def backward(self, grad_act, input_act):
"""
Backward propagation. This class is mostly wraps around _backward and does some extra
asserts. Child classes should overwrite _backward rather than this method.
Parameters
----------
grad_act : nump array, gradient of cost function with respect to the activations from this
layer (usually calculated in the layer above and passed down during backward
propagation), shape is self.output_shape or (NUMBER_OF_EXAMPLES,) + self.output_shape
input_act : numpy array, activations from the layer below; shape must either be the same as
self.input_shape, or (NUMBER_OF_EXAMPLES,) + self.input_shape
Returns
-------
grad_input_act : numpy array, gradient of cost function with respect to the input
activations this layer received, which is to be passed down to the layer below; shape
will be self.input_shape or (NUMBER_OF_EXAMPLES,) + self.input_shape, depending on the
input
grad_params : 1D numpy array of length self.num_params() (or None if self.num_params()==0),
gradient of cost function with respect to the params of this layer
"""
input_ndim = len(self.input_shape)
output_ndim = len(self.output_shape)
assert grad_act.shape[-output_ndim:] == tuple(
self.output_shape), 'wrong grad input shape'
assert input_act.shape[-input_ndim:] == tuple(
self.input_shape), 'wrong input shape'
assert ((grad_act.ndim == output_ndim and input_act.ndim == input_ndim)
or grad_act.shape[0]
== input_act.shape[0]), 'wrong number of samples'
many = (input_act.ndim > input_ndim)
input_act = atleast(input_act, input_ndim + 1)
grad_act = atleast(grad_act, output_ndim + 1)
grad_input_act, grad_params = self._backward(grad_act, input_act)
assert grad_input_act.shape[1:] == tuple(self.input_shape), \
'wrong input act grad shape'
if self.num_params() > 0:
grad_params = grad_params.ravel()
assert grad_params.size == self.num_params(
), 'wrong param grad shape'
return (grad_input_act if many else grad_input_act[0,
...], grad_params)
def forward(self, input_act):
"""
Forward propagation. This class is mostly wraps around _forward and does some extra
asserts. Child classes should overwrite _forward rather than this method.
Parameters
----------
input_act : numpy array, activations from the layer below; shape must either be the same as
self.input_shape, or (NUMBER_OF_EXAMPLES,) + self.input_shape
Returns
-------
output_act : numpy array, output activations from this layer; shape will be
self.output_shape or (NUMBER_OF_EXAMPLES,) + self.output_shape, depending on the input
"""
input_ndim = len(self.input_shape)
assert input_act.shape[-input_ndim:] == tuple(
self.input_shape), 'wrong input shape'
many = (input_act.ndim > input_ndim)
input_act = atleast(input_act, input_ndim + 1)
act = self._forward(input_act)
assert act.shape[1:] == tuple(self.output_shape), 'wrong output shape'
return act if many else act[0, ...]
def forward(self, input_act):
"""
Forward propagation. This class is mostly wraps around _forward and does some extra
asserts. Child classes should overwrite _forward rather than this method.
Parameters
----------
input_act : numpy array, activations from the layer below; shape must either be the same as
self.input_shape, or (NUMBER_OF_EXAMPLES,) + self.input_shape
Returns
-------
output_act : numpy array, output activations from this layer; shape will be
self.output_shape or (NUMBER_OF_EXAMPLES,) + self.output_shape, depending on the input
"""
input_ndim = len(self.input_shape)
assert input_act.shape[-input_ndim:] == tuple(self.input_shape), "wrong input shape"
many = input_act.ndim > input_ndim
input_act = atleast(input_act, input_ndim + 1)
act = self._forward(input_act)
assert act.shape[1:] == tuple(self.output_shape), "wrong output shape"
return act if many else act[0, ...]
def backward(self, grad_act, input_act):
"""
Backward propagation. This class is mostly wraps around _backward and does some extra
asserts. Child classes should overwrite _backward rather than this method.
Parameters
----------
grad_act : nump array, gradient of cost function with respect to the activations from this
layer (usually calculated in the layer above and passed down during backward
propagation), shape is self.output_shape or (NUMBER_OF_EXAMPLES,) + self.output_shape
input_act : numpy array, activations from the layer below; shape must either be the same as
self.input_shape, or (NUMBER_OF_EXAMPLES,) + self.input_shape
Returns
-------
grad_input_act : numpy array, gradient of cost function with respect to the input
activations this layer received, which is to be passed down to the layer below; shape
will be self.input_shape or (NUMBER_OF_EXAMPLES,) + self.input_shape, depending on the
input
grad_params : 1D numpy array of length self.num_params() (or None if self.num_params()==0),
gradient of cost function with respect to the params of this layer
"""
input_ndim = len(self.input_shape)
output_ndim = len(self.output_shape)
assert grad_act.shape[-output_ndim:] == tuple(self.output_shape), "wrong grad input shape"
assert input_act.shape[-input_ndim:] == tuple(self.input_shape), "wrong input shape"
assert (grad_act.ndim == output_ndim and input_act.ndim == input_ndim) or grad_act.shape[0] == input_act.shape[
0
], "wrong number of samples"
many = input_act.ndim > input_ndim
input_act = atleast(input_act, input_ndim + 1)
grad_act = atleast(grad_act, output_ndim + 1)
grad_input_act, grad_params = self._backward(grad_act, input_act)
assert grad_input_act.shape[1:] == tuple(self.input_shape), "wrong input act grad shape"
if self.num_params() > 0:
grad_params = grad_params.ravel()
assert grad_params.size == self.num_params(), "wrong param grad shape"
return (grad_input_act if many else grad_input_act[0, ...], grad_params)
