SmartEngine.NeuralNetworks Namespace Reference

Classes
class	A2CTrainer
	The A2C Trainer is a reinforcement learning trainer that is composed of two parts: an actor sub graph and a critic sub-graph. More...
class	ActivationNode
	Applies an activation function to the input More...
class	AddNode
	Component wise addition of two inputs. More...
class	Agent
	Agents are used to track the performance of one instance of a network. More...
class	AgentDataStore
	The agent data store keeps experience data for the purpose of training. Some RL trainers don't store long term data in the store, while others keep a long history of data. More...
class	BufferInput
	Input node into the AI graph. Allows for specifying a tensor input. More...
class	ChoiceNode
	Returns a random choice (as an integer) from the input node's probability distribution on each call to RetrieveOutput(). More...
class	ComponentInput
	Input node into a component. At the time of component construction, a node can be fed into this input. The incoming node must have the same dimension as this node. This node will then act as a passthrough for the incoming node. In this way, the component can be used at any point in the graph; not just the beginning. More...
struct	ComponentInputBinding
	Dynamically binds a node to a component input. More...
class	ConcatNode
	Merges two or more node streams into one. Each input stream must have the same number of rows. The output dimension is the sum of the input dimensions. More...
class	Constants
	Common constants More...
class	Context
	Every node in the AI graph must belong to the same context. More...
class	Conv2DInfo
class	Conv2DLayer
	2D convolution layer. Input is interpreted as a 2D grid laid in row major order. Filters are applied to blocks of the input grid at a time, producing another 2D grid as output. This 2D grid can be fed into neuron layers where it will be treated as a 1D array. More...
class	CuriosityModule
	A curiosity module is a way of rewarding an agent for behavior not yet seen. Rewards are given based on how expected the next set of observations are given the previous observations and the chosen actions. Unexpected results are given larger rewards than behavior the agent saw in the past, thus promoting the agent to be curious and explore the world. More...
class	D4PGTrainer
	The D4PGTrainer is a reinforcement learning trainer that is composed of two parts: an actor sub graph and a critic sub-graph. Unlike A2C and PPO, the critic graph is created and managed internally by SmartEngine. You only need to supply the actor graph. More...
class	DivideNode
	Component wise division of two inputs. More...
class	ExpNode
	Component wise natural exponentiation of X More...
class	ExtractNode
	Extracts a subset of columns from the input node. The start index + column count must not exceed the output dimension of the input node. More...
class	GeneSwapInfo
	GeneticTrainer gene swap info More...
class	GeneticAgentTrainer
	A genetic trainer that uses agents to automatically set the loss of each chromosome. Agents should be mapped to a chromosome after creation by calling MapAgentToChromosome() More...
class	GeneticTrainer
	Standard instance of IGeneticTrainer. More...
class	GeneticTrainerInitializationInfo
	Initializes the GeneticTrainer with a new population More...
class	GeneticTrainingInfo
	GeneticTrainer parameter info More...
class	GradientDescentTrainer
	Trains a set of networks using gradient descent. Training with GradientDescentTrainer requires a Loss structure and thus the output of the network must be known. If only a relative loss about the network is known, a GeneticTrainer needs to be used. More...
class	GradientDescentTrainingInfo
	GradientDescentTrainer training info More...
class	Graph
	A graph is a collection of buffers and nodes that together form a neural network. The graph is created from a json definition file. After creation, the contents of the graph can be loaded from or saved to disk. More...
class	GraphComponent
	A graph component allows graphs to embed other graphs. Nodes within components can be referenced using the following syntax: [Component Name]:[Component Graph Node Name] References can be chained together to dive to sub components. More...
class	GraphController
	Represents a use of a graph. Owners of the controller feed data to the controller at regular intervals. When the graph model is executed, the controller will have data that can be acted upon More...
class	GraphInput
	Base class for input nodes into the AI graph More...
class	GraphInputOutput
	Views on top of models deal with this class to inject and extract data from the underlying model. More...
class	GraphManager
	Connects graph controllers with graph models. Graph controllers define a single use of a graph. They pipe data to the model and later act on the output of the model. The model (usually a graph, but can be a dummy class) has the job of aggregating the input from one or more controllers and producing an output for each. More...
class	GraphModelOutput
	Used to return output back to controllers. More...
class	GraphNode
	A logical node in the AI graph. Some nodes, like NeuralNetwork, are composed of other nodes (neuron layers). More...
interface	IAgentFactory
	Trainers implement this to create agents. More...
interface	IGeneticTrainer
	Trains a NeuronLayer / NeuralNetwork using a genetic algorithm. With each step, networks are sorted by loss. The top networks are left unchanged, while the bottom networks are replaced by new ones. The new networks are formed by mixing the weights of the top networks and randomly mutating weights / neurons. Lastly, the bottom percent of the networks can be replaced with completely random weights, adding some variability to the pool. More...
interface	IGraphModel
class	LayerInfo
	Layer information More...
class	LengthNode
	Gets the length of a vector. Output is a one dimensional value equal to the sqrt of the square sums of the columns. More...
class	LogNode
	Component wise natural log of X More...
class	Loss
	The loss of a NeuralNetwork is computed using the formula (Expected Ouput - Actual Output)^2 The mean of all these values across the output tensor is used when crunching down the loss to a single value. More...
class	LossTrainer
	Base class for NeuralNetwork loss trainers More...
class	LossTrainingMethodInfo
class	MaximumNode
	Takes the maximum of the inputs. More...
class	MinimumNode
	Takes the minimum of the inputs. More...
class	MultiplexerNode
	Uses the value from the selector to decide which of the inputs to pass through. The selected input will be transmitted without adjustment. Each input should have the same dimension. The selector should have a dimension of 1 and a sigmoid activation. More...
class	MultiplyNode
	Component wise multiplication of two inputs. More...
class	MutationInfo
	GeneticTrainer mutation info More...
class	NegateNode
	Component wise negation of X More...
class	NeuronLayer
	A neuron layer is the trainable unit in the neural network graph. These can chained together to allow the network to learn and represent complex relationships. More...
class	NormalizeNode
	Treats spans of columns in the input as vectors to be normalized. Columns not covered by a span will be passed through unchanged. The output dimension will be equal to the input dimension. More...
class	NormalNode
	Returns a random value from the normal distribution given by the input nodes. NOTE: Input takes in the variance and not the standard deviation. More...
class	Parameter
	A parameter is similar to a NeuronLayer in that it is trainable. However, it takes no input and is simply a layer of weights that can be used as input into other parts of the graph. More...
class	PoolingLayer
	2D pooling layer. Typically used after a conv2d layer to reduce the output dimension. Fixed size of 2x2 blocks. More...
class	PPOTrainer
	The PPO Trainer is a reinforcement learning trainer that is composed of two parts: an actor sub graph and a critic sub-graph. One of the differences between PPO and A2C is that PPO works off of large batches of data. Instead of waiting for a small batch before training, PPO will wait until there is a mega batch, called the trajectory. When we've received that much data, we train over it in normal batch sizes for a couple of times and then wait for the next trajectory. More...
class	RegularizationLossInfo
	Regularization tries to keep the weight values from exploding during gradient descent by adding L1 and L2 loss of the weights. More...
class	RLTrainer
	Base class for all reinforcement learning trainers. More...
class	Rotate2DNode
	Rotation of a 2D vector. Output is also a 2D vector. Rotation value is expected to be in radians. More...
class	ScalarValueNode
	Returns a constant scalar value. More...
class	StandardGraphModel
class	StopGradientNode
	Stops the gradient from flowing backwards from this point. Only meaningful when training with gradient descent or reinforcement learning. More...
class	SubtractNode
	Component wise subtraction of two inputs. More...
class	UnpoolingLayer
	2D unpooling layer. Typically used before a conv2d transpose layer to expand the output dimension. Fixed size of 2x2 blocks. More...
class	UserGraphModel
	Takes input from a controller and produces an output. If using a graph as the model, this class does not need to be used or extended. However, you can provide a dervied instance of this class to create "dummy" models. More...
class	UserStandardGraphModel
	Allows for overriding of the standard graph model at the cost of another level of indirection. Can be used during training to modify the graph output before it is sent back to the controller. Consider fixing up data in the controller instead of using this class at runtime. More...

Enumerations
enum	ChoiceType { ChoiceType.Random, ChoiceType.Max }
	Defines the descision algorithm of the ChoiceNode More...
enum	DeviceType { DeviceType.Default, DeviceType.Cpu, DeviceType.Gpu }
	The type of context to create. More...
enum	Conv2DType { Conv2DType.Conv2D, Conv2DType.Conv2DWithBias, Conv2DType.Conv2DTranspose, Conv2DType.Conv2DTransposeWithBias }
enum	GeneticAgentScoringMethod : byte { GeneticAgentScoringMethod.Sum, GeneticAgentScoringMethod.Average }
	Defines how the trainer deals with chromosome scores from multiple agents. More...
enum	MutationTarget { MutationTarget.Auto, MutationTarget.Weight, MutationTarget.Neuron }
enum	GeneSwapTarget { GeneSwapTarget.Weight, GeneSwapTarget.Neuron }
enum	GradientDescentTrainingAlgorithm { GradientDescentTrainingAlgorithm.Adam, GradientDescentTrainingAlgorithm.AdaMax }
enum	LossTrainingMethod { LossTrainingMethod.WholeDataset, LossTrainingMethod.Stochastic }
enum	LayerType { LayerType.LinearNoBias, LayerType.Linear, LayerType.LSTM }
	The type of layer to create More...
enum	ActivationType {   ActivationType.None, ActivationType.Sigmoid, ActivationType.Tanh, ActivationType.Relu,   ActivationType.Relu6, ActivationType.LeakyRelu, ActivationType.Softmax, ActivationType.Elu,   ActivationType.Selu, ActivationType.Softplus }
	The activation function to apply to the output of the NeuronLayer More...
enum	PoolType : byte { PoolType.Max, PoolType.Average, Count }

Enumeration Type Documentation

ActivationType

enum SmartEngine.NeuralNetworks.ActivationType

strong

The activation function to apply to the output of the NeuronLayer

Enumerator
None	No activation
Sigmoid	Maps output into the range [0..1]
Tanh	Maps output into the range [-1..1]
Relu	max(x, 0)
Relu6	min(max(x, 0), 6)
LeakyRelu	Similar to Relu, but allows for scaled negative values
Softmax	Converts the outputs into weighted probabilities.
Elu	Similar to Relu, but uses an exponential for negative values
Selu	Similar to Relu, but uses an exponential for negative values. Positive values are scaled. Best when training with gradient descent. The inputs to the network should be be in the [-2..2] range.
Softplus	Maps output into a positive number

ChoiceType

enum SmartEngine.NeuralNetworks.ChoiceType

strong

Defines the descision algorithm of the ChoiceNode

Enumerator
Random	Treat the output as a probability distribution and sample from that distribution on each evaulation. Returns the index of the chosen output.
Max	Return the index of the output with the highest value.

Conv2DType

enum SmartEngine.NeuralNetworks.Conv2DType

strong

Enumerator
Conv2D	2D Convolution layer, no bias
Conv2DWithBias	2D Convolution layer, with bias
Conv2DTranspose	2D Transpose convolution layer (sometimes called deconvolution), no bias
Conv2DTransposeWithBias	2D Transpose convolution layer (sometimes called deconvolution), with bias

DeviceType

enum SmartEngine.NeuralNetworks.DeviceType

strong

The type of context to create.

Enumerator
Default	Chooses the best device for the current system.
Cpu	Train and evaulate only using the CPU.
Gpu	Train and evaulate using only the GPU.

GeneSwapTarget

enum SmartEngine.NeuralNetworks.GeneSwapTarget

strong

Enumerator
Weight	Swapping occurs at the individual weight level across all neurons.
Neuron	Swapping applies to all weights in a neuron.

GeneticAgentScoringMethod

enum SmartEngine.NeuralNetworks.GeneticAgentScoringMethod : byte

strong

Defines how the trainer deals with chromosome scores from multiple agents.

Enumerator
Sum	Sum the scores of all agents associated with the chromosome
Average	Average the scores of all agents associated with the chromosome

GradientDescentTrainingAlgorithm

enum SmartEngine.NeuralNetworks.GradientDescentTrainingAlgorithm

strong

Enumerator
Adam	Good default choice
AdaMax	Can be more stable in certain circumstances, such as RL training

LayerType

enum SmartEngine.NeuralNetworks.LayerType

strong

The type of layer to create

Enumerator
LinearNoBias	Linear with no bias (X * W)
Linear	Linear with bias (X * W + B)
LSTM	Long short term memory layer

LossTrainingMethod

enum SmartEngine.NeuralNetworks.LossTrainingMethod

strong

Enumerator
WholeDataset	Train against all data in the buffer in groups of size BatchSize.
Stochastic	Train by random sampling of the input data in groups of size BatchSize.

MutationTarget

enum SmartEngine.NeuralNetworks.MutationTarget

strong

Enumerator
Auto	Chooses the default value depending on the node type.
Weight	Mutation occurs at the individual weight level across all neurons.
Neuron	Mutation applies to all weights in a neuron.

PoolType

enum SmartEngine.NeuralNetworks.PoolType : byte

strong

Enumerator
Max	Outputs the maximum value of the block of inputs
Average	Takes the average value of the block of inputs