Implementation:Tensorflow Tfjs Activations Test

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

Overview

This test suite validates the correctness of all activation functions implemented in the TensorFlow.js Layers API. Each activation function is tested across multiple tensor dimensionalities (1D, 2D, 3D) to ensure consistent behavior regardless of input shape. The suite also includes memory leak checks to verify that activations do not leak tensors during execution. The activation functions tested include linear, elu, selu, relu, relu6, sigmoid, hardSigmoid, softplus, softsign, tanh, softmax, logSoftmax, swish, mish, gelu, and geluNew.

Code Reference

Source Location: tfjs-layers/src/activations_test.ts (463 lines)

Repository: GitHub

Test Describe Blocks

linear activation - Identity activation returning input unchanged
elu activation - Exponential Linear Unit
selu activation - Scaled Exponential Linear Unit
relu activation - Rectified Linear Unit
relu6 activation - ReLU capped at 6
sigmoid activation - Logistic sigmoid function
hardSigmoid activation - Piecewise linear approximation of sigmoid
softplus activation - Smooth approximation of ReLU
softsign activation - Softsign activation
tanh activation - Hyperbolic tangent
softmax activation - Normalized exponential
logsoftmax activation - Log of softmax
swish activation - Self-gated activation
mish activation - Mish self-regularized activation
gelu activation - Gaussian Error Linear Unit
gelu_new activation - Approximate GELU variant

I/O Contract

Inputs to tests:

Float32Array or number array values including negative, zero, and positive numbers (e.g., [-1, 2, 0, 4, -5, 6])
Tensors of varying dimensions: 1D (tensor1d), 2D (tensor2d), 3D (tensor3d)

Expected outputs/assertions:

Each activation produces mathematically correct output values matching the activation function formula
Output tensor shape matches the input tensor shape
Memory leak tests verify zero (or minimal expected) leaked tensors via expectNoLeakedTensors

Usage Example

describeMathCPUAndGPU('elu activation', () => {
  const initVals = [-1, 2, 0, 4, -5, 6];
  const expectedVals = initVals.map(x => x < 0 ? Math.exp(x) - 1 : x);
  const elu = new Elu().apply;
  it('1D', () => {
    const initX = tensor1d(initVals);
    expectTensorsClose(elu(initX), tensor1d(expectedVals));
  });
  it('Does not leak', () => {
    const initX = tensor1d(initVals);
    expectNoLeakedTensors(() => elu(initX), 1);
  });
});

Test Coverage Summary

Category	Count	Details
Activation Functions	16	linear, elu, selu, relu, relu6, sigmoid, hardSigmoid, softplus, softsign, tanh, softmax, logSoftmax, swish, mish, gelu, geluNew
Dimensionality Tests	3 per activation	1D, 2D, 3D tensor inputs
Memory Leak Tests	1 per activation	Validates no tensor leaks
Test Environment	CPU and GPU	All tests run via `describeMathCPUAndGPU`

Related Pages

Page Connections

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

Principle

Implementation

Heuristic

Environment