jiarunliu/swin-umamba

Swin-UMamba: Mamba-based UNet with ImageNet-based pretraining

swin-umamba's Issues

Zero-shot Inference

Excuse me, is your model a large-scale model? Can the program be used for zero-shot inference?

I want to ask why it keeping on "using pin_memory on device 0 using pin_memory on device 1"？

I want to ask why it keeping on "using pin_memory on device 0
using pin_memory on device 1"？I use two GPU and its Utilization rate is all 100%，My CPU has 32 cores but its Utilization rate is 400%。

The GPU and Training time

Your innovation with Mamba is amazing. I would like to ask you about your equipment and the time of a training session.

How to calculate the flops of mamba ?

请问网络模型在哪个文件夹下？

AbdomenCT dataset

Hi！May I know why you didn't include the AbdomenCT dataset in you experiments？

Why is 'Swin' included in the name of your model?

Thank you very much for your work!

I am a bit confused and don't understand why you wrote 'Swin' in the name of your model. I haven't found any aspects where it is connected to the original Swin model, or perhaps it is some sort of abbreviation?

Key error

Excellent work, very interesting.

While running the Swin-UMambaD from pre-trained weights, I faced the issue of key error in:

Swin-UMamba/swin_umamba/nnunetv2/nets/SwinUMambaD.py

Line 669 in 50eecc8

if "patch_embed" in k and ckpt["config"].MODEL.SWIN.IN_CHANS != num_input_channels:

.
Possibly the checkpoint doesn't have the "config" key?
It would be great to have your valuable feedback in this regard.

Best,
Atik