Image Classifier with TensorFlow2 and Google Cloud Platform

Goal

The goal of this repo is to learn two things:

1. How to train a neural network on the cloud.
2. How to serve the model on the cloud.

Hence, in this project, I will:

1. Create a training pipeline using transfer learning with a custom dataset.
2. Train the model on Google Cloud AI Platform
3. Create a web app that interacts with the model prediction
4. Deploy the web app through Google Cloud Run.

File Structure

The trainer folder contains scripts and dockerfile for the training pipeline. For more detail, checkout trainer/README.md.

The app folder contains the web app for classifing images using the model prediction. For more detail, checkout app/README.md.

Cloud Computing

Data

For simplicity, the dataset I used for this project is food-11, containing 16,643 food images grouped in 11 major food categories. You can download the dataset at Kaggle, which originates from EPFL-Food Image Dataset.

The dataset comes in three folders, Evaluate, Train, and Validation. There are, in total, 11 categories. Each image file has a naming format of <class_id>_<img_id>.jpg. For example, 0_1.jpg means the first image in class 0 (Bread). I further processed the dataset by grouping them into classes to simplify later operations.

Original Dataset	Preprocessed Dataset
/food-11 |_/evaluate | |_/0_1.jpg | /0_2.jpg | ... |_/training |_/validation	/food-11 |_/evaluate | |_/Bread | |_/0_1.jpg | /0_2.jpg | ... |_/training |_/validation