Hi, I am working on a Segmentation task with Unet. I am using a cucle step learning schedule with cycle length =50 and CrossEntropy Loss.
I am using IOU as a metric to evaluate performnace.
I obtained the following learning curve that is not a sign of overfitting in my opinion. But I can not understand the reason of difference between Train and Val loss. and if there is a problem in training how can I fix it.
Another question is that the leaning curve shows training is not finished yet and I should leave it to be trained for longer epochs, right?
My training set has 740 images and valset has 267. It is cityscapes dataset. Since the dataset is really large I chose a smaller subset of training and validation set for training.
thank you in advance.
