Raw data usually cannot go straight into a neural network. Preprocessing turns inputs—and, when necessary, targets—into tensors with values and shapes the model can use. The work depends on the data: a review needs a numeric representation, while an image may need rescaling and task-specific labels. section 6.2.1 section 12.2.3
From review to numbers
A text classifier can split a review into words, look up each word in a vocabulary, and use an [UNK] index for words the vocabulary lacks. Those indices are not the only possible model input: an IMDb sentiment example turns sequences of word indices into multi-hot float32 vectors. The packaged IMDb reviews already contain word indices; raw text requires the earlier steps too. section 14.2 section 6.2.1 section 4.1.1
Choose the transformations that solve a concrete input problem:
- Vectorize: turn data into tensors. Inputs and targets are typically floating-point tensors, with integers or strings used in specific cases. section 6.2.1
- Normalize: rescale image pixels from 0–255 to float32 values in 0–1, or scale each numeric feature independently when feature ranges differ. section 6.2.1 section 4.4
- Standardize text when differences such as capitalization or punctuation should not produce different representations. Lowercasing and removing punctuation is one simple choice. section 14.2
- Handle missing values where they occur; the appropriate preprocessing also depends on the data domain. section 6.2.1
Normalize house-price features
House-price inputs can mix small floating-point values with much larger integers. Before training, normalize each feature independently to mean 0 and standard deviation 1. Feeding large or heterogeneous values directly into a network is not a safe default. section 6.2.1
Preprocessing can also make labels match a model’s output. For a grid-based object detector, bounding boxes are translated into grid-relative coordinates; class and box arrays then assign labels, coordinates, and confidence values to grid locations. Images are decoded, passed through the matching image preprocessor, and batched with their labels. Streaming batches with tf.data avoids loading every decoded image into memory at once. section 12.2.3 section 12.2.2
Listing 12.9 Creating a dataset to train on
import tensorflow as tf def load_image(path): #1 x = tf.io.read_file(path) x = tf.image.decode_jpeg(x, channels=3) return preprocessor(x) images = tf.data.Dataset.from_tensor_slices([x["path"] for x in metadata]) images = images.map(load_image, num_parallel_calls=8) labels = {"box": box_array, "class": class_array} labels = tf.data.Dataset.from_tensor_slices(labels) dataset = tf.data.Dataset.zip(images, labels).batch(16).prefetch(2) #2 val_dataset, train_dataset = dataset.take(500), dataset.skip(500) #3
The loading pipeline pairs preprocessed images with prepared targets, then batches and prefetches the dataset. section 12.2.3
