def cal_max_question_representation(question_representation, atten_scores):
atten_positions = tf.argmax(
atten_scores, axis=2,
output_type=tf.int32) # [batch_size, passage_len]
max_question_reps = layer_utils.collect_representation(
question_representation, atten_positions)
return max_question_reps
def cal_max_node_2_representation(node_2_rep, relevancy_matrix):
# [batch_size, single_graph_1_nodes_size]
atten_positions = tf.argmax(relevancy_matrix, axis=2, output_type=tf.int32)
max_node_2_reps = layer_utils.collect_representation(
node_2_rep, atten_positions)
# [batch_size, single_graph_1_nodes_size, dim]
return max_node_2_reps
def cal_max_question_representation(self, question_representation,
atten_scores):
# question_representation: [batch_size, q_len, dim]
# atten_scores: [batch_size, passage_len, q_len]
atten_positions = tf.argmax(atten_scores, axis=2, output_type=tf.int32)
max_question_reps = layer_utils.collect_representation(
question_representation, atten_positions)
# [batch_size, passage_len, dim]
return max_question_reps
def onelayer_BiMPM_match(in_dim,
passage,
question,
passage_mask,
question_mask,
accum_dim=0,
passage_accum=None,
question_accum=None,
options=None,
scope_name='onelayer_BiMPM_match',
is_training=True,
dropout_rate=0.2,
reuse=False):
if passage_accum is None:
passage_accum = passage
question_accum = question
accum_dim = in_dim
match_results = []
match_dim = 0
QoP_reps = None
with tf.variable_scope(scope_name, reuse=reuse):
# attention passage over question
PoQ_atten = layer_utils.calcuate_attention(passage_accum,
question_accum,
accum_dim,
accum_dim,
scope_name="PoQ_atten",
att_type=options.att_type,
att_dim=options.att_dim,
remove_diagnoal=False,
mask1=passage_mask,
mask2=question_mask,
is_training=is_training,
dropout_rate=dropout_rate)
PoQ_reps = tf.matmul(
PoQ_atten,
layer_utils.dropout_layer(question,
dropout_rate,
is_training=is_training))
if options.with_QoP:
# attention question over passage
QoP_atten = layer_utils.calcuate_attention(
question_accum,
passage_accum,
accum_dim,
accum_dim,
scope_name="QoP_atten",
att_type=options.att_type,
att_dim=options.att_dim,
remove_diagnoal=False,
mask1=question_mask,
mask2=passage_mask,
is_training=is_training,
dropout_rate=dropout_rate)
QoP_reps = tf.matmul(
QoP_atten,
layer_utils.dropout_layer(passage,
dropout_rate,
is_training=is_training))
# attentive matching
(att_match_rep,
att_match_dim) = multi_perspective_match(in_dim,
passage,
PoQ_reps,
is_training=is_training,
dropout_rate=dropout_rate,
options=options,
scope_name='att_match')
match_results.append(att_match_rep)
match_dim += att_match_dim
# max attentive matching
PoQ_max_reps = layer_utils.collect_representation(
question, tf.argmax(PoQ_atten, axis=2, output_type=tf.int32))
(max_att_match_rep, max_att_match_dim) = multi_perspective_match(
in_dim,
passage,
PoQ_max_reps,
is_training=is_training,
dropout_rate=dropout_rate,
options=options,
scope_name='max_att_match')
match_results.append(max_att_match_rep)
match_dim += max_att_match_dim
match_results = tf.concat(axis=2, values=match_results)
return (match_results, match_dim, PoQ_reps, QoP_reps)
def multi_granularity_match(feature_dim,
passage,
question,
passage_length,
question_length,
passage_mask=None,
question_mask=None,
is_training=True,
dropout_rate=0.2,
options=None,
with_full_matching=False,
with_attentive_matching=True,
with_max_attentive_matching=True,
scope_name='mgm',
reuse=False):
'''
passage: [batch_size, passage_length, feature_dim]
question: [batch_size, question_length, feature_dim]
passage_length: [batch_size]
question_length: [batch_size]
'''
input_shape = tf.shape(passage)
batch_size = input_shape[0]
passage_len = input_shape[1]
match_reps = []
with tf.variable_scope(scope_name, reuse=reuse):
match_dim = 0
if with_full_matching:
passage_fw = passage[:, :, 0:feature_dim / 2]
passage_bw = passage[:, :, feature_dim / 2:feature_dim]
question_fw = question[:, :, 0:feature_dim / 2]
question_bw = question[:, :, feature_dim / 2:feature_dim]
question_fw = layer_utils.collect_final_step_of_lstm(
question_fw,
question_length - 1) # [batch_size, feature_dim/2]
question_bw = question_bw[:, 0, :]
question_fw = tf.expand_dims(question_fw, axis=1)
question_fw = tf.tile(
question_fw,
[1, passage_len, 1]) # [batch_size, pasasge_len, feature_dim]
question_bw = tf.expand_dims(question_bw, axis=1)
question_bw = tf.tile(
question_bw,
[1, passage_len, 1]) # [batch_size, pasasge_len, feature_dim]
(fw_full_match_reps, fw_full_match_dim) = multi_perspective_match(
feature_dim / 2,
passage_fw,
question_fw,
is_training=is_training,
dropout_rate=dropout_rate,
options=options,
scope_name='fw_full_match')
(bw_full_match_reps, bw_full_match_dim) = multi_perspective_match(
feature_dim / 2,
passage_bw,
question_bw,
is_training=is_training,
dropout_rate=dropout_rate,
options=options,
scope_name='bw_full_match')
match_reps.append(fw_full_match_reps)
match_reps.append(bw_full_match_reps)
match_dim += fw_full_match_dim
match_dim += bw_full_match_dim
if with_attentive_matching or with_max_attentive_matching:
atten_scores = layer_utils.calcuate_attention(
passage,
question,
feature_dim,
feature_dim,
scope_name="attention",
att_type=options.attn_type,
att_dim=options.attn_depth,
remove_diagnoal=False,
mask1=passage_mask,
mask2=question_mask,
is_training=is_training,
dropout_rate=dropout_rate)
# match_reps.append(tf.reduce_max(atten_scores, axis=2, keep_dims=True))
# match_reps.append(tf.reduce_mean(atten_scores, axis=2, keep_dims=True))
# match_dim += 2
if with_max_attentive_matching:
atten_positions = tf.argmax(
atten_scores, axis=2,
output_type=tf.int32) # [batch_size, passage_len]
max_question_reps = layer_utils.collect_representation(
question, atten_positions)
(max_att_match_rep, max_att_match_dim) = multi_perspective_match(
feature_dim,
passage,
max_question_reps,
is_training=is_training,
dropout_rate=dropout_rate,
options=options,
scope_name='max_att_match')
match_reps.append(max_att_match_rep)
match_dim += max_att_match_dim
if with_attentive_matching:
att_rep = tf.matmul(atten_scores, question)
(attentive_match_rep,
attentive_match_dim) = multi_perspective_match(
feature_dim,
passage,
att_rep,
is_training=is_training,
dropout_rate=dropout_rate,
options=options,
scope_name='att_match')
match_reps.append(attentive_match_rep)
match_dim += attentive_match_dim
match_reps = tf.concat(axis=2, values=match_reps)
return (match_reps, match_dim)