Freeze Cairo and Pango Versions

[ilkerkesen]

It seems that the current repository does not specify exact versions of pango, cairo, manimpango and pygobject packages. Based on our previous experience with PIXEL-based models, working with different versions of these packages can significantly affect performance.

For instance, consider a model pretrained with cairo==1.16.x. If we switch to cairo==1.18.x during downstream task fine-tuning, the model’s performance could degrade substantially. There are two reasons for this issue:

1. Different versions might produce slightly different visuals, which may be imperceptible to humans.
2. PIXEL-based models currently process 16×16 patches using 16×16 convolutional filters. Thus, even the slightest change in patch visual can disrupt input processing in PIXEL-based models. This would be the case for WeLT if one follows the same design choices.

So, the action items would be,

• Detect the versions of pango, cairo, manimpango and pygobject.
• Check the rendering whether it works without any issues.
• Fix the versions.

@AmitMY I can fix the versions if this is okay by you also as well.

[AmitMY]

Thanks for the clear and detailed explanation!
It makes sense to pin the versions, given the above scenarios.

Since these packages are not easily installed with pip, my preferred solution would be:

1. Specify exact versions in conda install -c conda-forge pycairo pygobject manimpango -y in the README.md, Dockerfile, and test.yaml
2. in renderer.py check the installed versions - and if they are not correct, raise an exception that also includes the link to this issue (to explain the motivation behind pinning versions)

See PR sign/WeLT#34 (made by copilot) - what do you think?

[AmitMY]

@ilkerkesen a better solution could be -
we can publish a single conda pakcage, called pixel-renderer that:

1. pins the dependencies
2. has a render function (whatever renderer.py does)

that way, we can pin a single dependency (e.g. pixel-renderer=0.1.0) and not manually write 5 package versions.

AND, more importantly, we can share code between this project and the rest of the pixel ecosystem (in terms of rendering)

[ilkerkesen]

@AmitMY I think that would be a better option, definitely.

[rggdmonk]

1. You don't need manimpango with latest versions pycairo and pygobject.
2. Should we get rid of conda-forge channel? And just move with pip install
3. Could you show examples?

For instance, consider a model pretrained with cairo==1.16.x. If we switch to cairo==1.18.x during downstream task fine-tuning, the model’s performance could degrade substantially. There are two reasons for this issue:

1. Different versions might produce slightly different visuals, which may be imperceptible to humans.
2. PIXEL-based models currently process 16×16 patches using 16×16 convolutional filters. Thus, even the slightest change in patch visual can disrupt input processing in PIXEL-based models. This would be the case for WeLT if one follows the same design choices.

They have stable API. The only it can happen:

1. (not sure thats possible), when you have different backend Darwin vs Linux
2. When you have not the same fonts (default Arial on MacOS != from web Arial)

[ilkerkesen]

@rggdmonk Let me try to replicate the problem minimally, but we experienced this issue on the same machine, same fonts etc. The only variation was in the mentioned packages' versions.

[ilkerkesen]

I could not replicate the slightly different visuals issue for the renderer currently implemented in this repo. However, the renderer implemented for PIXEL-M4 and PIXEL-BIGRAMS produces slightly different visuals across different versions, on the same machine with the same font (Go Noto Sans). This might be related to different antialiasing configurations across versions, although I am not 100% sure.

What I would do,

1. Freeze the versions.
2. Generate visuals using different writing systems on a predefined set of sentences.
3. In case of a version bump, compare the newly produced images against the previous set to ensure backward compatibility.

Here is the link if you want to see the visuals. The input text is ቱርክ በሶስት ጎኖች በባህር የተከበበች ናት-በምእራብ የኤጂያን ባህር ፣ በሰሜን በኩል ጥቁር ባህር እና በሜድትራንያን ባህር። (Amharic, Ge'ez script). If you zoom in and look at the images carefully, you can see that ች character is rendered differently in these images.

[elliottd]

We should definitely do something to freeze the currently used versions to prevent future problems because unfrozen versions have been a huge issue for some of our pixel projects.

[AmitMY]

Now only using pip 2c768e9