README

Final Report:

https://sheldonsebastian.github.io/vbd_cxr/

Directory Structure:

Path	Description
0_preprocessor	Code to convert DICOM to png and resize images.
1_eda	Code to perform EDA
2_data_split	Code to split the data into train-validation-holdout
3_trainer	Code to train classification models and object detection models.
4_saved_models	Saved models are stored here. Download trained models from here
5_inference_on_holdout_10_percent	Code to make predictions using classification, object detection, and ensemble models.
6_inference_on_kaggle_test_files	Utility files to make Kaggle submissions
7_deployment_files	Code related to Flask App
common	Utility files for making coding easier
archived	Contains Proof of Concepts and miscellaneous files for experimentation purposes
docs	files related to GitHub website
input_data	folder in which input data will be placed

Steps to replicate project:

1. Download processed data from here and download external data from here and place in root directory as "input_data" folder name.

2. To create data train-holdout split for classification and object detection models, run all scripts in 2_data_split in the order they appear.

3. Download trained models from here or run all the scripts in 3_trainer.

4. To make inference on holdout dataset using:

   a. classification models run all scripts in 5_inference_on_holdout_10_percent/1_classification_models folder.

   b. object detection model run all scripts in 5_inference_on_holdout_10_percent/2_object_detection_models folder.

   c. ensemble model run all scripts in 5_inference_on_holdout_10_percent/3_ensemble folder.

5. To make inference on kaggle test dataset run all scripts in 6_inference_on_kaggle_test_files folder.

Additional Packages required:

1. albumentations
2. pytorch
3. detectron2
4. ensemble-boxes
5. Mean-Average-Precision-for-Boxes
6. YOLOv5