TransFetch Knowledge Distillation

Dependencies

• Python 3.8
• CUDA 11.8+
• Pytorch 2.0.0

Running Program

• Import conda environment with conda env create -n transfetch-env --file transfetch-env.yml
• Change LoadTraces_ROOT, OUTPUT_ROOT, and GPU_ID in run.sh
• Set MODEL in run.sh to either "d", "r", or "v" indicating DenseNet, ResNet, or ViT
• ./run.sh
• Export cross-platform conda environment if packages changed with conda env export --from-history>transfetch-env.yml

Notes

• Hard coded dimension or DIM parameter for ResNet as 64, ResNet_tiny as 4

Potential Bugs

• Early stop is 10 for both the training of teacher and student