cdeeply-python

Python interface to C Deeply's neural network generators

Put cdeeply_neural_network.py into a reachable directory, then:

0. Create a class instance: myNN = cdeeply_neural_network.CDNN()
1. Call myNN.tabular_regressor(...) or myNN.tabular_encoder(...) to train a neural network in supervised or unsupervised mode. This step requires an internet connection! as the training is done server-side.
2. Call output = myNN.runSample(input) as many times as you want to process new data samples -- one sample per function call.

Function definitions:

trainingSampleOutputs = myNN.tabular_regressor(trainingSamples, sampleTableTranspose,
outputRowOrColumnList, importances=[],
maxWeights="NO_MAX", maxHiddenNeurons="NO_MAX", maxLayers="NO_MAX",
maxWeightDepth="NO_MAX", maxActivationRate=1.,
maxWeightsHardLimit=True, maxHiddenNeuronsHardLimit=True, maxActivationsHardLimit=True,
allowedAFs=[True,True,True,True,True],
ifQuantizeWeights=False, wQuantBits=0, wQuantZeroInt=0, wQuantRange=1.,
ifQuantizeActivations=False, yQuantBits=0, yQuantZeroInt=0, yQuantRange=1.,
sparseWeights=False, allowNegativeWeights=True, hasBias=True, allowIOconnections=True)

Generates a x->y prediction network using supervised training on trainingSamples.

• trainingSamples has dimensions numFeatures and numSamples, containing both inputs and target outputs.
  • Set sampleTableTranspose to "FEATURE_SAMPLE_ARRAY" for trainingSamples[feature, sample] array ordering, or "SAMPLE_FEATURE_ARRAY" for trainingSamples[sample, feature] array ordering.
  • The rows/columns in trainingSamples corresponding to the target outputs are specified by outputRowOrColumnList.
• The optional importances argument weights the cost function of the target outputs. If passed, it has dimensions numTargetOutputs and numSamples (ordered according to sampleTableTranspose).
• Optional integer parameters maxWeights, maxHiddenNeurons and maxLayers limit the size of the neural network, and maxWeightDepth limits the depth of layer-to-layer connections. The corresponding maxWeightsHardLimit and maxHiddenNeuronsHardLimit parameters should be either True or False.
• allowedAFs is a length-5 Boolean vector corresponding to the allowed activation functions: (step, ReLU, ReLU1, sigmoid, tanh). Set each Boolean according to whether the activation function should be considered for a given layer of the network.
• To quantize weights or neural activations, set ifQuantizeWeightsorifQuantizeActivationstoTrueand give values to the following three fields; otherwise set the respectiveifQuantizetoFalse`.
• Set sparseWeights to either True or False, depending on whether we are generating sparse weight matrices.
• The parameter maxActivationRate should be set to 1. unless sparsifying the neural activations, in which case this can be a lower positive value. Sparse activations do not work with sigmoid or tanh activation functions. Set maxActivationsHardLimit to either True or False.
• The allowNegativeWeights and ifNNhasBias parameters are either True or False. The bias is a constant input to each neuron.
• Set allowIOconnections to False to forbid the input layer from feeding directly into the output layer. (Outliers in new input data might cause wild outputs).
• trainingSampleOutputs has dimensions numTargetOutputs and numSamples, and stores the training output as calculated by the server. This is mainly a check that the data went through the pipes OK. If you don't care, ignore the return value.

trainingSampleOutputs = myNN.tabular_encoder(trainingSamples, sampleTableTranspose, importances=[],
doEncoder=True, doDecoder=True, numEncodingFeatures=1,
numVariationalFeatures=0, variationalDistribution="NORMAL_DIST",
maxWeights="NO_MAX", maxHiddenNeurons="NO_MAX", maxLayers="NO_MAX",
maxWeightDepth="NO_MAX", maxActivationRate=1.,
maxWeightsHardLimit=True, maxHiddenNeuronsHardLimit=True, maxActivationsHardLimit=True,
allowedAFs=[True,True,True,True,True],
ifQuantizeWeights=False, wQuantBits=0, wQuantZeroInt=0, wQuantRange=1.,
ifQuantizeActivations=False, yQuantBits=0, yQuantZeroInt=0, yQuantRange=1.,
sparseWeights=False, allowNegativeWeights=True, ifNNhasBias=True)

Generates an autoencoder (or an encoder or decoder) using unsupervised training on trainingSamples.

• trainingSamples has dimensions numFeatures and numSamples, in the order determined by sampleTableTranspose.
• sampleTableTranspose and importances are set the same way as for tabular_regressor(...).
• The size of the encoding is determined by numEncodingFeatures.
  • So-called variational features are extra randomly-distributed inputs used by the decoder, analogous to the extra degrees of freedom a variational autoencoder generates.
  • variationalDistribution is set to "UNIFORM_DIST" if the variational inputs are uniformly-(0, 1)-distributed, or "NORMAL_DIST" if they are normally distributed (zero mean, unit variance).
• Set doEncoder to False for a decoder-only network.
• Set doDecoder to False for an encoder-only network.
• The remaining parameters are set the same way as for tabular_regressor(...).

sampleOutput = myNN.runSample(sampleInput, sampleVariationalInput=[])

Runs the neural network on a single sample, and returns the network output.

• If it's an autoencoder (encoder+decoder), length(sampleInput) and length(sampleOutput) equal the number of features in the training sample space. If it's just an encoder, length(sampleOutput) equals numEncodingFeatures; if decoder only, length(sampleInput) must equal numEncodingFeatures.
• If it's a decoder or autoencoder network with variational features, sample the variational features from the appropriate distribution and pass them as a second argument.
• The return value is simply a reference to the last layer of the network myNN.y[myNN.numLayers]. Note that this may overwrite the output of a prior sample unless a copy is made of the output data.