Detail works please refer tianchi forum. This repo is a simplified seresnet-50 single model, which trained in 128 image size,achieved 0.87+ in four test set in average.
To achieve our final stage2-a 0.887, stage2-b 0.884 score, you may need add the following works:
- Extarct seresnet-50-128 model's fc feature of trainval data then use kmeans to get 5 folds, which can relieve heavy repeatness.(It seams 4 thousands up in single model)
- Train simplified res50 seres50 xcep incepres in scale [112,128,144,160] and ensemble.
- Predict softmax label of all testset, use threshold >0.85 to keep some semi-data, add to trainset and finetune or retrain the models.
mkdir ./data
mkdir ./preprocess_dir
(put the h5 raw file into ./data)
python preprocess.py
python train.py