1
$\begingroup$

I'm working on a undergraduate project with using deep learning. Currently, I'm trying to improve a model by modifying it. Model is Double U-Net and dataset that I'm using is DRIVE dataset. It consists from 20 images. My problem is validation score does not increase anymore. I think there is a overfit problem and also, I heard that is the max potential of network so I should try another dataset or model. Data separated 80% training and %10 validation. How can I improve result?

What I tried:

-Added one more U-Net network
-Using (128, 64, 32, 16) filter size
-Using (64, 32, 16, 8) filter size
-BinaryCrossEntropy loss function
-Added dropout exit of encoders with 0.5 value
-Added top, left, right and bottom symmetry augmentation for images, so total number of augmentation is 25

Augmentations (left-top is original image): enter image description here

Results:

enter image description here enter image description here enter image description here enter image description here enter image description here enter image description here

$\endgroup$

1 Answer 1

0
$\begingroup$

The U-Net is not overfitting it is just the training graphs are not smooth. This jittering in the training graphs can be removed by tweaking the learning rate of the model.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.