condensed-sparsity

Getting Started

This repository contains the source code for the Structured RigL (SRigL) a dynamic sparse training (DST) algorithm based on RigL. SRigL learns a sparse neural network with a hybrid fine-grained and structured (neuron) sparse topology. The fine-grained structure is obtained by enforcing a constant fan-in constraint for each layer in the neural network and is a particular case of N:M sparsity. The representation learned by SRigL is not only memory and parameter efficient, it also is amenable to acceleration on commodity hardware.

SRigL maintains generalization performance comparable with RigL and other unstructured sparse DST algorithms up to a high sparsity level. A key innovation required to match unstructured sparse performance is to enable neuron ablation; i.e., we fully mask select neurons based on the aggregated salience of its constituent weights.

We currently support the following dataset/model pairs:

• MNIST
  • Small four layer model w/ two convolutional layers and two fully connected.
• CIFAR-10:
  • Wide ResNet-22
  • ResNet-18, modified for CIFAR-10
• ImageNet
  • ResNet-50
  • MobileNet, Large and Small
  • ViT-B-16

Repository Structure

• /configs/* configuation .yaml files for running each of the dataset/model paris above. We use hydra to parse the configs in our main script.
• ./train_rigl.py is the main training script.
• ./src/rigl_torch contains the source code for SRigL.
  • ./src/rigl_torch/rigl_scheduler.py contains source code for RigLScheduler, an implementation of RigL in pytorch.
  • ./src/rigl_torch/rigl_constant_fan.py contains the source code for RigLConstFanScheduler aka: SRigL.
• ./src/condensed_sparsity/condensed_linear.py contains the source code for our naive pytorch GEMM implementation that leverages the constant fan-in structure learned by SRigL.
• ./configs/config.yaml contains the settings for the various hyperparameters and runtime options.

Example Commands

• ResNet50 trained on Imagenet using 4-GPUs on a single node:

python ./train_rigl.py \
    dataset=imagenet \
    model=resnet50

• Resnet18 trained on CIFAR-10:

python ./train_rigl.py \
  dataset=cifar10 \
  model=resnet18

• ViT/B-16 trained on CIFAR-10:

python ./train_rigl.py \
  dataset=imagenet \
  model=vit

Installation

This project was developed using Python version >=3.10 and uses poetry==1.6.1 to manage dependencies and build the project.

Installation instructions are provided for virtual environments, Compute Canada clusters, and Docker:

Virtual Env

    python -m venv .venv
    source .venv/bin/activate
    pip install --upgrade pip
    pip install poetry==1.6.1
    poetry install -vvv  # With install all dependency groups
    git submodule update # Pull in cocoapi for detection / segmentation
    cd ./src/cocoapi/PythonAPI
    make & make install
    cd ../../..  # back to workspace dir
    pre-commit install-hooks  # For development

Docker

For reproduction and instantiating replica's during training / inference, please use Dockerfile to build the image. Alternatively, you can pull the image from Docker Hub. A large shm-size is required for pytorch to train ImageNet as this directory is used by the dataloader workers

Replica / Reproduction Container

docker build --file ./Dockerfile -t rigl-agent --shm-size=16gb .
docker run -itd --env-file ./.env --mount source=/datasets/ILSVRC2012,target=/datasets/ILSVRC2012,type=bind --gpus all --shm-size 16G rigl-agent:latest

Development Container

For development, we recommend using vscode's devcontainer functionality to build and launch the development container. A devcontainer.json schema is provided in ./.devcontainer/ and if the project working directory is opened in vscode, the application will prompt the user to reopen in the development container. Please refer to the devcontainer.json schema and Dockerfile.dev for specifics on the development container environment and build process.

To get started, please complete the following steps before reopening the workspace in the devcontainer:

• Copy the .env.template file to your own .env environiment file and edit it to add environmental variables. Without the .env file the dev container will not start.
• Create a directory /datasets and place any datasets you want to use (except for CIFAR-10) in that location. Alternatively, edit the mount directories in ./.devcontainer/devcontainer.json
• Run git submodue update to pull from cocoapi third party repo. Migrate to ./src/cocoapi/PythonAPI and run make && make install

Pretrained Model Checkpoints:

A variety of model's pretrained with SRigL are available on huggingface.

Compute Canada

Compute Canada pre-builds many python packages into wheels that are stored in a local wheelhouse. It is best practice to use these wheels rather than use package distributions from PyPI. Therefore, the dependencies pinned in pyproject.toml have been carefully selected to ensure that the project enviornment can be replicated using the Compute Canada wheels that will match a local enviornment using PyPI package distributions.

For simplicity, a bash script for installing the project and dependencies is included, see: ./scripts/build_cc_venv.sh. Simply run this script from the project working directory after cloning the project from github.

Tests

This repository uses pytest.

Run tests using pytest