Problem with FP16 on TensorRT 3.0.2


I recently upgraded to the Drive with TensorRT 3.0.2, but it seems to have a side effect causing corrupted output with my caffe model. On the previous Drive release with TensorRT 3.0.0, the problem didn’t happen. I’m using the AutoCruise board.

I saw in the TensorRT 3.0.4 release notes a comment about “Fixed an accuracy issue in FP16 mode for NVCaffe models” and I’m wondering if maybe this is related? Do you know when the accuracy issue appeared?

When will the DPX2 support TensorRT 3.0.4?

Dear gordon1zrra,
The next Drive Release is June.

That’s quite a bit a ways. Could you confirm the “Fixed an accuracy issue in FP16 mode for NVCaffe models” issue is something introduced in 3.0.1/3.0.2 release or has it existed since 3.0.0?

Dear gordon1zrra,
Can you file a bug with sufficient information to reproduce it. We will look into it.
Please login to with your credentials. Please check MyAccount->MyBugs->Submit a new bug to file bug.

Please share ID here to follow up