Implementation:Tensorflow Tfjs Convolutional Test

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

Overview

This large test suite (1758 lines) validates all convolutional layer implementations in the TensorFlow.js Layers API. It covers 1D, 2D, and 3D convolutions including standard convolutions, transposed (deconvolution) layers, separable convolutions, cropping, and upsampling. Tests are organized by both low-level convolution functions (conv1dWithBias, conv2d, conv2dWithBiasActivation, conv3d, conv3dWithBias) and high-level Keras-style layer APIs. Each layer type is tested at both symbolic (shape inference) and tensor (numerical correctness) levels across data formats, padding modes, strides, and dilations.

Code Reference

Source Location: tfjs-layers/src/layers/convolutional_test.ts (1758 lines)

Repository: GitHub

Test Describe Blocks

conv1dWithBias - Low-level 1D convolution with bias across output channels, data formats, padding modes
conv1d - Low-level 1D convolution with strides and dilations
conv2d - Low-level 2D convolution across formats and padding
conv2dWithBiasActivation - 2D convolution with fused bias and activation
Conv2D Layers: Symbolic - 2D conv layer shape inference
Conv2D Layer: Tensor - 2D conv layer numerical correctness
conv3d - Low-level 3D convolution
Conv3D Layers: Symbolic - 3D conv layer shape inference
Conv3D Layer: Tensor - 3D conv layer numerical correctness
Conv2DTranspose: Symbolic - Transposed 2D conv shape inference
Conv2DTranspose: Tensor - Transposed 2D conv numerical correctness
Conv1D Layers: Symbolic - 1D conv layer shape inference
Conv1D Layer: Tensor - 1D conv layer numerical correctness
SeparableConv2D Layers: Symbolic - Depthwise separable 2D conv shape inference
SeparableConv2D Layer: Tensor - Depthwise separable 2D conv numerical correctness
Cropping2D Layer - Spatial cropping of 2D feature maps
UpSampling2D Layer: Symbolic - Upsampling shape inference (nearest, bilinear)
UpSampling2D Layer - Upsampling numerical correctness
Conv3DTranspose: Symbolic - Transposed 3D conv shape inference
Conv3DTranspose: Tensor - Transposed 3D conv numerical correctness

I/O Contract

Inputs to tests:

3D tensors for Conv1D: [batch, width, channels]
4D tensors for Conv2D: [batch, height, width, channels] or [batch, channels, height, width]
5D tensors for Conv3D: [batch, depth, height, width, channels]
Kernel tensors, bias tensors, and layer configurations
Parameters: outChannels, stride, dilations, padding (valid/same), dataFormat (channelsFirst/channelsLast)

Expected outputs/assertions:

Output tensor shapes match expected spatial dimensions after convolution
Numerical output values match hand-computed expected values (e.g., convolution of [10, 20, 40, 80] with kernel [1, -1] plus bias)
Symbolic shape inference produces correct output shapes
Transposed convolution reverses spatial dimension reduction
Separable convolution correctly decomposes into depthwise and pointwise stages

Usage Example

describeMathCPUAndGPU('conv1dWithBias', () => {
  it('outChannels=1, stride=1, valid, channelsLast', () => {
    let x: Tensor = tensor3d([10, 20, 40, 80], [1, 4, 1]);
    const kernel = tfc.transpose(
        tensor3d([1, -1], [1, 1, 2]), [2, 0, 1]);
    const bias = tensor1d([2.2]);
    const y = conv1dWithBias(x, kernel, bias, 1, 'valid', 'channelsLast');
    expectTensorsClose(y, tensor3d([-7.8, -17.8, -37.8], [1, 3, 1]));
  });
});

Test Coverage Summary

Category	Count	Details
Conv1D	30+	Low-level functions and layer API (symbolic + tensor)
Conv2D	40+	Standard, transposed, separable, with bias/activation
Conv3D	30+	Standard and transposed (symbolic + tensor)
Cropping2D	5+	Spatial cropping with various configurations
UpSampling2D	10+	Nearest and bilinear interpolation
Test Environment	Mixed	CPU, GPU, WebGL2

Related Pages

Page Connections

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

Principle

Implementation

Heuristic

Environment