def compress(cls, old_layer: Conv2D,
acceptable_error: float) -> Optional[FactorizedConv2DTucker]:
'''Compress layer's kernel 4D tensor using tucker decomposition under acceptable_error.
If it can't reduce the number of parameters, returns `None`.
:param old_layer:
:param acceptable_error:
:return:
'''
W = K.get_value(old_layer.kernel)
searcher = TuckerParamSearcher(W)
C, U_in, U_out = searcher.factorize_in_acceptable_error(
acceptable_error)
kernel = C
if U_in is None and U_out is None: # compression failed
return None
if U_in is None:
input_components = None
pre_kernel = None
else:
input_components = U_in.shape[1]
pre_kernel = U_in[np.newaxis, np.newaxis, :, :]
if U_out is None:
output_components = None
post_kernel = None
else:
output_components = U_out.shape[1]
post_kernel = U_out.T[np.newaxis, np.newaxis, :, :]
base_config = old_layer.get_config()
new_config = convert_config(base_config,
ignore_args=[
'kernel_constraint',
],
converts={
'kernel_regularizer': [
'pre_kernel_regularizer',
'kernel_regularizer',
'post_kernel_regularizer',
],
'kernel_initializer': [
'pre_kernel_initializer',
'kernel_initializer',
'post_kernel_initializer',
],
},
new_kwargs={
'input_components': input_components,
'output_components': output_components,
})
new_layer = FactorizedConv2DTucker(**new_config)
new_layer.build(
old_layer.get_input_shape_at(0)) # to initialize weight variables
K.set_value(new_layer.kernel, kernel)
if pre_kernel is not None:
K.set_value(new_layer.pre_kernel, pre_kernel)
if post_kernel is not None:
K.set_value(new_layer.post_kernel, post_kernel)
return new_layer
