Implementation:Tensorflow Tfjs Pooling Test

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

Overview

This test suite validates all pooling layer implementations in TensorFlow.js Layers. It covers low-level pooling functions (pool2d, pool3d) and high-level Keras-style pooling layers for 1D, 2D, and 3D inputs, as well as global pooling layers. Tests enumerate combinations of pool modes (max, avg), data formats (channelsFirst, channelsLast), strides, padding modes (same, valid), and pool sizes. Both symbolic (shape inference) and tensor (numerical correctness) tests are included.

Code Reference

Source Location: tfjs-layers/src/layers/pooling_test.ts (927 lines)

Repository: GitHub

Test Describe Blocks

pool2d - Low-level 2D pooling function (4x4, 5x5 inputs; max/avg; stride 1/2; same/valid)
pool3d - Low-level 3D pooling function
Pooling Layers 1D: Symbolic - MaxPooling1D and AveragePooling1D shape inference
Pooling Layers 1D: Tensor - 1D pooling numerical correctness
Pooling Layers 2D: Symbolic - MaxPooling2D and AveragePooling2D shape inference
Pooling Layers 2D: Tensor - 2D pooling numerical correctness
Pooling Layers 3D: Symbolic - MaxPooling3D and AveragePooling3D shape inference
Pooling Layers 3D: Tensor - 3D pooling numerical correctness
1D Global pooling Layers: Symbolic - GlobalMaxPooling1D and GlobalAveragePooling1D shape inference
1D Global Pooling Layers: Tensor - 1D global pooling numerical correctness
2D Global pooling Layers: Symbolic - GlobalMaxPooling2D and GlobalAveragePooling2D shape inference
2D Global Pooling Layers: Tensor - 2D global pooling numerical correctness

I/O Contract

Inputs to tests:

3D tensors for 1D pooling: [batch, steps, features]
4D tensors for 2D pooling: [batch, height, width, channels] or NCHW format
5D tensors for 3D pooling: [batch, depth, height, width, channels]
Pool configuration: poolSize, strides, padding, dataFormat, poolMode (max/avg)

Expected outputs/assertions:

Max pooling selects maximum value in each pool window
Average pooling computes mean of values in each pool window
Output spatial dimensions follow convolution output length formula
Padding mode 'same' preserves spatial dimensions, 'valid' reduces them
Global pooling reduces spatial dimensions to 1
Data format correctly handles channelsFirst vs channelsLast

Usage Example

describeMathCPUAndGPU('pool2d', () => {
  const x4by4Data = [[[
    [10, 30, 50, 70], [20, 40, 60, 80],
    [-10, -30, -50, -70], [-20, -40, -60, -80]
  ]]];

  it('4x4, stride=2, same, channelsFirst, max', () => {
    const x = tensor4d(x4by4Data, [1, 1, 4, 4]);
    const yExpected = tensor4d([[[[40, 80], [-10, -50]]]], [1, 1, 2, 2]);
    const y = pool2d(x, [2, 2], [2, 2], 'same', 'channelsFirst', 'max');
    expectTensorsClose(y, yExpected);
  });
});

Test Coverage Summary

Category	Count	Details
pool2d function	20+	Max/avg, stride 1/2, same/valid, data formats
pool3d function	10+	3D pooling across configurations
1D Pooling Layers	10+	Symbolic and tensor tests
2D Pooling Layers	15+	Symbolic and tensor tests
3D Pooling Layers	10+	Symbolic and tensor tests
Global Pooling	10+	1D and 2D global max/avg pooling
Test Environment	CPU and GPU	`describeMathCPUAndGPU`

Related Pages

Page Connections

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

Principle

Implementation

Heuristic

Environment