I worked through lessons 1 and 2. It was amazing to see how easy it is to create a dataset with google-images and some helper functions from fastai. I had to try that on my own:
In german, there is the proverb “… thats like comparing apples and pears”. Basically saying, that you should not compare the two. Since I’m not a huge fan of such rules, I will do exactly that! (Anarchy!!!)
Google-searching for “apfel” (german for apple, to sidestep the problem of the brand with the same name) and “birne” (german for pear) and hand-cleaning the dataset (FileDeleter-like widgets don’t work on colab), I ended up with 309 pictures.
From there on I followed the cookie-cutter recipe from lesson 1 to train a classifier.
- Training only the last layer brought me to 9% error rate in 2 iterations. No improvement from there on.
- Unfreezing and training all the layers sometimes gets the error to 8%, but it jumps back to 9% from time to time
9-10% error on my ~60 images validation-set means 6 images were wrong. Looking at those images, I have to say that I would only have gotten 3 right:
So all in all it was a fun little exercise, I’m really looking forward to the next lessons!