set precision to fp32

[cicoita]

Hi,

I'm using segnet to load a custom segmentation network and I was wondering if I can change precision to fp32?

Currently I am getting this when i run it
...
[TRT] desired precision specified for GPU: FASTEST
...
[TRT] selecting fastest native precision for GPU: FP16

Thanks for the help :)

[cicoita]

Oh and I am getting this error I haven't noticed before

[TRT] Registering tensor: output_34 for ONNX tensor: output
[TRT] Resize_261 [Resize] outputs: [output -> (1, 12, 1080, 1920)[FLOAT]],
[TRT] Marking output_34 as output: output
[TRT] Marking onnx::Resize_335_2 as output: onnx::Resize_335
[TRT] Marking input.1_3 as output: input.1
[TRT] Marking input.3_4 as output: input.3
[TRT] Marking output as output: output
�[0;31m[TRT] [network.cpp::markOutputHelper::1811] Error Code 4: Internal Error (Tensor output is already set as network output)

Do you know how this can be fixed?

[dusty-nv]

@cicoita the segmentation models in ros_deep_learning / jetson-inference are expected to be FCN-Resnet from PyTorch architecture. You might have better luck with the newer Isaac ROS Image Segmentation package: https://github.com/NVIDIA-ISAAC-ROS/isaac_ros_image_segmentation

And I don't believe I expose the desired precision as a ROS param, but you could hardcode that change or expose it if desired.