NPE

This module contains the functions about using Neural Posterior Estimation for self-calibration.

class npe.GripNPE(wl_scale, func_model, func_args=(), func_kwargs={}, use_cuda=True)

This class is a framework to use the Neural Posterior Estimation technique with GRIP.

The idea is to simulate data and train the neural network to map these data with some parameters. During inference, real data are fed to the neural network to infer the posterior of the parameters.

Parameters

wl_scale1d-array: Wavelength scale.
func_modelfunction or callable: Function formatting the data to be compared to a model (e.g. histogram or power density spectrum).
func_argslist, optional: Arguments used or transmitted by func_model. The default is ().
func_kwargsdict, optional: Dictionary of keywords used or transmitted by func_model. The default is {}.
use_cudabool, optional: Use cuda for training the neural network or inferring, providing CUDA is available. The default is True.

Returns

None.

create_train_set(nb_batch_per_epoch, batch_size, dist_type, dist_params_theta)

Generate data set to train the neural network. It creates several batched data to use in one training epoch.

Parameters

nb_batch_per_epochint: Number of batches to train the neural network during one epoch.
batch_sizeint: Number of simulated data sequence in a batch.
dist_typelist of strings: Define the distribution to choose for each parameter. Available keywords are normal (for normal distribution) or uniform (for uniform distribution).
dist_params_thetaList of tuples: List of parameters to use to generate random values from a distribution defined by dist_type.

Returns

train_datand-array: Dataset to train the neural network of shape (nb_batch_per_epoch, batch_size, data_size).
train_paramsnd-array: Associated parameters that have created train_data of shape (nb_batch_per_epoch, batch_size, number_of_params).

draw_params(dist_type, dist_params_theta)

Generate parameters sets to simulate data. These parameters are the one to be inferred by the NN.

Parameters

dist_typelist of strings: Define the distribution to choose for each parameter. Available keywords are normal (for normal distribution) or uniform (for uniform distribution). If the keyword is not recognised, a uniform distribution is considered.
dist_params_thetaList of tuples: List of parameters to use to generate random values from a distribution defined by dist_type.

Returns

paramsarray-like: Parameters generated.

Examples

>>> dist_type = ['uniform', 'normal', 'uniform']
>>> dist_params_theta = [(0, 0.01), (300, 100), (100, 20)]
>>> draw_params(dist_type, dist_params_theta)

generate_batch(batch_size, dist_type, dist_params_theta)

Generate batch of data to train de neural network and their associated parameters.

Parameters

batch_sizeint: Number of simulated data sequence in a batch.
dist_typelist of strings: Define the distribution to choose for each parameter. Available keywords are normal (for normal distribution) or uniform (for uniform distribution).
dist_params_thetaList of tuples: List of parameters to use to generate random values from a distribution defined by dist_type.

Returns

batched_dataarray: Batched simulated data of shape (batch_size, *data.shape).
batched_thetaarray: Batched parameters to infer of shape (batch_size, number of parameters).

inference_on_data(data_to_inf, nb_sample)

Infer data to deduce the posterior of the parameters with the trained neural network.

Parameters

data_to_infnd-array: Data to infer. They must be in the same shape and format as the simulated data used to train the neural network.
nb_sample: int: Number of samples to generate from the surrogate posteriors of the parameters.

Returns

samplesnd-array: Samples of the inferred parameters, which distributions are their posteriors. Shape is (nb elements, nb parameters).

set_nn(nb_transforms, nb_layers, hidden_features)

Create the normalising flow to infer data. The flow is a masked autoregressive flow.

Parameters

nb_transformsint: Number of transformation in the normalising flow.
nb_layersint: Number of layers in a Transform component.
hidden_featuresint: Number of features out of each layer.

Attributes

self.maf_estimator : neural network estimator flow

simulator(params)

Simulate the data to train the neural network.

Parameters

paramsarray-like: Parameters to infer of the model to fit on the data.

Returns

out1d-array: Modelled data.

train_nn(nb_epoch, show_plot=False, save_fig=False, save_path='cluster.png')

Train the neural network following the Neural Posterior Estimation flow.

Parameters

nb_epochint: Number of epochs to train the neural network.
show_plotbool, optional: Show the learning curve. The default is False.
save_figbool, optional: Save the learning curve. The default is False.
save_pathstring, optional: Path to save the figure. This path must end by the name of the file and its extension. The default is ‘cluster.png’. The figure is saved as a PNG with dpi=150.

Returns

training_loss1d-array: Values of the learning curve. The size of the array is the number of epochs.