Jiazhen Wang comments

Results 14 comments of


                                            Jiazhen Wang

modified code for using ROCm backened within the PyTorch framework

My env is gfx906, ROCm 4.0.0, PyTorch 1.8.0. `pytest tests/test_ops/` gives me `Aborted (core dumped)`.

[Fix] DDP duplicate logs handling

Hi @sleep3r , is there any progress?

auto_fp16 goes through torch.no_grad(), and the gradient used outside torch.no_grad() also disappears

It looks strange. Could you provide some codes that are easy to reproduce?

auto_fp16 goes through torch.no_grad(), and the gradient used outside torch.no_grad() also disappears

I tried a simple case ``` class MyModule1(nn.Module): def __init__(self): super().__init__() self.model = nn.Sequential(*list(resnet50().children())[:-2]) @auto_fp16(apply_to=('img'), out_fp32=True) def extract_img_feat(self, img, img_metas): return self.model(img) def forward(self, img, img_metas): with torch.no_grad(): feat1 =...

Move the `set_random_seed` function into mmcv.

Thank you for your suggestion. We will consider it.

The following error occurred when version 1.3.16 was compiled on AMD.

What are your compilation commands?

The following error occurred when version 1.3.16 was compiled on AMD.

It seems a bug of ROCm, still exists in mmcv master branch. After `.cu` files convert to `.hip` files, the files writen in `mmcv_full.egg-info/SOURCES.txt` become abspaths. If you just want...

The following error occurred when version 1.3.16 was compiled on AMD.

Probably no. These filenames written in `mmcv_full.egg-info/SOURCES.txt` are from `op_files` of setup.

Log config to wandb

It seems that there is no way to get access to the cfg object from `WandbLoggerHook`. Maybe you can modify cfg object in `train.py` like this ``` for h in...

Log config to wandb

> I am trying to achieve the same thing. While investigating, I realized that the `WandbLoggerHook` methods have access to the `runner` object. However, the `runner` has no `cfg` class...