Implementation:Tensorflow Tfjs Merge Test

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

Overview

This test suite validates the merge layers in TensorFlow.js Layers, which combine multiple input tensors into a single output. The tested layers include Add, Average, Multiply, Maximum, Minimum, Concatenate, and Dot. Tests cover symbolic shape inference (verifying output shapes from input shapes), functional API usage (calling layers directly with tensors), tensor-level numerical correctness, error handling for invalid inputs (single input, mismatched batch sizes), and serialization/deserialization. The Dot layer tests include various axes configurations and normalization options.

Code Reference

Source Location: tfjs-layers/src/layers/merge_test.ts (1011 lines)

Repository: GitHub

Test Describe Blocks

Merge Layers Except Concatenate: Symbolic - Shape inference for Add, Average, Multiply, Maximum, Minimum
Add-Functional - Functional API: symbolic and tensor usage, model prediction
Multiply-Functional - Element-wise multiplication functional API
Average-Functional - Element-wise averaging functional API
Maximum-Functional - Element-wise maximum functional API
Minimum-Functional - Element-wise minimum functional API
Concatenate-Functional - Concatenation along axes functional API
Concatenate Layer: Symbolic - Concatenation shape inference with axis configuration
Add Layer: Tensor - Numerical addition correctness (2D, 3D, broadcasting)
Multiply Layer: Tensor - Numerical multiplication correctness
Average Layer: Tensor - Numerical averaging correctness
Maximum Layer: Tensor - Numerical maximum correctness
Minimum Layer: Tensor - Numerical minimum correctness
Concatenate Layer: Tensor - Numerical concatenation correctness, masking support
Deserialize Merge Layers - Serialization round-trips for all merge layer types
Dot-Layer: Symbolic - Dot product shape inference with various axes and normalization
Dot-Layer: Tensor - Dot product numerical correctness

I/O Contract

Inputs to tests:

Multiple SymbolicTensors or concrete tensors (2-4 inputs per merge operation)
Layer configurations: axis for concatenation, axes and normalize for Dot
Various input shapes (2D, 3D) and number of inputs (2, 4)

Expected outputs/assertions:

Output shape matches input shape for element-wise operations (Add, Multiply, etc.)
Concatenate output shape grows along the specified axis
Numerical results match expected element-wise operations
Single input throws error ("at least 2 inputs")
Mismatched batch sizes throw error
Serialization preserves layer configuration

Usage Example

describeMathCPUAndGPU('Add-Functional', () => {
  it('Calling with tensors returns tensor', () => {
    const input1 = tensor2d([1, 2, 3, 4], [2, 2]);
    const input2 = tensor2d([10, 20, 30, 40], [2, 2]);
    const output = tfl.layers.add().apply([input1, input2]) as Tensor;
    expectTensorsClose(output, tensor2d([11, 22, 33, 44], [2, 2]));
  });
});

Test Coverage Summary

Category	Count	Details
Element-wise Merge	25+	Add, Multiply, Average, Maximum, Minimum
Concatenate	15+	Symbolic, tensor, masking, axis configuration
Dot	15+	Various axes, normalization, symbolic and tensor
Functional API	10+	Layer creation, apply with symbolic/concrete tensors
Serialization	10+	Round-trip for all merge layer types
Error Handling	3+	Single input, batch size mismatch
Test Environment	Mixed	CPU, GPU, WebGL2

Related Pages

Page Connections

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

Principle

Implementation

Heuristic

Environment