This repository is an official implementation of the paper Neural Adaptive Content-aware Internet Video Delivery.
Currently, we only provide NAS-MDSR, which is a super-resolution module of NAS.
- Python 3.6
- (p) PyTorch >= 1.0.0
- (p) numpy
- (p) skimage
- (p) scipy
- (p) cv2 (Use opencv package from here)
- (p) pillow
- (p) ffmpeg
- (b) MP4Box, x264 (Refer here for installing these binaries)
Download MPEG-DASH dataset from here and place like:
data
|------news
|------original.mp4
|------240p
|------360p
|------480p
|------720p
|------1080p
Or, if you want to use own 1080p video, first place it like:
data
|------[dataset name]
|------[video file]
Then, to generate MPEG-DASH content from the original video:
cd data/[dataset name]
../../dash_vid_setup.sh -i [video_file]
- Original video link for provided news dataset: here
To train NAS-MDSR:
python train_nas_awdnn.py --quality [quality level] --data_name [dataset name] --use_cuda --load_on_memory
NAS-MDSR provides total four quality levels as in the paper (e.g., low, medium, high, ultra-high). The higher the quality level, the bigger the model size, the higher the model quality.
Models will be saved like:
checkpoint
|------[dataset name]
|------[quality level]
|------epoch_[index].pth
|------ ...
|------DNN_chunk_1.pth
|------DNN_chunk_2.pth
|------DNN_chunk_3.pth
|------DNN_chunk_4.pth
|------DNN_chunk_5.pth
DNN chunks are save only for the last updated model. These chunks are used for streaming together with video chunks in NAS.
- Related code: model.py, dataset.py, option.py, trainer.py, train_nas_awdnn.py
To measure the quality of NAS-MDSR both in PSNR,SSIM:
python test_nas_quality.py --quality [quality level] --data_name [dataset name] --use_cuda --load_on_memory
Result will be saved like:
result
|------[dataset name]
|------[quality level]
|------result_quality_detail_0_[epoch].log
|------result_quality_detail_2_[epoch].log
|------result_quality_detail_4_[epoch].log
|------result_quality_detail_6_[epoch].log
|------result_quality_detail_8_[epoch].log
|------result_quality_summary_[epoch].log
To measure the inference time of NAS-MDSR:
python test_nas_runtime.py --quality [quality level] --data_name [dataset name] --use_cuda --load_on_memory
Result will be saved like:
result
|------[dataset name]
|------[quality level]
|------result_runtime.log
- Related code: model.py, dataset.py, option.py, tester.py, test_nas_quality.py, test_nas_runtime.py
NAS-MDSR can also process a video in which decoding, encoding, super-resolution are done parallely.
To apply NAS-MDSR to process video chunks:
python test_nas_video_process.py --quality [quality level] --data_name [dataset name] --use_cuda --load_on_memory
It will generate quality-enhanced video chunks like:
result
|------[dataset name]
|------[quality level]
|------[segment_[chunk index]_[resolution index]]
|------input.mp4
|------output.mp4 (quality-enhanced video chunk)
You can set chunk index and resolution index in test_nas_video_quality.py.
To measure the latency of NAS-MDSR to process video chunks:
python test_nas_video_runtime.py --quality [quality level] --data_name [dataset name] --use_cuda --load_on_memory
Result will be saved like:
result
|------[dataset name]
|------[quality level]
|------result_video_runtime.log
Refer process.py to understand detail procedures for processing video chunks.
- Related code: model.py, dataset.py, option.py, tester.py, test_nas_video_quality.py, test_nas_video_runtime.py
- Use the option 'load_on_memory' if you have enough memory since it highly affects on training speed.
- Use the option 'use_cuda' for using a GPU.
If you find paper useful for your research, please cite our paper.
Hyunho, et al. "Neural adaptive content-aware internet video delivery." 13th USENIX Symposium on Operating Systems Design and Implementation (OSDI 18). 2018. [Website]
@inproceedings{yeo2018neural,
title={Neural adaptive content-aware internet video delivery},
author={Yeo, Hyunho and Jung, Youngmok and Kim, Jaehong and Shin, Jinwoo and Han, Dongsu},
booktitle={13th $\{$USENIX$\}$ Symposium on Operating Systems Design and Implementation ($\{$OSDI$\}$ 18)},
pages={645--661},
year={2018}
}
If you plan to use the code for commercial usage, please contact to the following email.
Hyunho Yeo (PhD candidate at KAIST) / [email protected]
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent