Implementation:Tensorflow Tfjs Exports Initializers

Knowledge Sources	Tensorflow_Tfjs TensorFlow.js API
Domains	Deep_Learning, Layers_API, Initializers
Last Updated	2026-02-10 06:00 GMT

Overview

This module provides the public API factory functions for creating weight initializer instances in TensorFlow.js Layers. Initializers define the strategy for setting initial random or deterministic values for layer weights. These functions are exposed under the tf.initializers namespace and cover all standard initialization schemes including zeros, ones, constant, random distributions, variance-scaling methods (Glorot/He/LeCun), and orthogonal initialization.

Code Reference

Source Location

tfjs-layers/src/exports_initializers.ts (GitHub)

Imports

import {Constant, ConstantArgs, GlorotNormal, GlorotUniform, HeNormal, HeUniform,
        Identity, IdentityArgs, Initializer, LeCunNormal, LeCunUniform, Ones,
        Orthogonal, OrthogonalArgs, RandomNormal, RandomNormalArgs, RandomUniform,
        RandomUniformArgs, SeedOnlyInitializerArgs, TruncatedNormal,
        TruncatedNormalArgs, VarianceScaling, VarianceScalingArgs, Zeros}
    from './initializers';

Factory Functions

Function	Args Type	Description
`zeros()`	(none)	All weights initialized to 0
`ones()`	(none)	All weights initialized to 1
`constant(args)`	`ConstantArgs`	Weights initialized to a given constant value
`randomUniform(args)`	`RandomUniformArgs`	Uniform distribution between minval and maxval
`randomNormal(args)`	`RandomNormalArgs`	Normal distribution with given mean and stddev
`truncatedNormal(args)`	`TruncatedNormalArgs`	Truncated normal (values beyond 2 stddev re-drawn)
`identity(args)`	`IdentityArgs`	Identity matrix (square 2D only)
`varianceScaling(config)`	`VarianceScalingArgs`	Scale-adaptive initialization (configurable fan mode and distribution)
`glorotUniform(args)`	`SeedOnlyInitializerArgs`	Xavier uniform: limit = sqrt(6 / (fan_in + fan_out))
`glorotNormal(args)`	`SeedOnlyInitializerArgs`	Xavier normal: stddev = sqrt(2 / (fan_in + fan_out))
`heNormal(args)`	`SeedOnlyInitializerArgs`	He normal: stddev = sqrt(2 / fan_in)
`heUniform(args)`	`SeedOnlyInitializerArgs`	He uniform: limit = sqrt(6 / fan_in)
`leCunNormal(args)`	`SeedOnlyInitializerArgs`	LeCun normal: stddev = sqrt(1 / fan_in)
`leCunUniform(args)`	`SeedOnlyInitializerArgs`	LeCun uniform: limit = sqrt(3 / fan_in)
`orthogonal(args)`	`OrthogonalArgs`	Random orthogonal matrix

All functions return an Initializer instance.

I/O Contract

Input	Output
Initializer-specific config args	`Initializer` instance that generates weight tensors of a requested shape

Usage Example

import * as tf from '@tensorflow/tfjs';

const model = tf.sequential();
model.add(tf.layers.dense({
  units: 128,
  kernelInitializer: tf.initializers.glorotNormal({seed: 42}),
  biasInitializer: tf.initializers.zeros(),
  inputShape: [256]
}));

Related Pages

Tensorflow_Tfjs_Exports_Constraints - Factory functions for weight constraints
Tensorflow_Tfjs_Exports_Regularizers - Factory functions for weight regularizers
Tensorflow_Tfjs_Constraints - Constraint class implementations

Page Connections

Double-click a node to navigate. Hold to expand connections.

Principle

Implementation

Heuristic

Environment