Function and module reference

The context manager

cadjn.init()

Returns a CADJN context manager object.

Returns:

CADJN context manager object.

Return type:

CADJN

class cadjn.CADJN

CADJN context manager.

This class loads the CADJN C library and creates handles for the necessary functions.

The most important methods are init, which creates CAdjNet neural net objects, load, which creates CAdjNet neural net objects from files, and save, which saves CAdjNet neural net objects.

init(init)

Returns a CADJN context manager object.

Parameters:

init (str) – Neural net definition in JSON format.

Returns:

CAdjNet neural net object.

Return type:

CAdjNet

load(fname)

Returns a CADJN context manager object.

Parameters:

fname (str) – File to load.

Returns:

CAdjNet neural net object.

Return type:

CAdjNet

save(net, fname)

Saves CAdjNet neural net to file.

Parameters:

  • net (CAdjNet) – Neural net to save

  • fname (str) – File to save to.

Returns:

CAdjNet neural net object.

Return type:

CAdjNet

The neural network

class cadjn.cadjn.CAdjNet(cadjn, net, data)

The CADJN neural net object.

Hv(d_w, w, data, specs)

Compute a product of the Hessian of the scalar objective function objw given weights, data, and specs, with d_w. Also returns the gradient as with g and the scalar objective value.

Parameters:

  • d_w (numpy.ndarray) – The derivative direction(s). A 2D array of shape [N, NDD], where N is the number of weights (numel) and NDD is some integer > 0. A single vector with shape [N] is also accepted.

  • w (numpy.ndarray) – The network weights. The gradient returned is the derivative of the scalar objective function w.r.t. this parameter.

  • data (numpy.ndarray) – Data and specs to process, as for objw.

  • specs (numpy.ndarray) – Data and specs to process, as for objw.

Returns:

Product of the Hessian of the scalar objective function with d_w, a 2D array of shape [N, NDD] (N is numel), gradient of the scalar objective function, a vector of shape [N], and the scalar objective value.

Return type:

(numpy.ndarray, numpy.ndarray, float)

Jv(d_w, w, data, specs)

Compute a product of the Jacobian of the vector objective function objwv given weights, data, and specs, with d_w. Also returns the vector objective value.

Parameters:

  • d_w (numpy.ndarray) – The derivative direction(s). A 2D array of shape [N, NDD], where N is the number of weights (numel) and NDD is some integer > 0. A single vector with shape [N] is also accepted.

  • w (numpy.ndarray) – The network weights. The gradient returned is the derivative of the scalar objective function w.r.t. this parameter.

  • data (numpy.ndarray) – Data and specs to process, as for objw.

  • specs (numpy.ndarray) – Data and specs to process, as for objw.

Returns:

The Jacobian product with d_w, a 2D array of shape [M*J, NDD], where M is the number of samples (the leading dimension of data) and J is numout, and the vector objective value.

Return type:

(numpy.ndarray, numpy.ndarray)

compute(data)

Compute the neural network response from data, which is a 2D array of input data, one per row.

Parameters:

data (numpy.ndarray) – Data to process, must be an 2D array of shape [M, I], where M is the number of samples (batch size) and I is the number of inputs (numin).

Returns:

Network response, a 2D array of shape [M, J], where J is the number of outputs (numout).

Return type:

numpy.ndarray

g(w, data, specs)

Compute the gradient of the scalar objective function objw given weights, data, and specs. Also returns the scalar objective value.

Parameters:

  • w (numpy.ndarray) – The network weights. The gradient returned is the derivative of the scalar objective function w.r.t. this parameter.

  • data (numpy.ndarray) – Data and specs to process, as for objw.

  • specs (numpy.ndarray) – Data and specs to process, as for objw.

Returns:

Gradient of the scalar objective function, a vector of shape [N] (N is numel), and the scalar objective value.

Return type:

(numpy.ndarray, float)

getweights()

Set the weights of the neural network. The vector is a sequence of bias vectors and transfer matrixes according to the network configuration. Use setweights to set the weights.

Returns:

Vector of network weights, a 1D array of shape [N], where N is the number of weights (numel).

Return type:

numpy.ndarray

numel()
Returns:

The number of weights of the network.

Return type:

int

numin()
Returns:

The number of inputs of the network.

Return type:

int

numout()
Returns:

The number of outputs of the network.

Return type:

int

obj(data, specs)

Compute the scalar objective (error) function of the neural network response from data, which is a 2D array of input data, one per row, given the specs, which is a 2D array of the intended response, one per row. The result is the result of residual evaluated by the objective function configured in the network.

Parameters:

  • data (numpy.ndarray) – Data to process, must be an 2D array of shape [M, I], where I is the number of inputs (numin).

  • specs (numpy.ndarray) – Data to process, must be an 2D array of shape [M, J], where J is the number of outputs (numout).

Returns:

Value of the network scalar objective function.

Return type:

float

objv(data, specs)

Compute the vector objective (error) function of the neural network response from data, which is a 2D array of input data, one per row, given the specs, which is a 2D array of the intended response, one per row. The result is the result of residual evaluated by the objective function configured in the network.

Parameters:

  • data (numpy.ndarray) – Data to process, must be an 2D array of shape [M, I], where I is the number of inputs (numin).

  • specs (numpy.ndarray) – Data to process, must be an 2D array of shape [M, J], where J is the number of outputs (numout).

Returns:

Value of the network vector objective function.

Return type:

numpy.ndarray

objw(w, data, specs)

Like obj, but apply setweights before computation.

Parameters:

  • w (numpy.ndarray) – Vector of network weights, must be an 1D array of shape [N], where N is the number of weights (numel).

  • data (numpy.ndarray) – These parameters are passed on to obj.

  • specs (numpy.ndarray) – These parameters are passed on to obj.

Returns:

Value of the network scalar objective function.

Return type:

float

objwv(w, data, specs)

Like objv, but apply setweights before computation.

Parameters:

  • w (numpy.ndarray) – Vector of network weights, must be an 1D array of shape [N], where N is the number of weights (numel).

  • data (numpy.ndarray) – These parameters are passed on to objv.

  • specs (numpy.ndarray) – These parameters are passed on to objv.

Returns:

Value of the network vector objective function.

Return type:

numpy.ndarray

residual(data, specs)

Compute the residual (error) of the neural network response from data, which is a 2D array of input data, one per row, given the specs, which is a 2D array of the intended response, one per row. The result is the result of compute minus the specs.

Parameters:

  • data (numpy.ndarray) – Data to process, must be an 2D array of shape [M, I], where M is the number of samples (batch size) and I is the number of inputs (numin).

  • specs (numpy.ndarray) – Data to process, must be an 2D array of shape [M, J], where J is the number of outputs (numout).

Returns:

Residual of network response, a 2D array of shape [M, J], where J is the number of outputs (numout).

Return type:

numpy.ndarray

setweights(w)

Set the weights of the neural network. The vector is a sequence of bias vectors and transfer matrixes according to the network configuration. Use getweights to get the current weights vector.

Parameters:

w (numpy.ndarray) – Vector of network weights, must be a 1D array of shape [N], where N is the number of weights (numel).

Returns:

Status, 0 means no error

Return type:

int

vJ(d_w, w, data, specs)

Compute a product of d_w with the Jacobian of the vector objective function objwv given weights, data, and specs. Also returns the vector objective value.

Parameters:

  • d_w (numpy.ndarray) – The derivative direction(s). A 2D array of shape [NDD, M*J], where M is the number of samples (the leading dimension of data), J is numout and NDD is some integer > 0. A single vector with shape [M*J] is also accepted.

  • w (numpy.ndarray) – The network weights. The gradient returned is the derivative of the scalar objective function w.r.t. this parameter.

  • data (numpy.ndarray) – Data and specs to process, as for objw.

  • specs (numpy.ndarray) – Data and specs to process, as for objw.

Returns:

The product of d_w with the Jacobian, a 2D array of shape [NDD, N], where N is numel and NDD is some integer > 0, and the vector objective value.

Return type:

(numpy.ndarray, numpy.ndarray)