We are trying to access tensormeta from deepstream using nvinfer plugin. we got Not a Number in tensormeta always. our triton inference server uses tensorflow saved model backend. we also tried inference with triton client results are fine. Tensorrt and onnx too give the same results
.
TensorRT Version: 8.0.1
GPU Type: dGPU
CUDA Version: CUDA 11.3
All the files needed to reproduce issue are shared inthis link
https://we.tl/t-WBRnLQ5gFZ
Hi,
This looks more related to Deepstream. We are moving this post to the Deepstream forum to get better help.
Thank you.
Sorry for the late response, is this still an issue to support? Thanks
Sorry! what do you mean " we got Not a Number in tensormeta always" ? And, where did you get the tensormeta?
Did you enable “output_tensor_meta=true”?
https://docs.nvidia.com/metropolis/deepstream/dev-guide/text/DS_plugin_gst-nvinferserver.html
yes i had set output_tensor_meta to true
can you check this?
when i read tensormata inside python code the returned numpy array contains full of np.nan as values
sorry, the repo files are invalid, could you share me the repo file again?
when i read tensormata inside python code ===> I don’t mean python or C++, I mean if it’s add in nvinfer sink/src pad or any other places, can you share the piece of code?
shared the files again https://we.tl/t-Xot24rfHgL
