This is an exemplary project on how to solve an image classification problem but in a proper manner.
The project includes following features:
- Batch data preprocessing based on Python data generators
- Use of Python data generators to read and preprocess data
- Employment of tf.Data.Dataset API for providing data for model training
- Use of Jupyter Notebooks for both data exploration and model evaluation (incl. class activation maps)
- DVC for pipeline designing and execution, see dvc.yaml
- Parameter definition in an external file params.yaml
- Storage of metrics in an external file metrics.json
Thus, this project after some adaptation is supposed to be a template for simple ML problems in the field of CV.
The project expects raw data in a certain format. Each instance should consist of an image and a json file with metadata (incl. label). Currently, a publicly available dogs vs cats dataset is used (copyrights: https://www.microsoft.com/en-us/download/details.aspx?id=54765). To download already properly formatted data please contact @vkhaydarov
The downloaded files need to be put in folder 'data' (this folder must be created by the user).
To be able to run the pipeline specified in dvc.yaml you need first install DVC. After that you can run the entire pipeline executing the command:
dvc reproDVC expects the user to have the project folder as a working directory. Therefore, if you need to run single scripts in Python, please change the working directory of the interpreter to be the project directory.