Deep Learning Models Collection

This repository contains a curated collection of deep learning models implemented in PyTorch and organized by architecture family. It is designed as a learning and experimentation playground for classic and modern neural network models on common benchmark datasets.

Project Structure

• CNN/: Convolutional Neural Networks for image classification

  • Classic/
    • AlexNet.ipynb: AlexNet-style CNN for CIFAR-10 (with training, evaluation, and accuracy reporting)
    • LeNet.ipynb: LeNet-style CNN on a simple image dataset (e.g., MNIST-like)
    • VGG.ipynb: VGG-style deep CNN for image classification
  • Modern/
    • DenseNet.ipynb: DenseNet implementation and training pipeline on CIFAR-10
    • EfficientNet.ipynb: EfficientNet-based classifier (torchvision model) for CIFAR-10
    • ResNet.ipynb: ResNet (e.g., ResNet-18) implementation and training on CIFAR-10

• MLP/: Fully-connected feed-forward networks for tabular data

  • MLP.ipynb: End-to-end notebook for training an MLP on a CSV dataset
  • SimpleMLP.py: Minimal MLP example (script form)
  • HyperparameterTuning/
    • GridSearch.py: Hyperparameter search using grid search for MLP
    • RandomSearch.py: Hyperparameter search using random search for MLP
  • db.csv: Example tabular dataset used in MLP examples

• RNN/: Recurrent Neural Networks for sequence data

  • GRU.ipynb: GRU-based text classification model on IMDB reviews (sentiment analysis)
  • LSTM.ipynb: LSTM-based time-series forecasting on airline passengers data
  • data/
    • airline-passengers.csv: Time-series data for the LSTM example

• .gitignore: Git ignore rules including data folders (e.g., CNN/Classic/data/, CNN/Modern/data/, and other generated artifacts).

Requirements

All examples are implemented in Python using PyTorch and common scientific Python libraries. A typical environment will include:

• Python 3.10+ (or similar modern version)
• PyTorch
• torchvision
• numpy, pandas
• scikit-learn
• matplotlib
• jupyter / jupyterlab
• transformers and datasets (for the GRU IMDB example)

You can manage dependencies using a virtual environment (e.g., venv, conda, or poetry).

Datasets

• CIFAR-10: Automatically downloaded via torchvision.datasets.CIFAR10 in CNN notebooks.
• Airline Passengers: Loaded from RNN/data/airline-passengers.csv or via direct URL in some examples.
• IMDB: Loaded using datasets.load_dataset("imdb") from the Hugging Face Datasets library.
• Tabular MLP Data: Provided in MLP/db.csv.

Some data directories (such as raw downloads or intermediate artifacts) are ignored by Git, as configured in .gitignore.

Usage Notes

• The focus of this repository is educational: each notebook is structured to be readable and step-by-step, including data loading, model definition, training loop, and evaluation.
• Training settings (epochs, batch size, learning rate) are chosen to provide a reasonable trade-off between runtime and accuracy, and you are encouraged to modify them for further experiments.
• GPU support is enabled when available via torch.cuda.is_available(); otherwise, models fall back to CPU.

Results

CNN models

For all CIFAR-10 experiments, the number of epochs is intentionally kept low to save time, because the primary goal is to demonstrate model implementations rather than to achieve state-of-the-art accuracy. As a result, the reported accuracies should not be used for serious performance comparison.

Notebook	Dataset	Test Accuracy
CNN/Classic/LeNet.ipynb	MNIST	98.60%
CNN/Classic/AlexNet.ipynb	CIFAR-10	74.24%
CNN/Classic/VGG.ipynb	CIFAR-10	77.45%
CNN/Modern/ResNet.ipynb (e.g., ResNet-18)	CIFAR-10	81.19%
CNN/Modern/DenseNet.ipynb	CIFAR-10	81.60%
CNN/Modern/EfficientNet.ipynb	CIFAR-10	64.72%

RNN models

Notebook	Dataset	RMSE	MAE
RNN/GRU.ipynb	Time-series style evaluation	19.99	17.00
RNN/LSTM.ipynb	Airline Passengers time series	42.64	34.96

MLP model

Notebook	Dataset	Test Accuracy
MLP/MLP.ipynb	MLP/db.csv (tabular)	83.49%