This repository houses a set of utilities to help with training a neural network for image classification with TensorFlow 2 mainly utilizing transfer learning. The main focus of the project is retrain.py; this script, as the name suggests is used for the process of training itself. The goal of the file is to provide as much customization of the training procedure as possible from the comfort of the command-line interface.
dataset_formator servers the purpose of reformating the directory structure in which files are formated. When given a folder where each subfolder represents a class in a dataset, the script links the files into new directory structures split into separate folders for different phases in training (eg. training and validation) according to the ration given by the user. The program is also capable of oversampling underrepresented classes and similarly to retrain.py, this script attempts to CLI to its advantage.
The project also contains mute_tensorflow_warnings.py which is gets imported before Tensorflow in retrain.py. This file just executes the necessary command to mute the annoying warnings and information that TensorFlow likes to clutter the console with. If you wish to debug feel free to just remove the file and its import (this should be the first file imported).
You simply execute the file and add appropriate arguments. To see the list of arguments please use the -h flag on any of the scripts.
Make sure you have Python 3 and pip installed. Then run # python3 -m pip -r requirements.txt. Consider doing this in a virtual environment so that you don't have version conflicts. Then make sure you have the nescessary software to allow your GPU to be used by TensorFlow. If you don't have a dedicated GPU, uninstall tensorflow-gpu and install tensorflow-cpu. Then you should be good to go.
I am not interested in receiving contributions of any kind at the moment, but you are welcome to fork the project and do whatever you want with it.
The whole repository is licensed under the MIT License.