RuntimeError: derivative for _thnn_fused_gru_cell_backward is not implemented

autograd
1

Have no idea about this error.
_thnn_fused_gru_cell_backward can be found in pytorch/tools/autograd/derivatives.yaml, so I assumed it had been implemented. Am I right?
Are there any other possible reasons for this error to occur?

Thanks,
Yann

2

No, it only occurs as the backward of _thnn_fused_gru_cell.
If it had a line with - name: _thnn_fused_gru_cell_backward(...) then it would be differentiable itself.

Best regards

Thomas

3

Hi Thomas,

Thanks for your reply.
Does this mean I have to implement the backward function?

Thanks,
Yann

4

Well, so we don’t have RNN double backwards at the moment.
If you wanted that, you would have to implement it yourself or implement an RNN cell yourself.

Best regards

Thomas

5

Do you mean I implement a cell as regular nn.module then it would automatically work?

6

Yes, exactly.

Best regards

Thomas