Kaggle-Bengali-2020

75th place solution (Top 4%) that I didn't choose it as a final score with a private score of 0.9361 for Bengali.AI Handwritten Grapheme Classification

Contents

• Setup
• Overview
• Solutions
• Training
• Submission

Setup

Python 3.6+

To install the virtual environment:

$ pipenv install -r requirements.txt --python 3.6

or

$ pip install -r requirements.txt

Optional: To install apex, please check this github.

Overview

This repo contains many different approaches of implementing classification models for Bengali. For each approach, it's been briefed in this following table:

Experiments	Backbone	LB Score	Private Score
Grapheme	ResNext50
Vowel + Consonant	STN+ResNext50	0.9665	0.9361
Grapheme + Vowel + Consonant	SeResNext50	0.9776	0.9242
Grapheme + Vowel + Consonant	SeResNext50 + DenseNet169	0.9770	0.9261

The highest private score is the experiment of separated training between Grapheme and Vowel + Consonant.

bengali/grapheme: is grapheme model
bengali/vowel_consonant: is vowel and grapheme model with Spatial Transformer Network
bengali/grapheme_vowel_consonant: is the model with 3 heads

The reason behind training them separately is to avoid overfitting of Vowel and Consonant, also to tweak the accuracy of Grapheme independently.

Solutions

Preprocessing

Use this public kernel to process the images size of 224x224

Augmentation

For Spatial Transformer Network, I didn't do any extra augmentations.

import albumentations as A

# for other experiments
# A.CoarseDropout or Cutmix
# A.ShiftScaleRotate
# A.RandomBrightnessContrast
A.Compose(
    A.Normalize(mean=(0.0692, 0.0692, 0.0692),
                std=(0.2052, 0.2052, 0.2052)),
    ToTensorV2(),
)

Grapheme Model

• Backbone: Resnet50
• Loss: CrossEntropy
• Optimizer: SGD
• Scheduler: Cosine Annealing
• Epochs: 80 if Cutmix else 40-60

Vowel+Consonant Model

• Backbone: STN+ResNext50
• Loss: CrossEntropy
• Optimizer: SGD
• Scheduler: Cosine Annealing
• Epochs: 80-100

Training

Before training a model, you might need to customise the arguments in run_grapheme.sh and run_vowel_consonant.sh to fit with your system.

To train models:

$ bash /path/to/project/run_grapheme.sh
$ bash /path/to/project/run_vowel_consonant.sh

Submission

Code for submission has been written in this kernel