Method: Filter Decomposition

Model Compression

The goal of model compression is to achieve a model that is simplified from the original without performance deterioration. By compressing the large model, the user can reduce the storage and computational cost and allow to use in real-time applications.

NetsPresso supports the following compression methods.

  • Structured Pruning
  • Filter Decomposition

This page describes for Filter Decomposition.


What is "Filter Decomposition"?

Filter decomposition is to approximate original weights into lightweight representations via low-rank approximations to reduce the computational cost.

Supported method

Model Compressor supports the three following methods, Tucker Decomposition, Singular Value Decomposition (SVD), and Canonical Polyadic (CP) Decomposition.

1) Tucker Decomposition
Tucker decomposition decomposes the convolution with a 4D kernel tensor into two factor matrices and one small core tensor.

  • In Channel : The number of input channel in each layer.
  • Out Channel : The number of output channel in each layer.
  • In Rank : The number of input channel of core tensor that represent relation level of low-rank factor matrix.
  • Out Rank : The number of output channel of core tensor that represent relation level of low-rank factor matrix.

2) Singular Value Decomposition (SVD)

Singular Value Decomposition (SVD) decomposes the pointwise convolution or fully-connected layer into two pointwise or fully-connected layers.

  • In Channel: The number of input channels of the target layer
  • Out Channel: The number of output channels of the target layer
  • Rank: The condition number of weight matrix W

3) Canonical Polyadic (CP) Decomposition
CP Decomposition replaces an original tensor with a linear combination of rank one tensors.

  • In Channel : The number of input channel in each layer.
  • Out Channel : The number of output channel in each layer.
  • Rank : A sum of N-way outer products of rank-one tensor for estimating original convolution filter.

If you want to know more about filter decomposition methods:

"Recommendation" in Model Compressor

The "Recommendation" allows the user to set proper compress hyperparameter of the filter decomposition, in-rank, and out-rank.

  • "Recommendation" calculates the ranks based on the Variational Bayesian Matrix Factorization (VBMF)
  • Users can manipulate the compression amount by using the calibration ratio of the Recommendation.
  • "Recommendation" is only available for Tucker Decomposition and Singular value decomposition methods. (CP Decomposition will be available soon.)

Detailed information of the Calibration ratio

  • The default value of the calibration ratio is 0.0, which returns the same rank as VBMF
    • The smaller the calibration ratio(>= -1, <=1), the smaller the model will be returned.
  • Example of how the rank is decided by the calibration ratio.

What you can do with Model Compressor

  • User can compress the model using one of the filter decomposition methods.
  • After selecting the layer to be compressed, user fills in the rank value in the range. (available rank values are different in each layer)
  • To use Tucker Decomposition, user have to input two values, in/out rank
  • To use CP Decomposition or Singular Value Decomposition, only one value, rank is needed.
  • If user is hard to input proper values, you might as well use recommendation function.
  • 'Recommendation' function based on VBMF algorithm and user can control the compression level through 'calibration ratio' additionally.

๐Ÿšง

Not Supported Layers

  • Group convolutional layer is currently not supported and will be updated in the near future.