Channel-based Structured Pruning Adapters Experiments

This repository contains experiments for the channel-based SPA, Structured Pruning Low-rank Adapter (SPLoRA), proposed in Structured Pruning Adapters using various structured pruning methods, for weight-based transfer learning on image classification tasks.

Installation ⚙️

• Clone this repository and enter it:

  git clone https://github.com/LukasHedegaard/channel-spa-experiments.git
  cd channel-spa-experiments

• (Optionally) create conda environment:

  conda create --name channel-spa-experiments python=3.10

• Install as editable module

  pip install -e .[dev]

Run training + pruning 🏃‍♂️

Please see the scripts folder for the scripts used in our experiments. Below, you will find the command-line arguments for selecting splora (argument: "--splora") and each of the explored pruning methods (argument:"--method-type", choices: ["weight", "grad", "taylor", "lrp"]).

For a complete list of arguments, see:

python main.py --help

Structured Pruning Low-rank Adapter (SPLoRA)

python main.py --prune --splora --splora-rank 8 --splora-init-range 1e-3

If "--splora" is not specified, fine-tuning will be used. The commands below are used to specify the pruning methods.

Layer-wise Relevance Propagation Pruning

Yeom et al., "Pruning by explaining: A novel criterion for deep neural network pruning", in: Pattern Recognition 2021

python main.py --train --prune --method-type lrp\\
    --arch resnet50 --lr 0.0025 --batch-size 64 --epochs 20 --recovery-epochs 20

Grad Pruning + L2 norm

X. Sun et al., "Meprop: sparsified back propagation for accelerated deep learning with reduced overfitting", in: International Conference on Machine Learning (ICML), 2017

python main.py --train --prune --method-type grad --norm

Taylor Pruning + L2 Norm

P. Molchanov et al., "Pruning convolutional neural networks for resource efficient transfer learning", in: International Conference on Learning Representations (ICLR), 2017

python main.py --train --prune --method-type taylor --norm

Weight Pruning + L2 norm

python main.py --train --prune --method-type weight --norm

Training options

python main.py --help

Lint project

Lint using black and isort:

make lint

Results

Below you find some results achieved using this repository.

Compared to regular structured pruning with fine-tuning, our channel-SPA improves accuracy by 6.9% on average while using half the parameters at 90% pruned weights. Alternatively, it can learn adaptations with 17x fewer parameters at 70% pruning with 1.6% lower accuracy.

Please see the paper for more details.