Are pretrained models available?

[mattragoza]

I would love to use this module within liGAN instead of depending on the full gnina platform. Have the caffe model weights been converted to pytorch somewhere publicly available?

[RMeli]

Unfortunately, I never converted the original Caffe models weights to PyTorch. The examples folder contains some scripts to reproduce part of the results of your and Paul's work, but are meant mostly for testing.

However, I would be very interested in doing so (or incorporate PRs that will allow to do that). It would be a great feature, and would serve as double check that everything works as expected. Is there any recommended way of doing this Caffe to PyTorch conversion?

PS: For easier integration, I could push the package to PyPI and conda-forge, if it is of interest.

[drewnutt]

I actually have these weights files for all of the built-in models easily available.

The built-in model weights were pulled directly from the source code of GNINA and converted to a .caffemodel. Then https://github.com/vadimkantorov/caffemodel2pytorch was utilized to convert the .caffemodel files to .pt for PyTorch.

I have created PR #34 to add these weights files.

[Kerro-junior]

@drewnutt Hi, would you please tell me where can I find the caffemodel for built-in *default2018.model? I have generated types file ready for training, as format <label> <pk> <rmsd> <rec> <lig>:

0 1.9966 3.0868 pdb2019_refi_train_gninatypes/185l/185l_rec.gninatypes first_pdbbind_v2019_docked_gninatypes/185l_docked_0.gninatypes
1 2.8809 1.3182 pdb2019_refi_train_gninatypes/185l/185l_rec.gninatypes first_pdbbind_v2019_docked_gninatypes/185l_docked_1.gninatypes

I've searched for days but couldn't find their .caffemodel weights to use in train.py to use it as a pre-trained task, when I searched 'caffemodel' there is just the following corresponding results:

[RMeli]

@Kerro-junior this question is related to GNINA, not gnina-torch. There are no Caffe weights in this repository, only their corresponding PyTorch version.

You can find the model definitions for GNINA in the GNINA repo here: https://github.com/gnina/gnina/tree/master/gninasrc/lib/models

And the mode weights are here: https://github.com/gnina/gnina/tree/master/gninasrc/lib/models/weights

[drewnutt]

@Kerro-junior I have uploaded the weights files for all of the default2018 models to gnina/models/crossdocked_paper with (gnina/models/pull/31)

[dkoes]

I'll add that there is now a dumpmodel binary built when you build gnina that will export the built-in caffe model weights (but not that actual caffe model file as you can get this easily from the source code).

[RMeli]

That's fantastic, thanks @drewnutt ! I'll incorporate PR #34 and start adding the functionality to use them easily (including model ensemble, which is currently not implemented).

[mattragoza]

Awesome! Thanks, @drewnutt.

[RMeli]

I opened #35 as a follow-up. I'll link relevant PRs there. Please feel free to comment with specific needs and missing features from your point of view.

[RMeli]

@mattragoza you can now easily load the pre-trained default2017 and *default2018* model as follows:

from gninatorch.gnina import load_gnina_model

model = load_gnina_model(MODEL_NAME)

where MODEL_NAME corresponds to name of the .pt files in gninatorch/weights (without the .pt extension); this is essentially equivalent to GNINA's --cnn option (but without model ensembles, which will follow).

The dense model is not yet supported because I haven't managed to obtain the same output as GNINA.

[RMeli]

An ensemble of models is now also available:

from gninatorch.gnina import load_gnina_models

ensemble = load_gnina_models([MODEL_NAME_1, MODEL_NAME_2, ...])

The ensemble of models returns log_CNNscore, CNNaffinity and CNNvariance.

Still no dense model for the time being (I suspect that there might be an issue with BatchNorm layer weights, but I'm still investigating), so it is not currently possible to reproduce GNINA's default model (since it contains two dense models). However, *_default2018_ensemble work as expected.

[RMeli]

@mattragoza Thanks once again to @drewnutt , who manually converted the weights of the dense model (see #43), the dense_ models and the dense_ensemble are now supported. The same goes for GNINA's default model ensemble.

See PR #44 for details. Related PRs: #38, #36, #34.

You can now easily load GNINA's pre-trained models as follows:

from gninatorch import gnina

model, ensemble = setup_gnina_model("MODEL")

where MODEL corresponds to the string you would use as --cnn argument in GNINA, while ensemble is a boolean flag that tells you if model is a single model (ensemble == False) or if it is an ensemble of models (ensemble == True).

[RMeli]

Now pip install gninatorch should do the trick, and by default will pull the dependencies needed to load and use GNINA models.