This notebook explains the code for this tool. In short, it uses a deep convolutional network, with following properties:
Choice between 120 possible dog breeds. Top 1 accuracy is 87.0%, top 3 97.1% and top 5 98.1%
Transfer learned on the
Xception model, a deep convnet trained on ImageNet.
With additional convolutional and dense layers to customize the problem on dog breed classification.
Trained on Google Colab, with data augmentation on the fly - e.g. skewing, rotating, flipping the dog images.