ivanpanshin/VinBigData_MMDetection

Inference

In order to get final predictions of cascade models for VinBigData Chest X-ray Abnormalities Detection competition on Kaggle, run the following:

Install requirements

python3 -m venv venv
source venv/bin/activate
./setup.sh

Download model weights + test annotations for inference
```
./download_weights.sh
./download_test_data.sh
```
Run inference
```
./inference_cascade.sh
```
In the end, you'll end up with 2 submissions under final_subs directory. If you don't wish to run the inference yourself, you can simply download these submissions here.

In order to run the training from the beginning, run the following:

Download checkpoints + train annotations for training
```
./download_checkpoints.sh
./download_train_data.sh
```
Run 1 stage of training
```
./train_1stage.sh
```
Select the best weights for 5 folds, based on AP@0.4 (AP@0.5 should work as well), and place them into weights/cascade_r50_augs_with_empty/fold$i.pth directory.
Run 2 stage of training
```
./train_2stage.sh
```
Select the best weights for 5 folds, based on AP@0.4 (AP@0.5 should work as well), and place them into weights/cascade_r50_augs_rare_with_empty/fold$i.pth directory.

In the end, you'll end up with weights that you can use for ./inference_cascade.sh.