Transfer learning: Food classification using EffNet
Transfer learning is a powerful technique to improve accuracy of models. Instead of training the whole model with limited hardware, why not use a model trained on similar data and only re-train a part of it. Here we use a trained latent space that is able to extract various features from a photo, and this latent space is then used to train a much simpler classifier with much fewer trainable parameters.
I used a pretrained EfficientNet B0 and trained it on custom dataset of 3 food types: pizza, sushi, and steak. This resulted in an accuracy of approx. 95%, which is way higher than using a much simpler network from scratch.
