Implement DeepDDS #53

kkaris · 2022-01-19T03:13:35Z

Closes #19

Adds the DeepDDS model implementation

Code passes all tests
Unit tests provided for these changes
Documentation and docstrings added for these changes

Changes

Add DeepDDS
Add new file, deepdds_examples.py containing an example

kkaris · 2022-01-19T03:25:47Z

This branch obviously need more work, but for now I have this question:

There are two version of this model, one using GAT and one using GCN for the drug feature embedding. Is the idea to implement both of them in the same module and also duplicate examples, tests etc? Or do we pick one of the implementations?

Another comment: it would be nice to have the shapes/sized of the input data either available somewhere as importable constants or automatically set by the pipeline or some wrapper to the pipeline. I think @cthoyt is working on something along these lines in #52.

benedekrozemberczki · 2022-01-19T09:30:36Z

What could be done is a parameter which deploys gat/gcn in the main model. That way you can test the model with various parameter settings in the tests and a single example enough. What do you think?

benedekrozemberczki

Looks good! Thank you for your contribution. The molecular encoders could be parametrically deployed and that way you could switch between GCN and GAT.

benedekrozemberczki · 2022-01-19T09:33:05Z

chemicalx/models/deepdds.py

+        :param dropout:
+        """
+        super(DeepDDS, self).__init__()
+        self.cell_mlp = MLP(input_dim=context_feature_size, hidden_dims=[2048, 512, context_output_size])


Should be parametrized.

chemicalx/models/deepdds.py

benedekrozemberczki · 2022-01-19T12:01:44Z

@kkaris thanks to @cthoyt now we have the default atomic feature count for the datasets.

examples/deepdds_examples.py

cthoyt · 2022-01-19T15:11:43Z

I think the reason this is broken is the PackedGraph object has a shape of torch.Size([149868, 149868, 4]) instead of being a list of batch_size graphs. I think the solution is to apply a MaxReadout() at the end (which I think is the same as max pooling, but not sure)

benedekrozemberczki · 2022-01-19T15:14:18Z

Max readout pools the vectors and creates something that is batch_size X dim. It is a scattering transform - a way to think about it is a group by aggregate where nodes are grouped together by the source graph.

https://github.com/rusty1s/pytorch_scatter

The figure here is pretty great.

benedekrozemberczki

LGTM. That failure of the CI on the default values is weird.

benedekrozemberczki · 2022-02-03T17:04:05Z

chemicalx/models/deepdds.py


-from .base import UnimplementedModel


This is beautifully done.

benedekrozemberczki · 2022-02-03T17:04:27Z

chemicalx/models/deepdds.py


 __all__ = [
    "DeepDDS",
 ]


-class DeepDDS(UnimplementedModel):
+class DeepDDS(Model):


Looks good to mention both.

benedekrozemberczki · 2022-02-03T17:06:44Z

chemicalx/models/deepdds.py

+    ):
+        """Instantiate the DeepDDS model.
+
+        :param context_feature_size:


This is very detailed.

benedekrozemberczki · 2022-02-03T17:07:00Z

examples/deepdds_examples.py

+    dataset = DrugCombDB()
+    model = DeepDDS(
+        context_feature_size=dataset.context_channels,
+    )


Looks good.

cthoyt · 2022-02-03T18:01:29Z

The failure is because there are only a subset of special parameters (i.e., drug_channels, context_channels) that are allowed to not have a default. Fixed in 2a8e7ef

benedekrozemberczki · 2022-02-03T18:11:03Z

@cthoyt Can I merge?

kkaris requested review from cthoyt and benedekrozemberczki January 19, 2022 03:14

benedekrozemberczki reviewed Jan 19, 2022

View reviewed changes

cthoyt reviewed Jan 19, 2022

View reviewed changes

chemicalx/models/deepdds.py Outdated Show resolved Hide resolved

kkaris force-pushed the deepdds-19 branch from c6de28a to 2425639 Compare January 19, 2022 12:17

cthoyt reviewed Jan 19, 2022

View reviewed changes

examples/deepdds_examples.py Show resolved Hide resolved

kkaris force-pushed the deepdds-19 branch from 2425639 to 646cdad Compare February 1, 2022 14:08

cthoyt added the model label Feb 1, 2022

kkaris force-pushed the deepdds-19 branch from 61938d6 to b2aba72 Compare February 2, 2022 21:46

kkaris added 16 commits February 3, 2022 09:24

WIP First iteration of DeepDDS model

9c0d311

WIP Add reminder implement GAT version of model as well

4392065

WIP implement example

e9034c0

WIP Flake 8 fixes

293d9db

Import and use constant

116f007

Flake8

64e4014

WIP Update example

06bd77e

Default in_channels to TORCHDRUG_NODE_FEATURES

d8533f4

Add MaxReadout layer

a2c176f

Comment/doc spelling/grammar

a879ce8

Some cleanup

477676a

Update code, comments, where they diverge follow code

e663a61

Update example

68f738a

Rename: in_channels -> drug_channels

7326974

Update docstring

0213e39

Parameterize hidden layers, set defaults

f1c3886

kkaris and others added 6 commits February 3, 2022 09:25

Update class docstring

1e17c27

Update test

669c5cc

Tox lint did this

c340735

WIP: half working test

f97303d

Make test and example run

e4b29dc

Code cleanup

29bc649

kkaris force-pushed the deepdds-19 branch from dcd203c to 29bc649 Compare February 3, 2022 14:25

Small refactoring, big lolz

03c8e5d

cthoyt marked this pull request as ready for review February 3, 2022 16:42

benedekrozemberczki approved these changes Feb 3, 2022

View reviewed changes

Fix name

2a8e7ef

cthoyt changed the title ~~Implementation of the DeepDDS model~~ Implement DeepDDS Feb 3, 2022

cthoyt merged commit 400a983 into AstraZeneca:main Feb 3, 2022

kkaris deleted the deepdds-19 branch February 13, 2022 20:45

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Implement DeepDDS #53

Implement DeepDDS #53

kkaris commented Jan 19, 2022 •

edited by cthoyt

Loading

kkaris commented Jan 19, 2022 •

edited

Loading

benedekrozemberczki commented Jan 19, 2022

benedekrozemberczki left a comment

benedekrozemberczki Jan 19, 2022

benedekrozemberczki commented Jan 19, 2022

cthoyt commented Jan 19, 2022

benedekrozemberczki commented Jan 19, 2022

benedekrozemberczki left a comment

benedekrozemberczki Feb 3, 2022

benedekrozemberczki Feb 3, 2022

benedekrozemberczki Feb 3, 2022

benedekrozemberczki Feb 3, 2022

cthoyt commented Feb 3, 2022 •

edited

Loading

benedekrozemberczki commented Feb 3, 2022

Implement DeepDDS #53

Implement DeepDDS #53

Conversation

kkaris commented Jan 19, 2022 • edited by cthoyt Loading

Changes

kkaris commented Jan 19, 2022 • edited Loading

benedekrozemberczki commented Jan 19, 2022

benedekrozemberczki left a comment

Choose a reason for hiding this comment

benedekrozemberczki Jan 19, 2022

Choose a reason for hiding this comment

benedekrozemberczki commented Jan 19, 2022

cthoyt commented Jan 19, 2022

benedekrozemberczki commented Jan 19, 2022

benedekrozemberczki left a comment

Choose a reason for hiding this comment

benedekrozemberczki Feb 3, 2022

Choose a reason for hiding this comment

benedekrozemberczki Feb 3, 2022

Choose a reason for hiding this comment

benedekrozemberczki Feb 3, 2022

Choose a reason for hiding this comment

benedekrozemberczki Feb 3, 2022

Choose a reason for hiding this comment

cthoyt commented Feb 3, 2022 • edited Loading

benedekrozemberczki commented Feb 3, 2022

kkaris commented Jan 19, 2022 •

edited by cthoyt

Loading

kkaris commented Jan 19, 2022 •

edited

Loading

cthoyt commented Feb 3, 2022 •

edited

Loading