MiniML

This repository contains some of my experiments and trials with Deep Learning models that do not fit well in their own repositories. They are implemented using Pytorch Lightning to simplify training and remove boilerplate.

Installation

Clone the repository using:

git clone https://github.com/williamcorsel/miniml

Install dependencies using:

pip install -r requirements.txt

Usage

Training, evaluating, and testing of the models is implemented using the Pytorch Lighting CLI. This can be used to easily train arbitrary models (implemented as ModelModules) on arbitrary datasets (implemented using DataModules). The basic recipe is as follows:

python miniml.py {fit,validate,test,predict} --model model_name --data data_name

It also allows you to set hyperparameters using the CLI or yaml files. Some example configurations are provided in the configs folder. The can be used as follows:

python miniml.py {fit,validate,test,predict}  --config path_to_config.yaml

Capabilities

Image Classification

Contains simple versions of the LeNet and ResNet models to train on the CIFAR10 dataset. The datasets and predictions of the models can be visualised using FiftyOne by using:

python visualise_image_classification.py --ckpt_path path_to_checkpoint.ckpt --data_dir path_to_data --name fiftyone_dataset_name

3D Model Generation

Proof of concept of a 3D-GAN model to generate 3D Minecraft houses from the Craft3D dataset