Shape of the accuracy curve

Hello, I was just wondering as to the reason why the accuracy curve has a sudden drop in the shape? Is this a random phenomena, or can it be reproduced by retraining the U-net. Unfortunately, the time constraints forbids me from doing the latter. Could it be that the loss function got stuck in a local optima?

Thanks for your replies.

Hi soumdtt,

The accuracy curve I get looks like this:

It looks like something interfered with your results.

Hi reinoudbosch. I think that may be the most reasonable cause. Thanks for the answer.

Sharing my encounter for the future reference of the others.

I got the similar drop in my first run as well. I’m rather sure my code was correct so I restarted the kernel and reran a few times. I ended up with a “normal looking” curve afterwards.

The notebook (v2) doesn’t standardize a seed, including when shuffling the training set. So it may as well be an unlucky instance of randomization.