Implementation:Tensorflow Tfjs Losses Test

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

Overview

This test suite validates all loss function implementations in the TensorFlow.js Layers API. Loss functions measure the discrepancy between predicted and true values during model training. The tested functions include regression losses (MSE, MAE, MAPE, MSLE), hinge losses (squared hinge, hinge, categorical hinge), cross-entropy losses (categorical, sparse categorical, binary, with and without logits), and other losses (logcosh, Kullback-Leibler divergence, Poisson, cosine proximity). The suite also tests the loss function lookup utility and L2 normalization.

Code Reference

Source Location: tfjs-layers/src/losses_test.ts (443 lines)

Repository: GitHub

Test Describe Blocks

meanSquaredError - MSE for 1D and 2D inputs
meanAbsoluteError - MAE for 1D and 2D inputs
meanAbsolutePercentageError - MAPE for 1D and 2D inputs
meanSquaredLogarithmicError - MSLE for 2D inputs
squaredHinge - Squared hinge loss
hinge - Hinge loss
categoricalHinge - Categorical hinge for multi-class
logcosh - Log-cosh loss (smooth L1 approximation)
categoricalCrossentropy - Multi-class cross-entropy (two blocks: function and loss wrapper)
sparseCategoricalCrossentropy - Sparse categorical cross-entropy (two blocks)
sigmoidCrossEntropyWithLogits - Binary cross-entropy from logits
binaryCrossentropy - Binary cross-entropy
kullbackLeiblerDivergence - KL divergence
poisson - Poisson loss
cosineProximity - Cosine similarity loss
losses get - Loss function lookup by string name
l2Normalize - L2 normalization utility

I/O Contract

Inputs to tests:

True label tensors (yTrue) and predicted tensors (yPred) of matching shapes
1D tensors for per-sample loss, 2D tensors for batched loss
Various numeric ranges including zeros, negatives, and boundary values

Expected outputs/assertions:

Loss values match hand-computed formulas (e.g., MSE: (1^2 + 2^2 + 3^2) / 3)
2D inputs produce per-sample loss vectors
Cross-entropy losses handle edge cases (predictions near 0 or 1)
Loss lookup returns correct function for string identifiers
L2 normalization produces unit-length vectors along specified axis

Usage Example

describeMathCPUAndGPU('meanSquaredError', () => {
  it('1D', () => {
    const yTrue = tfc.zeros([3]);
    const yPred = tensor1d([1, 2, 3]);
    const expectedVal = scalar((1 * 1 + 2 * 2 + 3 * 3) / 3);
    const result = losses.meanSquaredError(yTrue, yPred);
    expectTensorsClose(result, expectedVal);
  });
  it('2D', () => {
    const yTrue = tfc.zeros([2, 2]);
    const yPred = tensor2d([[1, 2], [3, 4]], [2, 2]);
    const expectedVal = tensor1d([(1 + 4) / 2, (9 + 16) / 2]);
    const result = losses.meanSquaredError(yTrue, yPred);
    expectTensorsClose(result, expectedVal);
  });
});

Test Coverage Summary

Category	Count	Details
Regression Losses	8	MSE, MAE, MAPE, MSLE (1D and 2D)
Hinge Losses	3	Hinge, squared hinge, categorical hinge
Cross-Entropy	8	Categorical, sparse categorical, binary, sigmoid with logits
Other Losses	4	Logcosh, KL divergence, Poisson, cosine proximity
Utilities	2	losses.get lookup, l2Normalize
Test Environment	Mixed	CPU, GPU, WebGL2

Related Pages

Page Connections

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

Principle

Implementation

Heuristic

Environment