v2.1.5

Caution

Known critical issues in this version

• Incorrect results from se_e3 compressed model.

See #2866 for more information.

New features:

• Attention based model DPA-1 (#1866 #1886 #1923 #1943)

Enhancements:

• remove white space from train_attr/training_script (#1870)
• replace FastGFile with GFile. FastGFile throws a deprecated warning. (#1874)
• docs: make warnings more prominent (#1879)
• add examples to cli help (#1887)
• remove get_platform in setup.py (#1897)
• bump to C++ 17 for TF 2.10 (#1898)
• compare converted lr with almost equal (#1901)
• add tutorial and publication links to docs (#1904)
• LAMMPS: Use of “override” instead of “virtual” (#1915)
• highlight LAMMPS codes in doc (#1921)

Bug fixings:

• support initilize parameters from a fitting with suffix (#1885)
• fix DeprecationWarning of imp (#1896)
• find protobuf headers from extra paths (#1910)
• Fix bugs when init_frz_model using tebd. (#1891)
• Use GeluCustom as operator name (#1918)
• handle float point error in sys_probs (#1919)
• fix model conversion of 0.12 (#1941)