I did not see an uber thread for cats and dogs redux Kaggle competition, so here it is.
Question: Looks like my model is very over confident. I see this when I draw a histogram of probabilities. I have not tried updated VGG (with batch norm) and ensembling yet. But, wondering if you all also see an overly confident model?
I’m not sure that it is overly confident - it is genuinely a very easy dataset if you start with an imagenet model. Have you tried clipping it to, say [0.001,0.999] and submitting it to Kaggle?
@jeremy: Actually, I knew it is overconfident because confusion matrix says so (There are false predictions with very high probability)