my model involves 6 blocks of convolution2D layer + relu layer + maxpool, followed by 3 dense layers, with the last dense layer having 2 outputs (classfication problem). My model was written in keras and trained using channels first.
I froze my model to a protobuf. When i imported a graph def from the protobuf and fed in an image, i get the correct results ([0.99, 0.01]). However, when i try to do the same with the UFF model generated from the protobuf, I get very strange and wrong results ([0.22, 0.78]). I followed the exact flow from the tensorRT guide, can someone advice as to why the results are vastly different?
Is it due to some bug in the tensorRT optimisation? Thanks!