spacecowboy / aneuraln Goto Github PK

Should plot the performance of the network during training

Matlab: Press the performance button on the neural network window

Should be possible to plot the boundary of the network in graphs.

Matlab function: boundary()

Implement equivalent of stat-function in matlab.

function [spec, sens, tot, None, Nzero, miss] = stat(Y, T);

Nzero = max(1,length(find(T<0.5)));
None = max(1,length(find(T>0.5)));

specerr = length(find((T-Y)<-0.5));
spec = 100.0*(Nzero-specerr)/Nzero;

senserr = length(find((T-Y)>0.5));
sens = 100.0*(None-senserr)/None;

miss = sum(abs(round(Y)-T));
tot = 100.0*(ntarget-miss)/ntarget;

Divide a data file into a test and a validation set. Train on test, then test on validation set when training is complete

Input field where the number of iterations for the training is specified.

E.g. if the training should use online, batch or block updating for the weights

Which activation functions which should be used for the hidden and output nodes, e.g. tanh, logh, step, linear etc

Specify the config of the network, basically specify number of
inputs
hidden nodes
number of outputs

Convert network to use numpy arrays for input/output instead of python standard []

Make networks threadsafe by removing values from input nodes. Instead update/sim operations will work by sending the input values back from the output-node. E.g.:
network.update(input) -> sum of output_nodes.output(input) -> sum of hidden_nodes.output(input) -> sum of input_nodes.output(input) -> return input

The call ends with input node since their output() simply returns the value that was passed to it.

Plot data according to classification

Matlab function: plot2d2c()

implement loadsyn functions to be able to compare with Matlab

that is, keep a fixed amount of each class in the two splitted versions

Load a previously saved network state.

CherryPy to serve the page to the client.
Ajax for continous updating on the page: http://sv.wikipedia.org/wiki/AJAX

Hinton-diagram with matplotlib!!!
http://www.scipy.org/Cookbook/Matplotlib/HintonDiagrams

How this should be used is self-explanatory...

Instead of hold-out.

Using backward elimination, find the optimal set of variables.
Choose which to remove by first ranking the variables, replace a given variable by its average, then train the network. Base the training on an already trained network (using all variables)

Change the radio buttons for block/batch/online to a single text field preceeded by a label "Block Size:" with default value "1" (which yields Online)

Allow the file containing input values (in a specified format!) to be specified.

Implement a global process-pool. Use it to separate training from GUI and then to maximise performance of training.

Specifiy which error function should be used during the training and evalutation.

as the matlab function, to plot classification problems correctly

Choose which traning algorithm should be used, e.g. Gradient descent, some second order method, genetic algorithm etc

Button to save network state with weights and structure to a file in XML-format