How to get it to work on windows?

[simonmb]

I am running on windows. I followed the instructions for installation:

git clone https://github.com/DS4SD/MolGrapher.git
conda create -n molgrapher python=3.11
conda activate molgrapher
cd MolGrapher
pip install -e .["cpu"]

However, when I run:

from molgrapher.models.molgrapher_model import MolgrapherModel

model = MolgrapherModel()
images_or_paths = ["./data/benchmarks/default/images/image_0.png"] 
annotations = model.predict_batch(images_or_paths) 

I get the error:

[lucas-morin]

Hello @simonmb,

I don't manage to reproduce your error.
I followed the steps you listed. Then, I additionally downloaded the models with:

wget https://huggingface.co/ds4sd/MolGrapher/resolve/main/models/graph_classifier/gc_gcn_model.ckpt -P ./data/models/graph_classifier/
wget https://huggingface.co/ds4sd/MolGrapher/resolve/main/models/graph_classifier/gc_no_stereo_model.ckpt -P ./data/models/graph_classifier/
wget https://huggingface.co/ds4sd/MolGrapher/resolve/main/models/graph_classifier/gc_stereo_model.ckpt -P ./data/models/graph_classifier/
wget https://huggingface.co/ds4sd/MolGrapher/resolve/main/models/keypoint_detector/kd_model.ckpt -P ./data/models/keypoint_detector/

(But it should not be related to the downloading of models).
Running the example code snippet gave me the expected output.
Are you sure to be running your test in the correct environment?

[simonmb]

I forgot to mention that I changed setup.py to make it work on windows like so:

def install_torch(package: str, version: str = '', cpu: bool=True):
    """
    Creates path to download PyTorch (example: torch @ https://download.pytorch.org/whl/cpu/torch-2.1.2%2Bcpu-cp311-cp311-linux_x86_64.whl).
    Packages can be found at: https://download.pytorch.org/whl/.
    """
    cuda = "arm" not in platform.platform()
    if cpu:
        python_version = ''.join(platform.python_version().split('.')[:2])
        return ''.join([
            f'{package} @ https://download.pytorch.org/whl/',
            f'cpu/',
            f'{package}',
            f'-{version}' if version else '',
            '%2Bcpu',
            f'-cp{python_version}-cp{python_version}',
            'm' if int(python_version) <= 37 else '',
            '-linux_x86_64.whl' if platform.system() == 'Linux' else '-win_amd64.whl',
        ])
    else:
        python_version = "311"
        cuda_version = "117"
        return ''.join([
            f'{package} @ https://download.pytorch.org/whl/',
            f'cu{cuda_version}/' if cuda else '',
            f'{package}',
            f'-{version}' if version else '',
            f'%2Bcu{cuda_version}' if cuda else '',
            f'-cp{python_version}-cp{python_version}',
            'm' if int(python_version) <= 37 else '',
            '-linux_x86_64.whl' if platform.system() == 'Linux' else '-win_amd64.whl',
        ])

Other than that, I tried a second machine and repeated the same steps as stated above.

I get the exact same error. Yes I am using the correct env, Did you check on windows?

[lucas-morin]

Ah ok. To me, what you changed in the setup.py does not seem related to your error message.

I never tried to run MolGrapher on Windows, maybe you could try to:

• Add __init__.py files in the repositories under ./molgrapher/*
• Add the path to molgrapher to your system path

[simonmb]

Hey @lucas-morin , thanks, I had to add __init__.py in ./molgrapher and all of the subdirectories. On one machine I also had to install the GTK runtime.

Now I get to the point where the actual predictions happen, after Caption Removal Preprocessing. I am getting this error in molgrapher_model.py:

To me it sounds like this class is not picklable, whereby pickling is necessary for multiprocessing. If I set nb_workers to zero in the config and allow for some changes in the data_module.py to turn off multiprocessing, I get the model to work.

Do you know what the problem could be?

[lucas-morin]

I just pushed a small update—you might want to give it a try, though I doubt it will make a change. (I also added the init.py files.)

The parameter force_no_multiprocessing, which is set to True by default, here:

MolGrapher/molgrapher/models/molgrapher_model.py

Line 43 in 3d1ec15

"force_no_multiprocessing": True, # Disable PaddleOCR multiprocessing for abbreviation detection

controls whether multiprocessing is enabled for PaddleOCR during abbreviation recognition.

Unfortunately, I don't know how to get the multiprocessing of PaddleOCR to work on Windows.

However, to speed up the abbreviation recognition process, you can set to True the filter parameter, defined here:

MolGrapher/molgrapher/models/molgrapher_model.py

Line 219 in 3d1ec15

abbreviations_list = self.abbreviation_detector.mp_run(images_filenames, predictions["graphs"], bonds_sizes, filter=False)

This ensures that abbreviation recognition runs only on images where the node classifier has identified a node as an abbreviation.

[simonmb]

I'll have a look. I have 115k images I'd like to give it a go. Maybe I'll run it on a linux machine instead. How does the model compare to MolScribe?

[lucas-morin]

115k images should be doable yes! (We processed more than 100 million images to build PatCID.)

You can check the numbers on standard benchmarks on the two publications. Overall, MolScribe has a bit better accuracy on these. MolGrapher may be better for some of the large molecules or some molecules with abbreviations.

[geoHeil]

see #18