hpml-project: Pruning for model compression

Group project component for High Performance Machine Learning done in Columbia University for Fall 2023.
Work done by Jit Soon Foo, Tri Le

This project looks at applying pruning to a Mask-RCNN network to reduce the model size.

Dataset: Penn-Fudan Database

This repository looks at application and effect of pruning on a Mask-RCNN network for pedestrian detection.

Setting up -

1. Google colab

Studies performed -

We studied two main pruning algorithms, PyTorch's pruning and Torch-pruning.

Steps to reproduce experiment-

1. Follow tutorial in (revised)torchvision_finetuning_instance_segmentation.ipynb to perform transfer learning from Faster RCNN to Mask RCNN. Then follow up with running Pytorch's pruning on the network.

2. Perform pruning with Torch-Pruning in Torch-pruning_noRetrain.ipynb and Torch-pruning_retrain.ipynb. Results are logged in wandb (https://wandb.ai/hpml-f23-proj/hpml-project-torchpruning_retrain?workspace=user-jf3482 and https://wandb.ai/hpml-f23-proj/hpml-project-torchpruning?workspace=user-jf3482).