Skip to content

Latest commit

 

History

History

README.md

Experiment Examples

This directory contains 5 complete, runnable examples that can reproduce key experiments from the paper. Each example is self-contained with its own shell script and can be executed independently to verify the pipeline works correctly. We have added experiments for the MIDOGpp dataset as its the latest mitotic figure dataset and can be downloaded easily. The examples use the ViT-S model because it is available to everyone and is computationally more feasible. If you want to try out LoRA fine-tuning with the foundation models, make sure you have access to them via HuggingFace and that it is supported in src/classifier.py.

Overview of Examples

Example Dataset Method Training Sizes No. repititions Purpose Runtime
01_MIDOG2022_vit_mini_example_linear_probing MIDOG2022 (debug subset) Linear Probing 100% 1 Quick sanity check ~Minutes
02_MIDOG2022_vit_full_example_linear_probing MIDOG2022 (full training) Linear Probing 0.1%, 1%, 10%, 100% 5 Full linear probing experiment ~Hours
03_MIDOGpp_vit_full_example_linear_probing MIDOGpp Linear Probing 0.1%, 1%, 10%, 100% 5 Full linear probing experiment ~Hours
04_MIDOGpp_vit_example_lora MIDOGpp LoRA Fine-tuning 0.1%, 1%, 10%, 100% 5 Standard LoRA fine-tuning Long
05_MIDOGpp_vit_example_cross_domain_lora MIDOGpp LoRA Cross-Domain Full dataset 5 Tumor-specific LoRA fine-tuning Longest

Quick Start

Each example follows this pattern:

  1. Make executable: chmod +x run_*.sh
  2. Run: ./run_*.sh
  3. Check results: Look for results.pkl in the output directory

More detailed descriptions can be found at each subfolder.

Prerequisites

  • Check the prerequisited at the main page
  • Download the required datasets using the google-drive link or the python scripts

Expected Outputs

Each example produces a single results.pkl file containing aggregated metrics suitable for plotting/comparison.