cuda.amp.autocast doesn't work for this module

[giladsharir]

Hi
Thanks for the great work.
I tried running this module with mixed precision (autocast) and got an error indicating the the weights we not converted to half.
It would be great if this can be resolved.

[gungui98]

Could you please post some code snippets!

[giladsharir]

the following code

import torch
from depthwise_conv3d import DepthwiseConv3d
from torch.cuda.amp import autocast

depth_multiplier=2
conv = DepthwiseConv3d(2, 2 * depth_multiplier, kernel_size=3, groups=2).cuda()
input = torch.randn(2, 2, 6, 6, 6, device="cuda", dtype=torch.half).div_(2).requires_grad_()
with autocast():
  output = conv(input)

produces the error

Traceback (most recent call last):
File "", line 2, in
File "/opt/conda/lib/python3.8/site-packages/torch/nn/modules/module.py", line 744, in _call_impl
result = self.forward(*input, **kwargs)
File "/opt/conda/lib/python3.8/site-packages/depthwise_conv3d-1.0.3-py3.8-linux-x86_64.egg/depthwise_conv3d.py", line 93, in forward
return depthwise_conv3d(x, self.weight, self.bias, self.stride, self.padding, self.dilation,
File "/opt/conda/lib/python3.8/site-packages/depthwise_conv3d-1.0.3-py3.8-linux-x86_64.egg/depthwise_conv3d.py", line 26, in forward
output = DWCONV_CUDA.conv_depthwise3d_cuda(
RuntimeError: expected scalar type Half but found Float

[gungui98]

I just add a handful fix to the module. In fact, the module supports both full and mixed-precision, please test it in your model and let me know if it works

[giladsharir]

looks like it works now
Thanks