Implementation:Tensorflow Tfjs Initializers Test

Knowledge Sources	Tensorflow_Tfjs
Domains	Testing, Layers_API
Last Updated	2026-02-10 06:00 GMT

Overview

This test suite validates all weight initializer implementations in TensorFlow.js Layers. Initializers are responsible for setting the initial values of layer weights before training. The suite tests deterministic initializers (Zeros, Ones, Constant, Identity), random initializers (RandomUniform, RandomNormal, TruncatedNormal), variance-scaled initializers (GlorotUniform, GlorotNormal, HeNormal, HeUniform, LecunNormal, LecunUniform), and the Orthogonal initializer. Tests verify output shapes, dtypes, value ranges, statistical properties, serialization round-trips, and memory safety.

Code Reference

Source Location: tfjs-layers/src/initializers_test.ts (851 lines)

Repository: GitHub

Test Describe Blocks

Zeros initializer - All-zeros initialization (1D, 2D, case-insensitive naming)
Ones initializer - All-ones initialization
Constant initializer - Constant value initialization from config dict and builder
Identity initializer - Identity matrix initialization
RandomUniform initializer - Uniform random values within range
RandomNormal initializer - Normal distribution random values
HeNormal initializer - He normal variance scaling
HeUniform initializer - He uniform variance scaling
LecunNormal initializer - Lecun normal variance scaling
LeCunUniform initializer - Lecun uniform variance scaling
TruncatedNormal initializer - Truncated normal distribution
Glorot uniform initializer - Xavier/Glorot uniform
VarianceScaling initializer - General variance scaling with configurable fan mode and distribution
Glorot normal initializer - Xavier/Glorot normal
initializers.get - Initializer lookup by string name
Invalid initializer identifier - Error handling for unknown initializers
checkFanMode - Validation of fan mode values (fanIn, fanOut, fanAvg)
checkDistribution - Validation of distribution values (normal, uniform, truncatedNormal)
Orthogonal Initializer - Orthogonal matrix initialization

I/O Contract

Inputs to tests:

Shape arrays (e.g., [3], [2, 2], [4, 6])
Data types (float32)
Initializer configuration dicts and string identifiers
Seed values for reproducible random initializations

Expected outputs/assertions:

Output tensors have correct shape and dtype
Deterministic initializers produce exact values (all zeros, all ones, constant, identity)
Random initializers produce values within expected statistical ranges (mean, stddev, min, max)
Variance scaling respects fan-in/fan-out calculations
Orthogonal initializer produces matrices with orthogonal columns
Memory leak checks via expectNoLeakedTensors

Usage Example

describeMathCPU('Zeros initializer', () => {
  it('1D', () => {
    const init = getInitializer('zeros');
    const weights = init.apply([3], 'float32');
    expect(weights.shape).toEqual([3]);
    expect(weights.dtype).toEqual('float32');
    expect(weights.dataSync()).toEqual(new Float32Array([0, 0, 0]));
  });
  it('Does not leak', () => {
    expectNoLeakedTensors(() => getInitializer('zeros').apply([3]), 1);
  });
});

Test Coverage Summary

Category	Count	Details
Deterministic Initializers	4	Zeros, Ones, Constant, Identity
Random Initializers	3	RandomUniform, RandomNormal, TruncatedNormal
Variance Scaling	7	Glorot (uniform/normal), He (normal/uniform), Lecun (normal/uniform), general VarianceScaling
Orthogonal	1	Matrix orthogonality tests
Utility	4	get, checkFanMode, checkDistribution, invalid identifier
Memory Leak Tests	Per initializer	`expectNoLeakedTensors`
Test Environment	CPU (mostly)	`describeMathCPU`, `describeMathCPUAndWebGL2` for Orthogonal

Related Pages

Page Connections

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

Principle

Implementation

Heuristic

Environment