Chapter 6: Image Datasets – Getting Data Into the Network

“It’s not just about images—it’s about structure, batching, and flow. A CNN needs data, not chaos.”

---

Why This Chapter Matters

You’ve learned how to preprocess a single image, but deep learning thrives on batches of thousands of images—efficiently streamed, augmented, shuffled, and labeled.

If your dataset isn’t:

• Properly structured

• Mapped to class labels

• Efficiently loaded and shuffled

… then even the best model won't help you. You’ll face:

• GPU idle time

• Overfitting from data leaks

• Mislabeling issues

This chapter shows how to transform folders of images into usable model inputs, with complete control over batching, shuffling, label handling, and augmentation.

---

Conceptual Breakdown

🔹 The Canonical Folder Structure

Most frameworks expect images arranged like this:

dataset/
├── train/
│   ├── class1/
│   │   ├── img1.jpg
│   │   ├── img2.jpg
│   ├── class2/
│   │   ├── img3.jpg
│   │   ├── img4.jpg
├── val/
│   ├── class1/
│   ├── class2/

Each folder name becomes the class label, and all images inside are automatically mapped.

This structure is:

• Easy to understand

• Supported natively in both PyTorch and TensorFlow

• Flexible for augmentation and splitting

---

🔹 Dataset Terms You Must Know

Term	Definition
Dataset	Collection of images and labels (can be lazy-loaded)
Dataloader	Handles batching, shuffling, and parallel loading
Transform	Applied per sample to augment/normalize
Batch	A group of N samples fed to the model at once
Epoch	One full pass through the dataset

---

🔹 Why Shuffling Matters

Shuffling prevents:

• Learning label order bias (e.g., all cats, then all dogs)

• Memorizing batch patterns

• Overfitting on dataset structure

📌 Always shuffle during training, not validation.

---

🔹 Visualizing a Batch

Understanding shapes is critical.

Description	PyTorch Shape	TensorFlow Shape
One RGB image	[3, H, W]	[H, W, 3]
One batch (N images)	[N, 3, H, W]	[N, H, W, 3]
One batch + label	(images, labels)	(images, labels)

PyTorch Implementation

🔸 Folder-based Dataset Loading

from torchvision import datasets, transforms
from torch.utils.data import DataLoader

# Define transforms
transform = transforms.Compose([
    transforms.Resize((224, 224)),
    transforms.ToTensor(),
    transforms.Normalize([0.485, 0.456, 0.406],
                         [0.229, 0.224, 0.225])
])

# Load dataset
train_dataset = datasets.ImageFolder("data/train", transform=transform)
train_loader = DataLoader(train_dataset, batch_size=32, shuffle=True)

# Check class mapping
print(train_dataset.class_to_idx)  # e.g., {'cat': 0, 'dog': 1}

🔸 Iterating Through a Batch

for images, labels in train_loader:
    print(images.shape)  # [32, 3, 224, 224]
    print(labels)        # Tensor of shape [32]
    break

---

TensorFlow Implementation

🔸 Load from Folder with image_dataset_from_directory

import tensorflow as tf

train_ds = tf.keras.preprocessing.image_dataset_from_directory(
    "data/train",
    image_size=(224, 224),
    batch_size=32,
    label_mode="int",  # or 'categorical' for one-hot
    shuffle=True
)

# Check output
for images, labels in train_ds.take(1):
    print(images.shape)  # (32, 224, 224, 3)
    print(labels.shape)  # (32,)

🔸 Apply Preprocessing via .map()

def preprocess(img, label):
    img = tf.cast(img, tf.float32) / 255.0
    mean = tf.constant([0.485, 0.456, 0.406])
    std = tf.constant([0.229, 0.224, 0.225])
    img = (img - mean) / std
    return img, label

train_ds = train_ds.map(preprocess)

---

Framework Comparison Table

Feature	PyTorch	TensorFlow
Dataset Loader	datasets.ImageFolder	tf.keras.preprocessing.image_dataset_from_directory
Custom Transforms	transforms.Compose([...])	.map(preprocess_fn)
Shuffling	shuffle=True in DataLoader	shuffle=True in loader or .shuffle(buffer)
Label format	class_to_idx dictionary	Auto-mapped, label mode configurable
Image shape in batch	[B, 3, H, W]	[B, H, W, 3]

---

Mini-Exercise

Prepare a small dataset with 2–3 image classes (e.g., cat, dog, bird). Then:

1. Use both PyTorch and TensorFlow to:

   • Load dataset from folder

   • Apply resizing and normalization

   • Create dataloaders/batched datasets

2. Iterate through 1 batch and print:

   • Image tensor shape

   • Label mapping (class-to-index)

3. Try visualizing one batch using matplotlib.

   • Bonus: Add a simple augmentation (flip or crop) and re-inspect the batch output.

---