aten::cudnn_convolution_backward_weight not found

[iszihan]

Hello, I was trying to run the training file but got this error after following some fixes with this #91

aten::cudnn_convolution_backward_weight not found when calling from ops/conv2d_grad_fix.py line 179. After searching around, it seems to suggest we need to replace this function with aten::convolution_backward (according to pytorch/pytorch#74437), which requires a shuffling of input arguments and some additional arguments such as output mask, and bias size. I'm wondering if you know another fix and if not, how should one modify the arguments to convolution_backward() for this call? Thank you!

[SteveJunGao]

Hi @iszihan,

I guess this is related to the version of pytorch, are you using pytorch 1.9.0?

You can also try the fix from Stylegan3 codebase if you're using higher pytorch version

[Bathsheba]

You're on the right track, keep following

[SteveJunGao]

Hi @iszihan @Bathsheba,

any followups about this issue?

[iszihan]

I'm on pytorch 1.13.0, and currently the code does run with the following modification though I'm not 100% sure if this does what the old code do.

`
name = 'aten::convolution_backward'

#flags = [torch.backends.cudnn.benchmark, torch.backends.cudnn.deterministic, torch.backends.cudnn.allow_tf32]

op, _ = torch._C._jit_get_operation(name)

empty_weight = torch.tensor(0.0, dtype=input.dtype, device=input.device).expand(weight_shape)

output_mask = [False,True,False]

_, grad_weight, _ = op(grad_output, input, empty_weight, None, stride, padding, dilation, transpose, (0,0), groups, output_mask) #*flags)

return grad_weight
`