Skip to content

Latest commit

 

History

History
executable file
·
46 lines (34 loc) · 1.57 KB

Readme.md

File metadata and controls

executable file
·
46 lines (34 loc) · 1.57 KB
Raw

Sports Gesture Classification on Docker

Overview

  • This is an API for sports gesture image classification

  • Features

    • The model use Swin Transformer (accuracy = 80% on validation dataset)
    • Run training and flask API on Docker container
    • Provide onnx (CPU device)

  • Demo https://youtu.be/2DWbemtVgis

Run

  • Build a docker based on DockerFile
  • Train: bash train.sh
  • Flask api: bash api.sh
    • Browser: http://localhost:12000/

Dataset

Kaggle Sports Gesture

  • Classes: 100
  • Size: 224 X224 X 3
  • Format: .jpg
  • Number of images: 14572 (train: 4572 , valid: 500, test: 500)

Train

The training code is from Kaggle Sports Gesture Competition: SwinTransformer from Timm

python3 train.py --config {config_file}
  • Check the process on Tensorboard: tensorboard --logdir = sports_api

Run on Docker

Steps about Run on Docker

Refer

  1. Inspired from sportsnoma classification: https://www.youtube.com/watch?v=Kzrfw-tAZew https://github.com/abhishekkrthakur/sportsnoma-deep-learning
  2. Kaggle: Sports Gesture Classification
  3. Kaggle Sports Gesture Competition: SwinTransformer from Timm
  4. Pytest on Flask API