Summary of approach can be found under https://www.kaggle.com/c/hpa-single-cell-image-classification/discussion/238898
Run pip install -r requirements.txt for dependencies
To download necessary data you can use the kaggle API
kaggle competitions download hpa-single-cell-image-classification
unzip hpa-single-cell-image-classification.zip -d ./input/
As a preprocessing step its necessary to create single cell masks using the HPA-Cell-Segmentation, which is available under https://github.com/CellProfiling/HPA-Cell-Segmentation I made them public in the dataset https://kaggle.com/christofhenkel/hpa-single-cell-3rd-place-dieter-masks
so you can download using
kaggle datasets download christofhenkel/hpa-single-cell-3rd-place-dieter-masks
unzip hpa-single-cell-3rd-place-dieter-masks.zip -d ./input/
the folder structure should look like
hpa-single-cell-3rd-place-dieter
├── input
│ ├── train
│ ├── 000a6c98-bb9b-11e8-b2b9-ac1f6b6435d0_blue.png
│ └── ...
│ ├── test
│ ├── cell_masks_v5
│ ├── train_4folded_cells.csv
│ └── train_4folded.csv
├── models
├── data
├── configs
└── ...
as a minimum working example run
python train.py -C cfg_simple
After that the 4 basemodels can be trained by running
python train.py -C cfg_ch36_otf_ext1b_rerun
python train.py -C cfg_ch54_ext1
python train.py -C cfg_img_4_ext7
python train.py -C cfg_img_mask_4_1024_ext1_19
Weighting the 4 models out-of-fold labels can be generated and a new small model can be trained using
python train.py -C cfg_ch62_sc_ext7_clus
which has the same performance as the 4 individual models.