A PyTorch CV model based on a pretrained EfficientNetB2 model to classify 101 types of food.
Project directory:
├── demos
│ ├── foodvision-big
│ ├── foodvision-mini
|
├── foodvision
│ ├── 09_PyTorch_Model_Deployment.ipynb
│ ├── data_setup.py
│ ├── engine.py
│ ├── get_data.py
│ ├── model_builder.py
│ ├── models
│ │ └── foodvision_tinyvgg_model.pth
│ ├── train.py
│ └── utils.py
├── LICENSE
├── README.md
└── requirements.txt
In foodvision/foodvision, there are functions to set up data and train a PyTorch model. Additionally, in model_builder.py, I replicate the architechture of the TinyVGG model, another famous CV model, although this architechture isn't used later on.
These functions are used in foodvision/foodvision/09_PyTorch_Model_Deployment.ipynb to train an EfficientNetB2 model using transfer learning. In this notebook, I also train a ViT model, but the foodvision demo ultimately uses EfficientNet because it's faster, and only slightly less accurate.
Finally, foodvision/demos has the code to build the apps demos, which are available at huggingface. foodvision-mini can only classify three different foods (sushi, pizza and steak), while foodvision-big can classify 101.