Transfer Learning: Rose vs Tulip

The tf_flowers dataset has mainly top view roses photos. Since the petals of a rose and a tulip are similar, classifier confuses the side-bottom Rose views. But it is very easy to train it on procedural models.

Tulips with 5 vs 6 petals remake. The script generates 40 pairs of randomly rotated procedural roses and tulips (corresponding predictions 0 or 1). MobileNet (~15 MB of weights) is used to get a 1024 features map (embeddings) from a 224x224 image. The script collects embeddings in a dataset and trains one perceptron (just 1024 weights) for 20 epochs. It all takes a few seconds (HP laptop with Ryzen 5 3500U). See console for loss dynamics.