edge_dim added as argument to GATv2Conv

[allaffa]

This PR applies two major changes:

• allows to pass edge_dim as argument to GATStack because the original PyG implementation allows to use edge attributes GATv2Conv
• parametrizes the examples qm9, md17, and LennardJones over all the message passing layers

These changes are crucial to support the future integration go graph transformer architectures and allow a robust testing of those capabilities that should flexibly switch across all message passing layers.

[RylieWeaver]

@allaffa @ArCho48

Max, the fixes you put in place look good to me. For brevity here's my understanding:

It looks like you identified the following Errors and Fixes:
(1) Some MP-layers may not inherently support the existence of positional gradients. In this case, you exclude them from the LJ-example. This fixed the errors with: PNA, SAGE, GIN, GAT, MFC
(2) The examples did not allow you to choose the MP-layer via a command-line-argument. In this case, you added this optionality. This fixed the errors with PNAPlus, PNAEq

Additionally, I added the following in my commits:

(1) Fixes to do with object device placement during degree aggregation in distributed data processing and message-passing. This was not causing errors in only CPU-environments like these tests, but would cause errors in GPU environments and was identified when testing on the DGX.
(2) Squeezed the energy predictions in Base to to avoid the shape mismatch warning.
(3) Add a comment explaining which types of MP-layers are allowed in the grad_energy example in test_examples.py
(4) Added MACE to the grad_energy example because it will propagate a positional gradient.

Lastly, I wasn't able to reproduce the following warning to fix. Do we still want to tackle it?:

• /home/runner/work/HydraGNN/HydraGNN/hydragnn/utils/model/model.py:178: UserWarning: To copy construct from a tensor, it is recommended to use sourceTensor.clone().detach() or sourceTensor.clone().detach().requires_grad_(True), rather than torch.tensor(sourceTensor).

[allaffa]

@allaffa @ArCho48

Max, the fixes you put in place look good to me. For brevity here's my understanding:

It looks like you identified the following Errors and Fixes: (1) Some MP-layers may not inherently support the existence of positional gradients. In this case, you exclude them from the LJ-example. This fixed the errors with: PNA, SAGE, GIN, GAT, MFC (2) The examples did not allow you to choose the MP-layer via a command-line-argument. In this case, you added this optionality. This fixed the errors with PNAPlus, PNAEq

Additionally, I added the following in my commits:

(1) Fixes to do with object device placement during degree aggregation in distributed data processing and message-passing. This was not causing errors in only CPU-environments like these tests, but would cause errors in GPU environments and was identified when testing on the DGX. (2) Squeezed the energy predictions in Base to to avoid the shape mismatch warning. (3) Add a comment explaining which types of MP-layers are allowed in the grad_energy example in test_examples.py (4) Added MACE to the grad_energy example because it will propagate a positional gradient.

Lastly, I wasn't able to reproduce the following warning to fix. Do we still want to tackle it?:

• /home/runner/work/HydraGNN/HydraGNN/hydragnn/utils/model/model.py:178: UserWarning: To copy construct from a tensor, it is recommended to use sourceTensor.clone().detach() or sourceTensor.clone().detach().requires_grad_(True), rather than torch.tensor(sourceTensor).

@RylieWeaver
Thank you for helping me out in this PR.

[ArCho48]

Thank you both!

…

On Wed, Nov 27, 2024 at 12:26 Massimiliano Lupo Pasini < ***@***.***> wrote: Merged #310 <#310> into main. — Reply to this email directly, view it on GitHub <#310 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AMSKZ5POILL7MIXK7XZRVCT2CX6F5AVCNFSM6AAAAABSPOAGM6VHI2DSMVQWIX3LMV45UABCJFZXG5LFIV3GK3TUJZXXI2LGNFRWC5DJN5XDWMJVGQ2TKOBQGM4DANI> . You are receiving this because you were mentioned.Message ID: ***@***.***>