I’d like to comment on this after a long pause. The updates by jessegrabowski have versions of conv2d that work on GEMM in the “tests” directory. I import them using, for example:
from tests.tensor.conv.test_abstract_conv import conv2d_corr as conv2d
from tests.tensor.conv.test_abstract_conv import conv2d_corr_gi
I have tried these out, and works SUPER! for neural networks and I have tested them extensively . This is a huge help, THANKS!!. They’re orders of magnitude faster that the old abstract_conv.py. But, they are not yet part of standard PYTENSOR. What are the plans for these functions?