MUSCO: MUlti-Stage COmpression of neural networks

This repository contains supplementary code for the paper Automated Multi-Stage Compression of Neural Networks. It demonstrates how a neural network with convolutional and fully connected layers can be compressed using iterative tensor decomposition of weight tensors.

Requirements

numpy
scipy
scikit-tensor-py3
absl-py
flopco-pytorch
tensorly==0.4.5
pytorch

Installation

pip install musco-pytorch

Quick Start

from torchvision.models import resnet50
from flopco import FlopCo
from musco.pytorch import CompressorVBMF, CompressorPR, CompressorManual

model = resnet50(pretrained = True)
model_stats = FlopCo(model, device = device)

compressor = CompressorVBMF(model,
                            model_stats,
                            ft_every=5, 
                            nglobal_compress_iters=2)
while not compressor.done:
    # Compress layers
    compressor.compression_step()

    # Fine-tune compressed model.

compressed_model = compressor.compressed_model

# Compressor decomposes 5 layers on each iteration.
# Compressed model is saved at compressor.compressed_model.
# You have to fine-tune model after each iteration to restore accuracy.

Please, find more examples in musco/pytorch/examples folder

Compress the model

You can compress the model using diffrenet strategies depending on rank selection method.

• Using any of the below listed compressors, you can optionally specify:

  • which layers will NOT be compressed (ranks = {lname : None for lname in noncompressing_lnames})
  • how many layers to compress before next model fine-tuning (ft_every = 3, i.e. compression schedule is as follows: compress 3 layers, fine-tine, compress another 3 layers, fine-tune, ... )
  • how many times to compress each layer (nglobal_iters = 2, by default 1)

• CompressorVBMF: ranks are determined by aglobal analytic solution of variational Bayesian matrix factorization (EVBMF)

  • Tucker2 decomposition is used for nn.Conv2d layers with kernels (n, n), n > 1
  • SVD is used for nn.Linear and nn.Conv2d with kernels (1, 1)
  • You can optionally specify:
    • weakenen factor for VBMF rank(vbmf_weakenen_factors = {lname : factor for lname in lnames})

• CompressorPR: ranks correspond to chosen fixed parameter reduction rate (specified for each layer, default: 2x for all layers)

  • Tucker2/CP3/CP4 decomposition is used for nn.Conv2d layers with kernels (n, n), n > 1
  • SVD is used for nn.Linear and nn.Conv2d with kernels (1, 1)
  • You can optionally specify:
    • which decomposition to use for nn.Conv2d layers with kernels (n, n), n > 1 (conv2d_nn_decomposition = cp3)
    • parameter reduction rate (param_reduction_rates argument), can be different for each layer

• CompressorManual: manualy specified ranks are used

  • Tucker2/CP3/CP4 decomposition is used for nn.Conv2d layers with kernels (n, n), n > 1
  • SVD is used for nn.Linear and nn.Conv2d with kernels (1, 1)
  • You can optionally specify:
    • which decomposition to use for nn.Conv2d layers with kernels (n, n), n > 1 (conv2d_nn_decomposition = tucker2)
    • which ranks to use (ranks = {lname : rank for lname in lnames}, if you don't want to compress layer set None instead rank value)

Citing

If you used our research, we kindly ask you to cite the corresponding paper.

@inproceedings{gusak2019automated,
  title={Automated Multi-Stage Compression of Neural Networks},
  author={Gusak, Julia and Kholiavchenko, Maksym and Ponomarev, Evgeny and Markeeva, Larisa and Blagoveschensky, Philip and Cichocki, Andrzej and Oseledets, Ivan},
  booktitle={Proceedings of the IEEE International Conference on Computer Vision Workshops},
  pages={0--0},
  year={2019}
}

License

Project is distributed under Apache License 2.0.