def align_tensors(scope, options, primary_tensor, secondary_tensor, apply_tensor,
batch_size):
"""Runs alignment. Softmax probabilities are computed using the
primary and secondary tensors, and they are applied to the third
tensor.
Input:
primary_tensor: A tensor of shape [batch_size, M, d]
secondary_tensor: A tensor of shape [batch_size, N, d]
apply_tensor: A tensor of shape [batch_size, N, d]
Output:
An output tensor of shape [batch_size, M, d]
"""
with tf.variable_scope(scope):
coattention_matrix = tf.matmul(primary_tensor, tf.transpose(
secondary_tensor, perm=[0, 2, 1])) # size = [batch_size, M, N]
if primary_tensor == secondary_tensor:
sh = tf.shape(primary_tensor)
M = sh[1]
eye = tf.eye(M, batch_shape=[batch_size]) # size = [batch_size, M, M]
ones = tf.fill([batch_size, M, M], 1.0) # size = [batch_size, M, M]
mult = ones - eye # size = [batch_size, M, M]
coattention_matrix *= mult # size = [batch_size, M, M]
softmax = tf.nn.softmax(coattention_matrix, dim=-1) # size = [batch_size, M, N]
attention = tf.matmul(softmax, apply_tensor) # size = [batch_size, M, d]
apply_tensor_times_attention = primary_tensor * attention # size = [batch_size, M, d]
apply_tensor_minus_attention = primary_tensor - attention # size = [batch_size, M, d]
return semantic_fusion(primary_tensor, primary_tensor.get_shape()[-1].value,
[attention, apply_tensor_times_attention, apply_tensor_minus_attention],
"semantic_fusion") # size = [batch_size, M, d]
def _self_alignment(options, ctx, ctx_dim, sq_dataset):
"""Runs self alignment on the passage.
Input:
ctx: The (question-aware) passage of shape [batch_size, M, d]
ctx_dim: d, from above
Output:
A self-aligned representation of shape [batch_size, M, d]
"""
with tf.variable_scope("self_alignment"):
sh = tf.shape(ctx)
batch_size, M = sh[0], sh[1]
coattention_matrix = tf.matmul(ctx, tf.transpose(
ctx, perm=[0, 2, 1])) # size = [batch_size, M, M]
eye = tf.eye(M, batch_shape=[batch_size]) # size = [batch_size, M, M]
ones = tf.fill([batch_size, M, M], 1.0) # size = [batch_size, M, M]
mult = ones - eye # size = [batch_size, M, M]
B = coattention_matrix * mult # size = [batch_size, M, M]
softmax = tf.nn.softmax(B, dim=2) # size = [batch_size, M, M]
attention = tf.matmul(softmax, ctx) # size = [batch_size, M, d]
ctx_times_attention = ctx * attention # size = [batch_size, M, d]
ctx_minus_attention = ctx - attention # size = [batch_size, M, d]
return semantic_fusion(
ctx, ctx_dim,
[attention, ctx_times_attention, ctx_minus_attention],
"self_aligned_semantic_fusion") # size = [batch_size, M, d]
def _interactive_alignment(options, ctx, qst, ctx_dim, sq_dataset, keep_prob):
"""Runs interactive alignment on the passage and question.
Input:
ctx: The passage of shape [batch_size, M, d]
qst: The question of shape [batch_size, N, d]
ctx_dim: d, from above
Output:
A question-aware representation of shape [batch_size, M, d]
"""
with tf.variable_scope("interactive_alignment"):
coattention_matrix = tf.matmul(qst, tf.transpose(
ctx, perm=[0, 2, 1])) # size = [batch_size, N, M]
softmax = tf.nn.softmax(coattention_matrix,
dim=1) # size = [batch_size, N, M]
q_attention = tf.matmul(
tf.transpose(softmax, perm=[0, 2, 1]) # size = [batch_size, M, N]
,
qst) # size = [batch_size, M, d]
c_times_q_attention = ctx * q_attention # size = [batch_size, M, d]
c_minus_q_attention = ctx - q_attention # size = [batch_size, M, d]
return semantic_fusion(
ctx, ctx_dim,
[q_attention, c_times_q_attention, c_minus_q_attention],
"interactive_alignment_semantic_fusion"
) # size = [batch_size, M, d]
def _update_memory(memory, ctx, probs, ctx_dim, batch_size, M, scope):
evidence = tf.reshape(tf.matmul(
tf.transpose(ctx, perm=[0, 2, 1]) # size = [batch_size, d, M]
, tf.reshape(probs, [batch_size, M, 1]))
, [batch_size, ctx_dim]) # size = [batch_size, d]
memory_reshaped = tf.reshape(memory, [batch_size, ctx_dim])
updated_memory = semantic_fusion(memory_reshaped, ctx_dim, [evidence],
scope) # size = [batch_size, d]
return tf.reshape(updated_memory, [batch_size, 1, ctx_dim]) # size = [batch_size, 1, d]