The project aims at reproducing the results of the Soft Masking for Cost-Constrained Channel Pruning (SMCP) method, proposed by Humble et al. (2022) in the paper "Soft Masking for Cost-Constrained Channel Pruning" on a scaled-down architecture, specifically: a ResNet-18 model trained and tested on the CIFAR10 dataset. We utilized the code for the pruning developed by the authors, available at this Github repo.

The main goals of our project are:

• Apply the SCMP method on a scaled-down architecture, and on a smaller dataset, and compare the results to those obtained in the paper.
• Test different hyperparameters settings, specifically for pruning-ratio, rewiring frequencies, and learning rate schedule, compare the results.
• Running an ablation study by removing the warmup epochs.

For more information, check the full blog post at: https://francescoopiccoli.github.io/smcp-DL-project/

Marcus Plesner

Francesco Piccoli (ID: 5848474)

Aranya Sinha

Download the training.ipynb Jupyter Notebook and upload it on Colab, alternatively run it in your editor by removing colab-specific code cells, download also the repository of the authors and place it in the same folder as the notebook, in case you are running from colab, upload the original authors' repository folder on your google drive.

Results obtained for different pruning ratios.