Mallesham Dasari, Kumara Kahatapitiya, Samir R. Das, Aruna Balasubramanian, Dimitris Samaras.
NSDI 2022 (Conference on Networked Systems Design and Implementation).
[Paper][Slides][Video]

The repository contains an implementation of the above paper. It has three independent codebases: 1) Motion flow extracter, 2) Layered neural codec located in the codec folder, 3) Adaptive video streaming pipeline using the proposed layered neural codec, located in streamer folder.

Swift's layered coding is built on top of Pytorch-VCII, and hence the running instructions are mostly similar. To run the code, go to src directory and run train.sh 2 (the argument (0, 1, or 2) specifies the level of hierarchy).

• Python3
• Pytorch
• FFMPEG

• We used 64x64 patches of images with video samples of 10 seconds each to train the model.
• Prepare the data using the instructions from the flows directory in the root folder.
• To train the codec end-to-end, run train.sh with all three options (0, 1, or 2) to have a 3-level heirarchical coding.
• To test the codec, there is an eval function within train.py code, it will run if you pass max iters argument more than what used during the training.

Note that this is only for P-frame training and testing. The I-frames are encoded using pytorch-image-comp-rnn, and the corresponding instructions can be seen inside codec/icodec folder.

Once you train and test the codec, prepare the videos in DASH fomat and run the following streamer code to test end-to-end.

• Modify the "python_path" in grad.ini to the path of the executable python in your environment
• Run command python train_grad.py
• The model generated from training is kept in ./best_model/ folder
• Run command python test_grad.py [model path] [trace path] [0|1 use real trace infomation or not] e.g python test_grad.py ./example_model/nn_model_ep_48500.ckpt(the model we trained) ./datasets/test/ 0

There are also baseline codecs from MPEG for layered coding: SVC and SHVC scalable extensions of H.264 and H.265 codecs.

If you find this useful, please use the following citation.

@inproceedings {278366,
title = {Swift: Adaptive Video Streaming with Layered Neural Codecs},
author={Dasari, Mallesham and Kahatapitiya, Kumara and Das, Samir R. and Balasubramanian, Aruna and Samaras, Dimitris},
booktitle = {19th USENIX Symposium on Networked Systems Design and Implementation (NSDI 22)},
year = {2022},
address = {Renton, WA},
url = {https://www.usenix.org/conference/nsdi22/presentation/dasari},
publisher = {USENIX Association},
month = apr,
}

The layered codec is implemented largely on top of Pytorch-VCII originally developed by Chao-Yuan Wu and Biao Zhang. The streaming implementation follows Grad: Learning for Overhead-aware Adaptive Video Streaming with Scalable Video Coding [MM'20] paper.