Description
I want to convert fcos3d.onnx into tensorrt model with datatype of int8, which has three inputs ([batch,3,w,h],[3,3],[3,3]). How can I build a multi-input calibrator? How can I overwrite calibrator.get_batch()?
Hi,
Hope following are helpful to you.
opened 06:45AM - 16 Apr 21 UTC
closed 05:10AM - 21 May 21 UTC
Release: 7.x
triaged
Topic: INT8 Calibration
I'm attempting to build an int8 engine with dynamic batch sizes from an ONNX net… work. The documentation indicates that I should specify an optimisation profile for calibration using the IOptimizationProfile class. However, it appears that the config builder only takes one profile for calibration, and the profile only takes one input. How does one handle multiple inputs? Or is the string specifying the input in `set_shape` arbitrary for the calibration profile?
> ## [IBuilderConfig](https://docs.nvidia.com/deeplearning/tensorrt/api/python_api/infer/Core/NetworkConfig.html#ibuilderconfig)
> ``` python
>
> set_calibration_profile(self: tensorrt.tensorrt.IBuilderConfig, profile: tensorrt.tensorrt.IOptimizationProfile)
>```
> ## [IOptimizationProfile](https://docs.nvidia.com/deeplearning/tensorrt/api/python_api/infer/Core/OptimizationProfile.html#ioptimizationprofile)
>``` python
> set_shape(self: tensorrt.tensorrt.IOptimizationProfile, input: str, min: tensorrt.tensorrt.Dims, opt: tensorrt.tensorrt.Dims, max: tensorrt.tensorrt.Dims)
>````
CUDA 10.2
TensorRT 7.2.3
CUDNN 8
Thank you.