Donggeun Yu comments

Results 22 comments of


                                            Donggeun Yu

Conv2d stride=2 accuracy mismatch between PyTorch and TensorRT

@schegde did you solve it? There is no problem when padding=0 is used. Different results from PyTorch only when padding=1 or higher.

difference between DCNv2 and deform_conv2d in torchvision

What's the difference?

Can i using FGPU?

@sakjain92 FGPU evaluation.sh worked successfully, but I couldn't confirm the GPU split with nvidia-smi. As in TODO.md, GPU virtualization seems to be able to split the GPU into containers. But...

Face_Enhancement model converting to onnx failure

@zhangmozhe I'm not sure how to utilize identity matrices. I tried using torch.Identity and torch.Conv2d, but ONNX ignores those layers. As before, the channel size is *.

Why is DeformableDetrForObjectDetection slower with bfloat16 than float32?

@qubvel [profile_torch.bfloat16.txt](https://github.com/huggingface/transformers/files/15131446/profile_torch.bfloat16.txt) [profile_torch.float32.txt](https://github.com/huggingface/transformers/files/15131447/profile_torch.float32.txt)

Why is DeformableDetrForObjectDetection slower with bfloat16 than float32?

> Based on profiling results it seems like conv2d is slow with `bfloat16` > > I found the following information regarding this issue: > > Comment 1 [pytorch/pytorch#57707 (comment)](https://github.com/pytorch/pytorch/issues/57707#issuecomment-1166656767) >...