf.variable_scope("bi-directional_rnn"):
                    cell_fw = GRUCell(self.cfg.num_units) if self.cfg.cell_type == "gru" else \
                        LSTMCell(self.cfg.num_units)
                    cell_bw = GRUCell(self.cfg.num_units) if self.cfg.cell_type == "gru" else \
                        LSTMCell(self.cfg.num_units)
                    bi_outputs, _ = bidirectional_dynamic_rnn(cell_fw, cell_bw, source_emb, dtype=tf.float32,
                                                              sequence_length=self.enc_seq_len)
                    source_emb = tf.concat(bi_outputs, axis=-1)
                    print("bi-directional rnn output shape: {}".format(source_emb.get_shape().as_list()))
            input_project = tf.layers.Dense(units=self.cfg.num_units, dtype=tf.float32, name="input_projection")
            source_emb = input_project(source_emb)
            print("encoder input projection shape: {}".format(source_emb.get_shape().as_list()))
            enc_cells = self._create_encoder_cell()
            self.enc_outputs, self.enc_states = dynamic_rnn(enc_cells, source_emb, sequence_length=self.enc_seq_len,
                                                            dtype=tf.float32)
            print("encoder output shape: {}".format(self.enc_outputs.get_shape().as_list()))

        with tf.variable_scope("decoder"):
            self.max_dec_seq_len = tf.reduce_max(self.dec_seq_len, name="max_dec_seq_len")
            self.dec_cells, self.dec_init_states = self._create_decoder_cell()
            # define input and output projection layer
            input_project = tf.layers.Dense(units=self.cfg.num_units, name="input_projection")
            self.dense_layer = tf.layers.Dense(units=self.cfg.vocab_size, name="output_projection")
            if self.mode == "train":  # either "train" or "decode"
                # for training
                target_emb = input_project(target_emb)
                train_helper = TrainingHelper(target_emb, sequence_length=self.dec_seq_len, name="train_helper")
                train_decoder = BasicDecoder(self.dec_cells, helper=train_helper, output_layer=self.dense_layer,
                                             initial_state=self.dec_init_states)
                self.dec_output, _, _ = dynamic_decode(train_decoder, impute_finished=True,
                                                       maximum_iterations=self.max_dec_seq_len)
                print("decoder output shape: {} (vocab size)".format(self.dec_output.rnn_output.get_shape().as_list()))

                # for decode
                start_token = tf.ones(shape=[self.batch_size, ], dtype=tf.int32) * self.cfg.target_dict[GO]
                end_token = self.cfg.target_dict[EOS]

                def inputs_project(inputs):
                    return input_project(tf.nn.embedding_lookup(self.target_embs, inputs))

                dec_helper = GreedyEmbeddingHelper(embedding=inputs_project, start_tokens=start_token,
                                                   end_token=end_token)
                infer_decoder = BasicDecoder(self.dec_cells, helper=dec_helper, initial_state=self.dec_init_states,
                                             output_layer=self.dense_layer)
                infer_dec_output, _, _ = dynamic_decode(infer_decoder, maximum_iterations=self.cfg.maximum_iterations)
                self.dec_predicts = infer_dec_output.sample_id