Hi,
I have completed the assignment for Week 3. However, when I trained the U-Net for 40 epochs, as suggested in the notebook for better results, I saw a shocking drop in the model accuracy. I am attaching the screenshot of the graph plotting the model accuracy.
Is there any specific reason why this may have happened?
Does this accuracy correspond to training or validation set?
If shuffling would give new images to the model that the model was not trained on before for some reason, it would drop its accuracy or there may be some other reason (in the pipeline of data feed, maybe some broken file) that for that epoch the accuracy dropped so much. But I see its picking up again coming to a similar level!
I ran into this issue/phenomenon too-- You might wish to reference this post, there is quite a bit of discussion about it there:
