def __init__(self, input_node, output_format, filter_shape=None):
assert_equal(output_format.axes, ('b', 'c', '0', '1'))
if filter_shape is None:
filter_shape = (1, 1)
else:
assert_equal(len(filter_shape), 2)
filter_shape = numpy.asarray(filter_shape)
assert_equal(input_node.output_format.axes, ('b', 'c', '0', '1'))
input_shape = numpy.asarray(input_node.output_format.shape)[2:]
assert_equal(tuple(output_format.shape[2:]),
tuple(input_shape - filter_shape + 1))
channel_axis = 'c'
def get_num_filters():
c_index = output_format.axes.index(channel_axis)
return output_format.shape[c_index]
# def get_filter_shape():
# fmt = input_node.output_format
# assert_equal(fmt.axes, ('b', 'c', '0', '1'))
# return fmt.shape[2:]
# if filter_shape is None:
# filter_shape = get_filter_shape()
self.conv2d_node = CuDnnConv2d(input_node=input_node,
filter_shape=filter_shape,
num_filters=get_num_filters(),
pads='valid')
non_channel_axes = tuple(axis for axis
in self.conv2d_node.output_format.axes
if axis != channel_axis)
self.bias_node = Bias(self.conv2d_node,
output_format=self.conv2d_node.output_format,
input_to_bf_map={non_channel_axes: 'b',
channel_axis: 'f'},
bf_to_output_map={'b': non_channel_axes,
'f': channel_axis})
# input_to_bf_map={('c', '0', '1'): 'f'},
# bf_to_output_map={'f': ('c', '0', '1')})
assert_equal(self.bias_node.output_format.axes, output_format.axes)
assert_equal(self.bias_node.output_format.shape, output_format.shape)
super(Conv11AffineTransform, self).__init__(
input_node,
self.bias_node.output_symbol,
output_format)
class Conv11AffineTransform(Node):
'''
Equivalent to an AffineTransform, except the output is a 1x1 feature map.
A sequence of conv2d -> channel-wise bias
Unlike a Conv2dLayer, this does not pool the ReLU output.
Output format is bc01, (B, C, 1, 1), where C is the number of output
'''
def __init__(self, input_node, output_format, filter_shape=None):
assert_equal(output_format.axes, ('b', 'c', '0', '1'))
if filter_shape is None:
filter_shape = (1, 1)
else:
assert_equal(len(filter_shape), 2)
filter_shape = numpy.asarray(filter_shape)
assert_equal(input_node.output_format.axes, ('b', 'c', '0', '1'))
input_shape = numpy.asarray(input_node.output_format.shape)[2:]
assert_equal(tuple(output_format.shape[2:]),
tuple(input_shape - filter_shape + 1))
channel_axis = 'c'
def get_num_filters():
c_index = output_format.axes.index(channel_axis)
return output_format.shape[c_index]
# def get_filter_shape():
# fmt = input_node.output_format
# assert_equal(fmt.axes, ('b', 'c', '0', '1'))
# return fmt.shape[2:]
# if filter_shape is None:
# filter_shape = get_filter_shape()
self.conv2d_node = CuDnnConv2d(input_node=input_node,
filter_shape=filter_shape,
num_filters=get_num_filters(),
pads='valid')
non_channel_axes = tuple(axis for axis
in self.conv2d_node.output_format.axes
if axis != channel_axis)
self.bias_node = Bias(self.conv2d_node,
output_format=self.conv2d_node.output_format,
input_to_bf_map={non_channel_axes: 'b',
channel_axis: 'f'},
bf_to_output_map={'b': non_channel_axes,
'f': channel_axis})
# input_to_bf_map={('c', '0', '1'): 'f'},
# bf_to_output_map={'f': ('c', '0', '1')})
assert_equal(self.bias_node.output_format.axes, output_format.axes)
assert_equal(self.bias_node.output_format.shape, output_format.shape)
super(Conv11AffineTransform, self).__init__(
input_node,
self.bias_node.output_symbol,
output_format)
def _save_to_h5(self, group):
self.conv2d_node.save_to_h5(group.create_group('conv2d_node'))
self.bias_node.save_to_h5(group.create_group('bias_node'))
def _load_from_h5(self, group, **kwargs):
self.conv2d_node.load_from_h5(group['conv2d_node'], **kwargs)
self.bias_node.load_from_h5(group['bias_node'], **kwargs)
def _get_shared_vars(self):
return (self.conv2d_node.get_shared_vars() +
self.bias_node.get_shared_vars())
