About this repository

This is a personal project, mostly to learn some new technologies that I am not familiar with yet such as databases, HTML/CSS/Javascript and flask.

Code structure

├───notebooks # Contains notebook for development mainly, some quick tests
├───parameters # Contains the parameters file for the experiments
├───src # Contains all the core code
│   ├───models # (will) Contain custom models and helper function related to models
│   ├───optimizer #
│   ├───transforms
│   ├───scheduler #
│   ├───losses #
│   ├───trainer # Only contains the trainer class, that handle the training steps.
├───app # Contains code for the front end and some helpers
│   ├───static # Contains CSS and Javascript code for the UI
│   │   └───javascript # Contains javascript code for the UI
│   ├───templates # Contains HTML templates files (JINJA templating)
│   ├───utils # Contains python code with helper function for the front/back end interface
├───readme_files # Contains auxiliary input for this readme.
└───tests

How to use

I did not dockerize yet, I'll update the commands once it's done. in the mean time, dependancies are basically pytorch and flask.

As a CLI

python train.py <path-to-your-parameters-json>

You can check the existing parameters.json file to see how it is done, and also refer to the get_XXX.py files to see available methods, as well as the pytorch documentation to see the parameters each function takes.

As an App

While it's still in the development process, you can try and use the app by calling

python run.py

then head to the port 8000 of your machine and you should find this UI

Flow chart

Look something like that, might add details later

TODOs

• Find a good way to get weight using path of previous training
• Add a validation section
  • Give sample of images that are wrongly classified (csv format)
  • Give metrics graphs (Confusion matrix done, probably adding some bar graphs)
  • Add grad cam or alike
  • Add perturbation and see how results evolve (Gaussian Noise, Saturation, Compression artefact)
• Write comments
• Add tests
• Frontend
  • Create a parameters parser
  • make a link to tensorboard ?
  • Store training results in an object Database (mongo-db)
  • Make the previous training visible in the training section
  • Export validation section as HTML and make it accesible in the training section
• Add a flow chart in the readme