Goal

The goal of this program is to decompose each convolutional layers in a model to reduce the total number of floating-point operations (I'll use the shorthand flops) in the convolutions as well as the number of parameters in the model.

Contributions

This is an extension of the work https://github.com/jacobgil/pytorch-tensor-decompositions. In this implementation, everything, including finding the ranks and the actual CP/Tucker Decomposition, is done in PyTorch without switching to numpy.

CNN architecture decomposed

• [x] AlexNet
• [ ] VGG
• [x] ResNet50

Dataset used

• ImageNet ILSVRC2012 dataset

Usage

python3 scripts/decomp.py [-p PATH] [-d DECOMPTYPE] [-m MODEL] [-r CHECKPOINT] [-s STATEDICT] [-v]

• PATH specifies the path to the dataset
• DECOMPTYPE is either cp (default) or tucker
• If a model is already decomposed, it could be passed in as the MODEL parameter (By default, the Torchvision pretrained ResNet50 is used).
• If continue a fine-tuning from a checkpoint, pass in the checkpoint as CHECKPOINT
• To specify the parameters for the model, use STATEDICT
• [-v] option for evaluating the inference accuracy of the model without fine-tuning

Pre-decomposed and fine-tuned model

A pre-decomposed ResNet50 is included in the models directory as resnet50_tucker.pth.

The fine-tuned parameters for the model is the resnet50_tucker_state.pth in the models directory.

Results

It turn out that Tucker decomposition yields lower accuracy loss than CP decomposition in my experiments, so the results below are all from Tucker decomposition.

AlexNet

ResNet50

References

• CP-decomposition with Tensor Power Method for Convolutional Neural Networks Compression: https://arxiv.org/abs/1701.07148
• Compression of Deep Convolutional Neural Networks for Fast and Low Power Mobile Applications: https://arxiv.org/abs/1511.06530
• PyTorch CP and Tucker decomposition: https://github.com/jacobgil/pytorch-tensor-decompositions
• VBMF code: https://github.com/CasvandenBogaard/VBMF
• Tensorly: https://github.com/tensorly/tensorly

Any comments, thoughts, and improvements are appreciated