Caffe2 - Python API
A deep learning, cross platform ML framework
Public Member Functions | Public Attributes | List of all members
caffe2.python.rnn_cell.AttentionCell Class Reference
Inheritance diagram for caffe2.python.rnn_cell.AttentionCell:
caffe2.python.rnn_cell.RNNCell caffe2.python.rnn_cell.LSTMWithAttentionCell caffe2.python.rnn_cell.MILSTMWithAttentionCell

Public Member Functions

def __init__ (self, encoder_output_dim, encoder_outputs, encoder_lengths, decoder_cell, decoder_state_dim, attention_type, weighted_encoder_outputs, attention_memory_optimization, kwargs)
 
def get_attention_weights (self)
 
def prepare_input (self, model, input_blob)
 
def build_initial_coverage (self, model)
 
def get_state_names (self)
 
def get_output_dim (self)
 
def get_output_state_index (self)
 
- Public Member Functions inherited from caffe2.python.rnn_cell.RNNCell
def __init__ (self, name=None, forward_only=False, initializer=None)
 
def initializer (self)
 
def initializer (self, value)
 
def scope (self, name)
 
def apply_over_sequence (self, model, inputs, seq_lengths=None, initial_states=None, outputs_with_grads=None)
 
def apply (self, model, input_t, seq_lengths, states, timestep)
 
def apply_override (self, model, input_t, seq_lengths, timestep, extra_inputs=None)
 
def prepare_input (self, model, input_blob)
 
def get_output_state_index (self)
 
def get_state_names (self)
 
def get_state_names_override (self)
 
def get_output_dim (self)
 

Public Attributes

 encoder_output_dim
 
 encoder_outputs
 
 encoder_lengths
 
 decoder_cell
 
 decoder_state_dim
 
 weighted_encoder_outputs
 
 encoder_outputs_transposed
 
 attention_type
 
 attention_memory_optimization
 
 hidden_t_intermediate
 
 coverage_weights
 
- Public Attributes inherited from caffe2.python.rnn_cell.RNNCell
 name
 
 recompute_blobs
 
 forward_only
 

Detailed Description

Definition at line 1122 of file rnn_cell.py.

Member Function Documentation

def caffe2.python.rnn_cell.AttentionCell.build_initial_coverage (   self,
  model 
)
initial_coverage is always zeros of shape [encoder_length],
which shape must be determined programmatically dureing network
computation.

This method also sets self.coverage_weights, a separate transform
of encoder_outputs which is used to determine coverage contribution
tp attention.

Definition at line 1301 of file rnn_cell.py.


The documentation for this class was generated from the following file: