0.0.8

These changes are related to ML4Chem’s publication:

• Creation of atomistic module to comply with publication structure.

• models:

  • All models comply with base class inheritance using mixin.
  • models moved to atomistic.models. Scripts have to be changed as the following example
    from ml4chem.atomistic.models.neuralnetwork import NeuralNetwork.
  • SVM model refactoring.
  • training class can take checkpoint dictionary to save checkpoints during training.
  • NeuralNetwork and RetentionTime models support training with test error.
  • New get_activations function to get activations of neural network.

• features:

  • features moved to atomistic.features. Scripts have to be changed as the following example
    from ml4chem.atomistic.features.gaussian import Gaussian.

• data.visualization moved to .visualization.

  • read_log() function now supports data_only keyword argument to return only data instead of matplotlib plot.
  • plot_atomic_features supports now backend keyword arguments and preprocessing of features before doing PCA or T-SNE using make_pipeline. Example:

backend_kwargs = {"perplexity": 500}
dimension = 2
dot_size = 3

plot, df = plot_atomic_features(
    latent_space, 
    method="tsne", 
    preprocessor=StandardScaler(),
    dimensions=dimension, 
    backend="plotly", 
    dot_size=dot_size,
    backend_kwargs=backend_kwargs
)

• Restructured documentation.

• AtomicMSELoss supports penalization using a list of uncertainties.

• Potentials class moved to atomistic module. It has to be imported as
  from ml4chem.atomistic import Potentials.