The goal of this project is to provide a docker image to run fast-artistic-videos. It will give users a consistent, functional environment to build from.
Speed is good, so it uses flownet2, cudnn, and stnbhwd. You will need a machine with an nvidia GPU attached, as well as nvidia-docker.
You could host anywhere once the image is built, but for convenience-sake, there are commands provided to deploy to Kubernetes Engine on Google Cloud.
Run npm install to install packages from npm
Make a copy of .env-template and rename it to .env. There are defaults there, but you can download and self-host these files for drastic speed improvements. It can take upwards of 4 hours to build the flownet2 image using the default host.
Run gcloud init to log in and configure the gcloud CLI.
In flownet2-docker/ run ./build.sh to build the positlabs/flownet2 image.
In the root, run npm run build to build the main image. It will be tagged based on the current gcloud project. You can edit env.js to change the tag.
npm run gcloud:cluster:create: create a cluster
npm run gcloud:pool:delete-default: delete the default (non-gpu) pool
npm run gcloud:pool:create: create a new pool with gpu
npm run gcloud:cluster:drivers: install nvidia drivers
npm run deploy: build and deploy the image to your cluster
Connect to the container via ssh.
npm run gcloud:cluster:auth to connect kubectl to the cluster
kubectl get po to find the pod id
kucbectl exec -it <pod_id> bash to connect to the pod
The container will be running in the background when deployed via npm scripts. You can connect to it and run commands internally. (This is largely for development purposes. In production, there would be a REST API or something similar - which would run the docker command.)
Some useful paths:
-
/io: The volume for reads/writes. Thenpm startcommand will mount the current working directory to this volume so the container can read your input video, and save the output in the same directory. -
/flownet2: flownet2 installation directory -
/fast-artistic-videos: Slightly modified fork of fast-artistic-videos. Makes the prebuilt model host configurable, and outputs to the /io directory -
/app: Contents of this repo. There's a test video (mov_bbb.mp4), and a CLI tool (/app/fav) that is the default command for running the container.
Inside the container, you can fake the container mount behavior by copying the video into the /io directory
cp /app/mov_bbb.mp4 /io/mov_bbb.mp4
Then you can run the command and stylize the video
/app/fav stylize mov_bbb.mp4 checkpoint-candy-video.t7
Because I don't have a local nvidia card, I have been running locally just to poke around the filesystem and test out the CLI.
npm start will start the container and keep it running in the background
npm run shell will open a bash shell inside the container
If you do have a local nvidia card, you could do something like this, which would run the CLI directly. Note the bind mount, which is needed for the container to have access to the video file.
nvidia-docker run --mount type=bind,source="$(pwd)",target=/io $(./dev/env DOCKER_TAG) stylize ./mov_bbb.mp4 checkpoint-candy-video.t7
List basic help: docker run $(./dev/env DOCKER_TAG) --help
Get details on a specific command: docker run $(./dev/env DOCKER_TAG) stylize --help
List the prebuilt style models: docker run $(./dev/env DOCKER_TAG) list-models