https://github.com/pytorch/pytorch/issues/72175

https://github.com/pytorch/pytorch/issues/95377

The export only needs to run the model once so performance should not be an issue, and if everything is on the CPU device, then you shouldn’t run into this problem.

方案:network和输入images都放在cpu上,不放在cuda上,导出Onnx只使用cpu

Logo

Agent 垂直技术社区,欢迎活跃、内容共建。

更多推荐