Hi Andrea,

Thanks for the comment.

Re: image resolution — I actually did another technique here which I forgot to include in the article — progressive resizing. The images in the dataset are 308x168, so I started with 54x54 for the first half of training, then increased it to 108x108. I got 100% there so I never ended up running full size of 168x168 (if this was for production I would do that next).

The idea is to force the CNN to pick up on larger themes first by feeding it smaller images, and thus build a model that should generalize better on unseen data, and then increasing image size and thus allowing greater details to be included.

Hope that helps!