neurosynchro.impl: the neural-network implementation

The main neural network code.

This module contains PhysicalApproximator, the class that uses prebuild keras neural networks to rapidly approximate radiative transfer coefficients.

class neurosynchro.impl.PhysicalApproximator(nn_dir)

This class approximates the eight nontrivial RT coefficients using a physically-based parameterization.

See Tutorial, AKA, How to Use It for detailed documentation of how to prepare and train the neural networks used by this class.

Constructor argument

nn_dir

The path to a directory containing trained neural network data. This directory should contain the configuration file nn_config.toml and serialized neural network weights in files with names like rho_Q_sign.h5.

compute_all_nontrivial(nu, B, n_e, theta, **kwargs)

Compute the nontrivial radiative transfer coefficients.

Arguments

nu

The observing frequency, in Hz. This and all parameters may be scalars or arrays; they are broadcast to a common shape before performing the computations.

B

The local magnetic field strength, in G.

n_e

The local density of synchrotron-emitting particles, in cm^-3.

theta

The angle between the line of sight and the local magnetic field, in radians.

**kwargs

Other arguments to the synchrotron model; these can vary depending on which particle distribution was used.

Return values

A tuple of (coeffs, oos):

coeffs

The radiative transfer coefficients in the Stokes basis where the Stokes U axis is aligned with the magnetic field. This is an array of shape S + (8,) where S is the shape of the broadcasted input parameters. Along the inner axis of the array, the coefficients are: (j_I, alpha_I, j_Q, alpha_Q, j_V, alpha_V, rho_Q, rho_V).

oos

An array of integers reporting where the calculations encountered out-of-sample values, that is, inputs or outputs beyond the range in which the neural networks were trained. The shape of this array is the same as that of the broadcased input parameters, or a scalar if the inputs were all scalars. For each set of input parameters, the least significant bit is set to 1 if the first input parameter was out-of-sample, where “first” is defined by the order in which these parameters are listed in the nn_config.toml file. The next more significant bit is set if the second input parameter was out of sample, and so on. After all of the input parameters, there are 9 flag bits indicating whether any of the output results were out-of-sample, relative to the range of normalized values encountered in the training set. The order in which these parameters are processed is j_I, alpha_I, rho_Q, rho_V, j_frac_pol, alpha_frac_pol, j_V_share, alpha_V_share, rho_Q_sign. Therefore if the synchrotron model takes 4 input parameters and the rho_Q_sign output is the only one to have been out-of-sample, the resulting oos value will be 0x1000.