Input File

The XANESNET input file is a YAML configuration file that specifies all the necessary parameters for the training and prediction. The input file is divided into several sections, each corresponding to a different aspect of the configuration. This page details the syntax of each section, and provides guidance on how to customise your model.

dataset

The dataset section defines the dataset type, file locations, and preprocessing options used during model training. It specifies where the raw XYZ structures and XANES spectra are stored, where processed data should be saved, and how the data should be loaded and transformed. The Datasets page provides a detailed explanation of all available datasets and their configurations.

• type (str): Dataset type (see Datasets).

• root_path (str): Directory where the processed dataset will be stored or loaded from.

• xyz_path (str): Directory containing structure files.

• xanes_path (str): Directory containing XANES spectra files.

• params (dict, optional): Parameters for the chosen dataset type.

descriptors

The descriptors section defines the descriptor(s) used for feature extraction. A descriptor transforms atomic structures into fixed-size numerical vectors.

XANESNET provides a range of descriptor functions, each requiring its own set of parameters. See the Descriptors page for a detailed explanation of the available types and their configurations.

Multiple descriptors can be specified. When more than one descriptor is used, the resulting feature vectors are concatenated along the feature dimension.

• type (str): Type of descriptor (see Descriptors).

• params (dict, optional): Parameters for the chosen descriptor type.

Example:

descriptors:
  - type: wacsf

  - type: pdos
    params:
      code: pyscf

model

The model section defines the architecture and specific parameters of the neural network model. This section determines how the model is structured, including the type of neural network and the hyperparameters that control its training and operation.

XANESNET supports a variety of widely-used deep neural network architecture. Their architectures and parameters are further explained in the Models page.

• type (str): Type of model (see Models).

• params (dict, optional): Model-specific parameters.

• weights (dict, optional): Configuration of the model weight initialisation.

  • kernel (str, default): Weight initialization method for kernel parameters. Supported options: uniform, normal, xavier_uniform, xavier_normal, kaiming_uniform, kaiming_normal, default.

  • bias (str, zero): Initialisation method for bias parameters. Supported options: zeros, ones.

  • seed (int): Random seed.

• weights_params (dict, optional): Additional weight parameters.

Example:

model:
  type: mlp
  weights:
      kernel: kaiming_uniform
      bias: zeros
      seed: 2025
  weights_params:
      mode: fan_in
      nonlinearity: relu

hyperparameters

The hyperparameter section defines the settings that govern the training process of the neural network model.

• batch_size (int, default: 32): Number of samples per training batch.

• lr (float, 0.001): Learning rate.

• epochs (int, 100): Maximum number of training epochs.

• n_earlystop (int, 100): Number of consecutive epochs without improvement in validation loss before training is stopped.

• optimizer (str, adam): Optimization algorithm used for training. Supported options: adam, sgd, rmsprop, adamw, adagrad.

• loss (str, default: mse): Loss function. Supported options: mse, bce, l1, emd, cosine, wcc, hybrid, specplus, mw_ssim1d.

• loss_reg (str, none): Type of regularization loss. Supported options: l1, l2, none.

• loss_lambda (float, 0.0001): Weighting factor applied to the regularization loss.

• loss_params (dict, optional): Additional parameters specific to the selected loss function.

Example:

hyperparams:
  batch_size: 64
  lr: 0.0001
  epochs: 500
  optimizer: Adam
  seed: 2021
  loss: mse
  loss_reg: L1
  loss_lambda: 0.0001

lr_scheduler

The lr_scheduler section controls whether a learning rate scheduler is used to dynamically adjust the learning rate during training.

A learning rate scheduler modifies the learning rate over time, either according to a predefined schedule or in response to training progress.

• lr_scheduler (bool): Enables or disables the learning rate scheduler.

• scheduler_params (dict, optional):

  • type (str, step): Type of learning rate scheduler to use. Supported options: step or multistep or exponential or linear or constant

  • step_size (int, 10): Number of epochs between successive learning rate updates.

  • gamma (float, 0.5): Multiplicative factor applied to the learning rate at each decay step.

Example:

lr_scheduler: True
scheduler_params:
  type: step
  step_size: 100
  gamma: 0.5

kfold

The kfold section controls whether K-Fold cross-validation is used during model training. When enabled, the dataset is partitioned into k subsets (folds). The model is trained k times, each time using a different fold as the validation set and the remaining k-1 folds as the training set.

• kfold (bool): Enables or disables K-Fold cross-validation.

• kfold_params (dict, optional):

  • n_splits (int, 3): Number of folds (k).

  • n_repeats (int, 1): Number of times the K-Fold cross-validation is repeated.

  • seed (int): Random seed used for reproducible dataset splitting.

Example:

kfold: True
kfold_params:
  n_splits: 5
  n_repeats: 1
  seed: 2022

bootstrap

The bootstrap section controls whether bootstrap resampling is used during model training. When enabled, the original training dataset is randomly resampled with replacement to create new training datasets. The size of each resampled dataset is determined by multiplying the original dataset size by a user-defined factor. The model is trained multiple times using independently resampled datasets, each with a different random seed.

• bootstrap (bool): Enables or disables bootstrap resampling.

• bootstrap_params (dict, optional): Parameters controlling the bootstrap procedure.

  • n_boot (int, 3): Number of bootstrap training runs

  • n_size (float, 1.0): Scaling factor applied to the original dataset size when generating each bootstrap sample.

  • weight_seed (list): List of random seeds used to generate independent bootstrap samples.

Example:

bootstrap: True
bootstrap_params:
  n_boot: 4
  n_size: 1.0
  weight_seed: [97, 39, 22, 44]

ensemble

The ensemble section controls whether ensemble training is performed. When enabled, multiple models are trained independently using different random initializations of model parameters. Each ensemble member is initialized with a distinct random seed.

• ensemble (bool): Enables or disables ensemble training.

• ensemble_params (dict, optional):

  • n_ens (int, 3): Number of ensemble members to train.

  • weight_seed (list): List of random seeds used to initialise model weights and biases for each ensemble member.

Example:

ensemble: True
ensemble_params:
  n_ens: 3
  weight_seed: [97, 39, 22]