search

thu-media / deepladder

Code for 'Deep Reinforced Bitrate Ladders for Adaptive Video Streaming'
5 stars 2 forks source link

ABR module not used #1

Open sheepmao opened 1 year ago

sheepmao commented 1 year ago

First and foremost, we would like to express our gratitude for your diligent research efforts. After reviewing your published paper, I noticed that the entire training process is divided into two stages: the offline stage and the online stage, each with different rewards. The offline stage utilizes rewards based on predicted VMAF and size cost, among others, while the online stage relies on ABR algorithms to provide feedback on current metrics such as delay. However, it appears that the code provided in the repository does not include the online stage. I am wondering if it's possible to release the latest version of the code, as it would be extremely helpful for my ongoing research, which is similar in nature. Having access to the online stage portion would greatly serve as a valuable reference.

godka commented 1 year ago

Hi there, sorry for the long delay. In the NOSSDAV version, we didn't apply ABR algorithms for the online stage. Instead, we employ an `optimal ABR algorithm' -- which can be viewed as the upper bound of existing ABRs. For details please check https://github.com/thu-media/deepladder/blob/244c1575e4ae54d227f0cc466b2df06818a7698d/deepladder-cbr/network_env.py#L95

Regarding your question, it would be good to discuss the impact on different ABRs.

Feel free to let me know if you have any other questions.