`tltorch.factorized_conv`.CPConv

class tltorch.factorized_conv.CPConv(*args, **kwargs)[source]

Create a Factorized CP convolution of arbitrary order.

Parameters:

Parameters:	in_channelsint out_channelsint kernel_sizeint or int list if int, order MUST be specified if int list, then the conv will use order = len(kernel_size) rankint rank of the factorized kernel orderint, optional if kernel_size is a list see kernel_size implementation = {‘factorized’, ‘reconstructed’, ‘mobilenet’} Strategy to use for the forward pass. Options are: factorized : the CP conv is expressed as a series of 1D convolutions reconstructedfull kernel is reconstructed from the decomposition. the reconstruction is used to perform a regular forward pass mobilenet : the equivalent formulation of CP as a MobileNet block is used strideint, default is 1 paddingint, default is 0 dilationint, default is 0

in_channelsint

out_channelsint

kernel_sizeint or int list

if int, order MUST be specified if int list, then the conv will use order = len(kernel_size)

rankint

rank of the factorized kernel

orderint, optional if kernel_size is a list

see kernel_size

implementation = {‘factorized’, ‘reconstructed’, ‘mobilenet’}

Strategy to use for the forward pass. Options are:

factorized : the CP conv is expressed as a series of 1D convolutions
reconstructedfull kernel is reconstructed from the decomposition.
the reconstruction is used to perform a regular forward pass
mobilenet : the equivalent formulation of CP as a MobileNet block is used

strideint, default is 1

paddingint, default is 0

dilationint, default is 0

See also

References

[1]	Vadim Lebedev, Yaroslav Ganin, Maksim Rakhuba, Ivan V.Oseledets, and Victor S. Lempitsky. Speeding-up convolu-tional neural networks using fine-tuned cp-decomposition. InICLR, 2015.

[2]	Jean Kossaifi, Antoine Toisoul, Adrian Bulat, Yannis Panagakis, Timothy M. Hospedales, Maja Pantic; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020, pp. 6060-6069

Attributes:	kernel_shapeint tuple shape of the kernel weight parametrizing the full convolution rankint rank of the CP decomposition

Methods

`get_decomposition`([return_bias])	Returns a CP Tensor parametrizing the convolution
`init_from_decomposition`(cp_tensor[, bias])	Transpose the factors from a CP Tensor to the factorized version
`init_from_random`([decompose_full_weight])	Initialize the factorized convolution’s parameter randomly
`init_from_tensor`(kernel_tensor[, bias, …])	Initialize the factorized convolutional layer from a full tensor
`transduct`(kernel_size[, mode, padding, stride])	Transduction of the factorized convolution to add a new dimension

init_from_random(decompose_full_weight=True)[source]

Initialize the factorized convolution’s parameter randomly

Parameters:	decompose_full_weightbool If True, a full weight is randomly created and decomposed to intialize the parameters (slower) Otherwise, the parameters are initialized directly (faster) so the reconstruction has a set variance.

init_from_decomposition(cp_tensor, bias=None)[source]

Transpose the factors from a CP Tensor to the factorized version

Parameters:	factorscp_tensor

init_from_tensor(kernel_tensor, bias=None, decomposition_kwargs={})[source]: Initialize the factorized convolutional layer from a full tensor

get_decomposition(return_bias=False)[source]

Returns a CP Tensor parametrizing the convolution

Parameters:	return_biasbool, default is False if True also return the bias
Returns:	weights, factors, bias

transduct(kernel_size, mode=0, padding=0, stride=1)[source]

Transduction of the factorized convolution to add a new dimension

Parameters:	kernel_sizeint size of the additional dimension modewhere to insert the new dimension, after the channels, default is 0 by default, insert the new dimensions before the existing ones (e.g. add time before height and width) paddingint, default is 0 strideint: default is 1
Returns:	self