Question about usage

[achel-x]

Dear author,

I hope this message find you well. Thank you for sharing the excellent project.
I have encountered a few problems while using DCNv4, and I would greatly appreciate your guidance.

1. When processing features with 128 channels, the training seems fine, but during inference, I observed some values are NaN.
   Specifically, my work is related to learning-based image codec. The results of my experiment are rate (bits per pixel) and distortion(PSNR and MS-SSIM). The PSNR and MS-SSIM value are NaN as shown in bellow.
   

Additionally, when handling 192 channels feature, NaN is appeared, which seems to be a gradients vanishing. I am very confused about these phenomenon. Any insights on how to effectively use DCNv4 in other works?

2. I also noticed that DCNv4 has many parameters, and I'm not sure how to set each one. If you could provide more detailed explanations, I would be very grateful.
   

3. Lastly, I would like to confirm whther DCNv4 can only run on GPU and is not currently supported on CPU. If so, are there any plans for CPU support in the future ?

Thank you so much for your time and assistance!

Best regards.