NVlabs/RADIO

Any normalization required for the input image?

SeanGuo063 opened this issue · 2 comments

Any normalization required for the input image?

Hello, the RADIO models expect their inputs to be in the range of [0, 1].

Normalization is taking place within the forward() method of the model in the input_conditioner module. There are examples in the /examples folder if you wish to take a look.

If you need to integrate the RADIO model within an existing pre-processing pipeline that already normalizes inputs to zero mean, unit variance, then one trick you can apply is to overwrite the input conditioner in the model and make it an identity layer:

radio_model.input_conditioner = torch.nn.Identity()

That's very helpful. Thank you very much!