Woah! Struggling to understand all the high level APIs of TensorFlow.
So, I tried writing a cat-or-dog classifier, using a CNN. I mostly copy pasted code from everywhere and tried stitching it with whatever I understand. Code runs! I've downloaded some 25,000 pics of cats and dogs and ran the model on some 1000 of them.
Now I have a model that overfits with test accuracy of 100% but while evaluating for non-test images, I get less than 50% accuracy :-/ Must understand how to tweak a model.
More than that, I don't really understand how high level APIs work. So, will spend a day or 5 on this.
Mood: code, and more code.