SevenNet parameters for training model¶
There are preset parameters:
See full full development parameters here.
To access energy, force, and stress, in extxyz dataset, the following keys are used:
base¶
# Example input.yaml for training SevenNet.
# '*' signifies default. You can check log.sevenn for defaults.
model:
chemical_species: 'Auto' # Chemical elements present in the dataset, guess them from load_dataset data if 'auto'
cutoff: 5.0 # Cutoff radius in Angstroms. If two atoms are within the cutoff, they are connected.
channel: 32 # The multiplicity(channel) of node features.
lmax: 2 # Maximum order of irreducible representations (rotation order).
num_convolution_layer: 3 # The number of message passing layers.
#irreps_manual: # Manually set irreps of the model in each layer
#- "128x0e"
#- "128x0e+64x1e+32x2e"
#- "128x0e+64x1e+32x2e"
#- "128x0e+64x1e+32x2e"
#- "128x0e+64x1e+32x2e"
#- "128x0e"
weight_nn_hidden_neurons: [64, 64] # Hidden neurons in convolution weight neural network
radial_basis: # Function and its parameters to encode radial distance
radial_basis_name: 'bessel' # Only 'bessel' is currently supported
bessel_basis_num: 8
cutoff_function: # Envelop function, multiplied to radial_basis functions to init edge featrues
cutoff_function_name: 'poly_cut' # {'poly_cut' and 'poly_cut_p_value'} or {'XPLOR' and 'cutoff_on'}
poly_cut_p_value: 6
act_gate: {'e': 'silu', 'o': 'tanh'} # Equivalent to 'nonlinearity_gates' in nequip
act_scalar: {'e': 'silu', 'o': 'tanh'} # Equivalent to 'nonlinearity_scalars' in nequip
is_parity: False # Pairy True (E(3) group) or False (to SE(3) group)
self_connection_type: 'nequip' # Default is 'nequip'. 'linear' is used for SevenNet-0.
conv_denominator: "avg_num_neigh" # Valid options are "avg_num_neigh*", "sqrt_avg_num_neigh", or float
train_denominator: False # Enable training for denominator in convolution layer
train_shift_scale: False # Enable training for shift & scale in output layer
train:
random_seed: 1
is_train_stress: True # Includes stress in the loss function
epoch: 200 # Ends training after this number of epochs
#loss: 'Huber' # Default is 'mse' (mean squared error)
#loss_param:
#delta: 0.01
# Each optimizer and scheduler have different available parameters.
# You can refer to sevenn/train/optim.py for supporting optimizer & schedulers
optimizer: 'adam' # Options available are 'sgd', 'adagrad', 'adam', 'adamw', 'radam'
optim_param:
lr: 0.005
scheduler: 'exponentiallr' # 'steplr', 'multisteplr', 'exponentiallr', 'cosineannealinglr', 'reducelronplateau', 'linearlr'
scheduler_param:
gamma: 0.99
force_loss_weight: 0.1 # Coefficient for force loss
stress_loss_weight: 1e-06 # Coefficient for stress loss (to kbar unit)
per_epoch: 10 # Generate checkpoints every this epoch
# ['target y', 'metric']
# Target y: TotalEnergy, Energy, Force, Stress, Stress_GPa, TotalLoss
# Metric : RMSE, MAE, or Loss
error_record:
- ['Energy', 'RMSE']
- ['Force', 'RMSE']
- ['Stress', 'RMSE']
- ['TotalLoss', 'None']
# Continue training model from given checkpoint, or pre-trained model checkpoint for fine-tuning
#continue:
#checkpoint: 'checkpoint_best.pth' # Checkpoint of pre-trained model or a model want to continue training.
#reset_optimizer: False # Set True for fine-tuning
#reset_scheduler: False # Set True for fine-tuning
#use_statistic_values_of_checkpoint: False # Set True to use shift, scale, and avg_num_neigh from checkpoint or not
data:
batch_size: 4 # Per GPU batch size.
data_divide_ratio: 0.1 # Split dataset into training and validation sets by this ratio
shift: 'per_atom_energy_mean' # One of 'per_atom_energy_mean*', 'elemwise_reference_energies', float
scale: 'force_rms' # One of 'force_rms*', 'per_atom_energy_std', 'elemwise_force_rms', float
# ase.io.read readable data files or structure_list or .sevenn_data files can be used as dataset.
# .sevenn_data is preprocessed data set has edges connected (can be obtained by using sevenn_graph_build or by save_** options below)
data_format: 'ase' # One of 'ase', 'structure_list' (.sevenn_data is always readable)
data_format_args: # if `data_format` is `ase`, args will be passed to `ase.io.read`
index: ':' # see `https://wiki.fysik.dtu.dk/ase/ase/io/io.html` for more valid arguments
# If only load_dataset_path is provided, train/valid set is automatically decided by splitting dataset by divide ratio
# If both load_dataset_path & load_validset_path is provided, use load_dataset_path as training set.
load_dataset_path: ['../data/train.extxyz'] # Example of using ase as data_format, support multiple datasets and expansion(*)
#load_validset_path: ['./valid.sevenn_data']
#save_dataset_path: './total' # Save the preprocessed (in load_dataset_path) dataset
#save_by_train_valid: True # Save the preprocessed train.sevenn_data, valid.sevenn_data
fine_tune¶
# Example input.yaml for fine-tuning sevennet-0
# '*' signifies default. You can check log.sevenn for defaults.
model: # model keys should be consistent except for train_* keys
chemical_species: 'Auto'
cutoff: 5.0
channel: 128
is_parity: False
lmax: 2
num_convolution_layer: 5
irreps_manual:
- "128x0e"
- "128x0e+64x1e+32x2e"
- "128x0e+64x1e+32x2e"
- "128x0e+64x1e+32x2e"
- "128x0e+64x1e+32x2e"
- "128x0e"
weight_nn_hidden_neurons: [64, 64]
radial_basis:
radial_basis_name: 'bessel'
bessel_basis_num: 8
cutoff_function:
cutoff_function_name: 'XPLOR'
cutoff_on: 4.5
self_connection_type: 'linear'
train_shift_scale: False # customizable (True | False)
train_denominator: False # customizable (True | False)
train: # Customizable
random_seed: 1
is_train_stress: True
epoch: 100
optimizer: 'adam'
optim_param:
lr: 0.004
scheduler: 'exponentiallr'
scheduler_param:
gamma: 0.99
force_loss_weight: 0.1
stress_loss_weight: 1e-06
per_epoch: 10 # Generate checkpoints every this epoch
# ['target y', 'metric']
# Target y: TotalEnergy, Energy, Force, Stress, Stress_GPa, TotalLoss
# Metric : RMSE, MAE, or Loss
error_record:
- ['Energy', 'RMSE']
- ['Force', 'RMSE']
- ['Stress', 'RMSE']
- ['TotalLoss', 'None']
continue:
reset_optimizer: True
reset_scheduler: True
reset_epoch: True
checkpoint: 'SevenNet-0_11July2024'
# Set True to use shift, scale, and avg_num_neigh from checkpoint (highly recommended)
use_statistic_values_of_checkpoint: True
data: # Customizable
batch_size: 4
data_divide_ratio: 0.1
# ase.io.read readable data files or structure_list or .sevenn_data files can be used as dataset.
# .sevenn_data is preprocessed data set has edges connected (can be obtained by using sevenn_graph_build or by save_** options below)
data_format: 'ase' # One of 'ase', 'structure_list' (.sevenn_data is always readable)
data_format_args: # if `data_format` is `ase`, args will be passed to `ase.io.read`
index: ':' # see `https://wiki.fysik.dtu.dk/ase/ase/io/io.html` for more valid arguments
# If only load_dataset_path is provided, train/valid set is automatically decided by splitting dataset by divide ratio
# If both load_dataset_path & load_validset_path is provided, use load_dataset_path as training set.
load_dataset_path: ['fine_tune.extxyz'] # Support multiple files and expansion(*)
#load_validset_path: ['./valid.sevenn_data']
#save_dataset_path: './total' # Save the preprocessed (in load_dataset_path) dataset
#save_by_train_valid: True # Save the preprocessed train.sevenn_data, valid.sevenn_data
sevennet-0¶
# SevenNet-0
model:
chemical_species: 'auto'
cutoff: 5.0
channel: 128
is_parity: False
lmax: 2
num_convolution_layer: 5
irreps_manual:
- "128x0e"
- "128x0e+64x1e+32x2e"
- "128x0e+64x1e+32x2e"
- "128x0e+64x1e+32x2e"
- "128x0e+64x1e+32x2e"
- "128x0e"
weight_nn_hidden_neurons: [64, 64]
radial_basis:
radial_basis_name: 'bessel'
bessel_basis_num: 8
cutoff_function:
cutoff_function_name: 'XPLOR'
cutoff_on: 4.5
act_gate: {'e': 'silu', 'o': 'tanh'}
act_scalar: {'e': 'silu', 'o': 'tanh'}
conv_denominator: 'avg_num_neigh'
train_shift_scale: False
train_denominator: False
self_connection_type: 'linear'
train:
train_shuffle: False
random_seed: 1
is_train_stress : True
epoch: 600
loss: 'Huber'
loss_param:
delta: 0.01
optimizer: 'adam'
optim_param:
lr: 0.01
scheduler: 'linearlr'
scheduler_param:
start_factor: 1.0
total_iters: 600
end_factor: 0.0001
force_loss_weight : 1.00
stress_loss_weight: 0.01
error_record:
- ['Energy', 'RMSE']
- ['Force', 'RMSE']
- ['Stress', 'RMSE']
- ['Energy', 'MAE']
- ['Force', 'MAE']
- ['Stress', 'MAE']
- ['Energy', 'Loss']
- ['Force', 'Loss']
- ['Stress', 'Loss']
- ['TotalLoss', 'None']
per_epoch: 10
# continue:
# checkpoint: './checkpoint_last.pth'
# reset_optimizer: False
# reset_scheduler: False
data:
data_shuffle: False
batch_size: 128 # per GPU batch size, as the model trained with 32 GPUs, the effective batch size equals 4096.
scale: 'per_atom_energy_std'
shift: 'elemwise_reference_energies'
data_format: 'ase'
save_by_train_valid: False
load_dataset_path: ["path_to_MPtrj_total.sevenn_data"]
load_validset_path: ["validaset.sevenn_data"]
Full parameters¶
# Example input.yaml for training SevenNet.
# The underlying model is nequip (https://github.com/mir-group/nequip), but the names of hyperparameters might different.
# Defaults model parameter that works well of channel, lmax, and num_convolution_layer are 32, 1, 3 respectively.
# '*' signifies default. You can check log.sevenn.
model:
chemical_species: 'Auto' # Chemical elements present in the dataset, guess them from load_dataset data if 'auto'
cutoff: 5.0 # Cutoff radius in Angstroms. If two atoms are within the cutoff, they are connected.
channel: 4 # The multiplicity(channel) of node features.
lmax: 1 # Maximum order of irreducible representations (rotation order).
num_convolution_layer: 4 # The number of message passing layers.
#irreps_manual: # Manually set irreps of the model in each layer
#- "128x0e"
#- "128x0e+64x1e+32x2e"
#- "128x0e+64x1e+32x2e"
#- "128x0e+64x1e+32x2e"
#- "128x0e+64x1e+32x2e"
#- "128x0e"
weight_nn_hidden_neurons: [64, 64] # Hidden neurons in convolution weight neural network
radial_basis: # Function and its parameters to encode radial distance
radial_basis_name: 'bessel' # Only 'bessel' is currently supported
bessel_basis_num: 8
cutoff_function: # Envelop function, multiplied to radial_basis functions to init edge featrues
cutoff_function_name: 'poly_cut' # {'poly_cut' and 'poly_cut_p_value'} or {'XPLOR' and 'cutoff_on'}
poly_cut_p_value: 6
act_gate: {'e': 'silu', 'o': 'tanh'} # Equivalent to 'nonlinearity_gates' in nequip
act_scalar: {'e': 'silu', 'o': 'tanh'} # Equivalent to 'nonlinearity_scalars' in nequip
is_parity: False # Pairy True (E(3) group) or False (to SE(3) group)
self_connection_type: 'nequip' # Default is 'nequip'. 'linear' is used for SevenNet-0.
conv_denominator: "avg_num_neigh" # Valid options are "avg_num_neigh*", "sqrt_avg_num_neigh", or float
train_shift_scale: False # Enable training for shift & scale in output layer
train_denominator: False # Enable training for denominator in convolution layer
train:
random_seed: 1
is_train_stress: True # Includes stress in the loss function
epoch: 10 # Ends training after this number of epochs
#loss: 'Huber' # Default is 'mse' (mean squared error)
#loss_param:
#delta: 0.01
# Each optimizer and scheduler have different available parameters.
# You can refer to sevenn/train/optim.py for supporting optimizer & schedulers
optimizer: 'adam' # Options available are 'sgd', 'adagrad', 'adam', 'adamw', 'radam'
optim_param:
lr: 0.005
scheduler: 'exponentiallr' # One of 'steplr', 'multisteplr', 'exponentiallr', 'cosineannealinglr', 'reducelronplateau', 'linearlr'
scheduler_param:
gamma: 0.99
force_loss_weight: 0.1 # Coefficient for force loss
stress_loss_weight: 1e-06 # Coefficient for stress loss (to kbar unit)
per_epoch: 5 # Generate checkpoints every this epoch
# TotalEnergy, Energy, Force, Stress, Stress_GPa, TotalLoss
# RMSE, MAE, or Loss
error_record:
- ['Energy', 'RMSE']
- ['Force', 'RMSE']
- ['Stress', 'RMSE']
- ['TotalLoss', 'None']
# Continue training model from given checkpoint, or pre-trained model checkpoint for fine-tuning
#continue:
#reset_optimizer: False # Set True for fine-tuning
#reset_scheduler: False # Set True for fine-tuning
#checkpoint: 'checkpoint_best.pth' # Checkpoint of pre-trained model or a model want to continue training.
#use_statistic_values_of_checkpoint: False # Set True to use shift, scale, and avg_num_neigh from checkpoint or not
# If the dataset changed (for fine-tuning),
# setting 'use_statistic_value_of_checkpoint' to True roughly changes model's accuracy in the beginning of training.
# We recommand to use it as False, and turn train_shift_scale and train_avg_num_neigh to True.
data:
batch_size: 2 # Per GPU batch size.
data_divide_ratio: 0.1 # Split dataset into training and validation sets by this ratio
#shift: 'per_atom_energy_mean' # One of 'per_atom_energy_mean*', 'elemwise_reference_energies', float
#scale: 'force_rms' # One of 'force_rms*', 'per_atom_energy_std', 'elemwise_force_rms', float
# ase.io.read readable data files or structure_list or .sevenn_data files can be used as dataset.
# .sevenn_data is preprocessed data set has edges connected (can be obtained by using sevenn_graph_build or by save_** options below)
data_format: 'ase' # Default is 'ase'. Choices are 'ase', 'structure_list', '.sevenn_data'
data_format_args: # Paramaters, will be passed to ase.io.read
energy_key: 'TotEnergy' # Key for energy in extxyz file
force_key: 'force' # Key for force in extxyz file
# ASE tries to infer its type by extension, in this case, extxyz file is loaded by ase.
#load_dataset_path: ['../data/test.extxyz'] # Example of using ase as data_format
# If only load_dataset_path is provided, train/valid set is automatically decided by splitting dataset by divide ratio
# If both load_dataset_path & load_validset_path is provided, use load_dataset_path as training set.
load_dataset_path: ['./1_process_data/dataset_1593.xyz']
#load_validset_path: ['./valid.sevenn_data']
#save_dataset_path: 'total' # Save the preprocessed (in load_dataset_path) dataset
#save_by_train_valid: True # Save the preprocessed train.sevenn_data, valid.sevenn_data
#save_by_label: False # Save the dataset by labels specified in the structure_list