View source on GitHub

RNN decoder for the sequence-to-sequence model.



  • decoder_inputs: A list of 2D Tensors [batch_size x input_size].
  • initial_state: 2D Tensor with shape [batch_size x cell.state_size].
  • cell: rnn_cell.RNNCell defining the cell function and size.
  • loop_function: If not None, this function will be applied to the i-th output in order to generate the i+1-st input, and decoder_inputs will be ignored, except for the first element ("GO" symbol). This can be used for decoding, but also for training to emulate Signature -- loop_function(prev, i) = next * prev is a 2D Tensor of shape [batch_size x output_size], * i is an integer, the step number (when advanced control is needed), * next is a 2D Tensor of shape [batch_size x input_size].
  • scope: VariableScope for the created subgraph; defaults to "rnn_decoder".


A tuple of the form (outputs, state), where: outputs: A list of the same length as decoder_inputs of 2D Tensors with shape [batch_size x output_size] containing generated outputs. state: The state of each cell at the final time-step. It is a 2D Tensor of shape [batch_size x cell.state_size]. (Note that in some cases, like basic RNN cell or GRU cell, outputs and states can be the same. They are different for LSTM cells though.)