def predict(self, input):
input_variable = variable_from_sentence(self.lang1, input)
# Run through encoder
encoder_hidden = self.encoder.init_hidden()
encoder_outputs, encoder_hidden = self.encoder(input_variable,
encoder_hidden)
decoder_hidden = self._hidden_encoder_to_decoder(encoder_hidden)
query = beam_search(5, self.decoder, decoder_hidden, encoder_outputs,
self.lang2)
query = fix_parentheses(query)
return query
def _translate(self, process_id, input_item, models, sess):
"""
Actual translation (model sampling).
"""
# unpack input item attributes
k = input_item.k
x = input_item.batch
#max_ratio = input_item.max_ratio
y_dummy = numpy.zeros(shape=(len(x), 1))
x, x_mask, _, _ = prepare_data(x, y_dummy, maxlen=None)
sample = inference.beam_search(models, sess, x, x_mask, k)
return sample
def decode_at_test(self, enc_output, cross_attn_mask, batch_size,
beam_size, do_sample):
""" Returns the probability distribution over target-side tokens conditioned on the output of the encoder;
performs decoding via auto-regression at test time. """
def _decode_step(target_embeddings, memories):
""" Decode the encoder-generated representations into target-side logits with auto-regression. """
# Propagate inputs through the encoder stack
dec_output = target_embeddings
# NOTE: No self-attention mask is applied at decoding, as future information is unavailable
for layer_id in range(1, self.config.num_decoder_layers + 1):
dec_output, memories['layer_{:d}'.format(layer_id)] = \
self.decoder_stack[layer_id]['self_attn'].forward(
dec_output, None, None, memories['layer_{:d}'.format(layer_id)])
dec_output, _ = \
self.decoder_stack[layer_id]['cross_attn'].forward(dec_output, enc_output, cross_attn_mask)
dec_output = self.decoder_stack[layer_id]['ffn'].forward(
dec_output)
# Return prediction at the final time-step to be consistent with the inference pipeline
dec_output = dec_output[:, -1, :]
return dec_output, memories
def _pre_process_targets(step_target_ids, current_time_step):
""" Pre-processes target token ids before they're passed on as input to the decoder
for auto-regressive decoding. """
# Embed target_ids
target_embeddings = self._embed(step_target_ids)
signal_slice = positional_signal[:, current_time_step -
1:current_time_step, :]
target_embeddings += signal_slice
if self.config.dropout_embeddings > 0:
target_embeddings = tf.layers.dropout(
target_embeddings,
rate=self.config.dropout_embeddings,
training=self.training)
return target_embeddings
def _decoding_function(step_target_ids, current_time_step, memories):
""" Generates logits for the target-side token predicted for the next-time step with auto-regression. """
# Embed the model's predictions up to the current time-step; add positional information, mask
target_embeddings = _pre_process_targets(step_target_ids,
current_time_step)
# Pass encoder context and decoder embeddings through the decoder
dec_output, memories = _decode_step(target_embeddings, memories)
# Project decoder stack outputs and apply the soft-max non-linearity
step_logits = self.softmax_projection_layer.project(dec_output)
return step_logits, memories
with tf.variable_scope(self.name):
# Create nodes
self._build_graph()
positional_signal = get_positional_signal(
self.config.translation_max_len, self.config.embedding_size,
self.float_dtype)
if beam_size > 0:
# Initialize target IDs with <GO>
initial_ids = tf.cast(tf.fill([batch_size], 1),
dtype=self.int_dtype)
initial_memories = self._get_initial_memories(
batch_size, beam_size=beam_size)
output_sequences, scores = beam_search(
_decoding_function, initial_ids, initial_memories,
self.int_dtype, self.float_dtype,
self.config.translation_max_len, batch_size, beam_size,
self.embedding_layer.get_vocab_size(), 0,
self.config.length_normalization_alpha)
else:
# Initialize target IDs with <GO>
initial_ids = tf.cast(tf.fill([batch_size, 1], 1),
dtype=self.int_dtype)
initial_memories = self._get_initial_memories(batch_size,
beam_size=1)
output_sequences, scores = greedy_search(
_decoding_function,
initial_ids,
initial_memories,
self.int_dtype,
self.float_dtype,
self.config.translation_max_len,
batch_size,
0,
do_sample,
time_major=False)
return output_sequences, scores