automata4cps.denta

The module implements the novel DENTA algorithm for the learning of hybrid automata from data.

Author: Nemanja Hranisavljevic, hranisan@hsu-hh.de, nemanja@ai4cps.com

class automata4cps.denta.DENTA(num_signals, num_hidden_units, first_hidden_size=None, sigma=1.0, sigma_learnable=False, sparsity_weight=0.01, persistence=True, num_hidden_layers=1, window_size=1, window_step=1, sparsity_target=0.1, use_derivatives=0, device='cpu', log_mlflow=False)

Bases: Module, Automaton

anomaly_detection(s)

anomaly_score(s)

binary_vector_to_mode(h)

calculate_ad_threshold(d, quantile=0.95)

compute_purity(class_assignments)

Computes the purity between cluster and class assignments. Compare to https://nlp.stanford.edu/IR-book/html/htmledition/evaluation-of-clustering-1.html

Parameters:

• cluster_assignments (list) – List of cluster assignments for every point.

• class_assignments (list) – List of class assignments for every point.

Returns:

The purity value.

Return type:

float

contrastive_divergence(v0, h0, vk, hk)

decode(h, x_level=0, h_level=None)

dsm_loss(x, v, sigma=0.1)

DSM loss from A Connection Between Score Matching

and Denoising Autoencoders

The loss is computed as x_ = x + v # noisy samples s = -dE(x_)/dx_ loss = 1/2*||s + (x-x_)/sigma^2||^2 :param x: input samples :type x: torch.Tensor :param v: sampled noises :type v: torch.Tensor :param sigma: noise scale. Defaults to 0.1. :type sigma: float, optional

Returns:

DSM loss

encode(x, rounding=False, x_level=0, h_level=None)

encode_nominal(x, columns=None)

encode_ordinal(x, columns, order=None)

energy(x, h)

find_optimal_threshold_for_f1(data, search_every=1, plot=False)

forward(x)

Define the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

free_energy(x, level=1)

generate(num_examples, num_steps=10)

get_auroc(data)

gmm_model()

learn_denta_network(train_data, valid_data, max_epoch=10, min_epoch=0, weight_decay=0.0, batch_size=128, shuffle=True, verbose=True, early_stopping=False, early_stopping_patience=3, round_latent_during_learning=False)

learn_latent_automaton(train_data, valid_data)

mse(x, r, per_point=False)

num_h()

num_hidden_layers()

num_x()

plot_activation_probabilities(d, prepare_data=True)

plot_discretization(time, target, prediction, data=None, data_time=None)

plot_error_histogram(d, v=None)

plot_frequency_of_latent_combinations(d, prepare_data=True)

plot_input_space(data=None, samples=None, show_gaussian_components=False, data_limit=10000, xmin=None, xmax=None, ymin=None, ymax=None, figure_width=600, figure_height=600, show_axis_titles=True, show_energy_contours=False, showlegend=True, show_recon_error_contours=False, ncontours=None, plot_code_positions=True, show_recon_error_heatmap=False, plot_bias_vector=False, show_reconstructions=False, **kwargs)

plot_learning_curve()

predict_discrete_mode(data, prepare_data=True)

prepare_data(x, update_mean_std=False)

pretrain_dbn(train_data, valid_data, **kwargs)

recon(v, round=False, x_level=0, h_level=None)

recon_error(data, input=None, per_point=False, round=False)

sample_h(h)

sample_x(x)

score(x, sigma=None)

sparsity(activations, per_point=False)

sparsity_loss(activations)

train_rbm(train_data, valid_data, level=1, learning_rule='cd', min_epoch=0, max_epoch=10, weight_decay=0.0, batch_size=128, shuffle=True, num_k=1, verbose=True, early_stopping=False, early_stopping_patience=3, use_probability_last_x_update=False)

class automata4cps.denta.LinearRegression(input_dim, output_dim)

Bases: Module

forward(x)

Define the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

train_model(X, y, num_epochs=1000, lr=0.01)

Classes