This section of the manual describes all the details of the Swift enumerations used by the AIToolbox framework.
##Convolution2DMatrix The type of convolution done by a Convolution2D operator
| case | Description |
|---|
verticalEdge3 | A standard Sobel filter for vertical edges in a 3x3 matrix
horizontalEdge3 | A standard Sobel filter for horizontal edges in a 3x3 matrix
custom3 | A custom 3x3 matrix
learnable3 | A custom 3x3 matrix that can be learned from data
##DataSetType Defines the type of the output of a data set
| case | Description |
|---|
regression | for regression data sets (array of real numbers output)
classification | for classification data sets (single integer output)
realAndClass | for classification data sets that require real number outputs (like probabilities of each class)
##DeepNetworkOperatorType The type of operator in a deep-network channel operation list
| case | Description |
|---|
convolution2DOperation | Convolve a 2D matrix with a convolution matrix
poolingOperation | Pool a matrix (any dimension) into a smaller one
feedForwardNetOperation | Feed the incoming data into a neural network layer
##MLPlotSymbolShape The shape of the symbol to be used to identify a data point on an MLView
| case | Description |
|---|
circle | circle
rectangle | rectangle
plus | plus sign
minus | minus sign
##MLViewAxisSource The source of values for an MLView axis
| case | Description |
|---|
dataInput | the input vector value at the specified index is used
dataOutput | the output vector value at the specified index is used
classLabel | the class label is used. The index parameter is ignored
##MLViewLegendLocation The location of a legend object in an MLView
| case | Description |
|---|
upperLeft | the upper-left corner of the view
upperRight | the upper-right corner of the view
lowerLeft | the lower-left corner of the view
lowerRight | the lower-right corner of the view
custom | a custom location specified by the caller
##NeuronLayerType Defines the type of the layer in a neural network. Networks containing recurrent style layers should be trained with sequences, rather than individual data.
| case | Description |
|---|
simpleFeedForwardWithNodes | a feed-forward layer that has each node as an individual class. Easier to follow internally, but slower
simpleFeedForward | a feed-forward layer that has all the weights in one matrix. Faster than individual nodes
simpleRecurrentWithNodes | a recurrent layer (previous outputs are also used as additional inputs) that has each node as an individual class. Easier to follow internally, but slower
simpleRecurrent | a recurrent layer (previous outputs are also used as additional inputs) that has has all the weights in two matrices. Faster than individual nodes
lstm | a long-short-term-memory layer (recurrent with flow gates)
##NeuralActivationFunction Defines the function used to 'squash' the results of the dot-product of the inputs with the node's weights.
| case | Description |
|---|---|
| none | no change is made to the result of initial dot product result |
| hyperbolicTangent | the tanh function, which limits the result to the range [-1, 1] is used |
| sigmoid | the logistic function [1/(1+exp(-x)] is used, which limits the range to [0, 1] |
| sigmoidWithCrossEntropy | the sigmoid function is used, but cross-entropy loss function is preferred (this is not yet implemented fully, and may be removed or changed later) |
| rectifiedLinear | values below zero are clipped to zero, else no change to the value |
| softSign | x / (1.0 + abs(x)) |
| softMax | Only valid on output (last) layer. The function used is just exp(x), but each value is normalized by the sum of all the outputs of the nodes in the layer. This gives the output vector the look of a probability distribution (all values in the range [0,1], with the total equal to 1) |
##NeuralWeightUpdateMethod The method to use when updating the weights from (possibly accumulated) gradients with respect to the error term.
| case | Description |
|---|
normal | normal update (W -= learning_rate * gradient). No parameter required.
rmsProp | root-mean-square propogation (rmsprop_cache = decay_rate * rmsprop_cache + (1 - decay_rate) * gradient², W -= learning_rate * gradient / (sqrt(rmsprop_cache) + 1e-5) ). Requires decay rate as parameter.
##NonLinearRegressionConvergenceType The type of criteria used to determine if a non-linear regression model has converged
| case | Description |
|---|
smallGradient | the gradient of the model becomes small
smallAverageLoss | the average loss of the training data becomes small
smallParameterChange | the changes to the parameters becomes small. This is the only option for non-linear regression models using parameterDelta for the solution type
##NonLinearRegressionType The solution method for a non-linear regression model
| case | Description |
|---|
parameterDelta | small changes are made to the parameters of the model, and kept if the total error decreases
sgd | stochastic gradient descent
gaussNewton | a Gauss-Newton method
##PoolingType The type of operation used in a deep-network pooling operator
| case | Description |
|---|
average | average of the matrix volume being pooled
minimum | minimum of the matrix volume being pooled
maximum | maximum of the matrix volume being pooled
##StandardConstraintType The type of constraint in a standard constraint propagation problem, when not using custom constraint functions.
| case | Description |
|---|
cantBeSameValueInOtherNode | All nodes that connect to the current node in the graph cannot have the same value as the current node
mustBeSameValueInOtherNode | All nodes that connect to the current node in the graph must have the same value as the current node
cantBeValue | All nodes that connect to the current node in the graph cannot have the value specified in the constraing
mustBeGreaterThanOtherNode | All nodes that connect to the current node in the graph must have a larger value than the current node
mustBeLessThanOtherNode | All nodes that connect to the current node in the graph must have a lesser value than the current node
##SubtermFunction The type of function in a linear regression model subterm
| case | Description |
|---|
none | no function is added to the subterm
naturalExponent | the subterm is placed inside a natural exponent function
sine | the subterm is placed inside a sine function
cosine | the subterm is placed inside a cosine function
log | the subterm is placed inside a natural logarithm function
power | the subterm is the power value of the term raised to input
##SVMKernelType The type of an SVM kernel
| case | Description |
|---|
linear | a linear kernel
polynomial | a polynomial kernel of the specified polygon order
radialBasisFunction | radial basis functions with the given gamma
sigmoid | sigmoid function kernel
precomputed | (not implemented at this time)
##SVMType The type of SVM model. These came from the LIBSVM port
| case | Description |
|---|
c_SVM_Classification | standard classification
ν_SVM_Classification | ν solution method classification
oneClassSVM | single class classification
ϵSVMRegression | standard regression
νSVMRegression | ν solution method regression
##DataIndexError Inherits from Error. Defines exceptions thrown for errors in data indexing values
| case | Description |
|---|
negative | the index for the data point was negative
indexAboveDimension | the index for a sub-element was above the dimension bounds for the array
indexAboveDataSetSize | the index was greater than or equal to the number of points in the data set
##DataTypeError Inherits from Error. Defines exceptions thrown for errors with the type of data being added to a data set
| case | Description |
|---|
invalidDatatype | the data type of the data set is not correct for the operation requested (i.e. grouping classes on a regression data set)
dataWrongForType | the output data being added to a data set is not correct for the type the data set was created for
wrongDimensionOnInput | the vector passed as an input vector to a data set does not match the input dimension the data set was created for
wrongDimensionOnOutput | the vector passed as an output vector to a regression data set does not match the output dimension the data set was created for
##GaussianError Inherits from Error. Defines exceptions thrown for errors using gaussian (mostly muti-variate gaussians) objects
| case | Description |
|---|
dimensionError | the input vector did not match the dimension of the gaussian
zeroInVariance | the variance matrix had a zero determinant, causing a mathematical error
inverseError | the variance matrix could not be inverted
badVarianceValue | attempt to set a bad value (negative) for a variance, or set a variance outside the range of a multi-variate gaussian variance matrix dimensions
diagonalCovarianceOnly | attempt to set a variance value off the diagonal of a diagonal-only multi-variate gaussian
errorInSVDParameters | error in parameters passed to the eigen-value calculation routines in internal math
svdDidNotConverge | the eigen-value calculation routines in internal math did not converge
##LinearRegressionError Inherits from Error. Defines exceptions thrown for errors using linear regression models
| case | Description |
|---|
modelExceedsInputDimension | the linear regression model has more parameters than the input vector dimension
matrixSolutionError | the matrix in the regression calculation could not be inverted
negativeInLogOrPower | negative values in the input conflicted with logarithm or power functions in the model
divideByZero | the regression model attempted to divide by zero
##MachineLearningError Inherits from Error. Defines exceptions thrown for errors using many machine learning routines
| case | Description |
|---|
dataNotRegression | regression data was needed, but classification data was provided
dataNotClassification | classification data was needed, but regression data was provided
dataWrongDimension | the data passed in was the wrong dimension for the model
notEnoughData | not enough data was provided for the operation (for example, less data points than parameters in a regression learning)
modelNotRegression | the function is for use in regression models only
modelNotClassification | the function is for use in classification models only
notTrained | attempt to use a model that has not yet been trained
initializationError | there was an error initializeing the model
didNotConverge | the model did not converge on a solution
continuationNotSupported | the model only supports one-time training
operationTimeout | the operation took longer than expected and was terminated early
continueTrainingClassesNotSame | the classification data passed into a training continuation function contained classes that were not seen with the initial training
##MDPErrors Inherits from Error. Defines exceptions thrown for errors using many Markov-Decision-Process routines
| case | Description |
|---|
mdpNotSolved | attempt to get an action for an MDP that has not yet been solved
failedSolving | the MDP failed in one-step solving
errorCreatingSampleSet | there was an internal error creating sample episodes
errorCreatingSampleTargetValues | there was an internal error creating sample episodes
modelInputDimensionError | the learning model specified uses an input dimension different than the data
modelOutputDimensionError | the learning model specified has an output dimension different than the data
invalidState | the state specified is outside the MDP state set
noDataForState | no data has been supplied for the state, so Q or V values can not yet be determined
##MixtureOfGaussianError Inherits from Error. Defines exceptions thrown for errors using a MixtureOfGaussian model
| case | Description |
|---|
kMeansFailed | the K-Means grouping of the data failed
##MLViewError Inherits from Error. Defines exceptions thrown for errors using a MLView view
| case | Description |
|---|
dataSetNotRegression | attempted to add a regression data set object, but the data set was not of the regression type
dataSetNotClassification | attempted to add a classification data set object, but the data set was not of the classification type
inputVectorNotOfCorrectSize | the input vector passed did not match the size of data being displayed
inputIndexOutsideOfRange | the index to be displayed as an axis is outside of the input vector dimension
outputIndexOutsideOfRange | the index to be displayed as an axis is outside of the output vector dimension
allClassificationLabelsNotCovered | the classification label object did not cover all known class labels
##NeuralNetworkError Inherits from Error. Defines exceptions thrown for errors using a NeuralNetwork that are not covered by general machine learning errors
| case | Description |
|---|
expectedOutputNotSet | the expected output was not set before trying to gradient check a network
##NonLinearRegressionError Inherits from Error. Defines exceptions thrown for errors using a NonLinearRegression that are not covered by general machine learning errors
| case | Description |
|---|
convergenceTypeNotAllowed | the selected convergence type does not match the selected solution type
matrixSolutionError | an error occurred solving the matrix equations for the solution method (Gauss-Newton)
##PCAWriteErrors Inherits from Error. Defines exceptions thrown for errors writing Primary Component Analysis parameters to a file
| case | Description |
|---|---|
| failedWriting | error writing the parameters |
##SVMWriteErrors Inherits from Error. Defines exceptions thrown for errors writing Support Vector Machine parameters to a file
| case | Description |
|---|---|
| failedWriting | error writing the parameters |
##ValidationError Inherits from Error. Defines exceptions thrown for errors when using a validation set object
| case | Description |
|---|
computationError | overflow on the math determining relative loss on each model being validated