Pokemon Type Classifier

This project aims to classify the primary and secondary types of a given pokemon from its image using convolutional neural networks.

Dataset

The image dataset comes from veekun containing pokemon sprite images from games spanning Generation I to V, as well as icons, Global-Link artwork, and the spinoff Pokemon Conquest.

The pokemon stat and type data comes from kaggle and contains id, stats, type, and legendary status information on all the pokemon to date.

Additionally, in /data/type_labels are the type labels of all 18 types for use in visualization and the GUI. The labels are extracted from the main games, and are copyright of Nintendo Inc.

Package requirements

To run the code, use python3, and ensure the following packages are installed:

Tensorflow

Keras

Matplotlib

Numpy

Pandas

scipy

tkinter

scikit-learn

There have been some issues with loading the models in the anaconda environment due to serialization, but packages installed with pip3 should work fine. If an issue occurs, one can always retrain the models by running python3 cnn.py

Running the code

Preprocessing The Data

preprocess.py contains the necessary functions to preprocess the image and pokemon type data necessary for training. Running

python3 preprocess.py

each time will create new sets of training and test data taken from the data/ directory divided into primary and secondary types in type1_sorted and type2_sorted respectively. preprocess.py also contains an image resizing function by Hemagso taken from his repo here. The necessary functions to extract pokemon name and type information from the pokemon csv dataset from kaggle are all contained in this program as well.

Data Visualization

visualization.py contains the necessary functions to visualize the dataset as well as various model performance metrics. Running the program will print some same plots. However, these functions are intended for use in other parts of the project as well as for generating figures for the final writeup. Some sample plots are shown below:

Model Training

cnn.py contains the functions to build, train, and save the convolutional neural network models. Running

python3 cnn.py

will begin training the primary and secondary type models on the test and training data sets. The classifiers, after training for 20 epochs, will be saved to \model as .h5 files if s = True in the main method. Optional training accuracy and loss charts will be generated by matplotlib if plt = True in the mian function. Further, if plt = True, an optional .png representation of the model architecture is saved to \model such as:

All boolean values s and plt are set to True by default.

Running Predictions and Evaluations on the Models

run.py contains the necessary functions to load the model and preform predictions and evaluations. Running

python3 run.py

will perform load the saved models and if p = True in the main method, the program will perform primary and secondary type predictions on a random pokemon image from the dataset such as:

Then, if e = True in the main method, the program will then perform evaluations on the test sets using the evaluation_generator function from Keras to return the test accuracy and loss in the writeup. Furthermore, to further aid in assessing the performance, the program will compute the precision, recall, and f-score metrics on the entire test set and print to the terminal screen. Note that these statistics are computed using the metrics module from sklearn as keras does not support computing these values. Note that the prediction is done on each image once to generate a numpy array to be passed into the classification_report and accuracy_score functions from sklearn.metrics.

All boolean values p and e are set to True by default.

Extension: Fun Pokemon Type Classification Game!!!

To aid in visualizing the project, we created a GUI that will pit the player against the model in predicting the type of random pokemon. As an advantage for the player, there will only be four options to choose from whereas the model must infer the type from all 18 possiblities. We hope this can be a fun way to evaluate the model performance and serve as a homage to the pokemon games of our childhood.

The GUI and game was made using the Tkinter package in python3. To play, simply launch:

python3 GUI.py

Gotta classify 'em all!

Note that in some environments there were compatibility issues between matplotlib and tkinter causing the game to crash. Some lines have been added to GUI.py to fix these issues. If they cause any problems in other machines, consider removing the two lines denoted in the code.