mrgemy95/tensorflow-project-template

This repository is a starter template for building deep learning projects with TensorFlow. Rather than structuring each project from scratch, the template provides a consistent folder layout and a set of base Python classes you extend with your own code. The goal is to reduce the repetitive setup work that comes with starting a new neural network project so you can focus on the parts specific to your model. The template is organized around five components: a model class, a trainer class, a data loader, a logger, and a configuration file. Each has a base version you extend by inheriting from it in Python. The model base class handles saving and loading checkpoints, tracking the current training epoch and step, and defining the structure your model must fill in. The trainer base class wraps the training loop, and you override the training step and epoch logic with your own implementation. The data loader handles reading and preparing your training data. The logger connects to TensorBoard, a tool for visualizing training metrics like loss and accuracy over time as charts. The template also supports an optional integration with Comet.ml, a platform for tracking experiments, hyperparameters, and metrics across multiple runs from a web dashboard. Configuration is handled through JSON files, where you set values like the number of training epochs and the batch size. The main entry point wires all the components together, creates a TensorFlow session, and starts the trainer. The workflow throughout is the same: inherit a base class, override the specific methods with your logic, and connect the pieces in a main file. An example implementation for a convolutional neural network is included to show how the pattern works in practice.