A food vision project made by PyTorch and with implementation of MLOps.
The project structure is inspired by cookiecutter-data-science.
Right now, MLOPs is at level 0. The model is trained on
It's also hosted on gradio space
docker pull pascaldev/nourish:latest
docker run -d --name test -p 80:80 pascaldev/nourish
then access to http://127.0.0.1/docs and click on predict and try it on
You can simply send a POST request to http://127.0.0.1/predict with the body
{
"file": "LOCAL_IMAGE_FILE_PATH"
}{
"label": "PREDICTED_LABEL"
}- PyTorch (for model training) & Torchvision (for data loading)
- DVC (Data Version Control) for data versioning
- WandB (Weights and Biases) for experiment tracking
- Tox (for testing on multiple python versions)
- Docker (for containerization)
- Pillow (for image processing)
- Add CI/CD
- Add Streamlit
- Add Pytest
The dataset used is Food-101 dataset. It is a dataset of 101 food categories, with 101,000 images. For more information, visit here.
The model can differentiate between food and non-food images.
Right now the model is trained on 30% of the dataset, and the accuracy is 56% on the test set.