poor performance of ONNX model

[Alexyali]

hello, I am interested in AlphaRTC and did some network test, but the strange experiment results confused me very much.

Firstly, I compile AlphaRTC in win10 from scratch, and generated peerconnection_serverless.exe. Then I test ONNX Estimator and PyInfer Estimator separately in the local loopback network, which has no bandwidth limitation. I set the returned bandwidth to be 1Mbps in BandwidthEstimator.py for PyInfer.

The test results are shown below:

model	delay_95%/ms	avg_recv_rate/kbps
onnx	1	68
pyinfer (1M)	1	590

There are two questions that i am confused:

• why onnx model can only generate 68kbps data rate?
• why pyinfer(1M) cannot get nearly 1Mbps data rate? i guess that the origin data rate of video source is nearly 590kbps.

Thanks very much!

[Alexyali]

this is the recv_rate result of onnx model under unlimited bandwidth condition

[jeongyooneo]

Hi @Alexyali, sorry for the late response. The ONNX and PyInfer bandwidth estimator needs to be a converged RL-based bandwidth estimator. Default provided estimators are simple skeleton models, so the low performance. You can try an open-sourced AlphaRTC-testable checkpoint in the artifact of HRCC, published in MMSys 2021 (https://github.com/thegreatwb/HRCC). Happy training!