ai4materials.models.cnn_polycrystals module

ai4materials.models.cnn_polycrystals.normalize_images(images)

ai4materials.models.cnn_polycrystals.predict(x, y, configs, numerical_labels, text_labels, nb_classes=3, results_file=None, model=None, batch_size=32, conf_matrix_file=None, verbose=1, with_uncertainty=True, mc_samples=50, consider_memory=True, max_length=1000000.0)

ai4materials.models.cnn_polycrystals.predict_with_uncertainty(data, model, model_type='classification', n_iter=1000)

This function allows to calculate the uncertainty of a neural network model using dropout.

This follows Chap. 3 in Yarin Gal’s PhD thesis: http://mlg.eng.cam.ac.uk/yarin/thesis/thesis.pdf

We calculate the uncertainty of the neural network predictions in the three ways proposed in Gal’s PhD thesis,

as presented at pag. 51-54:

• variation_ratio: defined in Eq. 3.19
• predictive_entropy: defined in Eq. 3.20
• mutual_information: defined at pag. 53 (no Eq. number)

ai4materials.models.cnn_polycrystals.reshape_images(images, target_shape)

Reshape images according to the target shape

ai4materials.models.cnn_polycrystals.train_neural_network(x_train, y_train, x_val, y_val, configs, partial_model_architecture, batch_size=32, nb_epoch=5, normalize=True, checkpoint_dir=None, neural_network_name='my_neural_network', training_log_file='training.log', early_stopping=False, data_augmentation=True)

Train a neural network to classify crystal structures represented as two-dimensional diffraction fingerprints.

This model was introduced in [1].

x_train: np.array, [batch, width, height, channels]

[1]	A. Ziletti, A. Leitherer, M. Scheffler, and L. M. Ghiringhelli, “Crystal-structure identification via Bayesian deep learning: towards superhuman performance”, in preparation (2018)

Code author: Angelo Ziletti <angelo.ziletti@gmail.com>