CIFAR100 WideResNet28x10 Classification Pipeline

1. Project Overview

This project builds a high accuracy image classification system for the CIFAR100 dataset using the WideResNet28x10 architecture. The goal is to achieve strong top 1 accuracy through advanced augmentation, stable training strategy, and detailed model analysis.

Key objectives: • Construct a reproducible training pipeline • Apply strong augmentations like MixUp, CutMix, and RandAugment • Train WideResNet28x10 to high accuracy • Evaluate performance using top 1 and top 5 metrics • Perform model interpretation using GradCAM and t SNE

---

2. Dataset Information

Dataset: CIFAR100

• Number of classes: 100 • Image size: 32 x 32 RGB • Train images: 50000 • Test images: 10000 • Class types: animals, vehicles, household items, foods

Preprocessing statistics: • Mean: [0.5071, 0.4867, 0.4408] • Std: [0.2675, 0.2565, 0.2761]

---

3. Exploratory Data Analysis

Key EDA steps: • Display per class sample images • Verify class labels and counts • Visualize per class mean image • Plot RGB histograms for selected classes • Identify visually similar class groups

Insights from EDA: • Several classes are highly similar in appearance • Contrast and color variation across images is moderate • Augmentation is necessary to improve generalization

---

4. Data Augmentation and Preprocessing

Train transforms: • Random crop (32 with padding 4) • Random horizontal flip • RandAugment • MixUp augmentation (in training loop) • CutMix (probability 0.5) • Normalize using dataset mean and std

Test transforms: • ToTensor • Normalize

Optional: • Random erasing

Purpose: • Improve robustness • Reduce overfitting • Expand training sample diversity

---

5. Model Architecture

Model: WideResNet28x10

Key attributes: • Depth: 28 • Width factor: 10 • Uses wide residual blocks • Dropout p = 0.3 • Final fully connected layer outputs 100 classes • Weight initialization using He Normal

Why WideResNet: • Wider layers improve feature richness • Performs especially well on CIFAR sized images

---

6. Training Strategy

Optimizer: SGD • Momentum: 0.9 • Weight decay: 5e 4

Learning rate policy: • Initial LR: 0.1 • Warmup: 5 epochs • Cosine annealing schedule

Hyperparameters: • Batch size: 128 • Epochs: 250 • Loss: Cross entropy with label smoothing (0.1)

Stability features: • Save last 5 checkpoints • Enable CuDNN benchmark • Fixed random seeds

---

7. Evaluation Method

Metrics: • Top 1 accuracy • Top 5 accuracy • Per class accuracy • Confusion matrix

Advanced evaluation: • Test time augmentation • Ensemble predictions from last checkpoints

Reporting: • Plot training and validation accuracy curves • Display worst performing classes

---

8. Model Interpretation and Analysis

Tools used: • GradCAM heatmaps • t SNE or UMAP feature visualization • Error inspection for misclassified images

Insights: • Identifying confusing class pairs • Understanding model attention regions • Evaluating feature cluster separation

---

9. Results Summary

Expected outcomes: • High accuracy WideResNet model for CIFAR100 • Strong generalization due to augmentations • Stable training curves due to cosine scheduling • Meaningful visualizations for deeper understanding

Recommended result slides: • Accuracy curves • Confusion matrix • GradCAM samples • t SNE cluster plot

---

10. Conclusion

The project delivers an end to end CIFAR100 classification system using a powerful WideResNet architecture and modern augmentation methods. It emphasizes reproducibility, robustness, and deep model analysis, making it suitable for academic and applied machine learning presentations.

---