I have a 2 Tensorflow concat node that concats on the 3rd axis, even the input tensors they are concatenating shapes are the same, but they are at different parts of the graph. One of them gets translated flawlessly by UFF to concat on 1st axis (I assume it did some HWC to CHW conversion) and it throws no error when parsing with UFF parser in C++ code. The other concat node, however, gets translated to a concat that concatenates in the first axis followed by an added transpose(2,0,1) in TensorRT. That completely breaks my TensorRT inference network because everything was in CHW up until this point.
The original Tensorflow graph has 4 dimension tensors with a batch dimension. My TensorRT tensors have 3 for simplicity, but it compiles and validates up until that second concat node.
Any idea for this behavior? I translate my Tensorflow frozen .pb file to .uff using convert-to-uff.
Thank you in advance.