Docker Release Policy

[pgr0ss]

Hi, I noticed that the docker image seems to be about 7 months old, even though there have been newer releases:

• https://hub.docker.com/r/danielgatis/rembg/tags
• https://github.com/danielgatis/rembg/releases

Is there a release policy around the docker image? Is there a new version isn't pushed when the project has a release?

Thanks!

[danielgatis]

There is a GitHub Action that should run every time a tag is created, but for some reason, this GitHub Action stopped working. I would appreciate it if someone had the time to figure out what happened.

Here is the action:
https://github.com/danielgatis/rembg/blob/main/.github/workflows/publish_docker.yml

[pgr0ss]

I wonder if it's related to the runner image update: https://github.blog/changelog/2024-09-25-actions-new-images-and-ubuntu-latest-changes/

Can you try changing runs-on: ubuntu-latest to runs-on: ubuntu-22.04? If that doesn't work, I can try to figure out how to test this in a PR, and then I can iterate on it.

[pgr0ss]

I did a little more investigation, and I think what's happening is:

1. v2.0.59 added new models, which dramatically increased the size of the downloads: Add BiRefNet-General and BiRefNet-Portrait models as available models #665
2. The Dockerfile downloads all the models (RUN rembg d), and GitHub actions seems to start downloading very quickly (~ 80.0MB/s) but then after a few minutes, drops significantly (~2MB/s). I'm guessing some kind of rate limiting?
3. At the reduced speed, the GitHub action eventually times out.

When I build the docker image locally, you can see the jump in size:

REPOSITORY                          TAG          IMAGE ID       CREATED          SIZE
danielgatis/rembg                   latest       60f4d2719f66   7 months ago     4.37GB
remgb-mine                          latest       cae87817b347   11 minutes ago   17.9GB

One idea would be to only bundle a subset of the models in the Docker image by default. For example, just the default model u2net. That would shrink the downloads and the docker image significantly. In my local testing:

REPOSITORY                          TAG          IMAGE ID       CREATED          SIZE
rembg-u2net                         latest       2d363d1e8ca2   39 seconds ago   2.8GB

[pgr0ss]

It seems like the main source of slowness is building the arm64 image. All of the operations are slower, including the model download. For example, in my testing, building the amd64 image with all of the models takes only about 9 minutes, whereas building the arm64 model with just the u2net image takes 30 minutes. And the arm64 image with all of the models times out after several hours without finishing.

So another option if you want to keep an image with all of the models is to just publish an amd64 image and drop the arm64 image. I think you can still use the amd64 image from arm64 computers (e.g. newer macbooks) with docker run --platform amd64 ....

I can open up a PR shortly with some concrete changes to review.

[pgr0ss]

Here are two options:

• Only download u2net model in docker image #742
• Stop publishing arm64 image #743

I think either will fix the publishing build, so it depends on what you want in the final image.

[pgr0ss]

Thanks for merging these. Do you think you could tag a new release now so we get an updated docker image?

[danielgatis]

@pgr0ss sure lets try

[pgr0ss]

Seems like everything is working now, thanks!

[danielgatis]

yep! thanks a lot!