RuntimeError: one of the variables needed for gradient computation has been modified by an inplace operation:

[zhaoxiaodong789]

When run lsnet_bbox_r50_fpn_1x_coco.py, I got the problem like this:

[Duankaiwen]

what's your pytorch version?

[zhaoxiaodong789]

1.7.1

[Duankaiwen]

Please try to use Pytorch 1.5.0

[zhaoxiaodong789]

ok

[eeric]

Installation instruction https://github.com/eeric/LSNet_supplement

[zy937085318]

I meet the same question, I just use 3090 GPU, which pytorch1.5 is not compatible with. Is it possible to run lsnet on pytorch1.7+?

[dongdongyee]

I meet the same question, I just use 3090 GPU, which pytorch1.5 is not compatible with. Is it possible to run lsnet on pytorch1.7+?

You can use a higher version of PyTorch, e.g. 1.8.0, in case you have a GPU that only works under CUDA 11+ As you can see, the error is caused by inplace operations. If you turn on torch.autograd.set_detect_anomaly, you will find out what you need to do is to simply change some code in code/mmdet/models/dense_heads/lsnet_head.py You can change all the variable1 *= variable2 into variable1 = variable1 * variable2

I've tested it myself with NVIDIA A100 with PyTorch 1.8.0 and it works.

[zhaoxiaodong789]

收到