Hello everyone! I have tried implementing the fastai code for CycleGAN here. My TIFF images are 1 channel and are medical images (one is PET, the other is MRI). I’ve managed to get training started, but I notice that the loss is astronomically high:
What could be the reasons behind this? I’m guessing it’s because the values in my images might not be normalised to be between -1 and 1. I noticed the documentation for the Custom ItemList says “Here the object is the tuple of images and the data their underlying tensors normalized between -1 and 1.” with the following code:
class ImageTuple(ItemBase):
def init(self, img1, img2):
self.img1,self.img2 = img1,img2
self.obj,self.data = (img1,img2),[-1+2img1.data,-1+2img2.data]
I’m not very sure how the normalisation process works – where do the values get normalized to between -1 and 1?
Also, as I mentioned earlier I am working with two different image modalities – PET and MRI. The image values in the PET are in the range of around 0 to 2 whereas the values in the MRI can go up to 300. Would the difference in values affect the normalization process? If yes, how can I scale the values of both images to be between -1 and 1?
Thank you very much for your help!