pd: support dpa1 #4414
pd: support dpa1 #4414
Conversation
for more information, see https://pre-commit.ci
…y to coverage newly added code
for more information, see https://pre-commit.ci
HydrogenSulfate
force-pushed
the
add_dpa1
branch
from
99aa33b to
d98c644
Compare
HydrogenSulfate
force-pushed
the
add_dpa1
branch
from
d98c644 to
e74d272
Compare
![coderabbitai[bot]](https://avatars.githubusercontent.com/in/347564?s=60&v=4){kind=small}
There was a problem hiding this comment.
Actionable comments posted: 0
🧹 Outside diff range and nitpick comments (1)
source/tests/consistent/model/test_dpa1.py (1)
248-255: Consider using ravel() for consistencyThe Paddle backend uses
flatten()while other backends useravel(). While both methods produce similar results, usingravel()would maintain consistency across all backends.- ret["energy"].flatten(), - ret["atom_energy"].flatten(), - ret["force"].flatten(), - ret["virial"].flatten(), - ret["atom_virial"].flatten(), + ret["energy"].ravel(), + ret["atom_energy"].ravel(), + ret["force"].ravel(), + ret["virial"].ravel(), + ret["atom_virial"].ravel(),
📜 Review details
Configuration used: CodeRabbit UI
Review profile: CHILL
📒 Files selected for processing (2)
source/tests/consistent/descriptor/test_dpa1.py(5 hunks)source/tests/consistent/model/test_dpa1.py(7 hunks)
🔇 Additional comments (11)
source/tests/consistent/model/test_dpa1.py (6)
17-17: LGTM: Import follows consistent pattern
The addition of INSTALLED_PD import follows the established pattern for backend availability flags.
41-45: LGTM: Paddle imports follow established pattern
The conditional import block for Paddle models follows the same structure as other backends, with proper fallback handling.
99-99: LGTM: Class attribute follows convention
The addition of pd_class follows the established pattern for backend class assignments.
112-113: LGTM: Reference backend selection properly extended
The addition of Paddle backend to reference selection follows the established priority order and error handling pattern.
131-132: LGTM: Model initialization properly handled
The Paddle model initialization in pass_data_to_cls follows the established pattern.
204-211: LGTM: Evaluation method follows convention
The eval_pd method correctly implements the standard evaluation interface used by other backends.
source/tests/consistent/descriptor/test_dpa1.py (5)
21-21: LGTM: Import changes follow established patterns
The import changes correctly follow the same pattern used for other backends, with proper conditional handling when Paddle is not installed.
Also applies to: 43-46
195-221: LGTM: Skip condition matches other backends
The skip_pd property correctly implements the same skip logic used by other backends, properly handling both common skip conditions and specific attention layer tests.
323-323: LGTM: Class assignment follows convention
The pd_class assignment follows the established pattern used for other backend implementations.
424-432: LGTM: Evaluation method matches other backends
The eval_pd method correctly implements the evaluation logic consistent with other backend implementations, properly passing all required parameters to the descriptor evaluation method.
43-46: Verify test coverage for Paddle backend
While the implementation looks correct, let's verify that all test cases will be properly executed for the Paddle backend.
Also applies to: 195-221, 323-323, 424-432
✅ Verification successful
Let me gather more information to verify the test coverage specifically for the Paddle backend.
Let me try one more verification to understand the test execution flow and Paddle backend handling.
Test coverage for Paddle backend is properly implemented
The verification shows that the Paddle backend test coverage is comprehensive and well-implemented:
- The test class
TestDPA1includes proper handling of the Paddle backend through:- Conditional import of
DescrptDPA1PD - A dedicated
skip_pdproperty that controls test execution - Implementation of
eval_pdmethod that properly evaluates Paddle descriptors
- Conditional import of
- The test infrastructure in
common.pyincludes complete Paddle-specific utilities:- Proper device handling with
PD_DEVICE - Paddle-specific neighbor list building
- Tensor conversion and evaluation methods
- Proper device handling with
- The test parameterization covers all relevant scenarios, and Paddle tests are skipped only when appropriate (e.g., when specific features are incompatible)
The implementation follows the same pattern as other backends (TF, PyTorch, JAX) ensuring consistent test coverage across all supported frameworks.
🏁 Scripts executed
The following scripts were executed for the analysis:
Script:
#!/bin/bash
# Description: Check test coverage for Paddle backend implementation
# Check if there are any test cases that might be inadvertently skipped
echo "Checking for potential gaps in test coverage..."
# Look for test methods that might not handle pd backend
rg -l "def test_.*" source/tests/consistent/descriptor/
rg "(?<!skip_)pd" source/tests/consistent/descriptor/
# Check for any TODO or FIXME comments related to pd implementation
rg -i "todo.*pd|fixme.*pd" source/tests/consistent/descriptor/
# Look for any inconsistencies in backend handling
rg "(tf|pt|jax|pd)_(class|obj)" source/tests/consistent/descriptor/
Length of output: 7067
Script:
#!/bin/bash
# Check test files for Paddle backend implementation and test methods
rg -A 5 "class.*Test" source/tests/consistent/descriptor/test_dpa1.py
rg "def test_" source/tests/consistent/descriptor/test_dpa1.py
# Check if there are any skip decorators or conditional test executions
rg "@.*skip" source/tests/consistent/descriptor/test_dpa1.py
rg "skipif|pytest.mark.skip" source/tests/consistent/descriptor/test_dpa1.py
# Check the test base class to understand test execution flow
rg -A 10 "class.*Base" source/tests/consistent/descriptor/common.py
Length of output: 514
Script:
#!/bin/bash
# Check the base test class implementation and test execution flow
cat source/tests/consistent/descriptor/common.py
# Check the complete test file for DPA1
cat source/tests/consistent/descriptor/test_dpa1.py
# Look for any conditional test execution patterns
rg -l "INSTALLED_PD" source/tests/consistent/descriptor/
Length of output: 20080
There was a problem hiding this comment.
Actionable comments posted: 0
🧹 Outside diff range and nitpick comments (2)
source/tests/consistent/model/test_dpa1.py (2)
112-113: Document the backend priority orderWhile the PD backend is correctly added to the reference backend selection, the priority order of backends (PT > TF > PD > JAX > DP) is not documented. Consider adding a comment explaining why this specific order was chosen.
248-255: Consider using ravel() for consistencyWhile the return value extraction is functionally correct, it uses
flatten()while other backends useravel(). Although both methods achieve similar results, consider usingravel()for consistency with other backend implementations.Apply this diff to maintain consistency:
elif backend is self.RefBackend.PD: return ( - ret["energy"].flatten(), - ret["atom_energy"].flatten(), - ret["force"].flatten(), - ret["virial"].flatten(), - ret["atom_virial"].flatten(), + ret["energy"].ravel(), + ret["atom_energy"].ravel(), + ret["force"].ravel(), + ret["virial"].ravel(), + ret["atom_virial"].ravel(), )
📜 Review details
Configuration used: CodeRabbit UI
Review profile: CHILL
📒 Files selected for processing (2)
source/tests/consistent/descriptor/test_dpa1.py(5 hunks)source/tests/consistent/model/test_dpa1.py(7 hunks)
🚧 Files skipped from review as they are similar to previous changes (1)
- source/tests/consistent/descriptor/test_dpa1.py
🔇 Additional comments (4)
source/tests/consistent/model/test_dpa1.py (4)
17-17: LGTM: Backend initialization follows established patterns
The PD backend initialization follows the same pattern as other backends, with proper conditional imports and fallback handling.
Also applies to: 41-45
99-99: LGTM: Class attribute follows convention
The pd_class attribute is correctly added, maintaining consistency with other backend class attributes.
131-132: LGTM: Model instantiation is consistent
The PD model instantiation follows the established pattern using the appropriate get_model_pd function.
204-211: LGTM: Evaluation method is properly implemented
The eval_pd method correctly follows the established pattern for backend evaluation methods, passing all required parameters.
Thanks for comments, and updated 2 consistency UT of dpa1, |
HydrogenSulfate
force-pushed
the
add_dpa1
branch
from
edea0aa to
63ccdbd
Compare
![github-advanced-security[bot]](https://avatars.githubusercontent.com/in/57789?s=60&v=4){kind=small}
There was a problem hiding this comment.
Actionable comments posted: 2
🧹 Outside diff range and nitpick comments (3)
source/tests/pd/model/water/multitask_sharefit.json (2)
11-15: Document the rationale behind selection parametersThe selection parameters
[46, 92, 4]seem arbitrary. Consider adding a comment explaining why these specific values were chosen and their significance in the context of water molecule simulations.
56-61: Consider adjusting the minimum learning rateThe
stop_lrvalue of 3.51e-08 is extremely small. Consider using a larger value (e.g., 1e-6) as training might not benefit from such small learning rates and could unnecessarily extend training time.source/tests/pd/test_multitask.py (1)
71-76: Document the conditions for parameter sharingThe conditions for parameter sharing in fitting networks exclude certain parameters (
bias_atom_eandcase_embd). Consider adding a comment explaining why these specific parameters are excluded from sharing.
📜 Review details
Configuration used: CodeRabbit UI
Review profile: CHILL
📒 Files selected for processing (2)
source/tests/pd/model/water/multitask_sharefit.json(1 hunks)source/tests/pd/test_multitask.py(6 hunks)
🧰 Additional context used
🪛 GitHub Check: CodeQL
source/tests/pd/test_multitask.py
[notice] 48-48: Unused global variable
The global variable 'multitask_sharefit_template' is not used.
🔇 Additional comments (2)
source/tests/pd/model/water/multitask_sharefit.json (1)
64-82: LGTM! Well-structured loss weighting strategy
The loss configuration implements a good strategy by:
- Initially prioritizing force accuracy (pref_f: 1000 → 1)
- Gradually increasing energy contribution (pref_e: 0.02 → 1)
- Maintaining identical configurations for both models
source/tests/pd/test_multitask.py (1)
43-48:
Remove or utilize the unused template variable
The multitask_sharefit_template variable is loaded but never used in the code. Either:
- Remove it if it's not needed, or
- Update the tests to utilize this shared fitting configuration
🧰 Tools
🪛 GitHub Check: CodeQL
[notice] 48-48: Unused global variable
The global variable 'multitask_sharefit_template' is not used.
|
@coderabbitai resolve |
This reverts commit e8167ce.
Support DPA-2 in paddle backend. This PR will be updated after #4414 is merged. ### Training curve:  ### Accuracy test(left: paddle, right: torch):  Ralated optimization of Paddle framework: - [x] PaddlePaddle/Paddle#69349 - [x] PaddlePaddle/Paddle#69333 - [x] PaddlePaddle/Paddle#69479 - [x] PaddlePaddle/Paddle#69515 - [x] PaddlePaddle/Paddle#69487 - [x] PaddlePaddle/Paddle#69661 - [x] PaddlePaddle/Paddle#69660 - [x] PaddlePaddle/Paddle#69596 - [x] PaddlePaddle/Paddle#69556 <!-- This is an auto-generated comment: release notes by coderabbit.ai --> ## Summary by CodeRabbit - **New Features** - Introduced new classes for molecular descriptors: `DescrptDPA2`, `DescrptBlockRepformers`, `DescrptSeTTebd`, and `DescrptBlockSeTTebd`. - Added new functions for tensor operations and descriptor management, enhancing the capabilities of the module. - Updated JSON configurations for multitask models to refine selection criteria and data paths. - **Bug Fixes** - Improved error handling and parameter validation across various descriptor classes. - **Documentation** - Enhanced test coverage for new descriptor functionalities and configurations. - **Tests** - Added new test classes to validate the functionality of `DescrptDPA2` and multitask training scenarios. - Expanded test capabilities for descriptor classes based on installed dependencies. - Updated existing tests to support new configurations and functionalities. <!-- end of auto-generated comment: release notes by coderabbit.ai --> --------- Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Summary of this PR:
softmax_grad,p_norm_grad,split_grad, andconcat_gradto the composite operator blacklist(https://github.com/deepmodeling/deepmd-kit/pull/4414/files#diff-e678abb052b278f8a479f8d13b839a9ec0effd9923478a850bc13758f918e1e9R134-R148) to significantly improve model execution speed (reducing the time taken from 100% more than PyTorch to about 10% to 15% more).related PR: lanpa/tensorboardX#728
Training curve:
Accuracy test(left: paddle, right: torch):
Ralated optimization of Paddle framework:
paddle.whereandpaddle.where_in eager mode PaddlePaddle/Paddle#69556Summary by CodeRabbit
Release Notes
New Features
DescrptDPA1,DescrptBlockSeAtten, andLayerNorm, enhancing the modeling capabilities for molecular simulations.DPAtomicModeland various descriptor classes.TestDenoiseModelDPA1andTestDenoiseModelDPA2.ModelWrapperclass to clarify the handling of model parameters and state management.Bug Fixes
Documentation
Tests