Haohao comments

Results 15 comments of


                                            Haohao

On which GPU did you test centerpoint TensorRT engine

> > Hi @CarkusL, can you give us the inference time for batch_size =1 of your TensorRT implementation including also the preprocess and postprocess? > > [09/15/2021-10:38:05] [I] PreProcess Time:...

On which GPU did you test centerpoint TensorRT engine

@CarkusL Thanks for your great work, I wrote a new project based on your code, where computations of pre-process && post-process are done with Cuda, it runs much faster. Here...

PostProcess time is too long

sorry for my late reply, it may be related to the difference in GPU architectures, all my samples are tested on RTX3080. Also, generally, the first sample will take some...

PostProcess time is too long

Haven't received responses for a long time

How to visualize the detection results, but with my own images which I have clicked ?

here is a [repo](https://github.com/Abraham423/CenterFusion) to run CenterFusion, we provide rviz visualization codes

error: 'class nvinfer1::IBuilder' has no member named 'buildSerializedNetwork'

vTensorRT : 8.0.1.6 vUbuntu : 18.04 check it out ~

export_onnx set to true

I skipped the preprocess part, it is achieved in preprocess.cpp/.cu

inference time about int8 pfe

I can only guess the following reasons 1) pfe nn computation graph is simple, which only includes 2 groups of linear-bn1d-relu, float mode can already do a good job. 2)...

inference time about int8 pfe

Yes it only supported torch ckpt if `--run_infer ` is enabled, if you want to compute evaluation metrics for trt engine results, you should previously run the cpp codes, and...

Why two subgraphs instead of one whole graph

1. TensorRT is developed to optimize nn inference, the connection part between pfe & rpn (voxel assigning) involves no nn computing, I don't think TRT would optimize that part of...