ngocphucck/Sleeping-posture-classification

This project includes multiple models, loss functions, optimizers and image augmentations for image classification task

IR-pose-classification

Requirements

• YACS (Yet Another Configuration System)
• Torch
• Opencv-python

Installation

• Clone this repository:

git clone https://github.com/ngocphucck/IR-pose-classification
cd IR-pose-classification

• Install dependencies: Please type the command pip install -r requirements.txt.

Implementations

• Firstly, you have to prepare your annotations. I recommend that you organize your labelling file into 2 files: train.json and val.json with the form image_path: label.

• After that, you can define some parameters in your method. There are 2 options for you to do that:

  • Change parameters in defaults.py.
  • Another way is to be more flexible. You'll create YAML configuration files; typically, you'll make one for each experiment. But, when actually implementing, you need to merge this .yaml file with defaults.py. The following code makes this action:

  cfg = get_cfg('path_to_file')
  cfg.merge_from_file("experiment.yaml")
  cfg.freeze()

• Train a model:

cd tools
python train.py

Components

Backbones	Loss functions	Optimizers	Augmentations
Efficientnet ConvNeXt Ghostnet Ghost-efficientnetv2	Cross Entropy	Adam	RandAug

Future works

• Multiple backbones
• Data augmentations
• Multiple loss functions
• Experiment managment
• Edge devices deployment
• UI demo/Docker resource
• Distributed computing