Keras · Class

layers.SeparableConv2D

SeparableConv2D applies a depthwise spatial convolution followed by a pointwise channel-mixing convolution. It optionally adds bias and applies an activation.

How it works

SeparableConv2D first applies a spatial depthwise convolution independently to each input channel, then a pointwise convolution combines those intermediate channels into the requested number of filters. It can add a bias and apply an activation after the convolution.

Syntax

keras.layers.SeparableConv2D(filters, kernel_size, strides=(1, 1), padding="valid", data_format=None, dilation_rate=(1, 1), depth_multiplier=1, activation=None, use_bias=True, depthwise_initializer="glorot_uniform", pointwise_initializer="glorot_uniform", bias_initializer="zeros", depthwise_regularizer=None, pointwise_regularizer=None, bias_regularizer=None, activity_regularizer=None, depthwise_constraint=None, pointwise_constraint=None, bias_constraint=None, **kwargs)

Parameters

Name	Type	Description
`filters`	`int`	Number of filters in the pointwise convolution and therefore the output channels.
`kernel_size`	`int or tuple/list of 2 integers`	Height and width of the depthwise convolution window; an integer or a pair of integers.
`strides`	`int or tuple/list of 2 integers`	Depthwise convolution stride; an integer or a pair of integers.
`padding`	`string`	Padding mode, typically `"valid"` or `"same"`.
`data_format`	`string or None`	Input and output channel layout.
`dilation_rate`	`int or tuple/list of 2 integers`	Dilation factor for the depthwise convolution.
`depth_multiplier`	`int`	Number of depthwise output channels generated per input channel.
`activation`	`string, callable, or None`	Optional activation function.
`use_bias`	`bool`	Whether to add a bias vector.
`depthwise_initializer`	`string or initializer`	Initializer for the depthwise kernel.
`pointwise_initializer`	`string or initializer`	Initializer for the pointwise kernel.
`bias_initializer`	`string or initializer`	Initializer for the bias.

Returns

A rank-4 tensor representing activation(separable_conv2d(inputs, kernel) + bias); its channel axis depends on data_format.

Key points

This is Keras's layer for separable convolutions over two spatial dimensions.
The operation first processes each channel independently, then mixes channels with a pointwise convolution.
Bias addition is optional and controlled by use_bias.
An activation can be applied after convolution and optional bias addition.
filters sets the number of output channels produced by the pointwise stage.
kernel_size determines the height and width of the depthwise spatial window.
For channels-last input, the layer returns a four-dimensional tensor with filters in the final axis.

Common mistakes

A stride greater than one cannot be combined with a dilation rate greater than one.

Example: A separable convolution block

import keras
from keras import layers

inputs = keras.Input(shape=(32, 32, 3))
x = layers.SeparableConv2D(32, 3, padding="same", activation="relu")(inputs)
outputs = layers.GlobalAveragePooling2D()(x)
model = keras.Model(inputs, outputs)
model.summary()

Creates a model whose separable convolution preserves the spatial dimensions while producing 32 output channels.

When to use it

Efficient spatial feature extraction

Image classification models

Lower-cost convolutional processing

Convolutional blocks with limited compute or parameters

Separable feature transformations

Vision models requiring channel mixing after per-channel spatial filtering

Interview questions

QHow does SeparableConv2D differ from a standard convolution?

It splits convolution into two operations: depthwise convolution processes channels independently, and pointwise convolution mixes the resulting channels. This reduces the structure of a standard convolution while still producing the requested number of output filters.

Follow-up: What controls depthwise expansion? The depth_multiplier sets how many depthwise output channels are generated for each input channel.

QWhat do filters and kernel_size control?

filters determines the output channel count because it is the dimensionality of the pointwise convolution's output space. kernel_size specifies the two-dimensional depthwise window.

Follow-up: How would you preserve height and width? Use padding="same" with the usual unit strides, subject to the layer's padding and shape rules.

QWhat output shape does the layer produce?

With channels-last data, the result is a rank-4 tensor shaped (batch_size, new_height, new_width, filters). With channels-first data, the result is (batch_size, filters, new_height, new_width).

Follow-up: What does the layer return conceptually? It returns the activation applied to the separable convolution result plus bias when bias is enabled.

SeparableConv1D
DepthwiseConv2D
Conv2D
keras.ops.separable_conv

Source: SeparableConv2D Layer. Checked 2026-10-15.