Very slow 2d convolution with Pytensor

Thanks, that would be helpful.
Also, not that there is a bug in the existing abstract_conv.py, where I had to add the cast in line 682:

filters = pytensor.tensor.cast(filters,dtype=input.dtype)

Without this, abstract_conv will upcast data to float64, even if the input and outout
want to have float32.

1 Like