Skip to content

Latest commit

 

History

History
113 lines (82 loc) · 3.72 KB

RUN.MD

File metadata and controls

113 lines (82 loc) · 3.72 KB
Raw
  1. For the base workflow for training a GPR or DKL at a fixed temperatures can be found in run_singleT_ECG.py. The usage is listed below:
$ python run_singleT_ECG.py -h

usage: run_singleT_ECG.py [-h] -json JSONFILENAME [-Nt] [-is] [-qw] [-s] [-ml]
                          [-k] [-ni] [-rs]

A python workflow to run active learning for Electronic Coarse Graining

optional arguments:
  -h, --help            show this help message and exit
  -json JSONFILENAME, --jsonfilename JSONFILENAME
                        Input JSON configuration filename (including path)
  -Nt , --ntrials       Number of AL trials to run
  -is , --initialseed   Number of initial random points to bootstrap the model
  -qw , --querywidth    AL query width
  -s , --strategy       AL strategy : 'random, uncertainty, emoc'
  -ml , --mlmethod      GP type: EXACT, DKL, MULTITASK
  -k , --kernel         GP Kernel: RBF, MATERN
  -ni , --niter         Number of iteration for converging GP
  -rs , --randomseed    Initialize the random seed

The mandatory configuration file has the following format with a path and specification of the single temperature csv file.

{
"path" : "/path/to/numpy/data/folder",
"train_file" : "train_rigid.npy",
"test_file" : "test_rigid.npy",
"x_slice" : [0, 153],
"y_slice" : -4 , 
"train_refT_Xslice" : [2, 3],
"jump"    : {"1000K":[0, 1000],
       "650K":[1000, 2000],
       "300K":[2000, 3000],
       "50K":[3000, 4000]}
}

The y_slice corresponds to HOMO. The last index of the numpy array has the reference temperature label (e.g. 1000 K). train_refT_Xslice is the row slices of the single temperature training dataset with in the numpy arrays.

train_refT_Xslice T(K)
[0,1] 1000
[1,2] 650
[2,3] 300
[3,4] 50

Example usage with a DKL is listed below

python -u  run_singleT_ECG.py   -json  config.json  -Nt 1500  -is  3   -qw  6995  -ml DKL   -s  emoc   -k  RBF  -ni 300   -rs  111   > ECG_trial.out

2. The DKL runs require a mandatory hyperparameter json file. This can be generated using Bayesian optimization

The DKL BO script can be found in run_DKL_Tuner.py

An example usage:

python -u run_DKL_Tuner.py -rs 0 -Nbo 50 100  > DKL_BO.out

This will create a hyperparameter json file with the following format:

{
    "lr": 0.024,
    "n_iter": 800.0,
    "l1out": 1000.0,
    "l2out": 300.0,
    "l3out": 150.0,
    "final": 21.0,
    "MAE": 0.7796717286109924
}
  1. For the model training utilizing the dataset picked at mixted temperatures one must use the run_mixedT_ECG.py. The usage is listed below:
python -u  run_mixedT_ECG.py   -json  config.json  -Nt 1500  -is  3   -qw 32495  -ml DKL   -s  emoc   -k  RBF  -ni 300   -rs 123   > ECG_trial.out

The format for the configuration file (--json) to be used with the mixed temperature run is listed below. The hyperparameter json is frozen for a given CG resolution. Further, the LHS dataset and example configurations can be found here.

{
"path" : "/please/change/to/your/path",
"train_file" : "train_rigid.npy",
"test_file" : "test_rigid.npy",
"x_slice" : [0, 153], # feature dimension 
"y_slice" : -4 , # We are only using HOMO
"jump"    : {"1000K":[0, 1000],
       "650K":[1000, 2000],
       "300K":[2000, 3000],
       "50K":[3000, 4000]}
}