def curvature_range(self):
# set up the curvature window
self._curv_win = tf.Variable(np.zeros([
self._curv_win_width,
]),
dtype=tf.float32,
name="curv_win",
trainable=False)
# we can use log smoothing for curvature range to follow trend faster
# self._curv_win = tf.scatter_update(
# self._curv_win, self._global_step % self._curv_win_width,
# tf.log(self._grad_norm_squared + EPS))
self._curv_win = tf.scatter_update(
self._curv_win, self._global_step % self._curv_win_width,
self._grad_norm_squared + EPS)
# note here the iterations start from iteration 0
valid_window = tf.slice(
self._curv_win, tf.constant([
0,
]),
tf.expand_dims(tf.minimum(tf.constant(self._curv_win_width),
self._global_step + 1),
dim=0))
if self._h_min_log_smooth:
self._h_min_t = tf.log(tf.reduce_min(valid_window) + EPS)
else:
self._h_min_t = tf.reduce_min(valid_window)
if self._h_max_log_smooth:
self._h_max_t = tf.log(tf.reduce_max(valid_window) + EPS)
else:
self._h_max_t = tf.reduce_max(valid_window)
curv_range_ops = []
with tf.control_dependencies([self._h_min_t, self._h_max_t]):
avg_op = self._moving_averager.apply(
[self._h_min_t, self._h_max_t])
with tf.control_dependencies([avg_op]):
if self._h_min_log_smooth:
self._h_min = tf.exp(
tf.identity(
self._moving_averager.average(self._h_min_t)))
else:
self._h_min = \
tf.identity(self._moving_averager.average(self._h_min_t))
if self._h_max_log_smooth:
self._h_max = tf.exp(
tf.identity(
self._moving_averager.average(self._h_max_t)))
else:
self._h_max = \
tf.identity(self._moving_averager.average(self._h_max_t))
if self._sparsity_debias:
self._h_min = self._h_min * self._sparsity_avg
self._h_max = self._h_max * self._sparsity_avg
curv_range_ops.append(avg_op)
return curv_range_ops
def masked_softmax(logits, mask, axis):
"""Compute softmax with input mask."""
e_logits = tf.exp(logits)
masked_e = tf.multiply(e_logits, mask)
sum_masked_e = tf.reduce_sum(masked_e, axis, keep_dims=True)
ones = tf.ones_like(sum_masked_e)
# pay attention to a situation that if len of mask is zero,
# denominator should be set to 1
sum_masked_e_safe = tf.where(tf.equal(sum_masked_e, 0), ones, sum_masked_e)
return masked_e / sum_masked_e_safe
def call(self, tensors):
"""Attention layer."""
left, right = tensors
len_left = left.shape[1]
len_right = right.shape[1]
tensor_left = tf.expand_dims(left, axis=2)
tensor_right = tf.expand_dims(right, axis=1)
tensor_left = tf.tile(tensor_left, [1, 1, len_right, 1])
tensor_right = tf.tile(tensor_right, [1, len_left, 1, 1])
tensor_merged = tf.concat([tensor_left, tensor_right], axis=-1)
middle_output = self.middle_layer(tensor_merged)
attn_scores = self.attn(middle_output)
attn_scores = tf.squeeze(attn_scores, axis=3)
exp_attn_scores = tf.exp(
attn_scores - tf.reduce_max(attn_scores, axis=-1, keepdims=True))
exp_sum = tf.reduce_sum(exp_attn_scores, axis=-1, keepdims=True)
attention_weights = exp_attn_scores / exp_sum
return tf.matmul(attention_weights, right)