Spin lmp #34

iProzd · 2024-09-26T13:33:23Z

No description provided.

Fix deepmodeling#4151.  ## Summary by CodeRabbit - **New Features** - Enhanced path filtering logic to include a broader range of keys when generating subpaths. - **Bug Fixes** - Improved the accuracy of path results returned by the `glob` method.  Signed-off-by: Jinzhe Zeng <jinzhe.zeng@rutgers.edu>

See deepmodeling#4147 and deepmodeling#4143. We can first make `state_dict` safe for `weights_only`, then make a breaking change when loading `state_dict` in the future.  ## Summary by CodeRabbit - **New Features** - Enhanced model saving functionality by ensuring learning rates are consistently stored as floats, improving type consistency. - **Bug Fixes** - Updated model loading behavior in tests to focus solely on model weights, which may resolve issues related to state dictionary loading.

…modeling#4153) Per discussion in deepmodeling#4142 (comment). It should not be a public API as it lacks maintainance.  ## Summary by CodeRabbit - **New Features** - Introduced a new `eval_model` function in the testing module to enhance model evaluation capabilities with various input configurations. - **Bug Fixes** - Removed the old `eval_model` function from the main module to streamline functionality and improve code organization. - **Refactor** - Consolidated the import of `eval_model` to a common module across multiple test files for better organization and reduced dependencies.  --------- Signed-off-by: Jinzhe Zeng <jinzhe.zeng@rutgers.edu> Co-authored-by: Han Wang <92130845+wanghan-iapcm@users.noreply.github.com>

…one` (deepmodeling#4155)  ## Summary by CodeRabbit - **Bug Fixes** - Improved error handling by ensuring that the output data is not `None` before processing, preventing potential runtime errors.

## Summary by CodeRabbit ## Release Notes - **New Features** - Introduced JAX support, enhancing functionality and compatibility with JAX library. - Added new `JAXBackend` class for backend integration with JAX. - New functions for converting between NumPy and JAX arrays. - **Bug Fixes** - Improved compatibility of neural network layers with array API standards. - **Tests** - Added tests for JAX functionality and consistency checks against reference outputs. - Enhanced testing framework for activation functions and type embeddings. - **Chores** - Updated dependency requirements to include JAX library.  --------- Signed-off-by: Jinzhe Zeng <jinzhe.zeng@rutgers.edu>

Fix failed uts in deepmodeling#4145 .  ## Summary by CodeRabbit - **New Features** - Added a `"seed"` property to multiple JSON configuration files, enhancing control over randomness in model training and evaluation. - Introduced a global seed parameter in various test functions to improve reproducibility across test runs. - **Bug Fixes** - Ensured consistent random number generation in tests by integrating a global seed parameter. - **Documentation** - Updated configuration files and test methods to reflect the addition of the seed parameter for clarity and consistency.

…nsional data causes error (deepmodeling#4145) Fix issue deepmodeling#4108 If a pretrained model is labeled with energy and the `out_bias` is one dimension. If we want to finetune a dos/polar/dipole/property model using this pretrained model, the `out_bias` of finetuning model is multi-dimension(example: numb_dos = 250). An error occurs: `RuntimeError: Error(s) in loading state_dict for ModelWrapper:` ` size mismatch for model.Default.atomic_model.out_bias: copying a param with shape torch.Size([1, 118, 1]) from checkpoint, the shape in current model is torch.Size([1, 118, 250]).` ` size mismatch for model.Default.atomic_model.out_std: copying a param with shape torch.Size([1, 118, 1]) from checkpoint, the shape in current model is torch.Size([1, 118, 250]).` When using new fitting, old out_bias is useless because we will recompute the new bias in later code. So we do not need to load old out_bias when using new fitting finetune.  ## Summary by CodeRabbit - **New Features** - Enhanced parameter collection for fine-tuning, refining criteria for parameter retention. - Introduced a model checkpoint file for saving and resuming training states, facilitating iterative development. - **Tests** - Added a new test class to validate training and fine-tuning processes, ensuring model performance consistency across configurations.  --------- Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

`dp --pt freeze` does not output whether the operation finishes successfully. This PR adds a log message on it.

## Summary by CodeRabbit - **New Features** - Introduced a function to transform electronic configuration data into a spin representation, enhancing how electronic states are represented. - Updated the electronic configuration embedding to include negative values, reflecting a new encoding scheme. - Added a normalization function for electronic configuration vectors to improve data consistency. - **Tests** - Added a new test for the spin representation functionality to ensure accuracy of the transformation for iron (Fe). - Updated existing tests to align with the new expected output format.  --------- Co-authored-by: Han Wang <wang_han@iapcm.ac.cn> Co-authored-by: Duo <50307526+iProzd@users.noreply.github.com>

updates: - [github.com/astral-sh/ruff-pre-commit: v0.6.5 → v0.6.7](astral-sh/ruff-pre-commit@v0.6.5...v0.6.7) - https://github.com/pylint-dev/pylint/: v3.2.7 → v3.3.0  --------- Signed-off-by: Jinzhe Zeng <jinzhe.zeng@rutgers.edu> Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com> Co-authored-by: Han Wang <92130845+wanghan-iapcm@users.noreply.github.com> Co-authored-by: Jinzhe Zeng <jinzhe.zeng@rutgers.edu>

source/api_cc/include/DeepPotTF.h

+  // std::vector<double> virtual_len;
+  // std::vector<double> spin_norm;


source/api_cc/include/DeepPotTF.h

+  std::vector<int> extend_numneigh;
+  std::vector<std::vector<int>> extend_neigh;
+  std::vector<int*> extend_firstneigh;
+  // std::vector<double> extend_dcoord;
