def __init__(
self,
memory,
memory_sequence_length=None,
cell=None,
cell_dropout_mode=None,
vocab_size=None,
output_layer=None,
#attention_layer=None, # TODO(zhiting): only valid for tf>=1.0
cell_input_fn=None,
hparams=None):
RNNDecoderBase.__init__(self, cell, vocab_size, output_layer,
cell_dropout_mode, hparams)
attn_hparams = self._hparams['attention']
attn_kwargs = attn_hparams['kwargs'].todict()
# Parse the 'probability_fn' argument
if 'probability_fn' in attn_kwargs:
prob_fn = attn_kwargs['probability_fn']
if prob_fn is not None and not callable(prob_fn):
prob_fn = utils.get_function(prob_fn, [
'tensorflow.nn', 'tensorflow.contrib.sparsemax',
'tensorflow.contrib.seq2seq'
])
attn_kwargs['probability_fn'] = prob_fn
attn_kwargs.update({
"memory_sequence_length": memory_sequence_length,
"memory": memory
})
self._attn_kwargs = attn_kwargs
attn_modules = ['tensorflow.contrib.seq2seq', 'texar.custom']
# Use variable_scope to ensure all trainable variables created in
# the attention mechanism are collected
with tf.variable_scope(self.variable_scope):
attention_mechanism = utils.check_or_get_instance(
attn_hparams["type"],
attn_kwargs,
attn_modules,
classtype=tf.contrib.seq2seq.AttentionMechanism)
self._attn_cell_kwargs = {
"attention_layer_size": attn_hparams["attention_layer_size"],
"alignment_history": attn_hparams["alignment_history"],
"output_attention": attn_hparams["output_attention"],
}
self._cell_input_fn = cell_input_fn
# Use variable_scope to ensure all trainable variables created in
# AttentionWrapper are collected
with tf.variable_scope(self.variable_scope):
#if attention_layer is not None:
# self._attn_cell_kwargs["attention_layer_size"] = None
attn_cell = AttentionWrapper(
self._cell,
attention_mechanism,
cell_input_fn=self._cell_input_fn,
#attention_layer=attention_layer,
**self._attn_cell_kwargs)
self._cell = attn_cell
def __init__(self,
cell=None,
cell_dropout_mode=None,
vocab_size=None,
output_layer=None,
hparams=None):
RNNDecoderBase.__init__(self, cell, vocab_size, output_layer,
cell_dropout_mode, hparams)
def __init__(self,
cell=None,
cell_dropout_mode=None,
vocab_size=None,
output_layer=None,
position_embedder=None,
hparams=None):
RNNDecoderBase.__init__(self, cell, vocab_size, output_layer,
cell_dropout_mode, hparams)
self.position_embedder = position_embedder
self.current_segment_id = -1