lazy.nntools {lazy.nntools}

R Documentation

lazynntools: Some neuralnet tools for lazy boys and girls
(Alpha Version)

Description

This package provides some useful tools for learning supervised feed forward neural networks, which can be regarded as a non-linear regression or supervised classification models.

 The first layer is the input matrix, X.
 The last layer is the output matrix, Yhat.
 In between the first and the last layers, there are hidden layers.

The size of the network is given as a vector of the # of layers, nn, where nn[k] is the # of neurons of layer k.
nlayer is the # of layers: nlayer=length(nn)

Using the list of output matrices, O, and the list of weight matrices, W,
the output of the k-th layer will be calculated as:

  O[[k]] <- Psi( cbind( 1, O[[k-1]] ) %*% W[[k]] )

where Psi is the logistic function.

The Weight Matrices are stored in a list W of length nlayer.

 W[[k]] is nn[k-1]+1 x nn[k].
 W[[1]] is NULL.
 W[[2]] is nn[1]+1 x nn[2].
 W[[k]] is nn[k-1] x nn[k].
 W[[nlayer]] is nn[nlayer-1] x nn[nlayer].
 

The Output of each layer are stored in O, a list of length nlayer.

 O[[k]] is n x nn[k].
 O[[1]] = X.
 O[[k]] is n x nn[k].
 O[[nlayer]] is referred to as Yhat.
 

At the beginning of the session,
initialize_nn
must be called to create the storage lists for the output O and the weight W.

Tools

• output Given W, calculates the List of Output Matrices for each layer
  It is assumed that O exists.

• on Given unlist(W), calculates Yhat.

• rss_nn_a Given unlist(W), calculates rss w.r.t unlist(W)

• drss_nn_a Derivative of rss w.r.t unlist(W)

• drss_nn_aJ Derivative of rss w.r.t unlist(W) by Jac_nn_a

• Jac_nn_a Jacobian of Yhat w.r.t unlist(W)

• Jac_nn_a2 Jacobian of -Yhat w.r.t unlist(W)

• ymyhat_nn Given unlist(W), calculates vec(Y)-vec(Yhat)

• yhatmyh_nn Given unlist(W), calculates vec(Yhat)-vec(Y)
  

• initialize_nn Initialize W and O.

• getW_from_neuralnet Convert the list of weights from neuralnet::neuralnet to ours.

• reshapeW Recovers W from unlist(W)

• plot_nn Plots the Network
  This function requires neuralnet package or NeuralNetTools packages with grid packages.

• summary_nn Given W, calculates various quantities including, rss, rmse, Yhat, gradient.

Test Data and Examples

• data_1_2 Nonlinear Regression Data

• data_2_3 Supervised Classification Data
  
  

• ExamplesOfnn Tow Example Codes

---