def write(self, data):
""" Do a filtered write given the data """
if not self.write_grid:
raise ValueError('Writing is not supported')
filter_x, filter_y, gamma = self.get_filter(data, self.write_grid, scope='write/filter')
filter_y_transpose = tf.transpose(filter_y, [0, 2, 1])
window = layers.linear(data, reduce_prod(self.write_grid.size))
window = tf.reshape(window, (-1, self.write_grid.size[1], self.write_grid.size[0]))
patch = tf.matmul(filter_y_transpose, tf.matmul(window, filter_x))
return tf.reciprocal(tf.maximum(gamma, self.epsilon)) * layers.flatten(patch)
def create_attention(attention_type, data_shape, read_size=None, write_size=None):
""" Create the appropriate attention based on the passed in config """
if attention_type == 'none':
return NoAttention(reduce_prod(data_shape))
elif attention_type == 'sigmoid':
return SimpleAttention(tf.sigmoid, scope='SigmoidAttention')
elif attention_type == 'softmax':
return SimpleAttention(tf.nn.softmax, scope='SoftmaxAttention')
elif attention_type == 'content':
return ContentAttention(tuple(data_shape[:2]))
elif attention_type == 'gaussian':
grid_attention_cls = GaussianAttention
elif attention_type == 'cauchy':
grid_attention_cls = CauchyAttention
elif attention_type == 'grid':
grid_attention_cls = GridAttention
else:
raise TypeError('Unknown attention type: "{0}"'.format(attention_type))
return grid_attention_cls(tuple(data_shape[:2]), read_size, write_size)
def read_size(self, data):
""" Return the read size of the given data """
return (reduce_prod(data.get_shape().as_list()[1:]), )
def read_size(self, data):
""" Return the read size of the given data """
return (reduce_prod(self.read_grid.size), )
