Implementation:Tensorflow Tfjs Core Layers Test

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

Overview

This test suite validates the core layer types in TensorFlow.js Layers: Dropout, SpatialDropout1D, Dense, Flatten, Activation, RepeatVector, Reshape, Permute, and Masking. These are the fundamental building blocks of most neural networks. Each layer is tested at both symbolic (shape inference, configuration) and tensor (numerical correctness) levels. The suite covers dropout behavior during training vs inference, Dense layer with various activations and initializers, Flatten for reshaping multi-dimensional tensors, and utility layers like Reshape and Permute.

Code Reference

Source Location: tfjs-layers/src/layers/core_test.ts (871 lines)

Repository: GitHub

Test Describe Blocks

Dropout Layer: Symbolic - Shape inference with various dropout rates and input shapes
Dropout Layer - Tensor-level dropout behavior (training vs inference, noise shapes, scaling)
SpatialDropout1D Layer - Spatial dropout for 1D feature maps
Dense Layer: Symbolic - Shape inference, activations, regularizers, constraints, initializers
Dense Layer: Tensor - Numerical output verification with known kernels and biases
Flatten Layer: Symbolic - Shape inference for various input shapes and data formats
Flatten Layer: Tensor - Tensor reshaping correctness
Activation Layer: Tensor - Standalone activation layer with various activation functions
RepeatVector Layer: Symbolic - Shape inference for vector repetition
RepeatVector Layer: Tensor - Numerical repetition correctness
Reshape Layer: Symbolic - Shape inference for reshape operations
Reshape Layer: Tensor - Tensor reshaping correctness
Permute Layer: Symbolic - Shape inference for dimension permutation
Masking Layer: Symbolic - Shape inference for masking layers
Permute Layer: Tensor - Dimension permutation correctness
Masking Layer: Tensor - Masking behavior with custom mask values

I/O Contract

Inputs to tests:

SymbolicTensors with known shapes for symbolic tests
Concrete tensors (ones, zeros, random) for tensor tests
Layer configurations: units, activation, dropout rate, noise shape, target shape, dims
Training flag to control dropout behavior

Expected outputs/assertions:

Dropout outputs equal inputs during inference (training=false)
During training, dropout zeros out elements and scales remaining by 1/(1-rate)
Dense layer produces correct matrix multiplication results
Flatten correctly reshapes to 2D preserving batch dimension
Reshape and Permute produce correct output shapes and values
Masking correctly masks timesteps matching the mask value

Usage Example

describe('Dropout Layer: Symbolic', () => {
  for (const rate of [0, 0.5]) {
    for (const symbolicInput of symbolicInputs) {
      it(`dropoutRate=${rate}; input shape=${JSON.stringify(symbolicInput.shape)}`, () => {
        const dropoutLayer = tfl.layers.dropout({rate});
        const output = dropoutLayer.apply(symbolicInput) as tfl.SymbolicTensor;
        expect(output.dtype).toEqual(symbolicInput.dtype);
        expect(output.shape).toEqual(symbolicInput.shape);
      });
    }
  }
});

Test Coverage Summary

Category	Count	Details
Dropout	15+	Symbolic and tensor tests, noise shapes, training/inference
Dense	20+	Activations, initializers, regularizers, constraints, serialization
Flatten	10+	Various input shapes, data formats
Activation	5+	Standalone activation layer
Reshape/Permute	10+	Shape transformation layers
Masking	5+	Mask value and output masking
Test Environment	Mixed	CPU, GPU, WebGL2

Related Pages

Page Connections

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

Principle

Implementation

Heuristic

Environment