So, I re-installed PyTensor with conda -c conda-forge, and yes, I see now that the c-compiler and all dependencies are properly installed including (I think) GEMM/Blas And, I have seen the speedup of the execution time for the case without 2D convolution. But, for 2D convolution, there is still a huge difference. I notice that in AESARA, there is aesara/tensor/nnet/conv.py which has optimized convolution drawn from
from scipy.signal._sigtools import _convolve2d
and implemented as an OpenMP Op.
But in pytensor, this has been apparently removed. The convolution appears to be done brute-force in abstract_conv.py
Am I missing something here?
Thanks in advance