Keras · Class

layers.Embedding

keras.layers.Embedding maps nonnegative integer indexes to dense vectors through a trainable embedding matrix. Configure the vocabulary capacity with input_dim and the vector width with output_dim.

How it works

Embedding stores a trainable matrix whose rows represent vocabulary entries. When called with integer indices, it looks up the corresponding rows and returns one vector per input index, adding an output-dimension axis to the input shape.

Syntax

keras.layers.Embedding(input_dim, output_dim, embeddings_initializer="uniform", embeddings_regularizer=None, embeddings_constraint=None, mask_zero=False, weights=None, lora_rank=None, lora_alpha=None, quantization_config=None, **kwargs)

Parameters

Name	Type	Description
`input_dim`	`Integer`	Vocabulary size; valid indexes run up to `input_dim - 1`.
`output_dim`	`Integer`	Width of each dense embedding vector.
`embeddings_initializer`	`Initializer`	Initializer used for the embeddings matrix.
`embeddings_regularizer`	`Regularizer`	Regularizer applied to the embeddings matrix.
`embeddings_constraint`	`Constraint`	Constraint applied to the embeddings matrix.
`mask_zero`	`Boolean`	Whether input value 0 is treated as padding and masked.
`weights`	`Array-like`	Initial embedding values supplied directly.
`lora_rank`	`Integer or None`	Rank of the low-rank adaptation applied to the embedding matrix.
`lora_alpha`	`Float or None`	Scaling factor associated with low-rank adaptation.
`quantization_config`	`Configuration or None`	Quantization configuration for the layer.
`**kwargs`	`Keyword arguments`	Additional keyword arguments passed to the base layer.

Returns

A 3D tensor with shape (batch_size, input_length, output_dim) for the documented 2D input.

Key points

Each integer index is converted into a dense vector with the configured fixed width.
Set input_dim to one more than the largest index that can occur.
output_dim controls how many numeric features each lookup returns.
The documented input is a batch of integer sequences represented by a two-dimensional tensor.
The layer appends the embedding width, producing one vector for every sequence position.
Use embeddings_initializer to choose how the embedding matrix is initialized.

Common mistakes

Inputs must be nonnegative integers, and every index must fall within the configured vocabulary range.

Example: Embed integer sequences

import keras
import numpy as np

layer = keras.layers.Embedding(input_dim=1000, output_dim=64)
inputs = np.array([[4, 20], [7, 3]], dtype="int32")
outputs = layer(inputs)
print(outputs.shape)  # (2, 2, 64)

Creates a 64-dimensional vector for each integer index in a batch of length-two sequences.

When to use it

Represent words or categorical IDs as dense vectors.

Integer-token sequences

Provide vector sequences to recurrent, convolutional, or attention-based layers.

Sequence models

Learn compact representations for a fixed range of discrete indexes.

Large lookup tables

Interview questions

QWhat does keras.layers.Embedding do?

The layer maps integer indexes to dense vectors using a trainable embedding matrix. input_dim sets the vocabulary size, while output_dim sets each vector's width.

Follow-up: What output shape results from a 2D input? A tensor shaped (batch_size, input_length, output_dim).

QHow do input_dim and output_dim differ?

input_dim is the maximum integer index plus one, so valid indexes range from zero through input_dim - 1. output_dim is the dimensionality of each returned embedding vector.

Follow-up: Why must input_dim include one beyond the largest index? The rows are zero-based, so the largest index needs its own row.

QWhat are the consequences of enabling mask_zero?

Setting mask_zero=True treats index zero as padding and propagates a mask for compatible downstream layers. Consequently, zero cannot represent a normal vocabulary entry, and downstream layers must support masking.

Follow-up: What input shape does the page specify? The documented input is 2D: (batch_size, input_length).

keras.layers.ReversibleEmbedding
keras.initializers
keras.regularizers

Source: Embedding layer - Keras. Checked 2026-10-15.