This is a project for the course "Machine (Deep) Learning for physicists 2023-24 " taught by Prof. Rudo Roemer at university of Warwick. It is a deep learning model for classifying generation one Pokémon from images.

Dependencies

A working local environment with the following libraries:

• Tensorflow for python.
• Standard python libraries: NumPy and Matplotlib
• The Kaggle API for downloading the data sets

User guide

Data set generation

To download the data set and split it into training, validation and testing data sets run python generate_dataset.py. A local installation of the Kaggle API is necessary. 1.31 Gb of data will be downloaded from Ben Hawks' data set. Additionally, a small data set of 4Mb from user Vishal Subbiah is downloaded to provide a clean image for each Pokémon, for presentation purposes.

Model generation

To train the model, run python generate_model.py. This can take some time if a GPU is not used so pre-trained models are already included in the models/ directory. Once the model is generated, the history plots of fitting will be saved in the generated plots/ directory.

A model based on transfer learning from the VGG16 model can also be generated with python generate_transfer_model.py.

The model generation takes a couple of minutes to run in an NVDIA RTX A1000.

Testing

Saved models can be tested against the testing dataset using the test_model.py script. To use the models for prediction run python predict_pokemon.py <path_to_image>. The script predict_random_pokemon.sh loops through random Pokémon in the testing dataset and uses the model indicated in predict_pokemon.py.

In these python scripts, the variable model_name might need to be changed to use different models. By default, it is set to 'final_cnn_model', corresponding to the default name of the convolution neural network model. See which models are available in the models directory.