I am fairly new to deep learning and I have been testing out several architectures for the segmentation task of clouds in satellite imagery. I am using a simple Unet as my benchmark, Unet++, Efficient Net, Unet with a Resnet Backbone. And for the sake of experimenting, I branched out to transfer learning using VGG16 and ResNet34 (by removing the classifier and adding a decoder). My problem is the learning (train and validation) and the loss curves, they are very similar for all architectures and showing a spiky behavior as in the figures bellow. For the loss functions I tried -log (Dice), BCE, DiceBCE, and it's always the spiky behaviour. For the Evaluation Metrics, I am using IoU, Dice coeff, Precision, and Recall. I tried using both the lr=1e-4 and lr=1e-5 as well as increasing the batch size. But nothing seems to work. My dataset consists of 10000 images of size (3, 384, 384). Any suggestions on how I can improve on this? or what could I have possibly done wrong?

enter image description here

enter image description here


Your Answer

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

Browse other questions tagged or ask your own question.