by Nuowen Kan, Yuankun Jiang, Chenglin Li, Wenrui Dai, Junni Zou, and Hongkai Xiong at Shanghai Jiao Tong University

This repository is the official Pytorch implementation of MERINA

This document illustrates how to obtain the results shown in the paper "Improving Generalization for Neural Adaptive Video Streaming via Meta Reinforcement Learning".

Anaconda is suggested to be installed to manage the test environments.

• Linux or macOS
• Python >=3.6
• Pytorch >= 1.6.0
• numpy, pandas
• tqdm
• seaborn
• matplotlib
• CPU or NVIDIA GPU + CUDA CuDNN

Install PyTorch. Note that the command of PyTorch intallation depends on the actual compute platform of your own computer, and you can choose appropriate version following the guide page. For example, if you have intalled CUDA 10.2, you can intall PyTorch with the latest version by running this Command:

conda install pytorch torchvision torchaudio cudatoolkit=10.2 -c pytorch

Or you can create a specific environment (many redundant dependencies will be installed) just via

conda env create -f torch.yaml

The main training loop for MERINA can be found in main.py, the trained models are in models/ and the corresponding meta RL algorithms in algos/. Besides, envs/ includes emulator codes which simulate the environment of ABR virtual player. Some of baseline algorithms are located in baselines/.

There's quite a bit of documentation in the respective scripts so have a look there for details.

Improvements to the codes and methods (including a journal version which is undergoing a major revision) are provided in the repository below:

Type python main.py -h for the instruction in the terminal, or read the description of ArgumentParser part.

For example, you could choose work mode by selecting the argument --test [evaluate the model], --adp [run the meta adaptation procedure] or use the default setting [run the meta training procedure];

The default QoE metric is the linear form and you can change it to logrithmic form by add the argument --log;

The network throughput dataset must be chosen in the test mode. The datasets shown in the paper are available here and you can select them following the mappings below:

[--tf FCC traces] [--tfh FCC and HSDPA traces] [--t3g HSDPA traces] [--to Oboe traces] [--tp Puffer-Oct.17-21 traces] [--tp2 Puffer-Feb.18-22 traces]

Also, you can rename the lable of the results by [--name "yourname"]

    |--models
    |--Results  # !!
       |--sim
       |--test
          |--lin
          |--log
    |--utils
       |--log_results # !!
    |--main.py

The public bandwidth traces are stored in this repository. Download and put them in the directory ./envs/traces/.

To evalute MERINA on the in-distribution throughput traces with the $QoE_{log}$ metric from the paper, run

python main.py --test --tf --log

python main.py --test --t3g --log

cd utils
python plt_v2.py --log --merina --bola --mpc

python main.py --log

tensorboard --logdir=./Results/sim