Implementation:Tensorflow Tfjs Training Test

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

Overview

This is one of the largest test suites (2991 lines), providing comprehensive validation of the core training engine in TensorFlow.js Layers. It tests the LayersModel.fit(), LayersModel.predict(), LayersModel.evaluate(), LayersModel.trainOnBatch(), and LayersModel.execute() methods. The suite also validates utility functions for data standardization (isDataTensor, isDataArray, isDataDict, standardizeInputData), array length checking, metric collection, batch slicing, and memory leak prevention during training loops.

Code Reference

Source Location: tfjs-layers/src/engine/training_test.ts (2991 lines)

Repository: GitHub

Test Describe Blocks

isDataTensor - Type guard for single tensor inputs
isDataArray - Type guard for tensor array inputs
isDataDict - Type guard for named tensor dict inputs
standardizeInputData - Normalizing input data to standard array format
checkArrayLengths - Validating matching lengths of input/output arrays
collectMetrics - Gathering metrics for single and multi-output models
sliceArraysByIndices - Extracting tensor slices by index arrays
makeBatches - Creating batch boundaries from dataset size
LayersModel.predict - Model inference with various configurations
LayersModel.fit long - Extended training tests
LayersModel.fit - Core training loop (largest block: compilation, metrics, callbacks, validation split, class weights, sample weights, regularizers, multi-output)
LayersModel.fit: No memory leak - Memory leak prevention during training
LayersModel.fit: yieldEvery - Yield control during training
LayersModel.trainOnBatch - Single-batch training
LayersModel.evaluate - Model evaluation
LayersModel trainable setter and getter - Trainability toggling
LayersModel.execute - Direct execution of model subgraphs

I/O Contract

Inputs to tests:

Tensor inputs in three forms: single tensor, array of tensors, or named tensor dict
Model architectures: sequential and functional with Dense, Reshape layers
Training config: loss functions, optimizers (SGD, Adam), metrics, epochs, batch size, validation split
Regularizers (L1, L2, L1L2) on kernel and bias

Expected outputs/assertions:

Data type guards correctly identify tensor, array, and dict formats
standardizeInputData normalizes all input forms to arrays, throwing on mismatches
Training reduces loss over epochs
Memory leak checks: tensor count stable before/after fit()
Metric values match reference Python Keras outputs
trainOnBatch returns correct loss/metric values
Evaluate returns correct metrics for test data

Usage Example

describeMathCPU('isDataTensor', () => {
  const x = tfc.tensor2d([[3.14]]);
  it('Positive case', () => {
    expect(isDataTensor(x)).toEqual(true);
  });
  it('Negative cases', () => {
    expect(isDataTensor([x, x])).toEqual(false);
    expect(isDataTensor({'Pie': x})).toEqual(false);
  });
});

describeMathCPU('standardizeInputData', () => {
  it('Singleton Tensor, Array of one name', () => {
    const outputs = standardizeInputData(getX(), ['Foo']);
    expect(outputs.length).toEqual(1);
    expectTensorsClose(outputs[0], getX());
  });
});

Test Coverage Summary

Category	Count	Details
Data Utilities	20+	Type guards, standardization, batching
Model.predict	10+	Single/multi output, batch predictions
Model.fit	50+	Loss, metrics, callbacks, validation, regularizers
Memory Leaks	15+	Tensor count verification across training
Model.trainOnBatch	5+	Single batch training
Model.evaluate	5+	Model evaluation metrics
Model.execute	5+	Subgraph execution
Test Environment	Mixed	CPU, GPU, WebGL2 variants

Related Pages

Page Connections

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

Principle

Implementation

Heuristic

Environment