Requirements

python 3.7, anaconda or miniconda, https://github.com/Kaggle/kaggle-api, https://github.com/trent-b/iterative-stratification

Dataset:

kaggle competitions download -c bengaliai-cv19

Conda Env

conda create --name bengali-ai python=3.7 pandas pillow
conda install -c trent-b iterative-stratification
conda install -c conda-forge pyarrow tqdm imgaug
conda install albumentations -c albumentations
pip install pretrainedmodels
## For CPU
conda install pytorch torchvision cpuonly -c pytorch
## For GPU with cuda 10.0
conda install pytorch torchvision cudatoolkit=10.0 -c pytorch

Reference

https://www.youtube.com/watch?v=8J5Q4mEzRtY

Steps

1. mkdir input and pip install --user kaggle and add path using export PATH="$HOME/.local/bin:$PATH" and add kaggle credential using https://github.com/Kaggle/kaggle-api/blob/master/README.md
2. Download dataset in input folder using kaggle competitions download -c bengaliai-cv19 command
3. unzip it using unzip bengaliai-cv19.zip
4. run create_folds.py
5. check the .parquet data file using check_dataframes.py, if data is readable then conintue
6. Create .pickle files of the dataset using create_image_pickles.py. This is because training will be faster with pickles
7. run run.sh file to train the model. Change the training configurations on run.sh
8. model will be saved in src folder
9. Use the notebook located in infer.zip for kaggle submission and upload your model to the notebook

Todo Features

1. AUC/ROC
2. T-SNE

Todo Experiments

1. Use different loss functions like center-loss
2. Use different feature extractors like arcFace