haoyulance / sc-gan Goto Github PK

Hi,

This implementation looks very good, but I'm having a bit of trouble getting the generator to train on an ADNI dataset for PET prediction. I had to turn attn off because of memory constraints on my GPU, but after around 8 epochs the generators loss spikes into the thousands and then quickly both models stop learning. What would you recommend I play around with to try and stabilize the training? The loss weights?

How do you run inference on a trained model?

In my preprocess procedure of MRI,the skull is stripped from the source image space.And the MRI images are aligned to MNI152_T1_1mm_brain with the FLIRT linear registration algorithm. How can I register the PET and MRI？

Hi,

On line 97 of your training.py script, why do you divide by 2 at the end? I understand dividing by three because you're adding up the means of your three slice planes, but where does the 2 come from? Also, couldn't you just calculate

np.sqrt(np.mean(np.square(masked_output-masked_output)))

or even better

np.sqrt(np.mean(np.square(temp_pet[mask]-output[mask])))

Hello, Your work is great! I have a question. My data size is (131,114,103),but you set the input size to 256, how should I determine the input image size and training size?
Looking forward to your reply！