Skip to content

Latest commit

 

History

History
163 lines (108 loc) · 9.56 KB

README.md

File metadata and controls

163 lines (108 loc) · 9.56 KB
Raw

Topaz Video AI dockerized

This repository provides a turn-key solution to running Topaz Video AI on Linux (AMD64) in a containerized manner.

Usage

  1. Build the container

    make build

  2. Get an auth.tpz file for the machine you'll run the container on:

    make login

    Follow the two-step instructions presented when invoking the make target.

  3. Use VAI via ffmpeg

    make test

    The command shown contains all the information needed to run the container.

If you want to process clips on the host, just mount the directory containing the files to /workspace in the container, and they will appear at the default directory the container uses (CWD). E.g.

docker run --rm -ti --gpus all -hostname $HOSTNAME --user $UID:$GID \
-v $PWD/models:/models \
-v $PWD/auth/auth.tpz:/opt/TopazVideoAIBETA/models/auth.tpz \
-v $PWD:/workspace topaz-vai \
ffmpeg -i clip-from-host.mp4 ...

Customizing which information to mount where

If you want to separate the workspace from the models and license key, just mount those other directories instead. In my personal setup, I've got the models in /tmp/models on the host, and the license key in my home directory, so the command I use to run the container is instead:

docker run --rm -ti --gpus all -hostname $HOSTNAME --user $UID:$GID \
-v /tmp/models:/models \
-v $HOME/.vai/auth.tpz:/opt/TopazVideoAIBETA/models/auth.tpz \
-v $PWD:/workspace topaz-vai \
ffmpeg -i clip-from-host.mp4 ...

Where the container looks for the respective files is controlled by the ordinary Topaz environment variables;

  • TVAI_MODEL_DATA_DIR for the model files (the gigabytes of tpz files downloaded). Defaults to /models in the image.
  • TVAI_MODEL_DIR where ffmpeg looks for the model metadata files (the json files, such as aaa.json). Defaults to the install directory of VAI (/opt/TopazVideoAIBETA/models) since the VAI deb package installs all the model metadata files in that exact directory.

So you can change where ffmpeg looks for these two things by overriding those two environment variables when you launch a container. E.g. docker .... -e TVAI_MODEL_DATA_DIR=/mnt/some-dir -v $MY_HOST_MODEL_DIR:/mnt/some-dir, but I really don't see a reason anyone would want to. But if you do, then that's how.

Build different versions of TVAI

To build a container with a different version than the default, you need to specify two things:

  1. The version of TVAI to use.
  2. The SHA256 hash of the deb file for that version as obtained from Topaz.

Currently Topaz does not publish the SHA256 digest separately, so you have to download the entire deb file first and compute the message digest yourself; sha256sum TopazVideoAIBeta_<version>_amd64.deb

Example for building with a different version than the default:

$ curl -LO https://downloads.topazlabs.com/deploy/TopazVideoAIBeta/3.3.9.0.b/TopazVideoAIBeta_3.3.9.0.b_amd64.deb

$ sha256sum TopazVideoAIBeta_3.3.9.0.b_amd64.deb
a7de7730e039a8542280c65734143b6382c72eaa81e6fd8c0f23432ca13c8ba2  TopazVideoAIBeta_3.3.9.0.b_amd64.deb

$ docker build --tag topaz-vai:3390b \
  --build-arg "VAI_VERSION=3.3.9.0.b" \
  --build-arg "VAI_SHA2=a7de7730e039a8542280c65734143b6382c72eaa81e6fd8c0f23432ca13c8ba2" \
  .

Here are a couple of versions with their corresponding hashes, so you don't have to download the debs yourself redundantly just to compute the hashes:

a7de7730e039a8542280c65734143b6382c72eaa81e6fd8c0f23432ca13c8ba2  TopazVideoAIBeta_3.3.9.0.b_amd64.deb
b0deb4c919b6543879070cada170be2df9740db0a1be9fb16630151e005a8701  TopazVideoAIBeta_3.4.3.0.b-1_amd64.deb
607e1b9ad497a353f5efe901a1640a7fe1f9dc7445bbad16f86bf0969f5b9083  TopazVideoAIBeta_3.4.4.0.b_amd64.deb
2ce5c76f07e97d5c16d483644d1fae276061bc39a3a5c8220a262495876efb57  TopazVideoAIBeta_3.5.1.0.b_amd64.deb
27382ab6c3f3d0c81f3f18f3f7eacf760f740561ef12b567bd74ff73273f8392  TopazVideoAIBeta_3.5.2.0.b_amd64.deb
4c60c46ce9c0f76314319e304d08c646529aeb554a95ceb33b76f09044f4655a  TopazVideoAIBeta_3.5.3.0.b_amd64.deb
4d3971be331fe12fcd0e11e1e4666ef0ff3868cf03c3262338fc68b263afebf2  TopazVideoAIBeta_4.0.3.2.b_amd64.deb
5bef82e774b70f74f040958e6516d703fab839f19a6bab5ccf72a1e1fc4ccae3  TopazVideoAIBeta_4.0.5.0.b_amd64.deb
e8567bf60e1dec961cf4b471cd93c7ac63629ab49e97aac5b9e561409224d990  TopazVideoAIBeta_4.0.7.0.b_amd64.deb

This set of debs were those I could find in the Beta release notes that had binaries published for ubuntu, as of 2024-01-02.

Streamlining building and using different versions

Building and testing different versions of VAI is a common task, and sort of the whole point of this repository. As such that process has been streamlined using the provided Makefile.

$ make build VAI_VERSION=4.0.7.0.b \
             VAI_SHA2=e8567bf60e1dec961cf4b471cd93c7ac63629ab49e97aac5b9e561409224d990

will create the necessary docker build commands and give your container a tag derived from the VAI version number you provided. In the example, it will run these commands for you.

docker build -t topaz-vai \
--build-arg "VAI_VERSION=4.0.7.0.b" \
--build-arg "VAI_SHA2=e8567bf60e1dec961cf4b471cd93c7ac63629ab49e97aac5b9e561409224d990" .
docker tag topaz-vai topaz-vai:4070b

Do note that two docker tags are always produced; one being tagged with the specific VAI version, so you can refer back to older builds you may have, in order to compare the results between them, track down bugs or whatever. The other tag assigned is :latest, and is what the other make targets like login, test and benchmark use. Consequently those targets will use the latest container you built.

If you want to use those targets with a different version, then just set the latest tag to some of your other VAI builds. E.g. docker tag topaz-vai:3390b topaz-vai:latest

FAQ

Q1: I have a headless setup, will this work in that context ?

Yes. This project was specifically created to facilitate that use case. All you need is a valid auth.tpz file mounted or copied to the model's directory in the container: /opt/TopazVideoAIBETA/models/auth.tpz

Q2: Will I ever need to refresh the auth.tpz file ?

Yes. The file is only valid for a limited time. How long exactly, Topaz has not disclosed (to my knowledge).

If you run into a problem where the watermark suddenly starts being introduced, it may be because the license file expired. Just redo step 1 in the usage instruction to get a refreshed license file.

Q3: The make test doesn't run the encoding pipeline, just exits, why ?

Use version 3.5.1.0.b or later.

Topaz seems to have fixed the issue. I've confirmed that version as well as the latest of this writing now both launch with 100% successrate on my 3090. The below is historical but may still be relevant in case of future release regressions.

Most likely because of CUDA out of memory crash. Different versions of the Topaz engine produce different level of error details. 3.3.9 beta was very helpful reporting exactly the cause, down to the CUDA function that triggered the OOM. 3.4.4 beta seems less helpful and just silently dies prematurely.

The problem seems non-deterministic, and the only way I've found to circumnavigate the issue is to just try the same command again and again. May take as many as 10 attempts, but seems to work about 1/5th of the time.

This is one bug I really hope this repo can help Topaz Labs nail down and fix.

To see if this is the issue for you, run "make test" a number of times. Sometimes it helps to run nvtop in a different terminal on the same machine. Might be that nvtop resets some GPU state that the Topaz code has corrupted (?). Seems to reduce the probability of failure, at least on my ubuntu 22.04.1 host with nvidia driver 535.54.03, CUDA 12.2 and a Geforce 3090.

Q4: Why this repo?

Two reasons.

  1. To make it as painless as possible for users to run VAI on Linux.
  2. To provide both users and Topaz Labs employees a common and reproducible environment that allows for easier reproduction of user reported errors on Linux. Avoids wasted time and unhelpful support responses from Topaz Labs support representatives or engineers, such as "Unable to reproduce on our bespoke in-house environment".

Q5: Why not publish this image to dockerhub or other popular docker registries ?

Because there may be copyright issues involved. Needs to be cleared with Topaz Labs first in order to determine if this community service can be offered. This has not yet been done.

Q6: Why not provide a complete VM setup to avoid all variance entirely ?

This is a potential future improvement that is being considered. Right now I haven't been able to make VAI work on any of the AWS GPU instance types, so finding a suitable environment that both Topaz Labs and users can leverage would need to be discussed first.

For now having at least a common VAI setup that both the provider and customers can use is a first step. That reduces the variance to just the specific hardware used, the host kernel version and the specific NVIDIA driver used on the host. The rest is standardized in the container image.

Q7: With the license key mounted, I still get the watermark on videos. Why?

It's most likely because you don't provide a static hostname for the container, or the same hostname you used when "logging in" (getting the license key). Part of the minted license key information is the hostname. As such it's critical to use the same hostname for logging in as when running the container. That's why the examples in this document provide the --hostname argument to docker.