model_devi | sort models in default&support user-defined order #610
Conversation
current behavior: 1）the list of models written in input.lammps for model_devi is created by "glob" without "sort" --> may give different (random) orders in different machine environments --> would randomly alter the model being used to sample configurations (always use the first model written in imput.lammps) ---> could lead to unexpected performance, such as, when using together with "model_devi_activation_func" that allows four models to be nonequivalent. 2）when preparing input.lammps from iuser-provided template, the line begin with "pair_style deepmd" will be overwritten by dpgen, thus overwrites the user defined order of models ---> could lead to unexpected performance, such as, when using together with "model_devi_activation_func" that allows four models to be nonequivalent and users indeed expected a specific order of models. changes: 1）sorted the list of models in default, thus always use graph.000.pb to sample configurations; 2）check weather user writes the full line of begin with "pair_style deepmd" (by checking the length), if yes, leave it be; if not, overwrites with the default settings (use graph.000.pb to sample); besides, the original error trigger is retained if key words "pair_style deepmd" are not provided in the template.
Codecov Report
@@ Coverage Diff @@
## devel #610 +/- ##
==========================================
+ Coverage 33.14% 33.16% +0.02%
==========================================
Files 86 86
Lines 14826 14837 +11
==========================================
+ Hits 4914 4921 +7
- Misses 9912 9916 +4
Continue to review full report at Codecov.
|
|
Could you please upload or update the examples for revmat-mode? Thanks! |
will be uploaded afterwards |
|
There can be any number of space between "pair_style" and "deepmd", so the length check may fail. |
Sorry for the misleading description of the pr. It actually checks the 【numbers of non-space words in the line】 by 【.split()】, The description of this pr is correspondingly revised. |
current behavior:
1）the list of models written in input.lammps for model_devi is created by "glob" without "sort" --> may give different (random) orders in different machine environments --> would randomly alter the model being used to sample configurations (always use the first model written in imput.lammps)
---> could lead to unexpected performance, such as, when using together with "model_devi_activation_func" that allows four models to be nonequivalent.
2）when preparing input.lammps from iuser-provided template, the line begin with "pair_style deepmd" will be overwritten by dpgen, thus overwrites the user defined order of models ---> could lead to unexpected performance, such as, when using together with "model_devi_activation_func" that allows four models to be nonequivalent and users indeed expected a specific order of models.
changes:
1）sorted the list of models in default, thus always use graph.000.pb to sample configurations;
2）check weather user writes the full line of begin with "pair_style deepmd" (by checking the number of non-space words in this line), if yes, leave it be; if not, overwrites with the default settings (use graph.000.pb to sample); besides, the original error trigger is retained if key words "pair_style deepmd" are not provided in the template.