def test_conv_kernel_mask_wrong_dims(self, *input_shape):
kernel_shape = 1
strides = 1
conv_utils.conv_kernel_mask(input_shape, kernel_shape, strides,
'valid')
ndims = len(input_shape)
kernel_shape = (2, ) * (ndims + 1)
self.assertRaises(ValueError, conv_utils.conv_kernel_mask, input_shape,
kernel_shape, strides, 'same')
strides = (1, ) * ndims
self.assertRaises(ValueError, conv_utils.conv_kernel_mask, input_shape,
kernel_shape, strides, 'valid')
kernel_shape = (1, ) * ndims
strides = (2, ) * (ndims - 1)
self.assertRaises(ValueError, conv_utils.conv_kernel_mask, input_shape,
kernel_shape, strides, 'valid')
strides = (2, ) * ndims
conv_utils.conv_kernel_mask(input_shape, kernel_shape, strides,
'valid')
def test_conv_kernel_mask_wrong_padding(self, *input_shape):
ndims = len(input_shape)
kernel_shape = (1, ) * ndims
strides = (1, ) * ndims
conv_utils.conv_kernel_mask(input_shape, kernel_shape, strides,
'valid')
conv_utils.conv_kernel_mask(input_shape, kernel_shape, strides, 'same')
self.assertRaises(NotImplementedError, conv_utils.conv_kernel_mask,
input_shape, kernel_shape, strides, 'full')
def test_conv_kernel_mask_rect_kernel(self, *input_shape):
padding = 'valid'
ndims = len(input_shape)
strides = (1,) * ndims
for d in range(ndims):
kernel_shape = [1] * ndims
kernel_shape[d] = input_shape[d]
output_shape = list(input_shape)
output_shape[d] = min(1, input_shape[d])
mask = np.identity(int(np.prod(input_shape)), np.bool)
mask = np.reshape(mask, input_shape * 2)
for p in itertools.product(*[range(input_shape[dim])
for dim in range(ndims)]):
p = list(p)
p[d] = slice(None)
mask[p * 2] = True
mask = np.take(mask, range(0, min(1, input_shape[d])), ndims + d)
self.assertAllEqual(
mask,
conv_utils.conv_kernel_mask(
input_shape,
kernel_shape,
strides,
padding
)
)
def test_conv_kernel_mask_rect_kernel(self, *input_shape):
padding = 'valid'
ndims = len(input_shape)
strides = (1, ) * ndims
for d in range(ndims):
kernel_shape = [1] * ndims
kernel_shape[d] = input_shape[d]
output_shape = list(input_shape)
output_shape[d] = min(1, input_shape[d])
mask = np.identity(int(np.prod(input_shape)), np.bool_)
mask = np.reshape(mask, input_shape * 2)
for p in itertools.product(
*[range(input_shape[dim]) for dim in range(ndims)]):
p = list(p)
p[d] = slice(None)
mask[p * 2] = True
mask = np.take(mask, range(0, min(1, input_shape[d])), ndims + d)
self.assertAllEqual(
mask,
conv_utils.conv_kernel_mask(input_shape, kernel_shape, strides,
padding))
def get_locallyconnected_mask(input_shape, kernel_shape, strides, padding,
data_format):
"""Return a mask representing connectivity of a locally-connected operation.
This method returns a masking numpy array of 0s and 1s (of type `np.float32`)
that, when element-wise multiplied with a fully-connected weight tensor, masks
out the weights between disconnected input-output pairs and thus implements
local connectivity through a sparse fully-connected weight tensor.
Assume an unshared convolution with given parameters is applied to an input
having N spatial dimensions with `input_shape = (d_in1, ..., d_inN)`
to produce an output with spatial shape `(d_out1, ..., d_outN)` (determined
by layer parameters such as `strides`).
This method returns a mask which can be broadcast-multiplied (element-wise)
with a 2*(N+1)-D weight matrix (equivalent to a fully-connected layer between
(N+1)-D activations (N spatial + 1 channel dimensions for input and output)
to make it perform an unshared convolution with given `kernel_shape`,
`strides`, `padding` and `data_format`.
Args:
input_shape: tuple of size N: `(d_in1, ..., d_inN)` spatial shape of the
input.
kernel_shape: tuple of size N, spatial shape of the convolutional kernel /
receptive field.
strides: tuple of size N, strides along each spatial dimension.
padding: type of padding, string `"same"` or `"valid"`.
data_format: a string, `"channels_first"` or `"channels_last"`.
Returns:
a `np.float32`-type `np.ndarray` of shape
`(1, d_in1, ..., d_inN, 1, d_out1, ..., d_outN)`
if `data_format == `"channels_first"`, or
`(d_in1, ..., d_inN, 1, d_out1, ..., d_outN, 1)`
if `data_format == "channels_last"`.
Raises:
ValueError: if `data_format` is neither `"channels_first"` nor
`"channels_last"`.
"""
mask = conv_utils.conv_kernel_mask(
input_shape=input_shape,
kernel_shape=kernel_shape,
strides=strides,
padding=padding)
ndims = int(mask.ndim / 2)
if data_format == 'channels_first':
mask = np.expand_dims(mask, 0)
mask = np.expand_dims(mask, -ndims - 1)
elif data_format == 'channels_last':
mask = np.expand_dims(mask, ndims)
mask = np.expand_dims(mask, -1)
else:
raise ValueError('Unrecognized data_format: ' + str(data_format))
return mask
def test_conv_kernel_mask_fc(self, *input_shape):
padding = 'valid'
kernel_shape = input_shape
ndims = len(input_shape)
strides = (1, ) * ndims
output_shape = _get_const_output_shape(input_shape, dim=1)
mask = np.ones(input_shape + output_shape, np.bool_)
self.assertAllEqual(
mask,
conv_utils.conv_kernel_mask(input_shape, kernel_shape, strides,
padding))
def test_conv_kernel_mask_diag(self, *input_shape):
ndims = len(input_shape)
kernel_shape = (1, ) * ndims
strides = (1, ) * ndims
for padding in ['valid', 'same']:
mask = np.identity(int(np.prod(input_shape)), np.bool_)
mask = np.reshape(mask, input_shape * 2)
self.assertAllEqual(
mask,
conv_utils.conv_kernel_mask(input_shape, kernel_shape, strides,
padding))
def test_conv_kernel_mask_wrong_padding(self, *input_shape):
ndims = len(input_shape)
kernel_shape = (1,) * ndims
strides = (1,) * ndims
conv_utils.conv_kernel_mask(
input_shape,
kernel_shape,
strides,
'valid'
)
conv_utils.conv_kernel_mask(
input_shape,
kernel_shape,
strides,
'same'
)
self.assertRaises(NotImplementedError,
conv_utils.conv_kernel_mask,
input_shape, kernel_shape, strides, 'full')
def test_conv_kernel_mask_full_stride(self, *input_shape):
padding = 'valid'
ndims = len(input_shape)
kernel_shape = (1, ) * ndims
strides = tuple([max(d, 1) for d in input_shape])
output_shape = _get_const_output_shape(input_shape, dim=1)
mask = np.zeros(input_shape + output_shape, np.bool_)
if all(d > 0 for d in mask.shape): # pylint: disable=not-an-iterable
mask[(0, ) * len(output_shape)] = True
self.assertAllEqual(
mask,
conv_utils.conv_kernel_mask(input_shape, kernel_shape, strides,
padding))
def test_conv_kernel_mask_wrong_dims(self, *input_shape):
kernel_shape = 1
strides = 1
conv_utils.conv_kernel_mask(
input_shape,
kernel_shape,
strides,
'valid'
)
ndims = len(input_shape)
kernel_shape = (2,) * (ndims + 1)
self.assertRaises(ValueError,
conv_utils.conv_kernel_mask,
input_shape, kernel_shape, strides, 'same')
strides = (1,) * ndims
self.assertRaises(ValueError,
conv_utils.conv_kernel_mask,
input_shape, kernel_shape, strides, 'valid')
kernel_shape = (1,) * ndims
strides = (2,) * (ndims - 1)
self.assertRaises(ValueError,
conv_utils.conv_kernel_mask,
input_shape, kernel_shape, strides, 'valid')
strides = (2,) * ndims
conv_utils.conv_kernel_mask(
input_shape,
kernel_shape,
strides,
'valid'
)
def test_conv_kernel_mask_fc(self, *input_shape):
padding = 'valid'
kernel_shape = input_shape
ndims = len(input_shape)
strides = (1,) * ndims
output_shape = _get_const_output_shape(input_shape, dim=1)
mask = np.ones(input_shape + output_shape, np.bool)
self.assertAllEqual(
mask,
conv_utils.conv_kernel_mask(
input_shape,
kernel_shape,
strides,
padding
)
)
def test_conv_kernel_mask_diag(self, *input_shape):
ndims = len(input_shape)
kernel_shape = (1,) * ndims
strides = (1,) * ndims
for padding in ['valid', 'same']:
mask = np.identity(int(np.prod(input_shape)), np.bool)
mask = np.reshape(mask, input_shape * 2)
self.assertAllEqual(
mask,
conv_utils.conv_kernel_mask(
input_shape,
kernel_shape,
strides,
padding
)
)
def test_conv_kernel_mask_almost_full_stride(self, *input_shape):
padding = 'valid'
ndims = len(input_shape)
kernel_shape = (1, ) * ndims
strides = tuple([max(d - 1, 1) for d in input_shape])
output_shape = _get_const_output_shape(input_shape, dim=2)
mask = np.zeros(input_shape + output_shape, np.bool_)
if all(d > 0 for d in mask.shape): # pylint: disable=not-an-iterable
for in_position in itertools.product(*[[0, d - 1]
for d in input_shape]):
out_position = tuple([min(p, 1) for p in in_position])
mask[in_position + out_position] = True
self.assertAllEqual(
mask,
conv_utils.conv_kernel_mask(input_shape, kernel_shape, strides,
padding))
def test_conv_kernel_mask_full_stride(self, *input_shape):
padding = 'valid'
ndims = len(input_shape)
kernel_shape = (1,) * ndims
strides = tuple([max(d, 1) for d in input_shape])
output_shape = _get_const_output_shape(input_shape, dim=1)
mask = np.zeros(input_shape + output_shape, np.bool)
if all(d > 0 for d in mask.shape):
mask[(0,) * len(output_shape)] = True
self.assertAllEqual(
mask,
conv_utils.conv_kernel_mask(
input_shape,
kernel_shape,
strides,
padding
)
)
def test_conv_kernel_mask_almost_full_stride(self, *input_shape):
padding = 'valid'
ndims = len(input_shape)
kernel_shape = (1,) * ndims
strides = tuple([max(d - 1, 1) for d in input_shape])
output_shape = _get_const_output_shape(input_shape, dim=2)
mask = np.zeros(input_shape + output_shape, np.bool)
if all(d > 0 for d in mask.shape):
for in_position in itertools.product(*[[0, d - 1] for d in input_shape]):
out_position = tuple([min(p, 1) for p in in_position])
mask[in_position + out_position] = True
self.assertAllEqual(
mask,
conv_utils.conv_kernel_mask(
input_shape,
kernel_shape,
strides,
padding
)
)
def get_locallyconnected_mask(input_shape,
kernel_shape,
strides,
padding,
data_format,
dtype):
"""Return a mask representing connectivity of a locally-connected operation.
This method returns a masking tensor of 0s and 1s (of type `dtype`) that,
when element-wise multiplied with a fully-connected weight tensor, masks out
the weights between disconnected input-output pairs and thus implements local
connectivity through a sparse fully-connected weight tensor.
Assume an unshared convolution with given parameters is applied to an input
having N spatial dimensions with `input_shape = (d_in1, ..., d_inN)`
to produce an output with spatial shape `(d_out1, ..., d_outN)` (determined
by layer parameters such as `strides`).
This method returns a mask which can be broadcast-multiplied (element-wise)
with a 2*(N+1)-D weight matrix (equivalent to a fully-connected layer between
(N+1)-D activations (N spatial + 1 channel dimensions for input and output)
to make it perform an unshared convolution with given `kernel_shape`,
`strides`, `padding` and `data_format`.
Arguments:
input_shape: tuple of size N: `(d_in1, ..., d_inN)`
spatial shape of the input.
kernel_shape: tuple of size N, spatial shape of the convolutional kernel
/ receptive field.
strides: tuple of size N, strides along each spatial dimension.
padding: type of padding, string `"same"` or `"valid"`.
data_format: a string, `"channels_first"` or `"channels_last"`.
dtype: type of the layer operation, e.g. `tf.float64`.
Returns:
a `dtype`-tensor of shape
`(1, d_in1, ..., d_inN, 1, d_out1, ..., d_outN)`
if `data_format == `"channels_first"`, or
`(d_in1, ..., d_inN, 1, d_out1, ..., d_outN, 1)`
if `data_format == "channels_last"`.
Raises:
ValueError: if `data_format` is neither `"channels_first"` nor
`"channels_last"`.
"""
mask = conv_utils.conv_kernel_mask(
input_shape=input_shape,
kernel_shape=kernel_shape,
strides=strides,
padding=padding
)
ndims = int(mask.ndim / 2)
mask = K.variable(mask, dtype)
if data_format == 'channels_first':
mask = K.expand_dims(mask, 0)
mask = K.expand_dims(mask, - ndims - 1)
elif data_format == 'channels_last':
mask = K.expand_dims(mask, ndims)
mask = K.expand_dims(mask, -1)
else:
raise ValueError('Unrecognized data_format: ' + str(data_format))
return mask
