A generic multi-layer neural network.
A Neural network is constructed using an array of CNeuralLayer objects. The NeuralLayer class defines the interface necessary for forward and backpropagation.
The network can be constructed as any arbitrary directed acyclic graph.
How to use the network:
The network can also be initialized from a JSON file using CNeuralNetworkFileReader.
Supported feature types: CDenseFeatures<float64_t> Supported label types:
The neural network can be trained using L-BFGS (default) or mini-batch gradient descent.
NOTE: LBFGS does not work properly when using dropout/max-norm regularization due to their stochastic nature. Use gradient descent instead.
During training, the error at each iteration is logged as MSG_INFO. (to turn on info messages call sg_io->set_loglevel(MSG_INFO)).
The network stores the parameters of all the layers in a single array. This makes it easy to train a network of any combination of arbitrary layer types using any optimization method (gradient descent, L-BFGS, ..)
All the matrices the network (and related classes) deal with are in column-major format
When implemnting new layer types, the function check_gradients() can be used to make sure the gradient computations are correct.
在文件 NeuralNetwork.h 第 110 行定义.
Protected 属性 | |
int32_t | m_num_inputs |
int32_t | m_num_layers |
CDynamicObjectArray * | m_layers |
SGMatrix< bool > | m_adj_matrix |
int32_t | m_total_num_parameters |
SGVector< float64_t > | m_params |
SGVector< bool > | m_param_regularizable |
SGVector< int32_t > | m_index_offsets |
int32_t | m_batch_size |
bool | m_is_training |
float64_t | m_max_train_time |
CLabels * | m_labels |
ESolverType | m_solver_type |
bool | m_store_model_features |
bool | m_data_locked |
友元 | |
class | CDeepBeliefNetwork |
CNeuralNetwork | ( | ) |
default constuctor
在文件 NeuralNetwork.cpp 第 43 行定义.
CNeuralNetwork | ( | CDynamicObjectArray * | layers | ) |
Sets the layers of the network
layers | An array of CNeuralLayer objects specifying the layers of the network. Must contain at least one input layer. The last layer in the array is treated as the output layer |
在文件 NeuralNetwork.cpp 第 49 行定义.
|
virtual |
在文件 NeuralNetwork.cpp 第 153 行定义.
apply machine to data if data is not specified apply to the current features
data | (test)data to be classified |
在文件 Machine.cpp 第 152 行定义.
|
virtual |
apply machine to data in means of binary classification problem
重载 CMachine .
在文件 NeuralNetwork.cpp 第 158 行定义.
|
virtualinherited |
apply machine to data in means of latent problem
被 CLinearLatentMachine 重载.
在文件 Machine.cpp 第 232 行定义.
Applies a locked machine on a set of indices. Error if machine is not locked
indices | index vector (of locked features) that is predicted |
在文件 Machine.cpp 第 187 行定义.
|
virtualinherited |
applies a locked machine on a set of indices for binary problems
被 CKernelMachine , 以及 CMultitaskLinearMachine 重载.
在文件 Machine.cpp 第 238 行定义.
|
virtualinherited |
applies a locked machine on a set of indices for latent problems
在文件 Machine.cpp 第 266 行定义.
|
virtualinherited |
applies a locked machine on a set of indices for multiclass problems
在文件 Machine.cpp 第 252 行定义.
|
virtualinherited |
applies a locked machine on a set of indices for regression problems
被 CKernelMachine 重载.
在文件 Machine.cpp 第 245 行定义.
|
virtualinherited |
applies a locked machine on a set of indices for structured problems
在文件 Machine.cpp 第 259 行定义.
|
virtual |
apply machine to data in means of multiclass classification problem
重载 CMachine .
在文件 NeuralNetwork.cpp 第 199 行定义.
|
virtualinherited |
applies to one vector
被 CKernelMachine, CRelaxedTree, CWDSVMOcas, COnlineLinearMachine, CLinearMachine, CMultitaskLinearMachine, CMulticlassMachine, CKNN, CDistanceMachine, CMultitaskLogisticRegression, CMultitaskLeastSquaresRegression, CScatterSVM, CGaussianNaiveBayes, CPluginEstimate , 以及 CFeatureBlockLogisticRegression 重载.
|
virtual |
|
virtualinherited |
apply machine to data in means of SO classification problem
被 CLinearStructuredOutputMachine 重载.
在文件 Machine.cpp 第 226 行定义.
|
inherited |
Builds a dictionary of all parameters in SGObject as well of those of SGObjects that are parameters of this object. Dictionary maps parameters to the objects that own them.
dict | dictionary of parameters to be built. |
在文件 SGObject.cpp 第 597 行定义.
Checks if the gradients computed using backpropagation are correct by comparing them with gradients computed using numerical approximation. Used for testing purposes only.
Gradients are numerically approximated according to:
\[ c = max(\epsilon x, s) \]
\[ f'(x) = \frac{f(x + c)-f(x - c)}{2c} \]
approx_epsilon | Constant used during gradient approximation |
s | Some small value, used to prevent division by zero |
在文件 NeuralNetwork.cpp 第 554 行定义.
|
virtualinherited |
Creates a clone of the current object. This is done via recursively traversing all parameters, which corresponds to a deep copy. Calling equals on the cloned object always returns true although none of the memory of both objects overlaps.
在文件 SGObject.cpp 第 714 行定义.
|
protectedvirtual |
Forward propagates the inputs and computes the error between the output layer's activations and the given target activations.
inputs | inputs to the network, a matrix of size m_num_inputs*m_batch_size |
targets | desired values for the network's output, matrix of size num_neurons_output_layer*batch_size |
在文件 NeuralNetwork.cpp 第 546 行定义.
Computes the error between the output layer's activations and the given target activations.
targets | desired values for the network's output, matrix of size num_neurons_output_layer*batch_size |
被 CDeepAutoencoder , 以及 CAutoencoder 重载.
在文件 NeuralNetwork.cpp 第 519 行定义.
|
protectedvirtual |
Applies backpropagation to compute the gradients of the error with repsect to every parameter in the network.
inputs | inputs to the network, a matrix of size m_num_inputs*m_batch_size |
targets | desired values for the output layer's activations. matrix of size m_layers[m_num_layers-1].get_num_neurons()*m_batch_size |
gradients | array to be filled with gradient values. |
在文件 NeuralNetwork.cpp 第 467 行定义.
|
virtual |
Connects layer i as input to layer j. In order for forward and backpropagation to work correctly, i must be less that j
在文件 NeuralNetwork.cpp 第 75 行定义.
Locks the machine on given labels and data. After this call, only train_locked and apply_locked may be called
Only possible if supports_locking() returns true
labs | labels used for locking |
features | features used for locking |
被 CKernelMachine 重载.
在文件 Machine.cpp 第 112 行定义.
|
virtualinherited |
Unlocks a locked machine and restores previous state
被 CKernelMachine 重载.
在文件 Machine.cpp 第 143 行定义.
|
virtualinherited |
A deep copy. All the instance variables will also be copied.
在文件 SGObject.cpp 第 198 行定义.
|
virtual |
Disconnects layer i from layer j
在文件 NeuralNetwork.cpp 第 88 行定义.
|
virtual |
Removes all connections in the network
在文件 NeuralNetwork.cpp 第 93 行定义.
Recursively compares the current SGObject to another one. Compares all registered numerical parameters, recursion upon complex (SGObject) parameters. Does not compare pointers!
May be overwritten but please do with care! Should not be necessary in most cases.
other | object to compare with |
accuracy | accuracy to use for comparison (optional) |
tolerant | allows linient check on float equality (within accuracy) |
在文件 SGObject.cpp 第 618 行定义.
Ensures the given features are suitable for use with the network and returns their feature matrix
在文件 NeuralNetwork.cpp 第 614 行定义.
Applies forward propagation, computes the activations of each layer up to layer j
data | input features |
j | layer index at which the propagation should stop. If -1, the propagation continues up to the last layer |
在文件 NeuralNetwork.cpp 第 439 行定义.
|
protectedvirtual |
Applies forward propagation, computes the activations of each layer up to layer j
inputs | inputs to the network, a matrix of size m_num_inputs*m_batch_size |
j | layer index at which the propagation should stop. If -1, the propagation continues up to the last layer |
在文件 NeuralNetwork.cpp 第 446 行定义.
|
virtual |
|
inherited |
|
inherited |
|
inherited |
|
virtualinherited |
|
protected |
returns a pointer to layer i in the network
在文件 NeuralNetwork.cpp 第 723 行定义.
returns a copy of a layer's parameters array
i | index of the layer |
在文件 NeuralNetwork.cpp 第 712 行定义.
CDynamicObjectArray * get_layers | ( | ) |
Returns an array holding the network's layers
在文件 NeuralNetwork.cpp 第 744 行定义.
|
virtual |
|
inherited |
|
inherited |
在文件 SGObject.cpp 第 498 行定义.
|
inherited |
Returns description of a given parameter string, if it exists. SG_ERROR otherwise
param_name | name of the parameter |
在文件 SGObject.cpp 第 522 行定义.
|
inherited |
Returns index of model selection parameter with provided index
param_name | name of model selection parameter |
在文件 SGObject.cpp 第 535 行定义.
|
virtual |
Returns the name of the SGSerializable instance. It MUST BE the CLASS NAME without the prefixed `C'.
重载 CMachine .
被 CDeepAutoencoder , 以及 CAutoencoder 重载.
在文件 NeuralNetwork.h 第 232 行定义.
int32_t get_num_inputs | ( | ) |
returns the number of inputs the network takes
在文件 NeuralNetwork.h 第 224 行定义.
int32_t get_num_outputs | ( | ) |
returns the number of neurons in the output layer
在文件 NeuralNetwork.cpp 第 739 行定义.
int32_t get_num_parameters | ( | ) |
returns the totat number of parameters in the network
在文件 NeuralNetwork.h 第 218 行定义.
return the network's parameter array
在文件 NeuralNetwork.h 第 221 行定义.
|
inherited |
|
virtual |
Initializes the network
sigma | standard deviation of the gaussian used to randomly initialize the parameters |
在文件 NeuralNetwork.cpp 第 98 行定义.
|
virtualinherited |
If the SGSerializable is a class template then TRUE will be returned and GENERIC is set to the type of the generic.
generic | set to the type of the generic if returning TRUE |
在文件 SGObject.cpp 第 296 行定义.
|
protectedvirtual |
check whether the labels is valid.
Subclasses can override this to implement their check of label types.
lab | the labels being checked, guaranteed to be non-NULL |
重载 CMachine .
在文件 NeuralNetwork.cpp 第 689 行定义.
converts the given labels into a matrix suitable for use with network
在文件 NeuralNetwork.cpp 第 630 行定义.
|
virtualinherited |
Load this object from file. If it will fail (returning FALSE) then this object will contain inconsistent data and should not be used!
file | where to load from |
prefix | prefix for members |
在文件 SGObject.cpp 第 369 行定义.
|
protectedvirtualinherited |
Can (optionally) be overridden to post-initialize some member variables which are not PARAMETER::ADD'ed. Make sure that at first the overridden method BASE_CLASS::LOAD_SERIALIZABLE_POST is called.
ShogunException | will be thrown if an error occurs. |
被 CKernel, CWeightedDegreePositionStringKernel, CList, CAlphabet, CLinearHMM, CGaussianKernel, CInverseMultiQuadricKernel, CCircularKernel , 以及 CExponentialKernel 重载.
在文件 SGObject.cpp 第 426 行定义.
|
protectedvirtualinherited |
Can (optionally) be overridden to pre-initialize some member variables which are not PARAMETER::ADD'ed. Make sure that at first the overridden method BASE_CLASS::LOAD_SERIALIZABLE_PRE is called.
ShogunException | will be thrown if an error occurs. |
被 CDynamicArray< T >, CDynamicArray< float64_t >, CDynamicArray< float32_t >, CDynamicArray< int32_t >, CDynamicArray< char >, CDynamicArray< bool > , 以及 CDynamicObjectArray 重载.
在文件 SGObject.cpp 第 421 行定义.
|
virtualinherited |
在文件 SGObject.cpp 第 262 行定义.
|
inherited |
prints all parameter registered for model selection and their type
在文件 SGObject.cpp 第 474 行定义.
|
virtualinherited |
|
virtual |
Connects each layer to the layer after it. That is, connects layer i to as input to layer i+1 for all i.
在文件 NeuralNetwork.cpp 第 81 行定义.
|
virtualinherited |
Save this object to file.
file | where to save the object; will be closed during returning if PREFIX is an empty string. |
prefix | prefix for members |
在文件 SGObject.cpp 第 314 行定义.
|
protectedvirtualinherited |
Can (optionally) be overridden to post-initialize some member variables which are not PARAMETER::ADD'ed. Make sure that at first the overridden method BASE_CLASS::SAVE_SERIALIZABLE_POST is called.
ShogunException | will be thrown if an error occurs. |
被 CKernel 重载.
在文件 SGObject.cpp 第 436 行定义.
|
protectedvirtualinherited |
Can (optionally) be overridden to pre-initialize some member variables which are not PARAMETER::ADD'ed. Make sure that at first the overridden method BASE_CLASS::SAVE_SERIALIZABLE_PRE is called.
ShogunException | will be thrown if an error occurs. |
被 CKernel, CDynamicArray< T >, CDynamicArray< float64_t >, CDynamicArray< float32_t >, CDynamicArray< int32_t >, CDynamicArray< char >, CDynamicArray< bool > , 以及 CDynamicObjectArray 重载.
在文件 SGObject.cpp 第 431 行定义.
|
protectedvirtual |
Sets the batch size (the number of train/test cases) the network is expected to deal with. Allocates memory for the activations, local gradients, input gradients if necessary (if the batch size is different from it's previous value)
batch_size | number of train/test cases the network is expected to deal with. |
在文件 NeuralNetwork.cpp 第 604 行定义.
|
inherited |
在文件 SGObject.cpp 第 41 行定义.
|
inherited |
在文件 SGObject.cpp 第 46 行定义.
|
inherited |
在文件 SGObject.cpp 第 51 行定义.
|
inherited |
在文件 SGObject.cpp 第 56 行定义.
|
inherited |
在文件 SGObject.cpp 第 61 行定义.
|
inherited |
在文件 SGObject.cpp 第 66 行定义.
|
inherited |
在文件 SGObject.cpp 第 71 行定义.
|
inherited |
在文件 SGObject.cpp 第 76 行定义.
|
inherited |
在文件 SGObject.cpp 第 81 行定义.
|
inherited |
在文件 SGObject.cpp 第 86 行定义.
|
inherited |
在文件 SGObject.cpp 第 91 行定义.
|
inherited |
在文件 SGObject.cpp 第 96 行定义.
|
inherited |
在文件 SGObject.cpp 第 101 行定义.
|
inherited |
在文件 SGObject.cpp 第 106 行定义.
|
inherited |
在文件 SGObject.cpp 第 111 行定义.
|
inherited |
set generic type to T
|
inherited |
|
inherited |
|
inherited |
|
virtual |
|
virtual |
Sets the layers of the network
layers | An array of CNeuralLayer objects specifying the layers of the network. Must contain at least one input layer. The last layer in the array is treated as the output layer |
在文件 NeuralNetwork.cpp 第 55 行定义.
|
inherited |
|
inherited |
|
virtualinherited |
Setter for store-model-features-after-training flag
store_model | whether model should be stored after training |
在文件 Machine.cpp 第 107 行定义.
|
virtualinherited |
A shallow copy. All the SGObject instance variables will be simply assigned and SG_REF-ed.
被 CGaussianKernel 重载.
在文件 SGObject.cpp 第 192 行定义.
|
protectedvirtualinherited |
Stores feature data of underlying model. After this method has been called, it is possible to change the machine's feature data and call apply(), which is then performed on the training feature data that is part of the machine's model.
Base method, has to be implemented in order to allow cross-validation and model selection.
NOT IMPLEMENTED! Has to be done in subclasses
被 CKernelMachine, CKNN, CLinearMulticlassMachine, CTreeMachine< T >, CTreeMachine< ConditionalProbabilityTreeNodeData >, CTreeMachine< RelaxedTreeNodeData >, CTreeMachine< id3TreeNodeData >, CTreeMachine< VwConditionalProbabilityTreeNodeData >, CTreeMachine< CARTreeNodeData >, CTreeMachine< C45TreeNodeData >, CTreeMachine< CHAIDTreeNodeData >, CTreeMachine< NbodyTreeNodeData >, CLinearMachine, CGaussianProcessMachine, CHierarchical, CDistanceMachine, CKernelMulticlassMachine , 以及 CLinearStructuredOutputMachine 重载.
|
virtualinherited |
被 CKernelMachine , 以及 CMultitaskLinearMachine 重载.
|
virtualinherited |
train machine
data | training data (parameter can be avoided if distance or kernel-based classifiers are used and distance/kernels are initialized with train data). If flag is set, model features will be stored after training. |
被 CRelaxedTree, CAutoencoder, CSGDQN , 以及 COnlineSVMSGD 重载.
在文件 Machine.cpp 第 39 行定义.
|
protectedvirtual |
trains the network using gradient descent
在文件 NeuralNetwork.cpp 第 261 行定义.
trains the network using L-BFGS
在文件 NeuralNetwork.cpp 第 357 行定义.
Trains a locked machine on a set of indices. Error if machine is not locked
NOT IMPLEMENTED
indices | index vector (of locked features) that is used for training |
被 CKernelMachine , 以及 CMultitaskLinearMachine 重载.
|
protectedvirtual |
|
protectedvirtualinherited |
returns whether machine require labels for training
被 COnlineLinearMachine, CHierarchical, CLinearLatentMachine, CVwConditionalProbabilityTree, CConditionalProbabilityTree , 以及 CLibSVMOneClass 重载.
|
virtual |
Applies the network as a feature transformation
Forward-propagates the data through the network and returns the activations of the last layer
data | Input features |
被 CAutoencoder , 以及 CDeepAutoencoder 重载.
在文件 NeuralNetwork.cpp 第 222 行定义.
|
inherited |
unset generic type
this has to be called in classes specializing a template class
在文件 SGObject.cpp 第 303 行定义.
|
virtualinherited |
Updates the hash of current parameter combination
在文件 SGObject.cpp 第 248 行定义.
|
friend |
在文件 NeuralNetwork.h 第 112 行定义.
float64_t dropout_hidden |
Probabilty that a hidden layer neuron will be dropped out When using this, the recommended value is 0.5
default value 0.0 (no dropout)
For more details on dropout, see paper [Hinton, 2012]
在文件 NeuralNetwork.h 第 375 行定义.
float64_t dropout_input |
Probabilty that a input layer neuron will be dropped out When using this, a good value might be 0.2
default value 0.0 (no dropout)
For more details on dropout, see this paper [Hinton, 2012]
在文件 NeuralNetwork.h 第 385 行定义.
float64_t epsilon |
convergence criteria training stops when (E'- E)/E < epsilon where E is the error at the current iterations and E' is the error at the previous iteration default value is 1.0e-5
在文件 NeuralNetwork.h 第 400 行定义.
float64_t gd_error_damping_coeff |
Used to damp the error fluctuations when stochastic gradient descent is used. damping is done according to: error_damped(i) = c*error(i) + (1-c)*error_damped(i-1) where c is the damping coefficient
If -1, the damping coefficient is automatically computed according to: c = 0.99*gd_mini_batch_size/training_set_size + 1e-2;
default value is -1
在文件 NeuralNetwork.h 第 444 行定义.
float64_t gd_learning_rate |
gradient descent learning rate, defualt value 0.1
在文件 NeuralNetwork.h 第 415 行定义.
float64_t gd_learning_rate_decay |
gradient descent learning rate decay learning rate is updated at each iteration i according to: alpha(i)=decay*alpha(i-1) default value is 1.0 (no decay)
在文件 NeuralNetwork.h 第 422 行定义.
int32_t gd_mini_batch_size |
size of the mini-batch used during gradient descent training, if 0 full-batch training is performed default value is 0
在文件 NeuralNetwork.h 第 412 行定义.
float64_t gd_momentum |
gradient descent momentum multiplier
default value is 0.9
For more details on momentum, see this paper [Sutskever, 2013]
在文件 NeuralNetwork.h 第 432 行定义.
|
inherited |
io
在文件 SGObject.h 第 369 行定义.
float64_t l1_coefficient |
L1 Regularization coeff, default value is 0.0
在文件 NeuralNetwork.h 第 365 行定义.
float64_t l2_coefficient |
L2 Regularization coeff, default value is 0.0
在文件 NeuralNetwork.h 第 362 行定义.
|
protected |
Describes the connections in the network: if there's a connection from layer i to layer j then m_adj_matrix(i,j) = 1.
在文件 NeuralNetwork.h 第 458 行定义.
|
protected |
number of train/test cases the network is expected to deal with. Default value is 1
在文件 NeuralNetwork.h 第 480 行定义.
|
inherited |
parameters wrt which we can compute gradients
在文件 SGObject.h 第 384 行定义.
|
inherited |
Hash of parameter values
在文件 SGObject.h 第 387 行定义.
|
protected |
offsets specifying where each layer's parameters and parameter gradients are stored, i.e layer i's parameters are stored at m_params + m_index_offsets[i]
在文件 NeuralNetwork.h 第 475 行定义.
|
protected |
True if the network is currently being trained initial value is false
在文件 NeuralNetwork.h 第 485 行定义.
|
protected |
network's layers
在文件 NeuralNetwork.h 第 453 行定义.
|
inherited |
model selection parameters
在文件 SGObject.h 第 381 行定义.
|
protected |
number of neurons in the input layer
在文件 NeuralNetwork.h 第 447 行定义.
|
protected |
number of layer
在文件 NeuralNetwork.h 第 450 行定义.
|
protected |
Array that specifies which parameters are to be regularized. This is used to turn off regularization for bias parameters
在文件 NeuralNetwork.h 第 469 行定义.
|
inherited |
parameters
在文件 SGObject.h 第 378 行定义.
array where all the parameters of the network are stored
在文件 NeuralNetwork.h 第 464 行定义.
|
protectedinherited |
|
protectedinherited |
|
protected |
total number of parameters in the network
在文件 NeuralNetwork.h 第 461 行定义.
float64_t max_norm |
Maximum allowable L2 norm for a neurons weights When using this, a good value might be 15
default value -1 (max-norm regularization disabled)
在文件 NeuralNetwork.h 第 392 行定义.
int32_t max_num_epochs |
maximum number of iterations over the training set. If 0, training will continue until convergence. defualt value is 0
在文件 NeuralNetwork.h 第 406 行定义.
ENNOptimizationMethod optimization_method |
Optimization method, default is NNOM_LBFGS
在文件 NeuralNetwork.h 第 359 行定义.
|
inherited |
parallel
在文件 SGObject.h 第 372 行定义.
|
inherited |
version
在文件 SGObject.h 第 375 行定义.