Changelogs » Megnet

PyUp Safety actively tracks 263,166 Python packages for vulnerabilities and notifies you when to upgrade.

Megnet

1.2.0

* Bug fix

1.1.9

* Add multi-fidelity model examples
  * Add sample weights for model training
  * Add default optimizer gradient norm clip

1.1.8

* Bug fix of megnet descriptors

1.1.7

* Update the model training mechanism and move Gaussian expansion to tensorflow, training speed up 100%

1.1.6

* minor fix to include more linear readout option
  * add data type control
  * refactor local_env

1.1.5

* Code refactor and reformat
  * Fix tensorflow and numpy type compatibility issues

1.1.4

* Update to tensorflow.keras API instead of using keras

1.1.3

* Download mvl_models from figshare if not present

1.1.2

* Add mvl_models in wheel release file

1.1.0

* Bug fix and version correction

1.0.3

* Fix bug brought by migrating to tensorflow 2.0
  * New elasticity models trained on 2019 MP data base
  * Add meg command line tools

1.0.2

* Add mypy typing hint for non-tensorflow codes
  * Update keras to 2.3.1 to fix thread-safety issues

1.0.1

* New find_points_in_spheres algorithm in pymatgen for graph construction

1.0.0

* Tensorflow 2.0 version.

0.3.4

* Change `convertor` to `converter` in all model APIs
  * Improve `ReduceLRUponNan` callback function
  * WardLT major contributions to the `MolecularGraph` class
  * Add serialization methods for `local_env` classes
  * delete `data/mp.py`

0.3.3

* GraphModel and MEGNetModel now supports a metadata tag, which is included in
  the JSON. (suggestion of mhorton).
  * Misc bug fixes for edge cases as well as improved error messages for
  mismatches in inputs.

0.3.2

* Implement the option for a scaler in models, which is used in efermi models at
  the moment but also can be helpful for extensive quantities.

0.3.1

* Minor fixes to setup.py and licenses.

0.3.0

* Proper fix to release on PyPi.

0.2.0

* Major refactoring to conform to OOP principles. Note that the
  changes are not backwards compatible, but many things are a lot
  simpler. We do not expect much disruption to existing users.
  * Added pre-trained models developed in our work for users who
  wish to simply use them for prediction.
  * Major improvements to README and documentation.

0.1.0

* Bug fix for dimension problem when only one atom in structure

0.0.1

* Initial release