I am training a 3D segmentation model using V-Net in keras (tf backend). The number of binary classes = 1 (foreground =1 and background = 0). Original VNet paper.
I experience a weird behavior in the training graph. It hardly learns anything for the first 800 epochs, but then shoots up and over fits. I find this weird. Also, after trying multiple different types of augmentations, the model hardly improves. Can anyone explain this behavior to me? And ways to improve my model?
I have tried drop put layers, different augmentation techniques, nothing seems to work.
Input image - 128 X 128 X 96
Optimizer - SGD
Lr = 0.01
loss - Dice coeff * -1
metric - Dice coeff
training batch size = 1
Number of training imgs = 16
No. of validation imgs = 3
The training graph:
