Python Tensorflow - tf.keras.layers.Conv2D() Function

Last Updated : 09 Feb, 2025

The tf.keras.layers.Conv2D() function in TensorFlow is a key building block of Convolutional Neural Networks (CNNs). It applies convolutional operations to input images, extracting spatial features that improve the model’s ability to recognize patterns.

The Conv2D layer applies a 2D convolution over an input image, performing the following operation:

\text{output=activation(convolution(input,kernel)+bias)}

where:

convolution(input, kernel): A sliding window operation (filter) applied over the input image.
kernel: A set of learnable weights (filters) that detect specific features.
bias: A bias vector added to the convolution output.
activation: An activation function applied element-wise.

Syntax of tf.keras.layers.Conv2D()

tf.keras.layers.Conv2D(
filters,
kernel_size,
strides=(1, 1),
padding='valid',
activation=None,
use_bias=True,
kernel_initializer="glorot_uniform",
bias_initializer="zeros"
)

Parameters Explained:

filters (Required): Number of filters (kernels) applied in the convolution. Determines the number of output channels.
kernel_size (Required): Size of the filter (e.g., (3,3) for a 3×3 kernel).
strides (Default: (1,1)): Step size for moving the filter across the input.
padding:
- 'valid' (default): No padding, output size is reduced.
- 'same': Zero-padding is added to keep the output size the same as the input.
activation (Optional): Activation function (e.g., 'relu', 'sigmoid').
use_bias (Default: True): Whether to include a bias term.
kernel_initializer (Default: glorot_uniform): Defines how the filters are initialized.
bias_initializer (Default: zeros): Defines how the bias is initialized.

Using Conv2D in a CNN Model

Below is the implementation of the CNN model.

Python

from tensorflow import keras
from tensorflow.keras import layers
from tensorflow.keras.utils import plot_model
from IPython.display import Image

# Define a simple CNN model
model = keras.Sequential([
    layers.Conv2D(32, (3, 3), activation='relu', input_shape=(28, 28, 1)),  # First conv layer
    layers.MaxPooling2D(pool_size=(2, 2)),  # Downsample the feature maps
    layers.Conv2D(64, (3, 3), activation='relu'),  # Second conv layer
    layers.MaxPooling2D(pool_size=(2, 2)),  # Downsample again
    layers.Flatten(),  # Flatten the output for the dense layer
    layers.Dense(128, activation='relu'),  # Fully connected layer
    layers.Dense(10, activation='softmax')  # Output layer (10 classes)
])

# Compile the model
model.compile(optimizer='adam', loss='categorical_crossentropy', metrics=['accuracy'])
model.summary()