Move AtomRef Fitting to numpy to avoid bug

[BowenD-UCB]

Summary

Change the fitting method of AtomRef from torch.linalg.lstsq to np.linalg.pinv mthod

Issue with previous implementation

for i in range(3):
    training_set, validation_set, test_set = split_dataset(
        dataset, 
        frac_list=[0.9, 0.05, 0.05], 
        random_state=42, 
        shuffle=True
    )
    train_graphs = []
    energies = []
    forces = []
    for (g, lat, l_g, attrs, lbs) in training_set:
        train_graphs.append(g)
        energies.append(lbs["energies"])
        forces.append(lbs['forces'])
    element_refs = AtomRef(torch.zeros(89))
    element_refs.fit(train_graphs, torch.hstack(energies))
    print(element_refs.property_offset)

The output for the above block gives:

tensor([ -3.3743,  -0.3157,  -2.5775,  -3.6737,  -7.1368,  -8.9913,  -8.6316,
         -8.3601,  -6.4603,   0.0000,  -3.3769,  -4.0194,  -6.7818,  -8.5984,
         -9.2705,  -8.9952,  -7.0834,   0.0000,  -5.5053,  -7.3827, -11.5063,
        -12.9402, -13.5851, -14.2839, -14.2125, -13.8046, -12.6301, -11.5235,
        -10.0505,  -8.2310, -11.4159, -13.4431, -15.1073, -15.3583, -13.2858,
        -12.5697, -13.3787, -14.9390, -20.1400, -22.2983, -23.6922, -24.7945,
        -24.2404, -24.4331, -24.0019, -22.8785, -20.6089, -19.5086, -22.2762,
        -24.5344, -25.7248, -25.8512, -24.8157,   0.0000, -24.7562, -26.4343,
        -29.8681, -30.8821, -30.2253, -29.7089, -29.5530, -30.3643, -39.8913,
        -41.8080, -31.4633, -32.3610, -33.5807, -34.2840, -35.7176, -36.9142,
        -38.3371, -44.4273, -47.3390, -49.7190, -50.7680, -51.4679, -51.8229,
        -51.4508, -50.5201, -48.9885, -52.5689, -55.8178, -58.1974, -68.5187,
        -73.4779, -76.3807, -79.9365, -82.4917, -86.1527])
tensor([ -3.3741,  -0.3157,  -2.5770,  -3.6740,  -7.1357,  -8.9899,  -8.6315,
         -8.3597,  -6.4601,   0.0000,  -3.3769,  -4.0186,  -6.7819,  -8.5977,
         -9.2708,  -8.9946,  -7.0839,   0.0000,  -5.5013,  -7.3825, -11.5063,
        -12.9395, -13.5878, -14.2835, -14.2104, -13.8024, -12.6283, -11.5223,
        -10.0490,  -8.2301, -11.4112, -13.4407, -15.1071, -15.3573, -13.2845,
        -12.5697, -13.3787, -14.9336, -20.1477, -22.3081, -23.6842, -24.7941,
        -24.2397, -24.4296, -24.0034, -22.8813, -20.6172, -19.5162, -22.2746,
        -24.5340, -25.7308, -25.8615, -24.8152,   0.0000, -24.7562, -26.4344,
        -29.8817, -30.8813, -30.2255, -29.7091, -29.5509, -30.3798, -39.8911,
        -41.8073, -31.4633, -32.3481, -33.5813, -34.2697, -35.7157, -36.9134,
        -38.3561, -44.4379, -47.3521, -49.7313, -50.7668, -51.4649, -51.8226,
        -51.4462, -50.5176, -48.9868, -52.5704, -55.8190, -58.1957, -68.5190,
        -73.4780, -76.3839, -79.9386, -82.4904, -86.1447])
tensor([-2.3409e+00,  0.0000e+00, -3.8910e+01, -3.3982e+01, -3.6587e+00,
        -2.2113e+00, -5.9076e+00, -9.2585e+00, -2.3856e+01,  0.0000e+00,
        -9.4978e-01, -1.9004e+00, -2.2413e+00, -4.4733e+00, -2.2184e+00,
        -1.5416e+00, -4.9846e-01,  0.0000e+00, -2.4982e-01, -1.4432e+00,
        -1.5095e+00, -3.0284e+00, -1.6905e+00, -1.0165e+00, -2.1386e+00,
        -4.3439e+00, -4.6343e+00, -2.4745e+00, -2.5669e+00, -1.1788e+00,
        -1.2643e+00, -2.8717e+00, -1.3953e+00, -6.4185e-01, -4.3861e-01,
         0.0000e+00, -5.6298e-01, -2.0991e+00, -2.1398e+00, -1.4777e+00,
        -1.3475e+00, -1.0482e+00, -6.0546e-01, -4.2556e+00, -4.3206e+00,
        -2.9284e+00, -1.3899e+00, -1.2171e+00, -1.9820e+00, -1.3938e+00,
        -1.5526e+00, -7.1934e-01, -2.0661e-01,  0.0000e+00, -4.1892e-01,
         0.0000e+00, -1.4509e+00, -8.1555e-01,  0.0000e+00,  0.0000e+00,
         0.0000e+00,  0.0000e+00, -3.1925e-01, -4.8536e-01,  0.0000e+00,
         0.0000e+00,  0.0000e+00,  0.0000e+00, -1.1423e-06,  0.0000e+00,
        -3.3690e-01, -1.6648e+00, -1.7475e+00, -1.3389e+00, -2.6458e+00,
        -4.1660e+00, -4.1146e+00, -5.2788e+00, -2.3409e+00, -1.6857e+00,
        -1.1703e+00, -1.3632e+00, -1.4343e+00, -8.8050e-01, -6.5525e-01,
        -2.6236e-01, -5.4861e-01, -1.6136e+00, -4.7265e-01])

The fitted parameters have super large deviations from several runs.
np.linalg.pinv should solve the issue

[coderabbitai]

Walkthrough

The changes in the codebase involve transitioning certain tensor operations from PyTorch to NumPy within the _AtomRef class. This update includes returning a NumPy array instead of a PyTorch tensor in the get_feature_matrix method and utilizing NumPy functions for matrix operations in the fit method. These modifications aim to enhance compatibility and potentially boost performance.

Changes

File Path	Change Summary
.../src/matgl/layers/_atom_ref.py	- Converted return value of get_feature_matrix method to NumPy array. - Updated fit method to use NumPy for matrix operations.

---

Recent Review Details

Configuration used: .coderabbit.yaml
Review profile: CHILL

Commits

Files that changed from the base of the PR and between 67d10ce and 8bf79d7.

Files selected for processing (1)

• src/matgl/layers/_atom_ref.py (2 hunks)

Files skipped from review as they are similar to previous changes (1)

• src/matgl/layers/_atom_ref.py

---

Thank you for using CodeRabbit. We offer it for free to the OSS community and would appreciate your support in helping us grow. If you find it useful, would you consider giving us a shout-out on your favorite social media?

Share

• X
• Mastodon
• Reddit
• LinkedIn

Tips

Chat

There are 3 ways to chat with CodeRabbit:

• Review comments: Directly reply to a review comment made by CodeRabbit. Example:
  • I pushed a fix in commit <commit_id>.
  • Generate unit testing code for this file.
  • Open a follow-up GitHub issue for this discussion.
• Files and specific lines of code (under the "Files changed" tab): Tag @coderabbitai in a new review comment at the desired location with your query. Examples:
  • @coderabbitai generate unit testing code for this file.
  • @coderabbitai modularize this function.
• PR comments: Tag @coderabbitai in a new PR comment to ask questions about the PR branch. For the best results, please provide a very specific query, as very limited context is provided in this mode. Examples:
  • @coderabbitai generate interesting stats about this repository and render them as a table.
  • @coderabbitai show all the console.log statements in this repository.
  • @coderabbitai read src/utils.ts and generate unit testing code.
  • @coderabbitai read the files in the src/scheduler package and generate a class diagram using mermaid and a README in the markdown format.

Note: Be mindful of the bot's finite context window. It's strongly recommended to break down tasks such as reading entire modules into smaller chunks. For a focused discussion, use review comments to chat about specific files and their changes, instead of using the PR comments.

CodeRabbit Commands (invoked as PR comments)

• @coderabbitai pause to pause the reviews on a PR.
• @coderabbitai resume to resume the paused reviews.
• @coderabbitai review to trigger a review. This is useful when automatic reviews are disabled for the repository.
• @coderabbitai resolve resolve all the CodeRabbit review comments.
• @coderabbitai help to get help.

Additionally, you can add @coderabbitai ignore anywhere in the PR description to prevent this PR from being reviewed.

CodeRabbit Configration File (.coderabbit.yaml)

• You can programmatically configure CodeRabbit by adding a .coderabbit.yaml file to the root of your repository.
• Please see the configuration documentation for more information.
• If your editor has YAML language server enabled, you can add the path at the top of this file to enable auto-completion and validation: # yaml-language-server: $schema=https://coderabbit.ai/integrations/schema.v2.json

Documentation and Community

• Visit our Documentation for detailed information on how to use CodeRabbit.
• Join our Discord Community to get help, request features, and share feedback.
• Follow us on X/Twitter for updates and announcements.

[JiQi535]

I think I also went into this before. This seems to be an issue only when using CPU, while it gives consistent results when using CUDA. Not sure why.

[kenko911]

@JiQi535 This may be the reason see the linkhttps://github.com/pytorch/pytorch/issues/71222. I also want to say that I didn't discover the problem when I used GPU. Anyway, I think it is better to have more consistent outputs with CPU and CUDA. Thanks Bowen for finding out the problem!