def _weights_parameter(output_channels, init, group_name):
dilation_kernel = [(k - 1) * d + 1 for k, d in zip(kernel, dilation)]
# expand kernel to simulate dilation
used_init = init.copy()
if dilation_kernel != kernel:
for axis in range(len(dilation)):
kernel_sequence = [x * dilation[axis] for x in range(kernel[axis])]
insert_lines = list(set([x for x in range(dilation_kernel[axis])]) ^
set(kernel_sequence))
for index in range(len(insert_lines)):
insert_lines[index] -= index
used_init = np.insert(used_init, insert_lines, 0, axis=len(init.shape) - axis - 1)
return ops.parameter(shape=(output_channels, cntk.InferredDimension) + ops.sanitize_shape(dilation_kernel),
init=used_init, name=group_name)
def _weights_parameter(output_channels, init, group_name):
dilation_kernel = [(k - 1) * d + 1
for k, d in zip(kernel, dilation)]
# expand kernel to simulate dilation
used_init = init.copy()
if dilation_kernel != kernel:
for axis in range(len(dilation)):
kernel_sequence = [
x * dilation[axis] for x in range(kernel[axis])
]
insert_lines = list(
set([x for x in range(dilation_kernel[axis])])
^ set(kernel_sequence))
for index in range(len(insert_lines)):
insert_lines[index] -= index
used_init = np.insert(used_init,
insert_lines,
0,
axis=len(init.shape) - axis - 1)
return ops.parameter(
shape=(output_channels, cntk.InferredDimension) +
ops.sanitize_shape(dilation_kernel),
init=used_init,
name=group_name)
def _conv_ops(weights, data):
return ops.convolution(weights, data, strides=(cntk.InferredDimension, ) + \
ops.sanitize_shape(stride), auto_padding=[False, pad, pad])
def _conv_ops(weights, data):
return ops.convolution(weights, data, strides=(cntk.InferredDimension, ) + \
ops.sanitize_shape(stride), auto_padding=[False, pad, pad])