My model from lesson 1 is too good, what could I have done wrong?

I’m a little behind in running through all the notebooks and writing code - but I just finished playing around with resnet50 from lesson 1 - and after about 8 epochs I get 0 error rate, my confusion matrix on the test set has 0 misclassifications.
For reference - I took some 400~ images of one class, 170 images of another and randomly pulled out 30% of each for a test set before I did anything. (randomly might be a stretch, I used the shuf command)
the notebok is here

Obviously a model can’t be perfect, so what mistakes am I making - is the test set just so small that I’m getting artifically good numbers?

1 Like

looks good to me

I do not see anything wrong in the notebook, and you are getting 0% error with resnet50 but not with resnet34, so it may just be luck?
Did you get the same perfect accuracy in multiple re-runs of the notebook, for different selections of the test set?
I like the classification problem you did. It looks like the algorithm has to distinguish between head-legs-legs-head (roughly how heelhooks look like to me) and legs-…-legs-head.
By the way, I did something similar, classifying judo matches vs taekwondo matches, and got about 5% error with resnet34 in 10 epochs. Your problem looks harder!