Implementation:Tensorflow Tfjs Variables Test

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

Overview

This test suite validates the LayerVariable class and related variable management functions in TensorFlow.js Layers. LayerVariables wrap TensorFlow.js tensors with additional metadata (name, dtype, trainability, shape) and are used to represent layer weights. The tests cover variable construction (with and without explicit names, name scopes, dtypes), reading and writing values, variable creation helpers (zerosVariable, onesVariable, zerosLike, onesLike, eye), variable updates, batch get/set operations, and gradient computation through variables.

Code Reference

Source Location: tfjs-layers/src/variables_test.ts (457 lines)

Repository: GitHub

Test Describe Blocks

Variable - Core LayerVariable tests:
- Constructor: no explicit name (auto-naming), explicit name, name with name scope
- Trainable property
- Null/undefined name handling
- Int32 and bool dtype support
- Read value, write value, dispose
- Constraint setting and retrieval
- Unique name generation across instances
Create Variable - Variable creation from concrete tensors
ZerosVariable - Creating variables initialized to zeros (various shapes, dtypes)
OnesVariable - Creating variables initialized to ones
ZerosLike - Creating zero-valued variables matching shape of another variable
OnesLike - Creating one-valued variables matching shape of another variable
eye (I-matrix builder) - Creating identity matrix variables
Variable update - Updating variable values with new tensors (element-wise operations)
batchGetValue - Batch retrieval of values from multiple variables
batchSetValue - Batch setting of values for multiple variables
gradients - Gradient computation through variable operations

I/O Contract

Inputs to tests:

Tensor values: zeros, ones, scalars, random uniform
Shape arrays for variable creation
Name strings and name scopes
Dtype specifications: float32, int32, bool
Gradient target expressions using variable operations

Expected outputs/assertions:

Variable names include class prefix or explicit name, with name scope path
Auto-generated names are unique across instances
Read returns correct tensor values
Write updates the internal tensor
Dtype is preserved correctly
zerosVariable/onesVariable create variables with correct values and dtypes
zerosLike/onesLike match the shape and dtype of the source
eye creates correct identity matrices
Gradient computation returns correct derivative values

Usage Example

describeMathCPU('Variable', () => {
  it('Variable constructor: no explicit name', () => {
    const v1 = new V.LayerVariable(zeros([2]));
    expect(v1.name.indexOf('Variable')).toEqual(0);
    expect(v1.dtype).toEqual('float32');
    expect(v1.shape).toEqual([2]);
    expect(v1.trainable).toEqual(true);
    expect(v1.read().dataSync()).toEqual(new Float32Array([0, 0]));
  });

  it('Variable constructor: explicit name with name scope', () => {
    let v1: V.LayerVariable;
    nameScope('barScope', () => {
      nameScope('bazScope', () => {
        v1 = new V.LayerVariable(scalar(0), undefined, 'foo');
      });
    });
    expect(v1.name.indexOf('barScope/bazScope/foo')).toEqual(0);
  });
});

Test Coverage Summary

Category	Count	Details
Constructor	8+	Auto-naming, explicit name, name scope, dtype, trainable
Read/Write	3+	Value access and mutation
Creation Helpers	15+	zeros, ones, zerosLike, onesLike, eye
Variable Update	3+	Element-wise update operations
Batch Operations	4+	batchGetValue, batchSetValue
Gradients	3+	Gradient computation through variables
Test Environment	Mixed	CPU, GPU

Related Pages

Page Connections

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

Principle

Implementation

Heuristic

Environment