def count_alpha(self, context, aspect, output, ctx_bitmap, alpha_adj=None):
'''
Count the attention weights.
alpha = softmax(tanh(wa*asp + wb*ctx))
Args:
:type context: tensor, shape = [batch_size, time_steps, edim]
:param context: the input context.
:type aspect: tensor, shape = [batch_size, edim]
:param aspect: the input aspect.
:type ctx_bitmap: tensorflow, shape like context.
:param ctx_bitmap: the context's bitmap, use to remove the influence of padding. \
Returns:
A tensor. The attention weights of the context.
'''
time_steps = tf.shape(context)[1]
aspect_3dim = tf.reshape(
tf.tile(aspect, [1, time_steps]),
[-1, time_steps, self.edim]
)
output_3dim = tf.reshape(
tf.tile(output, [1, time_steps]),
[-1, time_steps, self.edim]
)
res_asp = self.line_layer_asp.forward(aspect_3dim)
res_ctx = self.line_layer_ctx.forward(context)
res_output = self.line_layer_output.forward(output_3dim)
res_sum = res_asp + res_ctx + res_output
res_act = activer(res_sum, self.active)
batch_size = tf.shape(context)[0]
w_shp0 = tf.shape(self.wline_ca)[0]
w_shp1 = tf.shape(self.wline_ca)[1]
w_line_3dim = tf.reshape(
tf.tile(self.wline_ca, [batch_size, 1]),
[batch_size, w_shp0, w_shp1]
)
res_act = tf.reshape(
tf.matmul(res_act, w_line_3dim),
[-1, time_steps]
)
alpha = normalizer(self.norm_type ,res_act, ctx_bitmap, 1)
if alpha_adj is not None:
alpha += alpha_adj
return alpha
def count_alpha(self, context, aspect, ctx_bitmap):
'''
count the content attention (weight)
'''
mem_size = tf.shape(context)[1]
context = context
aspect = aspect
# adjust attention
asp_3dim = tf.reshape(aspect, [-1, self.edim, 1])
# gout = tf.matmul(context, asp_3dim)
res_act = self.line_layer.forward(context)
res_act = tf.reshape(
tf.matmul(res_act, asp_3dim),
[-1, mem_size]
)
# alpha = tf.nn.softmax(tf.reshape(gout, [-1, mem_size]))
alpha = normalizer(self.norm_type, res_act, ctx_bitmap, 1)
return alpha