Week 3 Assignment 2 (UNet) - 4.3 - Predicted mask

I see that the predicted masks for all distinct input images are the same (always) and do not resemble the true mask, after running show_predictions() and despite passing the assignment. Is there some step I might have missed?

Hi Shijo,

For the predicted mask to resemble the true mask, the model would have to be trained much, much, much longer. To limit the amount of time needed to finish the assignment, the model is undertrained. The predicted mask in the assignment only provides an indication of the outcome had the model been trained sufficiently long.

2 Likes

Thanks for the response, Reinoud!
I see, makes sense now. I was misled by the assignment statement - “With 40 epochs you get amazing results!” - then. With 40 epochs although I get ~70% training accuracy, all 6 predictions seem incorrect (maybe if I try more images I should get some correct results).
In any case, after training for 100 epochs instead, I get ~90% accuracy and when I run predictions for 20 images, all of them seem to be correct (meaning, resemble the True Mask), and that’s great!