add pytorch 1.5 supported #58
Conversation
| @@ -104,7 +104,7 @@ dcn_v2_cuda_forward(const at::Tensor &input, | |||
| const int block = 128; | |||
| const int grid = (batch + block - 1) / block; | |||
|
|
|||
| createBatchGemmBuffer<<<grid, block, 0, THCState_getCurrentStream(state)>>>( | |||
| createBatchGemmBuffer<<<grid, block, 0, c10::cuda::getCurrentCUDAStream()>>>( | |||
There was a problem hiding this comment.
Hi!
Could you please explain why the way getting a stream is changed? A link may be.
Thanks in advance.
There was a problem hiding this comment.
There was a problem hiding this comment.
Also I tried to use you branch with torch1.5 and https://github.com/xingyizhou/CenterTrack
CenterTrack couldn't import DCN.
Then I switched to torch1.4 and it worked just fine
There was a problem hiding this comment.
I tested this version on my other projects and also centertrack, my suggestion is remove all your codes and find & rm pkg you installed in your python/pip/conda env
There was a problem hiding this comment.
There was a problem hiding this comment.
Hi!
Could you please explain why the way getting a stream is changed? A link may be.
Thanks in advance.
you success,can you share something this!
There was a problem hiding this comment.
It works for my case, cuda: 10.2, pytorch: 1.5
|
@lbin Hi, thanks for this PR. I tested it with 1.5.1 and it worked for me. However I tried to use it with 1.6.0 (nightly 1.6.0.dev20200623+cu101') and was unable to compile and got I think several deprecation warnings in 1.5 turned into errors, like:
I'm a C++ novice myself, so if you happen to get any time to update this PR for future compatibility I'd really appreciate it. (By the way using 1.6.0 for torch tracing support of dict output types) |
There was a problem hiding this comment.
This worked for me on image buildt by this dockerfile, i.e., CUDA10.2 & CUDNN7 & PyTorch1.5.1
Thus consider merging it.
|
Pls follow https://github.com/lbin/DCNv2/tree/pytorch_1.6 for 1.6 support |
No description provided.