Request for new features: stage-specific parameters for explore and fp, finetune/iter-initial-modle distinguishing training parameters #216
Description
REQUEST1 :
expect the following parameters to have further structures to support exploration-stage specific assignment :
explore/convergence(all paramenters within)explore/max_numb_iterexplore/fatal_at_maxfp/task_max
e.g.,
-
in stage_0 :
...
"explore": {
"type": "lmp",
"config": {
"command": "lmp -var restart 0",
"impl": "pytorch"
},
"convergence": {
"type": "adaptive-lower",
"conv_tolerance": 0.005,
"rate_candi_f": 0.15,
"level_f_hi": 5.0,
"n_checked_steps": 3
},
"max_numb_iter": 2,
"fatal_at_max": false,
...
"fp": {
"task_max": 4000,
... -
in stage_3 :
...
"explore": {
"type": "lmp",
"config": {
"command": "lmp -var restart 0",
"impl": "pytorch"
},
"convergence": {
"type": "adaptive-lower",
"conv_tolerance": 0.005,
"numb_candi_f": 4000,
"level_f_hi": 5.0,
"n_checked_steps": 3,
},
"max_numb_iter": 20,
"fatal_at_max": True,
"fp": {
"task_max": 4000,
...
REQUEST 2
expect to support different ending_pref_e/f/v for initial_finetune from multi-task pre-train models and the successive init_model form the finetuned_initial_model.
Currently train/config supports only start but no end parameters, like only "init_model_start_pref_e" but no "init_model_end_pref_e". Instead, the end_prefs are inherited from the limit_prefs from one training scripr defined intrain/config/templated_script
Maybe need to support two scripts as by train/config/templated_script, or adding new init_model_end_pref_e/f/v parameters in train/config
scenario arising REQUEST 1
In practice of the pre-train models initiated DP-GEN2, multiple successive exploration stages are used to ehance the exploration efficiency on a complex sample space.
The sample sapce consists of derivatives from (1) many severely different initial configurations, (2) both trivial dynamics images and significant low-probability instances, and (3) successors after low-probability instances which is also trivial but as well severely different compared with their initial/parent configurations.
(1) will suffer from the species bias after pre-train (and finetune), then leads to over-sampling on full trajectories of specific far-away-from-pretrain configurations
(2) is our central target
(3) will suffer from the conformer bias after pre-train (and finetune), then leads to over-sampling on these trivial successors configurations
Thus stage_0 and stage_1 are used to debiasing (1) and (3) through randomly select candidates from a broader model_devi range.
No final exploration convergence is expected for these two stages.
stage_2 is the actually meant to be converged one for *2), and related parameters would be different from debiasing stages.
scenario arising REQUEST 2
tests showed different parameter preferences for trainings in two stages.