Pokemon Classifier

• Pokémon Classifier built with TensorFlow.js - Reognizes 50 Pokémon with 85% accuracy!

Setup Dependencies

• run npm install

Download the MobileNet Model:

• make sure, jq, gnu-sed, parallel are installed - otherwise, install them with brew
• run the following:

mkdir models
mkdir models/mobilenet
cd models/mobilenet
curl -o model.json https://storage.googleapis.com/tfjs-models/tfjs/mobilenet_v1_0.25_224/model.json
cat model.json | jq -r ".weightsManifest[].paths[0]" | sed 's/^/https:\/\/storage.googleapis.com\/tfjs-models\/tfjs\/mobilenet_v1_0.25_224\//' |  parallel curl -O

Add data

• create folders ./data/pokemon/test & ./data/pokemon/train and add subfolders with samples images for each class

Configure Pipeline

• open buildModel.js and add folder names of classes that should loaded into the model
• set hyperparameters:
  • for ~10 classes: Learning Rate = 0.0003, Epochs = 20, Batch Size = 32, Dense Units = 100
  • for ~50 classes: Learning Rate = 0.0005, Epochs = 20, Batch Size = 16, Dense Units = 500

Run Project

• node run buildModel.js to run train, evaluate and save model
• node run buildModel.js path/to/an/image to load saved model and predict class of image