cvignac / midi Goto Github PK

Congrats on this very nice work!!
Are you planning to release the final weights of the model trained on the GEOM dataset?

Thanks,
Octavio

python3 main.py dataset=qm0 dataset.remove_h=True +experiment=qm9_no_h

This line is not even runnable. There is no dataset called qm0 and no experiment called qm9_no_h.

There are some typos I guess.

Hello Clément,

I'm trying to use your code for the QM9 dataset, but I have got this error :

                                                             ...
File "D:\Users\antoine\Documents\Sorbonne\Stage_PS\CTDGG\src\datasets\abstract_dataset.py", line 23, in __getitem__
    return self.dataloaders['train'][idx]
AttributeError: 'QM9DataModule' object has no attribute 'dataloaders'

It seems that the dataloaders attribute is not defined anywhere, hence raising an error when the Mixin __getitem__ method is called. In Digress implementation, it was defined in the prepare_data method which is not present anymore in this implementation. How do you create this attribute now ?

Note that I'm not using ray, nor ray-lightning packages for now, could it interfer with the dataloader creation ?

Nice update of Digress code btw, it's nice that you added multi-gpu support !

Best regards,

Antoine

Hi Clément,

When trying to use the testing procedure, I encountered a bug related to input_dims. It seems that in the current version of the code, they are updated twice : once in get_resume/load_from_checkpoint and once when the model is created.

It don't know exactly where it comes from but the following fix worked for me (at least for testing, haven't tried resuming) :

    if cfg.general.test_only: 
        cfg, model = get_resume(cfg, dataset_infos, train_smiles, to_absolute_path(cfg.general.test_only), test=True) 
    elif cfg.general.resume is not None: 
        # When resuming, we can override some parts of previous configuration 
        print("Resuming from {}".format(to_absolute_path(cfg.general.resume))) 
        cfg, model = get_resume(cfg, dataset_infos, train_smiles, to_absolute_path(cfg.general.resume), test=False) 
    else: 
        model = DiffusionModel(cfg=cfg, dataset_infos=dataset_infos, train_smiles=train_smiles) 

Also, it seems that all runs share the same seed when num_final_sampling is greater than 1, hence yielding the same results ...

Cheers,

Antoine

Hi, thanks for this very nice work and code.

I noticed this project does not have an explicit license. Could you please add a license stating the terms of use?

When setting up the environment with conda, pyg can not be installed. To solve this, when setting up the environment specify the RDKit version:

conda create -c conda-forge -n MoleculeDiffusion rdkit=2022.03.5 python=3.9

Alternatively @cvignac it might be worth adding an environment.yml so we can directly install the same package versions you are using.

MiDi/midi_src/datasets/adaptive_loader.py", line 6, in
from torch_geometric.data import LightningDataset

ImportError: cannot import name 'LightningDataset' from 'torch_geometric.data'

Hello,

I'm trying to use MiDi to generate molecules based on the model trained on GEOM with explicit H.
The trained model requires the dataset_infos as input, which needs the datamodule to get the statistics. However, I currently don't have enough RAM on my machine to load the training set of GEOM in the pickle file you provide.
I was thinking that probably having the processed files for the GEOMDrugsDataset could avoid the process() function (that is run when the processed files don't exist) and these files could be lighter than the whole pickle file containing molecules? Can you provide those ?
Or if you see another workaround (i.e. separating the statistics/configuration required for the dataset_infos in other files that do not always require the datamodule), please let me know?

Thank you very much,

Best,
Benoit