POKE TYPER CNN

by Bryan Yee

HOW THE MODEL WAS MADE

Training Dataset Details:

- Small dataset of 721 images each with 18 possible types as one-hot vectors
- Random data augmentation with rotations, reflections, brightness, etc.
- Imbalanced distribution of Pokemon types
    - 118 water types (~16%)
    - 37 ice types (~5%)

Validation Dataset Details:

- Webscraped dataset of 280 fan-made pokemon (credits to phoenixsong)

v SOURCE LINKS v

[kaggle images][kaggle data]

[phoenixsong fakemon data]

Building The Model:

> Step 1: Base model = EfficientNetB0 (pretrained on ImageNet)
> Step 2: Base frozen (no training on convolutional layers)
> Step 3: Added GlobalAveragePooling -> Dropout -> Dense(18, sigmoid)
> Step 4: Loss = BinaryCrossEntropy (multi-label setup)
> Step 5: Primary metric = AUC (captures performance across thresholds)

Refer to: [Keras Transfer Learning Guide]