Documentation
¶
Index ¶
Constants ¶
const ( // None represents the absence of a specific gradient descent optimization method. None int = iota // SGD represents the SGD gradient descent optimization method. SGD // AdaGrad represents the AdaGrad gradient descent optimization method. AdaGrad // Adam represents the Adam gradient descent optimization method. Adam // RAdam represents the RAdam gradient descent optimization method. RAdam // RMSProp represents the RMSProp gradient descent optimization method. RMSProp )
Variables ¶
This section is empty.
Functions ¶
Types ¶
type BatchScheduler ¶
type BatchScheduler interface {
// IncBatch beats the occurrence of a new batch.
IncBatch()
}
BatchScheduler is implemented by any value that has the IncBatch method.
type EpochScheduler ¶
type EpochScheduler interface {
// IncEpoch beats the occurrence of a new epoch.
IncEpoch()
}
EpochScheduler is implemented by any value that has the IncEpoch method.
type ExampleScheduler ¶
type ExampleScheduler interface {
// IncExample beats the occurrence of a new example.
IncExample()
}
ExampleScheduler is implemented by any value that has the IncExample method.
type GradientDescent ¶
type GradientDescent struct {
// contains filtered or unexported fields
}
GradientDescent implements Gradients Descent (GD) optimization.
func NewOptimizer ¶
func NewOptimizer(method Method, paramsIterator nn.ParamsGetter, opts ...Option) *GradientDescent
NewOptimizer returns a new GradientDescent optimizer. The gradient clipper can be set to nil.
func (*GradientDescent) IncBatch ¶
func (o *GradientDescent) IncBatch()
IncBatch beats the occurrence of a new batch.
func (*GradientDescent) IncEpoch ¶
func (o *GradientDescent) IncEpoch()
IncEpoch beats the occurrence of a new epoch.
func (*GradientDescent) IncExample ¶
func (o *GradientDescent) IncExample()
IncExample beats the occurrence of a new example.
func (*GradientDescent) Optimize ¶
func (o *GradientDescent) Optimize()
Optimize optimize the params, applying the optional gradient clipping. After the optimization the params have zero gradients.
type Method ¶
type Method interface {
// Label returns the enumeration-like value which identifies this gradient descent method.
Label() int
// Delta returns the difference between the current params and where the method wants it to be.
Delta(param nn.Param) mat.Matrix
// NewSupport returns a new support structure with the given dimensions.
NewSupport(r, c int) *nn.Payload
}
Method is implemented by any optimization method.
type MethodConfig ¶
type MethodConfig interface{}
MethodConfig is an empty interface implemented by the configuration structures of AdaGrad, Adam, RMSProp and SGD.
type Option ¶
type Option func(*GradientDescent)
Option allows to configure a new GradientDescent with your specific needs.
func ClipGradByNorm ¶
ClipGradByNorm is an option to clip the gradients during the training by norm.
func ClipGradByValue ¶
ClipGradByValue is an option to clip the gradients during the training between -value and +value.
func ConcurrentComputations ¶ added in v0.3.0
ConcurrentComputations sets the maximum number of concurrent computations handled by the GradientDescent for heavy tasks such as the params update steps. The value 1 corresponds to sequential execution.
Source Files
Directories