Mashiro

@wanghonglie Hi~ it seems a bug in MMCV, just as you say  'model_zoo/torchvision_0.12.json' is a linux format path, which is not compatible with windows platform. It should be corrected...

Resolved in https://github.com/open-mmlab/mmengine/pull/820

> Does this bug apply if `by_epoch=True` for `param_scheduler`? Unfortunately, that's the case. We haven't identified a suitable solution yet, as tightly integrating the functions of `OptimWrapper` and `scheduler` isn't...

我想你之前是通过在配置文件里指定 workflow 来输出验证集 loss 的吧？在 OpenMMLab 1.0 里，Runner.val 是专门用于输出验证集 loss 的，而 EvalHook 是用来输出评价指标的。 而在 OpenMMLab 2.0 里，统一使用 ValLoop 来进行评测。设计之初我们也讨论过如何输出验证集的 loss，你也可以发现 model.val_step 会接受 `mode` 参数，进而把 mode 传给 `model.forward`，根据 mode 的值选择不同的分支。如果你想在验证阶段同时输出验证集 loss，你可以新定义一个...

You can try to launch your task by ```bash CUDA_VISIBLE_DEVICES="" mim train mmcls resnet18_b16x8_cifar10.py --work-dir tmp --launcher slurm --partition aide_dev ```

For me, the unexpected icon is `tide_python_icon`. Set it to `.` save me 🤣

Hi, have you followed the guide in [docs](https://mmcv.readthedocs.io/en/latest/get_started/build.html#build-on-linux) to install [ROCm](https://docs.amd.com/)？

Hi, thanks for your feedback, can you provide an example to reproduce the problem?