def _build_embed_encoder_and_attention(self, x, q,
context_lengths, question_lengths):
"""Build modeling encoder and return start/end logits.
Args:
x: context Tensor of shape [batch, max_context_length, dim].
q: question Tensor of shape [batch, max_question_length, dim].
context_lengths: length Tensor of shape [batch].
question_lengths: length Tensor of shape [batch].
Returns:
x_final: output Tensor of shape [batch, max_context_length, dim].
"""
input_keep_prob = self.config['input_keep_prob']
# Embedding encoder
encoder_emb = self.init_submodule(
self.config['encoder_emb'], name='xq_encoder')
x = encoder_emb(x, context_lengths)['outputs']
q = encoder_emb(q, question_lengths)['outputs']
x = squad_helper.dropout_wrapper(x, input_keep_prob, self.mode)
q = squad_helper.dropout_wrapper(q, input_keep_prob, self.mode)
# Context-Question Attention
with tf.variable_scope('attention'):
xq, qx = squad_helper.bi_attention_memory_efficient_dcn(
a=q,
b=x,
mask_a=question_lengths,
mask_b=context_lengths)
x_final = tf.concat([x, xq, x * xq, x * qx], 2)
return x_final
def _build_embed(self, features, training):
"""Build embedding model and return document/question representations.
Args:
features: dict; input data.
training: bool; whether we are under training mode.
Returns:
x: context Tensor of shape [batch, max_context_length, dim].
q: question Tensor of shape [batch, max_question_length, dim].
"""
config = self.config
tf.logging.info('\n\n\n\n\n')
tf.logging.info(config)
tf.logging.info('\n\n\n\n\n')
embedder = self.init_submodule(config['embedding'])
xc, qc, xw, qw, _, _, _, _ = embedder(features)
emb = embedder(features)
# Concatenate word and character embeddings
x = tf.concat([emb['xc'], emb['xw']], 2)
q = tf.concat([emb['qc'], emb['qw']], 2)
x = squad_helper.dropout_wrapper(x, config['input_keep_prob'], self.mode)
q = squad_helper.dropout_wrapper(q, config['input_keep_prob'], self.mode)
x = squad_helper.highway_layer(
x, num_layers=2, activation='relu', training=training,
dropout_rate=1-config['hw_keep_prob'])
q = squad_helper.highway_layer(
q, num_layers=2, activation='relu', training=training,
dropout_rate=1-config['hw_keep_prob'], reuse=True)
return x, q
def _build_model_encoder(self, x, context_lengths, gpu_id=0, num_gpus=1):
"""Build modeling encoder and return start/end logits.
Args:
x: input Tensor of shape [batch, max_length, dim].
context_lengths: length Tensor of shape [batch].
gpu_id: start GPU id.
num_gpus: number of GPUs available.
Returns:
logits_start: Tensor of shape [batch, max_length].
logits_end: Tensor of shape [batch, max_length].
modeling_layers: a list of modeling layers, from bottom to top,
each has shape [batch, max_length, dim].
"""
output_keep_prob = self.config['output_keep_prob']
with tf.device(misc_util.get_device_str(gpu_id, num_gpus)):
encoder_model = self.init_submodule(
self.config['encoder_model'], name='encoder_model')
x0 = encoder_model(x, context_lengths)['outputs']
x0 = squad_helper.dropout_wrapper(x0, output_keep_prob, self.mode)
x1 = encoder_model(x0, context_lengths)['outputs']
x1 = squad_helper.dropout_wrapper(x1, output_keep_prob, self.mode)
gpu_id += 1
with tf.device(misc_util.get_device_str(gpu_id, num_gpus)):
x2 = encoder_model(x1, context_lengths)['outputs']
x2 = squad_helper.dropout_wrapper(x2, output_keep_prob, self.mode)
x3 = encoder_model(x2, context_lengths)['outputs']
x3 = squad_helper.dropout_wrapper(x3, output_keep_prob, self.mode)
logits_start = squad_helper.exp_mask(
tf.squeeze(
tf.layers.dense(tf.concat([x1, x2], 2), 1, name='logits1'), 2),
context_lengths)
logits_end = squad_helper.exp_mask(
tf.squeeze(
tf.layers.dense(tf.concat([x1, x3], 2), 1, name='logits2'), 2),
context_lengths)
modeling_layers = [x0, x1, x2, x3]
return logits_start, logits_end, modeling_layers
