Skip to main content
Version: 1.0

Parameter details

This section introduces the user-definable parameters in all models. There are 2 types of parameters: gloabl and local.

Global parameters: Most (but not all) gloabl parameters can either be passed in when creating the trainer class instance, or be altered via memebr function .set_< name_of_parameter >(). Example: for parameter called mypara, it can be passed in when the trainer class is created,

my_trainer = trainer(
                        ...
                        mypara = val,
                        ...
                    )

or via a member function ( for example, set_mypara()):

...
my_trainer.set_mypara(val)
...

Unless otherwise noted, all global parameters listed below can be altered in both ways.

Local parameters: these parameters are only effective with memeber functions such as train(), evaluate(), .etc.

Linear Model​

Global parameters​

max_neigh_num: size of neighbor buffer, with default value 100. However, for some systems it is not enough to accommodate all the neighbors, and the feature generation fails. The following warning will pop up:

Error! maxNeighborNum too small

In this case increase the value.

Usage: trainer.set_neigh_num(val) or pass in at the instantiation

etot_weight: weight of total energy in fitting. Default value is 0.5.

Usage: trainer.set_etot_weight(val) or pass in at the instantiation

force_weight: weight of force in fitting. Default value is 0.5.

Usage: trainer.set_force_weight(val) or pass in at the instantiation

ei_weight: weight of atomic energy in fitting. Default value is 0.0.

Usage: trainer.set_ei_weight(val) or pass in at the instantiation

Local parameters: Training​

None.

Local parameters: Evaluation​

The complete list of parameters in evaluation is

evaluate(num_thread=1)

num_thread: number of threads for evaluation. Default is 1.

Local parameters: Prediction​

The complete list of parameters in prediction:

run_md(init_config = "atom.config", md_details = None, num_thread = 1, follow = False)

init_config: inital configuration file for MD. Default is atom.config

md_details: md_detail array. Must be passed in by user.

num_thread: number of threads for prediction. Default is 1.

follow: if continue previous MD run. Default is False.

Deep Neural Network​

Global parameters​

max_neigh_num: size of neighbor buffer, with default value 100. However, for some systems it is not enough to accommodate all the neighbors, and the feature generation fails. The following warning will pop up:

Error! maxNeighborNum too small

In this case increase the value.

Usage: my_trainer.set_neigh_num(val) or pass in at the instantiation

nn_layer_config: dimension of the nueral network. Default is [15,15,1].

Usage: pass in at the instantiation ONLY

is_trainForce: if force is used in training. Default is True

Usage: my_trainer.set_train_force(val) or pass in at the instantiation

is_trainEi: if atomic energy is used in training. Default is False

Usage: my_trainer.set_train_Ei(val) or pass in at the instantiation

is_trainEgroup: if group energy is used in training. Default is False

Usage: my_trainer.set_train_Egroup(val) or pass in at the instantiation

is_trainEtot: if total energy is used in training. Default is True

Usage: my_trainer.set_train_Etot(val) or pass in at the instantiation

kf_prefac_Etot: KF update prefactor for total energy. Default is 1.0. Can be understood as the "learning rate" for KF.

Usage: my_trainer.set_kf_prefac_Etot(val)

kf_prefac_Ei: KF update prefactor for atomic energy. Default is 1.0

Usage: my_trainer.set_kf_prefac_Ei(val)

kf_prefac_F: KF update prefactor for force. Default is 1.0

Usage: my_trainer.set_kf_prefac_F(val)

kf_prefac_Egroup: KF update prefactor for group energy. Default is 1.0

Usage: my_trainer.set_kf_prefac_Egroup(val)

session_dir: name of directory that saves the training log and models. Default is record

Usage: my_trainer.set_session_dir(val) or pass in at the instantiation

device: device for traning. Default is cpu

Usage: pass in at the instantiation

recover: if recover previous training. Default is False.

Usage: pass in at the instantiation

n_epoch: number of epoch. Default is 25

Usage: my_trainer.set_epoch_num(val) or pass in at the instantiation

Local parameters:Training​

Complete list of parameters in member function set_model():

set_model(start_epoch = 1, model_name = None)

start_epcoh: start epoch number. No need to vary.

model_name: the model name to be load when recover=True. Default value is latest.pt, which is the latest model.

Local parameters: Evaluation​

Complete list of parameters in member function evaluate():

evaluate(num_thread=1)

num_thread: number of threads for evaluation. Default value is 1.

Complete list of parameters in member function extract_model_para():

extract_model_para(model_name = "latest.pt")

model_name: the name of model to be extracted. Default values is latest.pt

Local parameters: Prediction​

The complete list of parameters in prediction:

run_md(init_config = "atom.config", md_details = None, num_thread = 1, follow = False)

init_config: inital configuration file for MD. Default is atom.config

md_details: md_detail array. Must be passed in by user.

num_thread: number of threads for prediction. Default is 1.

follow: if continue previous MD run. Default is False.

DP-torch Network​

Global parameters​

max_neigh_num: size of neighbor buffer, with default value 100. However, for some systems it is not enough to accommodate all the neighbors, and the feature generation fails. The following warning will pop up:

Error! maxNeighborNum too small

In this case increase the value.

Usage: my_trainer.set_neigh_num(val) or pass in at the instantiation

kf_prefac_Etot: KF update prefactor for total energy. Default is 1.0. Can be understood as the "learning rate" for KF.

Usage: my_trainer.set_kf_prefac_Etot(val)

kf_prefac_F: KF update prefactor for force. Default is 1.0

Usage: my_trainer.set_kf_prefac_F(val)

session_dir: name of directory that saves the training log and models. Default is record

Usage: my_trainer.set_session_dir(val) or pass in at the instantiation

device: device for traning. Default is cpu

Usage: pass in at the instantiation

recover: if recover previous training. Default is False.

Usage: pass in at the instantiation

n_epoch: number of epoch. Default is 25

Usage: my_trainer.set_epoch_num(val) or pass in at the instantiation

batch_size: batch size. Default is 1.

Usage: my_trainer.set_batch_size(val) or pass in at the instantiation

select_num: number of selected atoms for force update in KF. Default is 24.

Usage: my_trainer.set_select_num(val) or pass in at the instantiation

group_size: number of groups the selected atoms will be divided into. Default is 6

Usage: my_trainer.set_group_size(val) or pass in at the instantiation

block_size: block size in layerwise KF. Default is 5120

Usage: my_trainer.set_block_size(val) or pass in at the instantiation

embedding_net_config: configuration of the embedding network, i.e. number of nodes in each layer. Default is [25, 25, 25] with KF, and [25, 50, 100] without KF

Usage: my_trainer.set_embedding_net_config(val) or pass in at the instantiation

fitting_net_config: configuration of the fitting network, i.e. number of nodes in each layer. Default is [50, 50, 50, 1] with KF, and [240, 240, 240, 1] without KF

Usage: my_trainer.set_fitting_net_config(val) or pass in at the instantiation

Rmin: low cut-offin DP's smoothing function. Default is 3.0

Usage: my_trainer.set_Rmin(val) or pass in at the instantiation

Rmax: high cut-off in DP's smoothing function. Default is 5.4

Usage: my_trainer.set_Rmax(val) or pass in at the instantiation

Local parameters:Training​

Complete list of parameters in member function set_model():

set_model(start_epoch = 1, model_name = None)

start_epcoh: start epoch number. No need to vary.

model_name: the model name to be load when recover=True. Default value is latest.pt, which is the latest model.

Local parameters: Evaluation​

Complete list of parameters in member function evaluate():

evaluate(num_thread=1)

num_thread: number of threads for evaluation. Default value is 1.

Local parameters: Prediction​

The complete list of parameters in prediction:

run_md(init_config = "atom.config", md_details = None, num_thread = 1, follow = False)

init_config: inital configuration file for MD. Default is atom.config

md_details: md_detail array. Must be passed in by user.

num_thread: number of threads for prediction. Default is 1.

follow: if continue previous MD run. Default is False.

Graphic Neural Network​

Global parameters​

device: device to train. Default is "cuda".

Usage: my_trainer.set_device() or pass in at the instantiation

session_dir: directory that contains training tasks. Default is "record"

Usage: my_trainer.set_session_dir() or pass in at the instantiation

task_name: directory under session_dir that stores the inforrmation of each task. Default is "gnn"

Usage: my_trainer.set_task_name() or pass in at the instantiation

epoch_num: number of epoch. Default is 25.

Usage: my_trainer.set_epoch_num() or pass in at the instantiation

num_train_img: number of images for training. YOU MUST SPECIFY THIS BASED ON YOU DATASET

Usage: my_trainer.set_num_train_img() or pass in at the instantiation

num_valid_img: number of images for validation. YOU MUST SPECIFY THIS BASED ON YOU DATASET

Usage: my_trainer.set_num_valid_img() or pass in at the instantiation

num_train_batch_size: number of images for training. Default is 5, and 1 to 5 are reasonable choices.

Usage: my_trainer.set_train_batch_size() or pass in at the instantiation

num_valid_batch_size: number of images for training. Default is 10.

Usage: my_trainer.set_valid_batch_size() or pass in at the instantiation

learning_rate: learning rate. Default is 0.005.

Usage: my_trainer.set_learning_rate() or pass in at the instantiation

r_max: cutoff radius in Angstrom. Default is 4.0

Usage: my_trainer.set_r_max() or pass in at the instantiation

num_layers: number of interaction blocks. Default is 4

Usage: my_trainer.set_num_layers() or pass in at the instantiation

l_max: the maximum irrep order (rotation order) for the network's features. Default is 1, which is good enough for most cases.

Usage: my_trainer.set_l_max() or pass in at the instantiation

num_features: number of features. Default is 32, which is good enough for most cases.

Usage: my_trainer.set_num_features() or pass in at the instantiation

num_basis: number of features. Default is 32, which is good enough for most cases.

Usage: my_trainer.set_num_basis() or pass in at the instantiation

num_basis: number of basis functions used in the radial basis. Default is 8, which is good enough for most cases.

Usage: my_trainer.set_num_basis() or pass in at the instantiation

PolynomialCutoff_p: p-exponent used in polynomial cutoff function, smaller p corresponds to stronger decay with distance. Default is 6.

Usage: my_trainer.set_PolynomialCutoff_p() or pass in at the instantiation

invariant_layers: number of radial layers. Default is 2. 1 to 3 are reasonable.

Usage: my_trainer.set_invariant_layers() or pass in at the instantiation

Local parameters:Training​

Complete list of parameters in member function generate_data():

generate_data(xyz_output = "./PWdata/training_data.xyz")

xyz_output: name of .xyz file after coversion. Default values is ./PWdata/training_data.xyz

Complete list of parameters in member function train():

train(train_data = r"./PWdata/training_data.xyz")

train_data: .xyz to be used in training. Default values is ./PWdata/training_data.xyz

Local parameters:Evaluation​

Complete list of parameters in member function evaluate():

evaluate(
                train_dir = None,
                model = None,   
                batch_size = 50, 
                device = None, 
                use_deterministic_algorithms = False
        )

train_dir: path to the data set to be evaluated. Images that are NOT used in training and validation will be used

model: the model to be evaluated. MUST BE A DELOPYED MODEL

batch_size: batch size.

device: device for evaluation task

use_deterministic_algorithms: if a deterministic method is used in evaluation. Notice that if CUDA is used, algorithms are non-deterministic, and forcing it using deterministic method may induce error. Set device = cpu instead.

Complete list of parameters in member function deploy():

deploy( 
            model = None, 
            train_dir = None, 
            out_file = None
        )

model: model to be deployed. Default is best_model.pth in train_dir

train_dir: directory that contains the model to be deployed

out_file: name of the output file