So, in the top 27% with a weeks work, isn’t so bad, however it isn’t so good either. It did though give me some useful experience in convolutional hyperparameter optimization with a pre-trained weights ResNeXt 50 with a sigmoid (two-class logistic regression) replacing the softmax classifier (top=False). Pretty simple stuff,
The legendary (late 2014) K80 is dropping rapidly in price. Often at around UK 300 pounds and not much more in the US in dollars amazing to see a 24 GB card at the price of a GTX 1080, more so considering the street price back then was over $4,500.
I am going to be controversial here, because my advice runs contrary to most peoples, but it may be useful to some. The usual advice is to code up an MNIST CNN and see it work its magic, well it may well be magic to you, hence a step back