Convolution operation concept in TensorFlow

Introduction

Convolution helps computers find patterns in images or signals by sliding a small filter over data.

When you want to detect edges or shapes in pictures.

When analyzing sound waves to find specific sounds.

When building image recognition apps like face detectors.

When processing video frames to track movement.

When extracting features from data for machine learning models.

Syntax

TensorFlow

tf.nn.conv2d(input, filters, strides, padding, data_format='NHWC', dilations=None, name=None)

input: The input data, usually a 4D tensor (batch, height, width, channels).

filters: The convolution kernels (small matrices) to slide over the input.

Examples

This applies a convolution with stride 1 and keeps the output size same as input.

TensorFlow

output = tf.nn.conv2d(input_tensor, filter_tensor, strides=[1,1,1,1], padding='SAME')

This applies convolution with stride 2 and no padding, so output is smaller.

TensorFlow

output = tf.nn.conv2d(input_tensor, filter_tensor, strides=[1,2,2,1], padding='VALID')

Sample Model

This code creates a 5x5 input and a 3x3 filter, then applies convolution with stride 1 and padding SAME. It prints the output shape and values.

TensorFlow

import tensorflow as tf
import numpy as np

# Create a simple 5x5 input with 1 channel
input_data = np.array([[[[1],[2],[3],[4],[5]],
                        [[6],[7],[8],[9],[10]],
                        [[11],[12],[13],[14],[15]],
                        [[16],[17],[18],[19],[20]],
                        [[21],[22],[23],[24],[25]]]], dtype=np.float32)

# Create a 3x3 filter with 1 input and 1 output channel
filter_data = np.array([[[[1]],[[0]],[[1]]],
                        [[[0]],[[1]],[[0]]],
                        [[[1]],[[0]],[[1]]]], dtype=np.float32)

# Convert to tensors
input_tensor = tf.constant(input_data)
filter_tensor = tf.constant(filter_data)

# Apply convolution with stride 1 and SAME padding
output = tf.nn.conv2d(input_tensor, filter_tensor, strides=[1,1,1,1], padding='SAME')

print("Output shape:", output.shape)
print("Output values:")
print(output.numpy()[0,:,:,0])

Important Notes

Stride controls how much the filter moves each step; bigger stride means smaller output.

Padding='SAME' keeps output size same by adding zeros around input edges.

Filters learn to detect different features like edges or textures during training.

Summary

Convolution slides filters over input to find patterns.

Stride and padding affect output size and detail.

It is a key step in image and signal processing models.