DermaMNIST_Training

DermaMNIST_Training

Directory containing the code for reproducing our benchmark experiment results.

Directory Structure

• medmnist_corrected: The new MedMNIST package, updated to include DermaMNIST-C and DermaMNIST-E at both 28 × 28 and 224 × 224 spatial resolutions.
• models.py: Model definitions for ResNet-18 and ResNet-50 architectures. Same as the original MedMNIST implementation.
• train_and_eval_pytorch.py: Training and evaluation script for MedMNIST datasets. Same as the original MedMNIST implementation.
• train_and_eval_pytorch_corrected.py: Updated training and evaluation script to account for the new DermaMNIST-C and DermaMNIST-E datasets. This file is the same as train_and_eval_pytorch.py except for 3 changes:
  • L#7: medmnist changed to medmnist_corrected.
  • L#15: medmnist changed to medmnist_corrected.
  • L#54: Commented out.
• run_28.sh: Run script for all the benchmark experiments on DermaMNIST, DermaMNIST-C, and DermaMNIST-E datasets at 28 × 28 spatial resolution.
• run_224.sh: Run script for all the benchmark experiments on DermaMNIST, DermaMNIST-C, and DermaMNIST-E datasets at 224 × 224 spatial resolution.

From medmnist to medmnist_corrected

This code is built on medmnist v2.2.2. The only changes from medmnist to medmnist_corrected are as follows:

• __init__.py: Added L#6-L#7.
• __main__.py: Replaced all references of medmnist to medmnist_corrected.
• dataset.py: Added L#225-L#235.
• evaluator.py: Changed medmnist to medmnist_corrected on L#9.
• info.py: Changed URL on L#25 and added L#466-L#565.

To use the DermaMNIST-C and DermaMNIST-E datasets, please copy the corresponding .npz files from OneDrive (DermaMNIST-C, DermaMNIST-E) into your local ~/.medmnist/ directory. This directory should have been created when you installed medmnist.