model

type Layer ¶

type Layer struct {
	Name    string
	Dtype   string
	Weights *tensor.Dense
}

Layer keeps the Weights of a certain layer of the Neural Network the weights can be either the weights or bias indistinctly

type Model ¶

type Model struct {
	Name string

	// StateDict holds the layer names
	// and the layers of the model. Each
	// layer has a bias and a weight
	StateDict map[string]*Layer
	// contains filtered or unexported fields
}

Holds the Layers of the model

func NewModel(
	logger *zap.Logger,
	jobId string,
	task api.TrainRequest,
	layerNames []string,
	pool *redis.Pool) *Model

func (m *Model) Build() error

func (m *Model) Clear()

func (m *Model) Save() error

func (m *Model) Summary()

func (m *Model) Update(funcId int)

type ParallelSGD ¶

type ParallelSGD struct {
	// contains filtered or unexported fields
}

ParallelSGD simply averages the weights of the models trained independently. In this way, we get the freedom of using any optimizer in the functions.

Simply fetch all the model weights and average them

func MakeParallelSGD(logger *zap.Logger) ParallelSGD

func (psgd ParallelSGD) Average(m *Model, num int) error