4ce9a9d078d41af1a960f0e5bf16e373f69d5117,open_seq2seq/encoders/rnn_encoders.py,GNMTLikeEncoderWithEmbedding,_encode,#GNMTLikeEncoderWithEmbedding#Any#,344

Before Change

    if self.params["encoder_layers"] < 2:
      raise ValueError("GNMT encoder must have at least 2 layers")

    cell_params = copy.deepcopy(self.params)
    cell_params["num_units"] = self.params["encoder_cell_units"]

    with tf.variable_scope("Level1FW"):
      self._encoder_l1_cell_fw = create_rnn_cell(
        cell_type=self.params["encoder_cell_type"],
        cell_params=cell_params,
        num_layers=1,
        dp_input_keep_prob=1.0,
        dp_output_keep_prob=1.0,
        residual_connections=False,
      )
    with tf.variable_scope("Level1BW"):
      self._encoder_l1_cell_bw = create_rnn_cell(
        cell_type=self.params["encoder_cell_type"],
        cell_params=cell_params,
        num_layers=1,
        dp_input_keep_prob=1.0,
        dp_output_keep_prob=1.0,
        residual_connections=False,
      )

    if self._mode == "train":
      dp_input_keep_prob = self.params["encoder_dp_input_keep_prob"]
      dp_output_keep_prob = self.params["encoder_dp_output_keep_prob"]

After Change

      dp_output_keep_prob = 1.0

    with tf.variable_scope("UniDirLevel"):
      self._encoder_cells = [single_cell(
        cell_class=self.params["core_cell"],
        cell_params=self.params.get("core_cell_params", {}),
        dp_input_keep_prob=dp_input_keep_prob,
        dp_output_keep_prob=dp_output_keep_prob,
        residual_connections=False) for _ in range(self.params["encoder_layers"] - 1)]

      // add residual connections starting from the third layer
      for idx, cell in enumerate(self._encoder_cells):
        if idx > 0:

Instances