Improve Setup and Docs

[patchmeifyoucan]

Hello,

I am developing a similar application to yours and would like to include yours in a publication benchmark.
However, I was not able to get the repository running. Errors range from issues with OCR to segmentation
faults, which is pretty bad. The barrier to "just run" the benchmark on a few thousand images is quite high.

Would it be possible to streamline the whole process? Tools like DECIMER
provide a pip package which an easy-to-use Python interface (which works on CPU only on my machines, which
I find ok since it is quite an effort to have broad CUDA support as repos get older).
A minimalistic version would include one import, either the model itself of an inference function. Additionally,
there should be a simpler way to configure the model other than commenting code.
So far, there is no documentation about how the system works, what it does in terms of its code and flags are hidden
in scripts and issues. Behavior like this should ideally also be
documented such that someone looking at the repo the first time gets a sense of what different modes the
model supports.

I think something along these lines would help users of the model or those wanting to benchmark against it
since it is among the best models currently and should be used and included in OCSR benchmarks.

Thanks a lot.

[geoHeil]

see #18 - this might help you

[patchmeifyoucan]

Thank you.

[lucas-morin]

Hello @patchmeifyoucan, thanks for your interest in MolGrapher.
As you suggest, I will try to improve the setup and documentation.

Could you please share the exact error that you encountered? Which operating system are you using?

[patchmeifyoucan]

Hello and thanks to coming back to this.

Given the instructions in #18, I managed to get running what I need. The scripts work and if you only need to do things once then that's fine. But there could be improvements to the programmatic use of the model. In general, it would be nice if the model interface was written such that one could, from Python code, load data yourself in whatever way, and then pass batches of images to the model's forward method. This makes it more flexible since the current setup is quite tied to the project structure. As far as I can see, this would require a bit of work. Here are some more detailed suggestions:

• First and foremost, for writing code, it would be good if the package manager used would be supported by all major tooling. I did not manage to hack Pixi to get an interpreter running in PyCharm. It would be nice if the project would rather use uv. Or at least have a uv-compatible pyproject.toml. Otherwise, it's impossible to debug code.
• The files from refine packaging process  #18 should be merged to the main branch. Especially when looking at the repo for the first time, you want to see the dependencies which you need to install. They exist with a proper environment management but are hidden in issues. Most people would give up after an initial try I guess.
• There should somewhere be an object or method available that hides all of the setup and prediction complexities of the model. So far, I need to somewhere generate image files and image meta data, place them correctly, run the script, read the outputs and merge back the resulting json according to the meta data I curate, simply to know which prediction was which. A simply forward method that takes the images and returns the graphs or SMILES would be much nicer and also easier to integrate into more complex workloads.
• It might be a good idea to not use the location of the model's in the setup. Because of this, even if people would try to implement it themselves, they would be strictly tied to the project's organization and at this point, most would also give up.
  Ideally, everything should be loaded in a clear way outside the model and rather passed to it. That makes moving that code around much easier, which you need to do if you want to use it in own code. Maybe it also makes sense to document the overall system a bit. Maybe some of the problems I'm describing are already solved but not obvious how to achieve that.

Let me know if anything is unclear.

[lucas-morin]

Thanks for your detailed feedback.

1. As you suggest, I'll try to convert the setup.py to a pyproject.toml.
2. Currently the dependencies are accessible here:

   MolGrapher/setup.py

   Line 65 in 9333d6c

   install_requires=[

3. The README contains a minimal example to run predictions in Python (without the script):
   https://github.com/DS4SD/MolGrapher/tree/main?tab=readme-ov-file#python
4. As you propose, I'll try to get rid of the model explicit location.

Best,
Lucas

[patchmeifyoucan]

You are right, the interface is already there... sorry for that.

What about this pyproject.toml? In principle, this only has to be made compatible with uv.

[lucas-morin]

No problems. Ah yes I'll check if this one supports PaddleOCR with GPU. 🙂

[geoHeil]

given that https://github.com/DS4SD/MolGrapher/pull/18/files only uses pip packages it should be rather simple to port it over to uv.

I guess the critical piece might be translating this pixi notation to uv

paddlepaddle = { version = ">=3.0.0rc1,<3.1", index = "https://www.paddlepaddle.org.cn/packages/stable/cu123/" }

FYI pixi behind the scenes uses uv as well - but additionally gives you conda packages.

FYI there is a pycharm integration available https://pixi.sh/v0.20.1/ide_integration/pycharm/ in case you want to move forward and merge the MR.

Depending on the complexity of the other computer vision packages used in your setup potentially sticking to conda might be beneficial.

[patchmeifyoucan]

Hello again,

have you been able to port the dependencies? For me, the PyCharm shim does not work.

Attempting to run: ['run', '-n', 'pixi', '--no-capture-output', 'python', '-c', 'import sys; print(sys.executable)']
Attempting to run: ['run,', '-n,', 'pixi,', '--no-capture-output,', 'python,', '-c,', 'import sys; print(sys.executable)']
Traceback (most recent call last):
  File "/home/andre/code/repos/supervised-graphs/benchmarks/molgrapher/MolGrapher/.pixi/envs/default/libexec/conda", line 182, in <module>
    sys.exit(main(args))
             ~~~~^^^^^^
  File "/home/andre/code/repos/supervised-graphs/benchmarks/molgrapher/MolGrapher/.pixi/envs/default/libexec/conda", line 165, in main
    raise NotImplementedError(str(args))
NotImplementedError: ['run,', '-n,', 'pixi,', '--no-capture-output,', 'python,', '-c,', 'import sys; print(sys.executable)']
Attempting to run: ['run', '-n', 'pixi', '--no-capture-output', 'python', '-c', 'import sys; print(sys.executable)']

The interpreter does not show up properly. It just runs some commands and then simply does nothing. Any suggestions for this? It would be nice if I could run the code without the script as I need to benchmark on 10k images with several systems and I do not want to resort to the file-based script since I generate data on-the-fly and this would make stuff much more complicated than needed.

Or are there any issues regarding the conda packages? I do not have much time left, so if there are issues with porting it to uv then please let me know, I'll sort it out otherwise then.

Thank you.

[patchmeifyoucan]

Due to my time constraints, I implemented something that uses the scripts which work on CPU. But the GPU-env also did not work for me, unfortunately.