Implementation:Tensorflow Tfjs Constraints

Knowledge Sources	Tensorflow_Tfjs TensorFlow.js API
Domains	Deep_Learning, Layers_API, Constraints
Last Updated	2026-02-10 06:00 GMT

Overview

This module implements weight constraint classes for TensorFlow.js Layers, ported from Keras constraints.py. Constraints are applied to weight tensors during training to enforce properties such as maximum norm, unit norm, non-negativity, or bounded norms. The module provides the abstract Constraint base class and four concrete implementations: MaxNorm, UnitNorm, NonNeg, and MinMaxNorm. It also includes serialization/deserialization utilities and a registry for resolving constraint identifiers.

Code Reference

Source Location

tfjs-layers/src/constraints.ts (GitHub)

Key Imports

import * as tfc from '@tensorflow/tfjs-core';
import {serialization, Tensor, tidy} from '@tensorflow/tfjs-core';
import {epsilon} from './backend/common';
import {deserializeKerasObject, serializeKerasObject} from './utils/generic_utils';

Classes

Constraint (Abstract Base)

export abstract class Constraint extends serialization.Serializable {
  abstract apply(w: Tensor): Tensor;
  getConfig(): serialization.ConfigDict;
}

MaxNorm

Constrains weights so that the L2 norm of each weight vector does not exceed maxValue along the specified axis.

export class MaxNorm extends Constraint {
  static readonly className = 'MaxNorm';
  constructor(args: MaxNormArgs);  // maxValue?: number (default 2), axis?: number (default 0)
  apply(w: Tensor): Tensor;
}

UnitNorm

Constrains weights to have unit L2 norm along the specified axis.

export class UnitNorm extends Constraint {
  static readonly className = 'UnitNorm';
  constructor(args: UnitNormArgs);  // axis?: number (default 0)
  apply(w: Tensor): Tensor;
}

NonNeg

Constrains weights to be non-negative by applying ReLU.

export class NonNeg extends Constraint {
  static readonly className = 'NonNeg';
  apply(w: Tensor): Tensor;  // returns tfc.relu(w)
}

MinMaxNorm

Constrains weights so that the L2 norm falls between minValue and maxValue, with a configurable enforcement rate.

export class MinMaxNorm extends Constraint {
  static readonly className = 'MinMaxNorm';
  constructor(args: MinMaxNormArgs);
  // minValue?: number (default 0), maxValue?: number (default 1),
  // rate?: number (default 1), axis?: number (default 0)
  apply(w: Tensor): Tensor;
}

Utility Functions

export function serializeConstraint(constraint: Constraint): serialization.ConfigDictValue;
export function deserializeConstraint(config: serialization.ConfigDict,
    customObjects?: serialization.ConfigDict): Constraint;
export function getConstraint(identifier: ConstraintIdentifier |
    serialization.ConfigDict | Constraint): Constraint;

The getConstraint function resolves string identifiers ('maxNorm', 'minMaxNorm', 'nonNeg', 'unitNorm'), config dicts, or existing Constraint instances.

I/O Contract

Method	Input	Output
Constraint.apply	Weight Tensor	Constrained weight Tensor (same shape)
getConstraint	String identifier, config dict, or Constraint	Constraint instance

Usage Example

import * as tf from '@tensorflow/tfjs';

const model = tf.sequential();
model.add(tf.layers.dense({
  units: 64,
  kernelConstraint: tf.constraints.maxNorm({maxValue: 2, axis: 0}),
  inputShape: [128]
}));

Related Pages

• Tensorflow_Tfjs_Exports_Constraints - Public API factory functions for constraints
• Tensorflow_Tfjs_Regularizers - Weight regularizers (complementary to constraints)
• Tensorflow_Tfjs_Exports_Initializers - Weight initializer factory functions