PiTree is a conversion tool to automatically and faithfully convert complex adaptive bitrate algorithms into lightweight decision trees. This repository is the official release of the following paper:
Zili Meng, Jing Chen, Yaning Guo, Chen Sun, Hongxin Hu, Mingwei Xu. PiTree: Practical Implementations of ABR Algorithms Using Decision Trees. In Proceedings of ACM Multimedia 2019.
For more information, please refer to https://transys.io/pitree.
Tested with Python 3.7.4:
pip install -r requirements.txt
unzip traces.zip
unzip models.zip
python learn_dt.py -a pensieve -t fcc -i 500 -n 100 -q lin
| Parameter | Candidates | Explanation |
|---|---|---|
| -a | {robustmpc, pensieve, hotdash} | The ABR algorithm to convert. |
| -i | Integer (default=500) | Number of iterations during training. |
| -n | Integer (default=100) | Number of leaf nodes. |
| -q | {lin, log, hd} | QoE metrics. |
| -t | {fcc, norway, oboe} | Trained traces. |
| -v | {0,1} | Visualized the output decision tree. |
| -w | Integer (default=1) | Degree of parallelism of teacher.predict(). |
The converted decision tree could be found at tree/, in the pickle format.
If you want to test your own ABR algorithms with PiTree, you could
- Expose the predict function of your methods in the format of
$a=f(s)$ . - Put your model into
models/(if any). - Add your methods into the interfaces defined in
learn_dt.py.
(We will refactor the codes soon in a more user-friendly way and will update the repo soon.)
python main.py -a pensieve -t fcc -q lin -d
| Parameter | Candidates | Explanation |
|---|---|---|
| -a | {robustmpc, pensieve, hotdash} | The ABR algorithm to convert. |
| -d | {0,1} | Predict with the decision tree (1) or the original model (0). |
| -l | {0,1} | Log the states and bitrates. |
| -q | {lin, log, hd} | QoE metrics. |
| -t | {fcc, norway, oboe} | Trained traces. |
python server_tornado.py
For any questions, please post an issue or send an email to [email protected].
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent