sotiraslab / agileformer Goto Github PK

Thank you so much for doing such a good job, I did get 86.1 accuracy with your base_best weights, but I used 3090 graphics (unlike your V100) and the environment was the same. I directly used your commands to train the model, but there was an error of load weight. I modified your configuration file and trained Base from the beginning, and the accuracy was only 84.7. I am not sure if there is any error in your configuration file, or could you please tell me how you trained base 85.7, thank you very much

when will you release the code?

Dear author, I am very excited after reading your paper and feel that you have written it very well. However, when reproducing the code, I encountered this issue. My natten and torch are both 2.20 and cu118, and there are no other issues. However, the natten.functional function does not include na2d_qk_with'bias

File "/root/data1/zhangxuanxuan/AgileFormer/networks/nat_2d.py", line 30, in
from natten.functional import na2d_av, na2d_qk_with_bias
ImportError: cannot import name 'na2d_qk_with_bias' from 'natten.functional' (/root/anaconda3/envs/AF/lib/python3.10/site-packages/natten/functional.py)

/home/liubn/anaconda3/envs/swin_umamba/lib/python3.10/site-packages/torchvision/io/image.py:13: UserWarning: Failed to load image Python extension: '/home/liubn/anaconda3/envs/swin_umamba/lib/python3.10/site-packages/torchvision/image.so: undefined symbol: _ZN3c1017RegisterOperatorsD1Ev'If you don't plan on using image functionality from torchvision.io, you can ignore this warning. Otherwise, there might be something wrong with your environment. Did you have libjpeg or libpng installed before building torchvision from source?
warn(
tvdcn is installed, using it for deformable convolution
Traceback (most recent call last):
File "/home/liubn/anaconda3/envs/swin_umamba/lib/python3.10/site-packages/natten/functional.py", line 32, in
from natten import libnatten # type: ignore
ImportError: /home/liubn/anaconda3/envs/swin_umamba/lib/python3.10/site-packages/natten/libnatten.cpython-310-x86_64-linux-gnu.so: undefined symbol: _ZN2at4cuda19getDevicePropertiesEl

During handling of the above exception, another exception occurred:

Traceback (most recent call last):
File "/home/liubn/0-149liubaoning/43-AgileFormer-main/train.py", line 12, in
from networks.agileFormer_2d import AgileFormer2D as ViT_seg
File "/home/liubn/0-149liubaoning/43-AgileFormer-main/networks/agileFormer_2d.py", line 3, in
from .agileFormer_sys_2d import AgileFormerSys2D
File "/home/liubn/0-149liubaoning/43-AgileFormer-main/networks/agileFormer_sys_2d.py", line 9, in
from .nat_2d import NeighborhoodAttention2D
File "/home/liubn/0-149liubaoning/43-AgileFormer-main/networks/nat_2d.py", line 30, in
from natten.functional import na2d_av, na2d_qk_with_bias
File "/home/liubn/anaconda3/envs/swin_umamba/lib/python3.10/site-packages/natten/init.py", line 24, in
from .functional import (
File "/home/liubn/anaconda3/envs/swin_umamba/lib/python3.10/site-packages/natten/functional.py", line 34, in
raise ImportError(
ImportError: Failed to import NATTEN's CPP backend. This could be due to an invalid/incomplete install. Please uninstall NATTEN (pip uninstall natten) and re-install with the correct torch build: shi-labs.com/natten .

这个dice计算有点问题吧。

这是开源框架monai计算方式

I encountered an issue while installing the package with "RuntimeError: Couldn't load custom C++ops. Recompile C++extension with: Python setup.py build_ext -- in-place". I understand that the version of torch vision needs to be greater than 0.3 to use ops, but the torch vision version that is compatible with CUDA 12.1+PyTorch 2.2 is 0.17. Can you please provide the complete package and version? thanks