Qsingle

设置`MEGENGINE_INTERP_ASYNC_LEVEL=0`和`CUDA_LAUNCH_BLOCKING=1`，追查到如下报错日志: ```shell Traceback (most recent call last): File "cityscapes_train.py", line 300, in main() File "cityscapes_train.py", line 256, in main train(model, train_loader, val_loader, optimizer, loss_fn, gm, epochs, args, lr_scheduler=lr_sche) File "cityscapes_train.py",...

更新: 如果设定`align_corners=True`的话则不会出现这个问题

> 你好，可以补充一下关键的代码片段吗？便于复现问题 没有整理出一个比较好的最小复现，用我现在那个就太大了，也不太好定位问题,我也不确定以下代码是否就会出现这个问题。 ```python import megengine as mge import megengine.module as M import megengine.audodiff as ad import megengine.optimizer as optim import megengine.functional as F class Model(M.Module): def __init__(self): self.conv1 =...

Yeah, it also supports the RTX 40X GPU, you can download the latest version 1.13.0 for the Windows system through the [Baidu Cloud Driver](https://pan.baidu.com/s/1qHkv9slHvVSNvSfbjkXgWg&pwd=bwwn).

因为WHL打包的时候是会修改rpath(主要是官方为了让用户免去装cuda和cudnn的这个过程，所以做了CUDA和CUDNN的依赖打包和这个检测)，但是这个会和PATH里面的冲突，所以可以把PATH里面关于CUDA和CUDNN的部分去除，不然就会出现编译用的版本和装的版本的冲突。

目前来说他是会编进去的，所以不能那个。你可以直接安装官方的版本试试，那个不是该rpath而是直接静态链接了，所以不会出现这个问题，920m我不记得是不是cuda11好像不支持，可以查一下cuda11支持的架构。

> Since I'm not using my own training set at the moment and only have single-digit images for testing, could you please provide me with your pre-trained weights so that...

You can see the issue #3 to get the answer.

Thank you for your question. You can set the `--mask_path` for the AROI dataset to `patient/mask/number`. For example, `patient12/mask/number` and set the `--image` to `patient12/raw/labelled`.

Good issue, we use torch 1.13.1+CU117. It seems that the model or data can not be feed into the GPU. I suggest you check the version of CUDA.