Implementation:Tensorflow Tfjs Backend Operations Test

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

Overview

This test suite validates the TensorFlow.js core backend operations used by the Layers API. It covers fundamental tensor manipulation operations including type casting, reshaping, slicing, concatenation, mathematical operations (dot product, sign, square, pow), dropout, bias addition, and activation functions at the backend level. The tests verify correct behavior for backend configuration, dtype handling, parameter counting, and tensor memory management.

Code Reference

Source Location: tfjs-layers/src/backend/tfjs_backend_test.ts (895 lines)

Repository: GitHub

Test Describe Blocks

TensorMath - Backend setting and getting
dtype - Data type resolution for Tensor and SymbolicTensor
countParams - Parameter counting across tensor dimensions
cast - Type casting between float32, int32, and bool
expandDims - Dimension expansion operations
Repeat - Tensor repetition along axes
Flatten - Tensor flattening operations
batchFlatten - Batch-aware flattening
sliceAlongFirstAxis - Slicing along first dimension
sliceAlongLastAxis - Slicing along last dimension
sliceAlongAxis - General axis slicing
concatenate - Tensor concatenation along axes
concatAlongFirstAxis - Concatenation along first axis
tile - Tensor tiling operations
randomNormal - Random normal tensor generation
dot - Dot product operations
sign - Element-wise sign function
OneHot - One-hot encoding
Gather - Tensor gathering by indices
Square - Element-wise squaring
Pow - Element-wise power operations
softsign - Softsign activation at backend level
dropout - Dropout implementation with rate and noise shape
biasAdd - Bias addition with data format support
elu - ELU activation at backend level
Sigmoid - Sigmoid at backend level
hardSigmoid - Hard sigmoid at backend level
inTrainPhase - Training phase branching

I/O Contract

Inputs to tests:

Tensors of various ranks (scalar through 6D) with known numeric values
Backend identifiers (e.g., 'cpu', 'webgl')
Data types (float32, int32, bool)
Configuration parameters for dropout rates, noise shapes, data formats

Expected outputs/assertions:

Correct output dtype and shape after operations
Numerically accurate results matching hand-computed expected values
Proper parameter counts for tensors of different ranks
Memory safety verified through expectNoLeakedTensors

Usage Example

describeMathCPU('countParams', () => {
  it('Tensor2D', () => {
    const x = zeros([3, 2]);
    expect(K.countParams(x)).toEqual(6);
    expect(K.countParams(new LayerVariable(x).read())).toEqual(6);
  });
});

describeMathCPUAndGPU('cast', () => {
  it('float32 to int32', () => {
    const x = tensor2d([[-1.1, -1.6], [1.1, 2.2], [3.6, 4.7]], [3, 2], 'float32');
    const y = K.cast(x, 'int32');
    expect(y.dtype).toEqual('int32');
    expect(Array.from(y.dataSync())).toEqual([-1, -1, 1, 2, 3, 4]);
  });
});

Test Coverage Summary

Category	Count	Details
Backend Configuration	2	TensorMath, dtype
Tensor Manipulation	10	expand, repeat, flatten, slice, concat, tile, gather
Math Operations	6	dot, sign, square, pow, oneHot, randomNormal
Activations	5	softsign, elu, sigmoid, hardSigmoid, dropout
Utility	2	biasAdd, inTrainPhase
Test Environment	CPU and GPU	Mix of `describeMathCPU` and `describeMathCPUAndGPU`

Related Pages

Page Connections

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

Principle

Implementation

Heuristic

Environment