Q1:¶
Dataloader:¶
Training our ConvNet¶
Examine the results¶
This is the results of our neural net architecture (a variant of LeNet) on a subset of random faces of both the training and test datasets.
**Red is prediction of net and green is true label
We see on images from the eval dataset where the heads are angled performed poorly. We likeley need data augmentation to help the net generalize better.
These two that turned out poorly look quite different than the "average man" we calculated earlier and are very much so out of distribution. This is evaluation data so our classifier never saw these two before so the baldness and the wild messy hair probably threw our classifier for a loop.
Notice how the two with great performance look quite "normal" in the sense they are closer to the average man.