Help a Hematologist Out Challenge
This repository is based on the Image Classification and Regression Framework by Helmholtz Imaging! You can check it out here: https://github.com/MIC-DKFZ/image_classification
Similar frameworks are available for semantic segmentation and object detection!
For reproducing our model training you need to place the splits.json in the data directory and then run all 5 folds:
python main.py ResNet34 --data AcevedoMatek --num_classes 11 --augmentation bg --scheduler CosineAnneal --metrics f1 f1_per_class --batch_size 128 --num_workers 32 --epochs 500 --lr 0.1 --sampler weighted_and_random --balanced --random_prob 0.66 --fold 0 --save_model --chpt_name RN34_fold0
...
python main.py ResNet34 --data AcevedoMatek --num_classes 11 --augmentation bg --scheduler CosineAnneal --metrics f1 f1_per_class --batch_size 128 --num_workers 32 --epochs 500 --lr 0.1 --sampler weighted_and_random --balanced --random_prob 0.66 --fold 4 --save_model --chpt_name RN34_fold4
For inference run:
python predict.py --cp_dir path/to/checkpoint_dir --data_dir path/to/hematology_data --save_dir path/where/to/save/results --set val
If you want to predict the test data use --set test instead.
Acknowledgements
This Repository is developed and maintained by the Applied Computer Vision Lab (ACVL) of Helmholtz Imaging.