Procedure

The objective of this experiment is to implement a Convolutional Neural Network (CNN) for multi-class image classification and to analyze its performance on a real-world color image dataset. This experiment focuses on understanding convolutional feature extraction, pooling operations, data augmentation, training dynamics, and evaluation strategies using the CIFAR-10 dataset.

1. Import Required Libraries

• PyTorch (torch and torch.nn): It helps to build the CNN, define the loss, and train the model.
• Torchvision (torchvision.datasets, torchvision.transforms): load CIFAR-10 and apply image preprocessing/augmentation.
• NumPy (numpy): import basic numerical utilities
• Matplotlib (matplotlib.pyplot): visualize sample images and training/validation graphs.

2. Dataset Loading and Description

• Load CIFAR-10 with torchvision.datasets.CIFAR10
• Dataset info (important):
  • Total: 60,000 RGB images, size 32×32
  • Split: 50,000 train + 10,000 test
  • Classes: 10 categories (e.g., airplane, automobile, bird, cat, etc.)

3. Exploratory Data Analysis (EDA)

• Visualize: 1 sample image per class (10 classes) in a labeled grid.
• Verify: image size 32×32, RGB channels.
• Inspect: class diversity, background variation, and ambiguity between similar classes.

4. Data Preprocessing

a) Tensor conversion

Convert each image to a PyTorch tensor of shape (C, H, W) with values in [0, 1].

b) Normalization (CIFAR-10)

Per-Channel Statistics: Since CIFAR-10 images are in RGB (Red, Green, Blue) format, there are three separate values for both the mean and standard deviation, each corresponding to one of these color channels.

CIFAR-10 Specifics: For this dataset, the values are:

• Mean (R, G, B): (0.4914, 0.4822, 0.4465).
• Standard Deviation (R, G, B): (0.2470, 0.2435, 0.2616) or commonly (0.2023, 0.1994, 0.2010) depending on the pre-computation used.

It is critical for training because:

• Gradient Stability: Neural networks train much more effectively when input features are on a similar scale. This prevents some pixels from "overpowering" others during the learning process.
• Faster Convergence: By ensuring stable gradients, the model can reach the optimal solution (converge) much faster than it would with unscaled raw data.

5. Data loading and augmentation

• Augment and preprocess (train only): The random horizontal flip, random crop with padding, and AutoAugment increase diversity, reduce overfitting, and improve generalization.
• Create DataLoaders (train & test): load data in mini-batches for faster, GPU-friendly training and evaluation.
• Shuffling rule: shuffle training loader (avoid order bias), no shuffle for test loader (consistent evaluation).

6. CNN Model Architecture Design

• Input (32×32×3)
• (Conv → BatchNorm → ReLU) × N (extract features)
• MaxPool (periodically) (reduce H×W, keep important patterns
• Global Average Pooling (convert feature maps → single feature vector)

7. Training Configuration

• Loss: CrossEntropyLoss (for 10-class classification)
• Adam Optimizer with Regularization: use SGD/Adam + weight decay (L2) (optional: dropout) to reduce overfitting
• LR Scheduler: adjust learning rate during training (e.g., StepLR / CosineAnnealing) for better convergence

8. Model Training

The model is trained for a fixed number of epochs, where each epoch consists of performing a forward pass to generate predictions, computing the corresponding loss, and then applying backpropagation followed by an optimizer step to update the model parameters. Throughout the training process, both training loss and accuracy are recorded at the end of each epoch to monitor learning progress and convergence.

9. Model Evaluation

After training, the CNN is evaluated on unseen test data to assess its generalization performance. Test accuracy and additional error metrics, such as the confusion matrix and class-wise accuracy, are computed to analyze prediction behavior and identify specific classes or instances where the model fails.

10. Result Visualization

• Plot train vs test curves for loss and accuracy to check convergence/overfitting.
• Show a few sample predictions with true label + predicted label (optionally, confidence) for qualitative evaluation.

11. Performance Analysis

• Class-wise accuracy: identify best/worst performing classes.
• Confusion matrix: highlight frequently confused class pairs and error patterns.