Class OptimizerBase

• Defined in File optimizers.h

Inheritance Relationships

Base Types

• public marian::TrainingObserver (Class TrainingObserver)

• public marian::ExponentialSmoothing (Class ExponentialSmoothing)

Derived Types

• public marian::Adagrad (Class Adagrad)

• public marian::Adam (Class Adam)

• public marian::Sgd (Class Sgd)

Class Documentation

class OptimizerBase : public marian::TrainingObserver, public marian::ExponentialSmoothing

Base class for optimizers.

Subclassed by marian::Adagrad, marian::Adam, marian::Sgd

Public Types

typedef std::function<void(size_t, const char *, const char *)> ScatterStateSetFunc

typedef std::function<io::Item(size_t)> GatherStateGetFunc

typedef std::function<void(const io::Item&, const ScatterStateSetFunc&)> ScatterStateFunc

typedef std::function<io::Item(const GatherStateGetFunc&)> GatherStateFunc

Public Functions

OptimizerBase(Ptr<Options> options)

virtual ~OptimizerBase()

float update(Ptr<ExpressionGraph> graph, size_t mbSize, float costScaleFactor = 1.f)

float update(Tensor params, Tensor grads, size_t mbSize, float costScaleFactor = 1.f)

virtual void init(TrainingState &state)

virtual void actAfterLoaded(TrainingState &state)

virtual void actAfterEpoch(TrainingState &state)

virtual void actAfterBatches(TrainingState &state)

virtual void actAfterStalled(TrainingState &state)

virtual void setParams(const std::vector<float> &params) = 0

void load(std::vector<io::Item> &items, const std::vector<Ptr<OptimizerBase>> &opts, const std::vector<Ptr<Backend>> &backends, const ScatterStateFunc &scatterFn, bool isMainProcess)

void save(std::vector<io::Item> &items, const std::vector<Ptr<OptimizerBase>> &opts, const GatherStateFunc &gatherFn, bool isMainProcess)

void swapWithSmoothed(Tensor params)

virtual std::vector<Tensor> getShards()

Protected Functions

virtual void updateImpl(Tensor params, Tensor grads, size_t actualMBSize) = 0

virtual void resetStats() = 0

Protected Attributes

Ptr<Options> options_

float eta_

size_t refMBWordsParam_ = {0}

size_t batchesSeen_ = {0}

bool normalizedGradient_ = {false}

Type optimizerType_ = {Type::float32}

bool castOptimizerType_ = {false}

Ptr<Clipper> clipper_

Ptr<TensorAllocator> baseAlloc_

Ptr<Allocator> alloc_

Tensor avg_

Tensor pm_

Tensor gd_