DimeNet++ implementation

[arunppsg]

In this PR, I implemented DimeNet++ by subclassing DimeNet. The InteractionPPBlock and OutputPPBlock was based on this implementation.

Also, I depreciated OutputBlock, thereby DimeNet model will also be using OutputPPBlock. OutputPPBlock allows up-projection and down-projection of embeddings, thereby removing information bottlenecks.

Fixes #4427

[codecov]

Codecov Report

Merging #4432 (dd4790f) into master (5a6e826) will decrease coverage by 0.64%.
The diff coverage is 10.38%.

@@            Coverage Diff             @@
##           master    #4432      +/-   ##
==========================================
- Coverage   82.98%   82.33%   -0.65%     
==========================================
  Files         319      319              
  Lines       16968    17117     +149     
==========================================
+ Hits        14081    14094      +13     
- Misses       2887     3023     +136     

Impacted Files	Coverage Δ
torch_geometric/nn/models/dimenet.py	14.58% <9.80%> (-3.09%)	⬇️
torch_geometric/nn/models/__init__.py	100.00% <100.00%> (ø)

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 5a6e826...dd4790f. Read the comment docs.

[rusty1s]

Thanks for the PR. Can we add a basic test for DimeNetPlusPlus? Also I am interested in whether it is possible to make use of pre-trained weights for this model as well (similar to what we do for DimeNet), see here.

[arunppsg]

Sure, I will look into using pre-trained weights and also add tests.

[arunppsg]

@rusty1s regarding tests for DimeNetPlusPlus, I have added a test for checking output size and another one test for checking whether the model is training well or not - test_overfit. The downside of test_overfit is that it is sometimes flaky but otherwise it is a good test for checking whether the model is able to train or not.

[rusty1s]

Thank you @arunppsg! I am currently on vacation, will have a final look in the next week.

[arunppsg]

The failures in testing are due to not having sympy in requirements. It is used by DimeNet++ and DimeNet for basis computation.

[rusty1s]

@arunppsg Can we add the necessary dependencies to full_install_requires, see here?

[arunppsg]

@rusty1s added the necessary requirements.

[rusty1s]

Thanks @arunppsg. I think I fixed the dependency issue during testing. I also added a --use_dimenet_plus_plus option to the QM9 script. However, the results of the pre-trained model are far off, e.g.:

Target: 00, MAE: 4.85522 ± 3.42503
Target: 01, MAE: 15.25113 ± 6.21239
Target: 02, MAE: 603.06146 ± 484.24606
Target: 03, MAE: 994.61578 ± 703.35645
Target: 05, MAE: 977.33118 ± 332.41879
Target: 06, MAE: 594.56433 ± 317.66690
Target: 07, MAE: 3095.07080 ± 1889.53760

Do you know what's the reason for this?

[arunppsg]

Thanks @arunppsg. I think I fixed the dependency issue during testing. I also added a --use_dimenet_plus_plus option to the QM9 script. However, the results of the pre-trained model are far off, e.g.:

Target: 00, MAE: 4.85522 ± 3.42503
Target: 01, MAE: 15.25113 ± 6.21239
Target: 02, MAE: 603.06146 ± 484.24606
Target: 03, MAE: 994.61578 ± 703.35645
Target: 05, MAE: 977.33118 ± 332.41879
Target: 06, MAE: 594.56433 ± 317.66690
Target: 07, MAE: 3095.07080 ± 1889.53760

Do you know what's the reason for this?

Not sure about it, the results are far off from the dimenet++ paper. Can we create an issue and track it separately?

[rusty1s]

This sounds good to me. I will try to look into this as well.

[arunppsg]

Thank you!