fp16 for deform_conv2d missing?

[paapu88]

🐛 Bug

To Reproduce

This example below is copy pasted from https://pytorch.org/docs/stable/torchvision/ops.html
except dtypes are added.
The first part with float32 runs fine, but the second part with float16
produces error:

File "/home/hu-mka/venvs/pytorch/lib/python3.8/site-packages/torchvision/ops/deform_conv.py", line 67, in deform_conv2d
return torch.ops.torchvision.deform_conv2d(
RuntimeError: th_addmm not supported on CPUType for Half

Terveisin, Markus

input = torch.rand(5, 3, 10, 10, dtype=torch.float32)
kh, kw = 3, 3
weight = torch.rand(5, 3, kh, kw, dtype=torch.float32)
offset = torch.rand(5, 2 * kh * kw, 8, 8, dtype=torch.float32)
out = ops.deform_conv2d(input, offset, weight)
print(out)
input = torch.rand(5, 3, 10, 10, dtype=torch.float16)
kh, kw = 3, 3
weight = torch.rand(5, 3, kh, kw, dtype=torch.float16)
offset = torch.rand(5, 2 * kh * kw, 8, 8, dtype=torch.float16)
out = ops.deform_conv2d(input, offset, weight)
print(out)

Environment

(pytorch) hu-mka@humkaLenovo:~/Downloads$ python collect_env.py
Collecting environment information...
PyTorch version: 1.6.0
Is debug build: False
CUDA used to build PyTorch: 10.2
ROCM used to build PyTorch: N/A

OS: Ubuntu 20.04.1 LTS (x86_64)
GCC version: (Ubuntu 8.4.0-3ubuntu2) 8.4.0
Clang version: 10.0.0-4ubuntu1
CMake version: version 3.16.3

Python version: 3.8 (64-bit runtime)
Is CUDA available: True
CUDA runtime version: 10.1.105
GPU models and configuration: GPU 0: GeForce GTX 1650 with Max-Q Design
Nvidia driver version: 450.80.02
cuDNN version: Probably one of the following:
/usr/local/cuda-10.0/targets/x86_64-linux/lib/libcudnn.so.7.6.5
/usr/local/cuda-10.1/targets/x86_64-linux/lib/libcudnn.so.7.6.5
HIP runtime version: N/A
MIOpen runtime version: N/A

Versions of relevant libraries:
[pip3] numpy==1.19.0
[pip3] pytorch-lightning==0.9.0
[pip3] torch==1.6.0
[pip3] torchfunc==0.2.0
[pip3] torchsummary==1.5.1
[pip3] torchvision==0.7.0
[pip3] torchviz==0.0.1
[conda] numpy 1.19.1 pypi_0 pypi

[fmassa]

Thanks for opening the issue!

That's correct, for now we don't yet have mixed precision for deform_conv2d, but we should be adding support for it soon (together with the remaining ops that don't support it yet).

[paapu88]

Thank you guys!!! Perfect timing!