Implementation:Tensorflow Tfjs Activations

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

Overview

The activations.ts module defines the core activation function classes used throughout the TensorFlow.js Layers API. Each activation extends the abstract Activation base class (itself extending serialization.Serializable) and implements the apply(tensor, axis?) method. The module also provides serialization/deserialization utilities and a registry lookup function (getActivation) that resolves string identifiers to activation instances.

Supported activations: Elu, Selu, Relu, Relu6, Linear, Sigmoid, HardSigmoid, Softplus, Softsign, Tanh, Softmax, LogSoftmax, Gelu, GeluNew, Mish, and Swish. Each class is registered via serialization.registerClass() for model persistence.

Code Reference

Source Location

tfjs-layers/src/activations.ts (View on GitHub)

Key Signatures

export abstract class Activation extends serialization.Serializable {
  abstract apply(tensor: Tensor, axis?: number): Tensor;
  getConfig(): serialization.ConfigDict;
}

export function serializeActivation(activation: Activation): string;

export function deserializeActivation(
    config: serialization.ConfigDict,
    customObjects?: serialization.ConfigDict): Activation;

export function getActivation(
    identifier: ActivationIdentifier | serialization.ConfigDict | Activation
): Activation;

Import

import { Activation, getActivation, serializeActivation, deserializeActivation } from './activations';

Activation Classes

Class	className	Formula / Behavior
Elu	elu	alpha * (exp(x) - 1) for x < 0, else x
Selu	selu	Self-normalizing ELU (Klambauer et al., 2017)
Relu	relu	max(0, x)
Relu6	relu6	min(6, max(0, x))
Linear	linear	Identity (no-op)
Sigmoid	sigmoid	1 / (1 + exp(-x))
HardSigmoid	hardSigmoid	Piecewise linear approximation of sigmoid
Softplus	softplus	log(1 + exp(x))
Softsign	softsign	x / (abs(x) + 1)
Tanh	tanh	Hyperbolic tangent
Softmax	softmax	Normalized exponentials along axis
LogSoftmax	logSoftmax	Log of softmax
Gelu	gelu	x * Phi(x) using erf
GeluNew	gelu_new	Approximate GELU using tanh
Mish	mish	x * tanh(softplus(x))
Swish	swish	x * sigmoid(alpha * x)

I/O Contract

Input

• apply(tensor, axis?): Accepts a tf.Tensor of any shape. The optional axis parameter is used by Softmax and LogSoftmax (defaults to -1). Elu and Swish accept an alpha scaling parameter.

Output

• Returns a tf.Tensor of the same shape as the input, with the activation function applied element-wise.

getActivation

• Input: A string identifier (e.g., 'relu'), a ConfigDict, or an existing Activation instance. If null, defaults to 'linear'.
• Output: An Activation instance ready for use.

Usage Examples

import { getActivation } from './activations';

// Resolve an activation by string identifier
const relu = getActivation('relu');
const output = relu.apply(tf.tensor1d([-1, 0, 1]));
// output: [0, 0, 1]

// Use softmax with axis
const softmax = getActivation('softmax');
const probs = softmax.apply(tf.tensor2d([[1, 2, 3]]), -1);

// Serialize and deserialize
const serialized = serializeActivation(relu); // 'relu'

Related Pages

• Tensorflow_Tfjs_Advanced_Activations - Layer-based activation implementations (LeakyReLU, PReLU, ELU, ThresholdedReLU, Softmax as layers)
• Tensorflow_Tfjs_Backend_Operations - Low-level backend functions used by activations (e.g., K.elu, K.hardSigmoid, K.softsign)
• Tensorflow_Tfjs_Exports_Layers - Public API factory functions that create activation layers