autovmaf - A toolkit to automatically encode multiple bitrates and perform automated VMAF measurements on all of them.
By optimizing ABR-ladders for specific content, you will make sure to not have wasteful rungs and this has been shown to cut bandwidth usage in half.
Transcoding and VMAF analysis can either be run in AWS or locally. When running in aws, you will need a running ECS cluster with a task definition configured to run easyvmaf-s3.
npm install --save @eyevinn/autovmaf
A few environment variables can be set. These are:
LOAD_CREDENTIALS_FROM_ENV=true //Load AWS credentials from environment variables
AWS_REGION=eu-north-1
AWS_ACCESS_KEY_ID=ABCD...
AWS_SECRET_ACCESS_KEY=EFGH...
To generate VMAF measurements, you will need to define a job which can be created with the createJob()
-function.
const { createJob } = require('@eyevinn/autovmaf');
const vmafScores = await createJob({
name: 'MyVMAFmeasurements',
pipeline: 'pipeline.yml',
encodingProfile: 'profile.json',
reference: 'reference.mp4',
models: ['HD', 'PhoneHD'], // optional
resolutions: [
{
// optional
width: 1280,
height: 720,
range: {
// optional
min: 500000,
max: 600000
}
}
],
bitrates: [
// optional
500000, 600000, 800000
],
method: 'bruteForce' // optional
});
When creating a job, you can specify:
- Name
- This will name the folder in which to put the files.
- Pipeline
- Path to a YAML-file that defines the pipeline. See
examples/pipeline.yml
for an example AWS-pipeline. - When running locally, pipeline data can be inlined in the job definition.
- Path to a YAML-file that defines the pipeline. See
- Encoding Profile
- Path to a JSON-file that defines how the reference should be encoded. When using AWS, this is a MediaConvert configuration. See an example for AWS at
examples/aws/encoding-profile.json
. For local pipelines, this is key-value pairs that will be passed as command line arguments to FFmpeg. If pipeline data is inlined in the job definition, encodingProfile can be omitted and key-value pairs can instead be set in theffmpegOptions
property of the pipeline object.
- Path to a JSON-file that defines how the reference should be encoded. When using AWS, this is a MediaConvert configuration. See an example for AWS at
- Reference
- Path to the reference video to analyze. Normally a local path, but when using AWS, this can also be an S3-URI.
- Models (optional)
- A list of VMAF-models to use in evaluation. This can be HD, MobileHD and UHD. HD by default.
- Resolutions (optional)
- A list of resolutions to test. By default it will test all resolutions in the example ABR-ladder provided by Apple in the HLS Authoring Spec.
- Range (optional)
- A min and max bitrate for testing a specific resolution. Adding a range will filter out bitrates that are outside of the given range. It is disabled by default.
- Bitrates (optional)
- A list of bitrates to test. By default a list of bitrates between 150 kbit/s to 9000 kbit/s.
- Method (optional)
- The method to use when analyzing the videos. Either
bruteForce
orwalkTheHull
. By defaultbruteForce
. NOTE:walkTheHull
is not implemented at the moment.
- The method to use when analyzing the videos. Either
const { createJob } = require('@eyevinn/autovmaf');
const YAML = require('yaml');
const fs = require('fs');
const parseResolutions = (resolutions) => {
resolutions.map((resolutionStr) => ({
width: parseInt(resolutionStr.split('x')[0]),
height: parseInt(resolutionStr.split('x')[1])
}));
};
const jobFile = fs.readFileSync('job.yml', 'utf-8');
const jobData = YAML.parse(jobFile);
const job = {
...jobData,
resolutions:
jobData['resolutions'] !== undefined
? parseResolutions(jobData['resolutions'])
: undefined
};
createJob(job);
An example of creating a job from a YAML-file can be seen in the examples-folder
.
Using getVmaf()
, you can read VMAF-scores from a JSON-file or a directory of JSON-files. This works on both local paths as well as S3-URIs with a "s3://"-prefix.
Example:
const vmafFiles = await getVmaf('s3://path/to/vmaf/');
vmafFiles.forEach((file) => {
console.log(file.filename + ': ' + file.vmaf);
});
When running with the cli, all transcoding and vmaf analysis will be run locally.
Installing globally with npm -g
will make the autovmaf
command available in your path
npm install -g @eyevinn/autovmaf
These are only needed if you are running transcodes and VMAF measurements locally
EASYVMAF_PATH
- needs to point to the fileeasyVmaf.py
from your easyVmaf installation.FFMPEG_PATH
- only needs to be set if ffmpeg is not in your path.PYTHON_PATH
- only needs to be set if python is not in your path.
Available command line options for the cli can be listed with the --help
argument
autovmaf [source]
run transcode and vmaf analysis for videofile source
Commands:
autovmaf [source] run transcode and vmaf analysis for
videofile source [default]
autovmaf suggest-ladder <folder> Suggest bitrate ladder given vmaf results
autovmaf export-csv <folder> Export Vmaf results as csv
autovmaf vmaf Use pipeline to run VMAF for the given
distorted and reference files
Positionals:
source SOURCEFILE [string]
Options:
--version Show version number [boolean]
--help Show help [boolean]
--resolutions List of resolutions, ie 1920x1080,1280x720... [string]
--bitrates List of bitrates, ie 800k,1000k,... [string]
--name Name for this autovmaf run [string]
--models List of VMAF Models to use [string]
--job File with job definition [string]
--saveAsCsv Save VMAF measurements as a .csv file in addition to a JSON
file [boolean] [default: false]
--skipTranscode Skip transcode and run vmaf on already transcoded files
[boolean] [default: false]
--skipExisting Skip transcode for already transcoded files
[boolean] [default: true]
--probeBitrate Read bitrate of transcoded file with ffprobe
[boolean] [default: false]
--ffmpeg-options List of options to pass to ffmpeg, on the form
key1=value1:key2=value2 [string]
Output files will be stored in a folder corresponding to the argument given to the --name
option.
If resolutions and/or bitrates are not specified default values will be used, See above.
With the --job
option, a path to a yaml or json file with a job definition can be passed to to the cli. The values
defined in the file can be overridden with other commandline options. For instance the reference
video defined
in the job file can be overridden by passing a source file on the command line.
It is possible to iterate over other variables than bitrate and resolutions when running a local encode. For instance, to run transcode and vmaf analysis with x265 in CRF mode for a number of CRF values, a job definition like below can be used (also available in examples/local/local-job-crf.yaml)
models:
- HD
resolutions:
- width: 1920
height: 1080
bitrates:
- 0
pipeline:
ffmpegEncoder: libx265
singlePass: true
skipDefaultOptions: true
ffmpegOptions:
'-pix_fmt': 'yuv420p'
'-preset': 'veryslow'
'-x265-params': 'crf=${CRF}:scenecut=0:keyint=50:min-keyint=50:open-gop=0'
easyVmafExtraArgs:
'-threads': 20
pipelineVariables:
CRF:
- 22
- 26
- 30
- 34
This will run transcode and vmaf analysis for CRF values 22,26,30, and 34. Variables are used in the ffmpeg options
by inserting ${VARIABLENAME}
. This string will then be substituted with a value from the list of values from
pipelineVariables.VARIABLENAME
. Note that when running CRF encode or other non-ABR mode, skipDefaultOptions
must
be set to avoid injecting bitrate options to ffmpeg. Also note that the cli needs to be run with the --probe-bitrate
option to get the correct bitrate from the transcoded files.
It is also possible to use pipelineVariables with the AWSPipeline. The following example will run transcode and vmaf analysis using the AWS MediaCOnvert QVBR levels 6, 7, 8 and 9.
name: job-name
pipeline: pipeline.yml
encodingProfile: encoding-profile.json
reference: reference.mp4
models:
- HD
resolutions:
- width: 1920
height: 1080
bitrates:
- 0
pipelineVariables:
QVBR:
- 6
- 7
- 8
- 9
autovmaf --resolutions 1920x1080,1280x720,960x540 --bitrates 500k,800k,1200k,1600k,2000k,3000k,4000k --name my-autovmaf-test1 my-source-video.mp4
With the above command, when the run is finished transcoded files will be available in the folder my-autovmaf-test1
, and vmaf-data in the folder my-autovmaf-test1/HD
.
To export results to csv, use the export-csv
command.
autovmaf export-csv <folder>
Export Vmaf results as csv
Positionals:
folder Folder with vmaf measurement results [string] [required]
Options:
--version Show version number [boolean]
--help Show help [boolean]
--probeBitrate Read bitrate of transcoded file with ffprobe
[boolean] [default: false]
--variables List of variables to include as columns in csv [string]
If your job uses variables (See above), the variables that should
be included in the csv data should be specified with the --variables
option.
This command allows using a configured pipeline to run VMAF measurements on any files without a related autovmaf job.
In case of an aws pipeline, output will be written to the outputBucket
configured in the pipeline.
autovmaf vmaf
Use pipeline to run VMAF for the given distorted and reference files
Options:
--version Show version number [boolean]
--help Show help [boolean]
--pipeline Path to pipeline file [string] [required]
--reference Uri or path to reference file [string] [required]
--distorted Uri or path to distorted file [string] [required]
--model VMAF model to use [string] [default: "HD"]
npm test
Eyevinn Technology is an independent consultant firm specialized in video and streaming. Independent in a way that we are not commercially tied to any platform or technology vendor.
At Eyevinn, every software developer consultant has a dedicated budget reserved for open source development and contribution to the open source community. This give us room for innovation, team building and personal competence development. And also gives us as a company a way to contribute back to the open source community.
Want to know more about Eyevinn and how it is to work here. Contact us at work@eyevinn.se!