Image Pipeline

This repository contains an image machine learning pipeline and some of the projects i've worked on, either for fun/education or competition on Kaggle.
Each project have it's own readme containing information about the specific problematics of each.

I train the models locally on my pc using a Nvidia 1080 GPU.

Data

The data is not in the repository directly if you want to launch a model on one the projects in here you must download the data and change the config file in the task folder to be adequate.
Links to the datasets are in the tasks README's.

Projects

---

The projects folder contains the specific code about each project :

• config.py file containing most of the parameters necessary to train a model.
• augment.py that contain the specific augmentations you want to perform on the images for training, validation and testing

What if i want to add a new project ?

To add a new project you'll to create a few things :

• a new folder in the projects/ folder containing a __init__.py file, a config.py file and a augment.py file
• copy and paste the content of another config.py file and change the information to be adequate with your task
• Add the augment you require to the augment.py file

Training

To start training a model on any task use this command in terminal :

python -m train --project=AERIAL_CACTUS --run_note=test

You can replace the AERIAL_CACTUS with any folder in projects/. Default parameters train a RESNET18 model. You can change these parameters as such :

python -m train --project=LEAF_DISEASE --model_name=RESNET34 --run_note=test

The parameters can take different values : *- run_note : str, allow you to add a title to your training runs

• project : The project you want to train a model on, atm you can train a model on the aerial_cactus task, melanoma, blindness_detection & leaf_disease projects.
• model_name : You can choose any model that is in the models/ folder, name must be typed in MAJ like in the example above.

Inference

To start prediction on new data for a project you can use this :

python -m predict --project=AERIAL_CACTUS --model_name=RESNET18 --run_note=test

To do

---

• Add more models
• Add more loss functions available
• Add notebooks (for model evaluation - EDA - hyperparameter optimization etc...)
• add support for meta features