confiwent/neuralabr-pensieve-ppo-maml

Reimplementing existing learning-based ABR algorithms for dynamic video streaming. These algorithms were implemented with Pytorch and python3

PPO and A2C based adaptive bitrate algorithms (Variants of Pensieve, Pytorch version)

Re-implementation of existing neural ABR algorithms for Video-on-demand services with Pytorch.

User guide

To train the policy with a linear or logarithmic video quality metric, refer to ./main.py

To train the policy with a perceptual video quality metric, i.e., VMAF, refer to ./variant_vmaf/main_vmaf.py

In these files, you can run the Pensive training by python main.py --a2c or python ./variant_vmaf/main_vmaf.py --a2c. Please refer to ./script/train.sh for more details.

• We also have implemented two variants of Pensieve: Pensieve with A3C algorithm (a well-established DRL method), and Pensieve with MAML algorithm (a meta-reinforcement learning method). You can run their training processes by python ./variant_vmaf/main_vmaf.py --a2c and python ./variant_vmaf/main_vmaf.py --a2br, respectively.

implemented by pytorch and trained using GPU

Note that the original version of Pensieve using asynchronous advantage actor-critic algorithm (A3C) to train the policy, which can only implementated on CPU. Our A2C version removes the asynchronous setting and use GPU to accelerate the speed of NNs training.

Further improvements are ongoing...

neuralabr-pensieve-ppo-maml's Issues

请问怎么generate cooked_trace的？

您好~
我正在参考您的实验数据集。 cooked_traces/trace_866_http---www.amazon.com

我把ffc的rawdate下载下来之后，发现相同的uid和target比如uid=866, target=http://www.amazon.com的时间跨度很长大概7000多秒。
想请问是怎么格式化成间隔为5.0s的带宽数据集的？