search

nobodyplayer1 / VM-UNetV2

Apache License 2.0
81 stars 11 forks source link

Computation complexity of different models. #1

Closed godlin-sjtu closed 7 months ago

godlin-sjtu commented 7 months ago

To whom it may concern,

Nice job! And I have a small question about the results of Table 3 in your paper. It shows that the Flops and Params of VM-UnetV2 are lower than those of VM-Unet（I guess it refers to 'VM-UNet: Vision Mamba UNet for Medical Image Segmentation'). However, in my view, this paper seems to add an extra SDI module to VM-Unet, and why have these two metrics decreased?

I am looking forward to your response, many thanks.

nobodyplayer1 commented 7 months ago

Thanks for the question, it's a good one. Here is the answer: the parameter amount is smaller because we are not directly applying the decoder module in VM-Unet.As you can see in the open source code, our four layers output is directly from the encoder part of the four layers through an SDI module.（If there is still any doubt about this we will upload a file again in the near future to clarify this matter, in the meantime you can check our code in the file models/vmunet/vmunet_v2.py）

godlin-sjtu commented 7 months ago

Thanks for your answer, I have checked the paper and code again, and my issue is resolved.