Some layers in DiffusionModelUNet are detached and zeroed

[AbrightWay]

Hi folks,

I am encountering the output of Unet model from DiffusionModelUNet being a zero tensor no matter what input I feed into it. Therefore, I look into the source code of the model and found that the convolution of the final layer of the residual block of UNet is a zero module (

MONAI/monai/networks/nets/diffusion_model_unet.py

Line 393 in 57fdd59

self.conv2 = zero_module(

)
Also, the output convolution of UNet itself is a zero module too (

MONAI/monai/networks/nets/diffusion_model_unet.py

Line 1720 in 57fdd59

zero_module(

).

Could you please explain why zeroing out these layers, which makes the overal UNet output to be zero. This is quite counterintuitive to me.

Thank you in advanced!