Minimum number of structures to label

[omidshy]

Hi,

based on my understanding from the provided description for the fp_task_min key in PARAM file, I was expecting that the dpgen continues with the labeling (02.fp) step only if the number of candidate structures from the exploration (01.model_devi) step are larger than the value of the fp_task_min.
However, what actually happens is that the dpgen continues to the 02.fp step even if the number of candidates are smaller than the fp_task_min value, but stops at the the end of the current iteration (i.e. after labeling the candidates).

Is this an expected behavior from dpgen?
If so, doesn't it make more sense to stop the run after 01.model_devi, when the number of candidates are smaller than the fp_task_min value?

[HuangJiameng]

I was expecting that the dpgen continues with the labeling (02.fp) step only if the number of candidate structures from the exploration (01.model_devi) step are larger than the value of the fp_task_min.

You can refer to https://github.com/deepmodeling/dpgen/blob/master/dpgen/generator/run.py#L312-L314
Could you provide more information to help locate the problem?

[omidshy]

Parameters for the labeling step in my param.json file:

"fp_style": "cp2k",
"shuffle_poscar": false,
"fp_task_max": 900,
"fp_task_min": 200,
"fp_pp_path": "./cp2k_basis_pot_files/",
"fp_pp_files": ["BASIS_MOLOPT", "GTH_POTENTIALS", "dftd3.dat"],
"external_input_path": "./cp2k_inp_files/template.inp",

Here is the content of the dpgen.log file:

14:32:22,078 - INFO : start running
14:32:22,101 - INFO : =============================iter.000000==============================
14:32:22,101 - INFO : -------------------------iter.000000 task 00--------------------------
14:32:22,147 - INFO : -------------------------iter.000000 task 01--------------------------
03:50:06,313 - INFO : -------------------------iter.000000 task 02--------------------------
03:50:06,387 - INFO : -------------------------iter.000000 task 03--------------------------
03:50:09,967 - INFO : -------------------------iter.000000 task 04--------------------------
04:14:04,081 - INFO : -------------------------iter.000000 task 05--------------------------
04:14:04,106 - INFO : -------------------------iter.000000 task 06--------------------------
04:14:05,145 - INFO : system 000 candidate :      4 in   5010   0.08 %
04:14:05,146 - INFO : system 000 failed    :     10 in   5010   0.20 %
04:14:05,146 - INFO : system 000 accurate  :   4996 in   5010  99.72 %
04:14:05,158 - INFO : system 000 accurate_ratio:   0.9972    thresholds: 1.0000 and 1.0000   eff. task min and max   -1  900   number of fp tasks:      4
04:14:05,623 - INFO : -------------------------iter.000000 task 07--------------------------
05:28:54,147 - INFO : -------------------------iter.000000 task 08--------------------------
05:28:54,237 - INFO : failed tasks:      0 in      4    0.00 % 
05:28:55,133 - INFO : failed frame:      0 in      4    0.00 % 
05:28:55,145 - INFO : =============================iter.000001==============================
05:28:55,145 - INFO : -------------------------iter.000001 task 00--------------------------
05:28:55,197 - INFO : -------------------------iter.000001 task 01--------------------------
05:28:55,210 - INFO : -------------------------iter.000001 task 02--------------------------
05:28:55,223 - INFO : -------------------------iter.000001 task 03--------------------------
05:28:55,223 - INFO : finished

based on my understanding from the dpgen.log file; the value of the fp_task_min from the param.json file is apparently ignored by dpgen (0.10.6).

[poyi81]

Dear @omidshy ,
I am wondering how did you set up your training so that the iter.000000 would have 99.72% accuracy. If you can provide your param.json file I would be appreciated.
Thanks.

[HuangJiameng]

The value of the fp_task_min from the param.json file only works in the make_train step. In fp step, fp_task_min is '-1' in default. You can set it in model_devi_jobs part of the param.json file. See https://github.com/deepmodeling/dpgen/blob/master/dpgen/generator/run.py#L2465-L2471