Cheese Classification challenge

This codebase allows you to jumpstart the INF473V challenge. The goal of this channel is to create a cheese classifier without any real training data. You will need to create your own training data from tools such as Stable Diffusion, SD-XL, etc...

Instalation

Cloning the repo:

git clone git@github.com:nicolas-dufour/cheese_classification_challenge.git
cd cheese_classification_challenge

Install dependencies:

conda create -n cheese_challenge python=3.10
conda activate cheese_challenge

Install requirements:

Install Pytorch:

If CUDA>=12.0:

pip install torch torchvision

If CUDA == 11.8

pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118

Then install the rest of the requirements

pip install -r requirements.txt

Download the data from kaggle and copy them in the dataset folder The data should be organized as follow: dataset/val, dataset/test. then the generated train sets will go to dataset/train/your_new_train_set

Using this codebase

This codebase is centered around 2 components: generating your training data and training your model. Both rely on a config management library called hydra. It allow you to have modular code where you can easily swap methods, hparams, etc

Training

To train your model you can run

python train.py

This will save a checkpoint in checkpoints with the name of the experiment you have. Careful, if you use the same exp name it will get overwritten

to change experiment name, you can do

python train.py experiment_name=new_experiment_name

Generating datasets

You can generate datasets with the following command

python generate.py

If you want to create a new dataset generator method, write a method that inherits from data.dataset_generators.base.DatasetGenerator and create a new config file in configs/generate/dataset_generator. You can then run

python generate.py dataset_generator=your_new_generator

VRAM issues

If you have vram issues either use smaller diffusion models (SD 1.5) or try CPU offloading (much slower). For example for sdxl lightning you can do

python generate.py image_generator.use_cpu_offload=true

Create submition

To create a submition file, you can run

python create_submition.py experiment_name="name_of_the_exp_you_want_to_score" model=config_of_the_exp

Make sure to specify the name of the checkpoint you want to score and to have the right model config