Add a flag window_reversal as an argument to conv_transpose2d, when true the transformation of the weights will include reversing the data in the kernel window, when false it will not. Currently it always reverses.
This is being requested because some frameworks that compile through MLIR reverse the weights as part of the conv_transpose2d op (like PyTorch) and some do not (like JAX).
FYI: @nsmithtt you said we will likely need this flag for the regular conv2d op as well for training purposes.
window_reversalas an argument toconv_transpose2d, when true the transformation of the weights will include reversing the data in the kernel window, when false it will not. Currently it always reverses.This is being requested because some frameworks that compile through MLIR reverse the weights as part of the
conv_transpose2dop (like PyTorch) and some do not (like JAX).FYI: @nsmithtt you said we will likely need this flag for the regular conv2d op as well for training purposes.