def get_initial_histories(self) -> DecoderHistories:
histories = AutoregressiveDecoder.get_initial_histories(self)
rnn_histories = RNNHistories(
rnn_outputs=tf.zeros(
shape=[0, self.batch_size, self.rnn_size],
dtype=tf.float32,
name="hist_rnn_output_states"),
attention_histories=[a.initial_loop_state()
for a in self.attentions if a is not None])
return histories._replace(other=rnn_histories)
def get_initial_histories(self) -> DecoderHistories:
histories = AutoregressiveDecoder.get_initial_histories(self)
rnn_histories = RNNHistories(
rnn_outputs=tf.zeros(shape=[0, self.batch_size, self.rnn_size],
dtype=tf.float32,
name="hist_rnn_output_states"),
attention_histories=[
a.initial_loop_state() for a in self.attentions
if a is not None
])
return histories._replace(other=rnn_histories)
def get_initial_histories(self) -> DecoderHistories:
histories = AutoregressiveDecoder.get_initial_histories(self)
# TODO: record histories properly
tr_histories = tf.zeros([])
# tr_histories = TransformerHistories(
# self_attention_histories=[
# empty_multi_head_loop_state(self.batch_size,
# self.n_heads_self)
# for a in range(self.depth)],
# encoder_attention_histories=[
# empty_multi_head_loop_state(self.batch_size,
# self.n_heads_enc)
# for a in range(self.depth)])
return histories._replace(other=tr_histories)