Deep Learning Experiments: CNNs, Vision Transformers, and Framework Comparison

Collection of deep learning experiments exploring image classification, transfer learning, Vision Transformers, and differences between PyTorch and Keras workflows.

Overview

This repository groups multiple experiments into a single structured project to demonstrate progression from baseline models to more advanced architectures.

The goal is to systematically compare deep learning approaches across architectures (CNN vs Vision Transformers) and frameworks (Keras vs PyTorch), highlighting practical differences in performance, flexibility, and training workflows. This helps the understanding of:

• model training pipelines
• architectural differences (CNN vs ViT)
• framework trade-offs (Keras vs PyTorch)
• evaluation and performance comparison

---

Experiments

1. CNN Image Classification (Anime Dataset)

• Built a convolutional neural network for multi-class image classification
• Demonstrates end-to-end pipeline: preprocessing → training → evaluation

---

2. FashionMNIST Baseline

• Implemented baseline classification model
• Used as a controlled environment for testing architectures

---

3. Waste Classification (Transfer Learning)

• Applied transfer learning using pretrained CNN models
• Shows ability to adapt models to real-world datasets with limited data

---

4. Vision Transformers (ViT)

• Implemented in both:
  • Keras
  • PyTorch
• Demonstrates understanding of modern architectures beyond CNNs

---

5. CNN + ViT Hybrid Model

• Combined convolutional and transformer-based approaches
• Explores architectural integration and performance impact

---

6. Keras vs PyTorch Comparison

• Compared training workflows and performance
• Highlights differences in:
  • flexibility
  • control
  • implementation complexity

---

Key Skills Demonstrated

• Deep learning model development
• Transfer learning
• CNN and Vision Transformer architectures
• PyTorch and Keras workflows
• Model evaluation and comparison
• Data preprocessing and augmentation

---

Structure

deep-learning-experiments/
├── README.md
├── notebooks/
│ ├── anime_cnn_classifier.ipynb
│ ├── cnn_vit_hybrid.ipynb
│ ├── fashion_mnist_baseline.ipynb
│ ├── keras_vs_pytorch_comparison.ipynb
│ ├── vit_keras.ipynb
│ ├── vit_pytorch.ipynb
│ ├── waste_classification_transfer_learning.ipynb

---

Notes

These experiments are intentionally grouped into a single repository to present a coherent progression rather than isolated course exercises.

Key Takeaways

• CNNs remain strong baselines for structured image data with limited compute
• Vision Transformers provide flexibility but require more data and tuning
• PyTorch offers greater control for experimentation, while Keras enables faster prototyping
• Transfer learning significantly improves performance on smaller datasets