Challenge DLMI 2024

Code for the contribution of Diffused Burgers to the Kaggle challenge Classification of lymphocytosis from white blood cells (see https://www.kaggle.com/competitions/dlmi-lymphocytosis-classification).

Installation

• Install Python 3.11
• Install the package pip install -e .
• Unzip the raw data in data/raw

The main function is in src/dlmi/__init__.py

Use

• Start mlflow: run the command mlflow_run (you can access the web app at http://localhost:5001/)
• Launch the program (training+inference on test) from the main folder of the project: dlmi_train
  • You can specify the config file dlmi_train --config-name [train_mlp,train_moe]: some examples are in src/dlmi/conf
• Results are in the outputs folder, the prediction for the test set is in submission_test_final.csv

Tested on Windows + Nvidia RTX 4080 12GB and Linux + Nvidia V100S 32GB. Training time of less than 30 minutes on V100S.

Main models

• Simple MLP on clinical attributes
• Mixture Of Experts MLP on clinical attributes + CNN on images

Results

See report and leaderboard (#2 ex-aequo in the private leaderboard).

Troubleshooting

• MiniDataset has not been updated to the latest version of the code (with Stratified K-Fold). Please use only MILDataset in the configs.
• ResNet is the default backbone in the code. Switching to ViT can be done by uncommenting code in src/dlmi/models/moe_model.py. ViT was more computationnally-intensive and did not provide better results, hence our choice.