Hi,
Sorry for the late reply.
We have tested ResNet with TensorRT before. The result works fine.
But there are lots of variants, we are not sure if it is identical to yours.
From your comment, please remember that the flag ‘use_global_stats’ is not supported by TensorRT currently.
Reason is that this flag is more commonly used in training stage but TensorRT is designed for inferencing.
We are not sure if a refinement is needed when turning off the use_global_stats flag.
A possible test is to turn-off the flag in both Caffe and TensorRT interface and recheck it again.
If you still meet different results between Caffe and TensorRT, could you share the ResNet model with us?
Recently, we also have an example for ResNet but on TensorFlow frameworks:
Thanks.