vague chasm
Hello. I am currently training a UNet to segment areas of cancer (by Gleason grade) in histology images. I am stuck because my model is overfitting a lot and doesn't seem to make much progress. I have tried adding more data, checking the data that I currently have, and class balancing the data. None of this has improved the model, and I'm not sure what else to do. Does anyone have any advice? The graphs from my latest run are below.
bright mica
Maeby try some data augmentation. For example flip and rotations.
bright mica
You could also try lowering the learning rate
violet granite
^^ i like the second idea
flat scroll
Hello. I am currently training a UNet to segment areas of cancer (by Gleason gra...
what activation are you using ?
and are you sure the problem is indeed overfitting ?
vague chasm
what activation are you using ?
I'm using ReLU
vague chasm
and are you sure the problem is indeed overfitting ?
I assumed it is always overfitting when the training set does much better than validation. Is there something else it would be instead?
vague chasm
You could also try lowering the learning rate
Thanks for the advice! I will try it
flat scroll
I'm using ReLU
try normalizing the images and then use LeakyReLU
helps for faster convergence and better learning
another thing that might help is preprocessing your images. Idk what images you have, but some sort of filters / edge detection etc might help draw out the features more
which will also help the model learn
flat scroll
Maeby try some data augmentation. For example flip and rotations.
data augmentation for sure is a good solution to consider, but the current train accuracy itself is around 40%, the model is not even learning the train data well...
flat scroll
I assumed it is always overfitting when the training set does much better than v...
also you can be sure of overfitting when you're training acc is pretty high and testing acc is low.
in your case however, the train accuracy itself is low, so the model is finding it difficult to learn whatever you're asking it to learn. Once the model performs better on the train data, we can then go from there
that's a first ive heard lol
vague chasm
I’ve had that problem once before. ChatGPT said it could be normal and caused by things like dropout.
reef lily
what is "accuracy"
it would be helpful to know what your loss function is