Exact Compression

Exact-Compression is an approach to compress a rectifier network while exactly preserving its underlying functionality with respect to a given input domain if some of its neurons are stable.

Implementation

Requirements

1. Python 3
2. Pytorch
3. Torchvision
4. Cuda 8.0 or higher
5. Gurobi 7.51

Installation

We assume that cuda-8.0 is present on your machine.

Python

conda create -n lossless python=3.6.3
conda activate lossless

pip install -r requirements.txt

Gurobi

Gurobi is not available on the public PyPI server. The following command instructs the Python interpreter to download gurobipy from the private PyPI server

python -m pip install -i https://pypi.gurobi.com gurobipy

a. Visit Gurobi Website and request for an academic license here

c. Then, type the following

/opt/gurobi_server900/linux64/bin/grbgetkey xxxxxxxx-xxxx-xxxx-xxxx-xxxxxxxxxxxx

The grbgetkey program will prompt you to store the license file on your machine.

Directory structure

We need to make some extra directories to store the dataset models

cd $PROJECT_DIR
# For storing train datasets
mkdir data
# This directory stores the models of the training in its sub-directories
mkdir -p model_dir  

The directory structure should look like this

./LLC-NeurIPS-21-code/
├── data
│   ├── cifar-100-python
│   └── MNIST
├── model_dir
│   ├── CIFAR100-rgb
│   ├── CIFAR10-rgb
│   └── MNIST
│   ...

Training the models with a specific regularisation

python train_fcnn.py --arch fcnn2d --save-dir model_dir/CIFAR10-rgb/dnn_CIFAR10-rgb_400-400_0.000175_0003 --l1 0.000175 --dataset CIFAR10-rgb --eval-stable

Get stably active/inactive neurons with the proposed new approach

This will invoke the Gurobi optimizer to list out all stable neurons in a file stable_neurons.npy inside the respective model directory.

python get_activation_patterns.py -b --input model_dir/CIFAR10-rgb/dnn_CIFAR10-rgb_400-400_0.000175_0003/weights.dat --formulation network --time_limit 10800 --dataset CIFAR10-rgb --preprocess_all_samples

Get stably active/inactive neurons with the baseline old approach

python get_activation_patterns.py -b --input model_dir/CIFAR10-rgb/dnn_CIFAR10-rgb_400-400_0.000175_0003/weights.dat --formulation neuron --time_limit 10800 --dataset CIFAR10-rgb

Get a compressed model from the stable neuron file

This will generate a new checkpoint file pruned_checkpoint_120.tar of the updated weights and biases after compressing the network according to stable_neurons.npy inside the specified model directory.

python prune_network.py model_dir/CIFAR10-rgb/dnn_CIFAR10-rgb_400-400_0.000175_0003

Evaluate the uncompressed model

python train_fcnn.py --arch fcnn_prune --resume model_dir/CIFAR10-rgb/dnn_CIFAR10-rgb_400-400_0.000175_0003/pruned_checkpoint_120.tar -e --dataset CIFAR10-rgb

Evaluate the compressed model

python train_fcnn.py --arch fcnn2d --resume model_dir/CIFAR10-rgb/dnn_CIFAR10-rgb_400-400_0.000175_0003/checkpoint_120.tar -e --dataset CIFAR10-rgb

We provided the script 'llc2e.py' with the config of different combination of the netwrok architecure, regularization. Please feel free to play around it.

Our main contribution is the new approach to get the stable neurons for a linear network. We thus provided a few models to make it easy to play with pruning networks.

model_dir
├── CIFAR100-rgb
│   └── dnn_CIFAR100-rgb_400-400_7.500000000000001e-05_0001
├── CIFAR10-rgb
│   └── dnn_CIFAR10-rgb_400-400_0.000175_0003
└── MNIST
    ├── dnn_MNIST_100-100_0.0_0000
    └── dnn_MNIST_200-200_0.0_0001

Citing Exact-Compression

Please cite Exact-Compression in your publication if it helps your research:

@article{serra2021scaling,
  title={Scaling Up Exact Neural Network Compression by {ReLU} Stability},
  author={Serra, Thiago and Yu, Xin and Kumar, Abhinav and Ramalingam, Srikumar},
  journal={NeurIPS},
  year={2021}
}