I am exporting a simple model to ONNX and noticed that a [conv + BN + relu] block parameters were renamed. After stepping through the export code, I find the offending change happening within
torch._C._jit_pass_onnx_eval_peephole, and presumably during conv+bn fusion here.
parameters ‘18’ and ‘19’ are introduced after peephole, and later replace “conv_bn_and_relu” parameter names in the final graph. Is it possible to retain the “conv_bn_and_relu” parameter name in the graph?