Bengali.AI

This is the combined effort to build a pipeline in PyTorch to attempt Kaggle competition for Bengali.AI

In this competition, we are given the images of a handwritten Bengali grapheme and are challenged to separately classify three constituent elements in the image: grapheme root, vowel diacritics, and consonant diacritics.

Project writeup: https://julienbeaulieu.github.io/2020/03/16/building-a-flexible-configuration-system-for-deep-learning-models/

Implementation Highlights

YAML config files, Mixup & albumentation library augmentations, label smoothing, OHEM, label weights for class imbalance, progressive resizing, OneCycleLR, Tensorboard. Models: MobileNet_V2, DenseNet121, SE_ResNeXT50

Project Organization

Based on Facebook Research Detectron2 project and Cookie Cutter data science

├── README.md          
├── data
│   ├── external        <- Data from third party sources.
│   ├── interim         <- Intermediate data that has been transformed.
│   ├── processed       <- The final, canonical data sets for modeling.
│   └── raw             <- The original, immutable data dump.
│
├── notebooks           <- Jupyter notebooks for exploration and communication
│
├── experiments         <- Trained models, predictions, experiment configs, Tensorboard logs, backups
│   ├─ exp01
│   │  ├── model_backups
│   │  ├── results
│   │  └── tensorboard logs
│   │
│   ├─ exp02
│   ...
│
├── requirements.txt   <- The requirements file for reproducing the analysis environment,
│                         generated with `pip freeze > requirements.txt`
│
└── src                <- Source code for use in this project.
    ├── config
    │   ├── config.py  <- The default configs for the model. 
    │   └── densenet121
    │       ├── exp01_config.yaml <- Configs for a specific experiment. Overwrites default configs
    │       └── exp02_config.yaml
    │       
    ├── data                  
    │   ├── make_dataset.py  <- Script to generate data
    │   ├── bengali_data.py  <- Custom Pytorch Dataset & Collator class; build_dataloader function
    │   └── preprocessing.py <- Custom data augmentation class
    │
    ├── modeling         <- Scripts to create the model's architecture             
    │   ├── backbone     <- Model's backbone architecture
    │   │   └── densenet121.py
    │   │
    │   ├── layers       <- Custum layers specific to your project
    │   │   └── linear.py
    │   │
    │   ├── meta_arch    <- Scripts to combine and build backbone + layers + head
    │   │   ├── baseline.py
    │   │   └── build.py
    │   │
    │   ├── head         <- Build the head of the model - in our case a classifier
    │   │   ├── build.py
    │   │   └── simple_head.py
    │   │
    │   └── solver       <- Scripts for building loss function, evaluation and optimizer
    │       ├── loss
    │       │   ├── build.py
    │       │   ├── softmax_cross_entropy.py
    │       │   └── label_smoothing_ce.py
    │       ├── evaluation.py
    │       └── optimizer.py 
    │ 
    ├── tools            <- Training loop and custom tools for the project
    │   ├── train.py
    │   └── registry.py 
    │ 
    └── visualization  <- Scripts to create exploratory and results oriented visualizations or to store reports
           └── visualize.py