Running in a Jupyter notebook works nicely. Running as a .py causes my computer to freeze due to insane cpu usage.

[willfinnigan]

Running MolScribe in a Jupyter notebook works nicely for me. Running the same code as a .py causes my computer to freeze due to insane cpu usage.

Running this piece of code on M1 Mac

model = MolScribe(f'{folder}/data/swin_base_char_aux_1m.pth', device=torch.device('cpu'))

def mol_scribe_osr(filepath):
    output = model.predict_image_file(filepath, return_atoms_bonds=True, return_confidence=True)
    return output['smiles']

[thomas0809]

Hi,

Running a Python script should have similar performance to running a Jupyter notebook locally.

To limit CPU usage, please see https://discuss.pytorch.org/t/how-to-limit-the-number-of-cpus-used-by-pytorch/67254/2

[willfinnigan]

Hi Thomas,

Thanks for the quick reply.

I agree, it seems really odd that I see this difference.

I tried running torch.set_num_threads(2) directly before the function call and in a few other places but with no luck. Looking in the activity monitor it seems 16 python processes are being launched (I have 10 cores - but 2 are slower). Running the code in a Jupiter notebook only launches a single process.

[tobbber]

@willfinnigan This might be related to #14, at least molscribe.chemistry.convert_graph_to_smiles spawns 16 processes as well