Question regarding normalization input

Jan_Huiskes · February 10, 2020, 2:14pm

Hi,

When the Unet is trained, the input is normalized and with the mean and std of this input, the target is also normalized. Why do you do this? I am making a neural network that circumvents the Inverse fast Fourier transform (IFFT), so that you can just use one big neural network on the masked images. However, I don’t know what to do with the normalization of the images, since it happens right after the IFFT. Maybe you can help me out?

tullie · February 10, 2020, 6:58pm

Hi @Jan_Huiskes ,

The normalization helps stabilize training but is not required. When evaluating the reconstructions they need to be unnormalized to the original scale anyway, (see: https://github.com/facebookresearch/fastMRI/blob/master/models/unet/train_unet.py#L188).

Thanks,
Tullie