Implementation:Tensorflow Tfjs Recurrent 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 (3198 lines), providing exhaustive validation of all recurrent layer implementations in TensorFlow.js Layers. It covers the low-level rnn() function, the RNN base layer, SimpleRNN, GRU, and LSTM layers (including their cell variants), and StackedRNNCells for multi-layer recurrent networks. Each RNN type is tested at symbolic (shape inference), tensor (numerical correctness against Python Keras reference), and serialization levels. Tests cover return sequences, return state, go backwards, masking, dropout, recurrent dropout, stateful mode, bidirectional compatibility, and various activation/initializer/regularizer/constraint configurations.

Code Reference

Source Location: tfjs-layers/src/layers/recurrent_test.ts (3198 lines)

Repository: GitHub

Test Describe Blocks

rnn - Low-level RNN step function (3D inputs, 1 state, 2 states, goBackwards, masking)
RNN-Layer - RNN base layer symbolic tests (cell wrapping, stateSpec)
RNN-Layer-Math - RNN base layer tensor computation
SimpleRNN Symbolic - SimpleRNN shape inference (returnSequences, returnState)
SimpleRNN Tensor - SimpleRNN numerical correctness against reference values
SimpleRNN Serialization - Config getConfig/fromConfig round-trips
GRU Symbolic - GRU shape inference
GRU Tensor - GRU numerical correctness (hardSigmoid, resetAfter, goBackwards, masking, dropout)
GRU-deserialization - GRU model JSON deserialization
GRU Serialization - GRU config round-trips with initializers, regularizers, constraints
LSTM Symbolic - LSTM shape inference (returnSequences, returnState)
LSTM Tensor - LSTM numerical correctness (hardSigmoid, goBackwards, masking, stateful, dropout, initialState)
LSTM-deserialization - LSTM model JSON deserialization from Python Keras
LSTM Serialization - LSTM config round-trips
StackedRNNCells Symbolic - Multi-layer RNN shape inference
Stacked RNN serialization - Stacked cell config round-trips
StackedRNNCells Tensor - Multi-layer RNN numerical computation

I/O Contract

Inputs to tests:

3D input tensors: [batch, timesteps, features]
Initial state tensors for each recurrent cell
RNN configuration: units, activation, returnSequences, returnState, goBackwards, unroll, stateful
Dropout and recurrent dropout rates
Masks for variable-length sequence support

Expected outputs/assertions:

Last output, per-step outputs, and new state tensors have correct shapes
Numerical values match Python Keras reference values (verified to multiple decimal places)
goBackwards reverses sequence processing order
Masking correctly ignores masked timesteps
Stateful RNNs maintain state across batches
Serialization preserves all configuration including nested cell configs

Usage Example

describeMathCPUAndGPU('rnn', () => {
  it('Simple step function: 3D inputs, 1 state', () => {
    const inputs = tensor3d(
        [[[1, 2], [3, 4], [5, 6]], [[10, 20], [30, 40], [50, 60]]],
        [2, 3, 2]);
    const initialStates = [tfc.zeros([2, 4])];
    const rnnOutputs = rnn(
        rnnStepForTest, inputs, initialStates,
        false, null, null, false, true);
    const lastOutput = rnnOutputs[0];
    expectTensorsClose(lastOutput, tensor2d(
        [[-57.75, -57.75, -57.75, -57.75],
         [-57.75, -57.75, -57.75, -57.75]], [2, 4]));
  });
});

Test Coverage Summary

Category	Count	Details
rnn() function	10+	Step functions, states, backwards, masking
SimpleRNN	30+	Symbolic, tensor, serialization
GRU	40+	Symbolic, tensor, deserialization, serialization, resetAfter
LSTM	50+	Symbolic, tensor, deserialization, serialization, stateful
StackedRNNCells	10+	Multi-layer symbolic, tensor, serialization
Test Environment	Mixed	CPU, GPU, WebGL2

Related Pages

Page Connections

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

Principle

Implementation

Heuristic

Environment