Python attention_decoder.attention_decoder方法代码示例

# 需要导入模块: import attention_decoder [as 别名]
# 或者: from attention_decoder import attention_decoder [as 别名]
def _add_decoder(self, inputs):
        """Add attention decoder to the graph. In train or eval mode, you call this once to get output on ALL steps. In decode (beam search) mode, you call this once for EACH decoder step.

    Args:
      inputs: inputs to the decoder (word embeddings). A list of tensors shape (batch_size, emb_dim)

    Returns:
      outputs: List of tensors; the outputs of the decoder
      out_state: The final state of the decoder
      attn_dists: A list of tensors; the attention distributions
      p_gens: A list of scalar tensors; the generation probabilities
      coverage: A tensor, the current coverage vector
    """
        hps = self._hps
        cell = tf.contrib.rnn.LSTMCell(hps.hidden_dim, state_is_tuple=True, initializer=self.rand_unif_init)

        prev_coverage = self.prev_coverage if hps.mode == "decode" and hps.coverage else None  # In decode mode, we run attention_decoder one step at a time and so need to pass in the previous step's coverage vector each time

        outputs, out_state, attn_dists, p_gens, coverage = attention_decoder(inputs, self._dec_in_state,
                                                                             self._enc_states, self._enc_padding_mask,
                                                                             cell, initial_state_attention=(
                    hps.mode == "decode"), pointer_gen=hps.pointer_gen, use_coverage=hps.coverage,
                                                                             prev_coverage=prev_coverage)

        return outputs, out_state, attn_dists, p_gens, coverage 

# 需要导入模块: import attention_decoder [as 别名]
# 或者: from attention_decoder import attention_decoder [as 别名]
def _add_decoder(self, inputs):
    """Add attention decoder to the graph. In train or eval mode, you call this once to get output on ALL steps. In decode (beam search) mode, you call this once for EACH decoder step.

    Args:
      inputs: inputs to the decoder (word embeddings). A list of tensors shape (batch_size, emb_dim)

    Returns:
      outputs: List of tensors; the outputs of the decoder
      out_state: The final state of the decoder
      attn_dists: A list of tensors; the attention distributions
      p_gens: A list of tensors shape (batch_size, 1); the generation probabilities
      coverage: A tensor, the current coverage vector
    """
    hps = self._hps
    cell = tf.contrib.rnn.LSTMCell(hps.hidden_dim, state_is_tuple=True, initializer=self.rand_unif_init)

    prev_coverage = self.prev_coverage if hps.mode=="decode" and hps.coverage else None # In decode mode, we run attention_decoder one step at a time and so need to pass in the previous step's coverage vector each time

    outputs, out_state, attn_dists, p_gens, coverage = attention_decoder(inputs, self._dec_in_state, self._enc_states, self._enc_padding_mask, cell, initial_state_attention=(hps.mode=="decode"), pointer_gen=hps.pointer_gen, use_coverage=hps.coverage, prev_coverage=prev_coverage)

    return outputs, out_state, attn_dists, p_gens, coverage 

# 需要导入模块: import attention_decoder [as 别名]
# 或者: from attention_decoder import attention_decoder [as 别名]
def _add_decoder(self, emb_dec_inputs, embedding):
    """Add attention decoder to the graph. In train or eval mode, you call this once to get output on ALL steps. In decode (beam search) mode, you call this once for EACH decoder step.

    Args:
      emb_dec_inputs: inputs to the decoder (word embeddings). A list of tensors shape (batch_size, emb_dim)
      embedding: embedding matrix (vocab_size, emb_dim)
    Returns:
      outputs: List of tensors; the outputs of the decoder
      out_state: The final state of the decoder
      attn_dists: A list of tensors; the attention distributions
      p_gens: A list of tensors shape (batch_size, 1); the generation probabilities
      coverage: A tensor, the current coverage vector
    """
    hps = self._hps
    cell = tf.contrib.rnn.LSTMCell(hps.dec_hidden_dim, state_is_tuple=True, initializer=self.rand_unif_init)

    prev_coverage = self.prev_coverage if (hps.mode=="decode" and hps.coverage) else None # In decode mode, we run attention_decoder one step at a time and so need to pass in the previous step's coverage vector each time
    prev_decoder_outputs = self.prev_decoder_outputs if (hps.intradecoder and hps.mode=="decode") else tf.stack([],axis=0)
    prev_encoder_es = self.prev_encoder_es if (hps.use_temporal_attention and hps.mode=="decode") else tf.stack([],axis=0)
    return attention_decoder(_hps=hps,
      v_size=self._vocab.size(),
      _max_art_oovs=self._max_art_oovs,
      _enc_batch_extend_vocab=self._enc_batch_extend_vocab,
      emb_dec_inputs=emb_dec_inputs,
      target_batch=self._target_batch,
      _dec_in_state=self._dec_in_state,
      _enc_states=self._enc_states,
      enc_padding_mask=self._enc_padding_mask,
      dec_padding_mask=self._dec_padding_mask,
      cell=cell,
      embedding=embedding,
      sampling_probability=self._sampling_probability if FLAGS.scheduled_sampling else 0,
      alpha=self._alpha if FLAGS.E2EBackProp else 0,
      unk_id=self._vocab.word2id(data.UNKNOWN_TOKEN),
      initial_state_attention=(hps.mode=="decode"),
      pointer_gen=hps.pointer_gen,
      use_coverage=hps.coverage,
      prev_coverage=prev_coverage,
      prev_decoder_outputs=prev_decoder_outputs,
      prev_encoder_es = prev_encoder_es) 

# 需要导入模块: import attention_decoder [as 别名]
# 或者: from attention_decoder import attention_decoder [as 别名]
def _add_decoder(self, emb_dec_inputs, embedding):
    """Add attention decoder to the graph. In train or eval mode, you call this once to get output on ALL steps. In decode (beam search) mode, you call this once for EACH decoder step.

    Args:
      emb_dec_inputs: inputs to the decoder (word embeddings). A list of tensors shape (batch_size, emb_dim)
      embedding: embedding matrix (vocab_size, emb_dim)
    Returns:
      outputs: List of tensors; the outputs of the decoder
      out_state: The final state of the decoder
      attn_dists: A list of tensors; the attention distributions
      p_gens: A list of tensors shape (batch_size, 1); the generation probabilities
      coverage: A tensor, the current coverage vector
    """
    hps = self._hps
    cell = tf.contrib.rnn.LSTMCell(hps.dec_hidden_dim, state_is_tuple=True, initializer=self.rand_unif_init)

    prev_coverage = self.prev_coverage if (hps.mode=="decode" and hps.coverage) else None # In decode mode, we run attention_decoder one step at a time and so need to pass in the previous step's coverage vector each time
    prev_decoder_outputs = self.prev_decoder_outputs if (hps.intradecoder and hps.mode=="decode") else tf.stack([],axis=0)
    prev_encoder_es = self.prev_encoder_es if (hps.use_temporal_attention and hps.mode=="decode") else tf.stack([],axis=0)
    return attention_decoder(hps,
      self._vocab.size(),
      self._max_art_oovs,
      self._enc_batch_extend_vocab,
      emb_dec_inputs,
      self._target_batch,
      self._dec_in_state,
      self._enc_states,
      self._enc_padding_mask,
      self._dec_padding_mask,
      cell,
      embedding,
      self._sampling_probability if FLAGS.scheduled_sampling else 0,
      self._alpha if FLAGS.E2EBackProp else 0,
      self._vocab.word2id(data.UNKNOWN_TOKEN),
      initial_state_attention=(hps.mode=="decode"),
      pointer_gen=hps.pointer_gen,
      use_coverage=hps.coverage,
      prev_coverage=prev_coverage,
      prev_decoder_outputs=prev_decoder_outputs,
      prev_encoder_es = prev_encoder_es) 

# 需要导入模块: import attention_decoder [as 别名]
# 或者: from attention_decoder import attention_decoder [as 别名]
def _add_decoder(self, inputs):
        """Add attention decoder to the graph. In train or eval mode, you call this once to get output on ALL steps. In decode (beam search) mode, you call this once for EACH decoder step.

        Args:
          inputs: inputs to the decoder (word embeddings). A list of tensors shape (batch_size, emb_dim)

        Returns:
          outputs: List of tensors; the outputs of the decoder
          out_state: The final state of the decoder
          attn_dists: A list of tensors; the attention distributions
          p_gens: A list of scalar tensors; the generation probabilities
          coverage: A tensor, the current coverage vector
        """
        hps = self._hps
        if self._isrum in ['all', 'dec']:
            cell = RUMCell(
                self._hps.hidden_dim,
                eta_=self._time_norm,
                lambda_=self._lambda,
                update_gate=self._update_gate,
                use_layer_norm=self._layer_norm,
                use_zoneout=self._zoneout
            )
        else:
            cell = tf.contrib.rnn.LSTMCell(
                hps.hidden_dim, state_is_tuple=True, initializer=self.rand_unif_init)

        # In decode mode, we run attention_decoder one step at a time and so
        # need to pass in the previous step's coverage vector each time
        prev_coverage = self.prev_coverage if hps.mode == "decode" and hps.coverage else None

        outputs, out_state, attn_dists, p_gens, coverage = attention_decoder(inputs, self._dec_in_state, self._enc_states, self._enc_padding_mask, cell, initial_state_attention=(
            hps.mode == "decode"), pointer_gen=hps.pointer_gen, use_coverage=hps.coverage, prev_coverage=prev_coverage, isrum=self._isrum, lambda_=self._lambda)

        if self._isgrad:
            if self._isrum in ['none', 'enc']:
                tf.summary.scalar('out_state_c', tf.norm(out_state.c))
                tf.summary.scalar('out_state_h', tf.norm(out_state.h))
            else:
                tf.summary.scalar('out_state', tf.norm(out_state))

        return outputs, out_state, attn_dists, p_gens, coverage