from __future__ import absolute_import
from keras.engine import InputSpec
from keras.layers import Wrapper, Merge
class Residual(Wrapper):
"""This wrapper automatically applies a residual to a model.
For an input `x` and a model `F(x)`, the residual wrapper gives the output
`y = x + F(x)`. In this configuration, the output of F(x) must have the
same shape as x. Other merge modes are supported besides summation.
input = Input(shape=(5,))
# Apply the residual normally
output1 = Residual(Dense(5), merge_mode='sum')(input)
# Throws an exception due to mismatching shapes
output2 = Residual(Dense(3), merge_mode='sum')(input)
# Product: `y = x * F(x)`
output3 = Residual(Dense(5), merge_mode='mul')(input)
For more modes, see: https://keras.io/layers/core/#merge
Alternatively, a function which takes the input and the layer output
can be passed to define the merge:
from keras.layers import Merge
def diff_merge(): # x_fx = [x, fx]
diff = lambda x: x[1] - x[0]
return Merge(mode=diff, output_shape=lambda x: x[0])
# Difference: `y = F(x) - x`
output4 = Residual(Dense(5), merge_mode=diff_merge())(input)
Args:
layer: The layer to wrap
merge_mode: The merge operation
"""
def __init__(self, layer, merge_mode='sum', **kwargs):
self.merge_mode = merge_mode
self.supports_masking = True
super(Residual, self).__init__(layer, **kwargs)
def build(self, input_shape):
output_shape = self.layer.get_output_shape_for(input_shape)
if output_shape != input_shape:
raise Exception('Cannot apply residual to layer "{}": '
'mismatching input and output shapes'
'input="{}" and output="{}"'
.format(self.layer.name, input_shape, output_shape))
if not self.layer.built:
self.layer.build(input_shape)
self.layer.built = True
self.input_spec = [InputSpec(shape=input_shape)]
super(Residual, self).build()
def get_output_shape_for(self, input_shape):
return input_shape
def call(self, x, mask=None):
layer_output = self.layer.call(x, mask)
if isinstance(self.merge_mode, str):
self.merge_mode = Merge(mode=self.merge_mode)
output = self.merge_mode([x, layer_output])
return output
@classmethod
def from_config(cls, config):
from keras.utils.layer_utils import layer_from_config
merge_mode = layer_from_config(config.pop('merge_mode'))
residual = super(Residual, cls).from_config(config)
residual.merge_mode = merge_mode
return residual
def get_config(self):
config = {'merge_mode': {'class_name': 'Merge',
'config': self.merge_mode.get_config()}}
base_config = super(Residual, self).get_config()
return dict(list(base_config.items()) + list(config.items()))
