How can Tensorflow be used to standardize the data using Python?

TensorFlow provides powerful tools for data preprocessing, including standardization of image data. The flowers dataset contains thousands of flower images across 5 classes, making it perfect for demonstrating data normalization techniques using TensorFlow's preprocessing layers.

Data standardization is crucial for neural networks as raw pixel values (0-255) can cause training instabilities. We'll use TensorFlow's Rescaling layer to normalize pixel values to the [0, 1] range.

Setting Up the Environment

We are using Google Colaboratory to run the code. Google Colab provides free access to GPUs and requires zero configuration, making it ideal for TensorFlow projects.

Creating the Normalization Layer

The Rescaling layer divides each pixel value by 255, converting the range from [0, 255] to [0, 1] ?

import tensorflow as tf
from tensorflow.keras import layers
import numpy as np

# Assume train_ds is already loaded (flowers dataset)
print("Normalization layer is created")
normalization_layer = layers.Rescaling(1./255)

print("This layer is applied to dataset using map function")
normalized_ds = train_ds.map(lambda x, y: (normalization_layer(x), y))

# Check the normalized values
image_batch, labels_batch = next(iter(normalized_ds))
first_image = image_batch[0]
print(np.min(first_image), np.max(first_image))

Normalization layer is created
This layer is applied to dataset using map function
0.0 1.0

Alternative: Including Layer in Model

You can also include the rescaling layer directly in your model definition for simpler deployment ?

model = tf.keras.Sequential([
    layers.Rescaling(1./255, input_shape=(img_height, img_width, 3)),
    layers.Conv2D(32, 3, activation='relu'),
    layers.MaxPooling2D(),
    # ... other layers
])

Key Benefits of Data Standardization

Why Standardization Matters

• RGB channel values in range [0, 255] are not ideal for neural networks
• Large input values can cause gradient instability during training
• Standardized values [0, 1] ensure better convergence and training stability
• The Rescaling layer provides an efficient way to normalize data
• Including the layer in the model simplifies deployment and inference

Conclusion

TensorFlow's Rescaling layer provides an efficient way to standardize image data by normalizing pixel values from [0, 255] to [0, 1]. This preprocessing step is essential for stable neural network training and can be applied either during data loading or within the model architecture.