AlexNet Legacy Implementation

This repository contains a Hydra-configured, PyTorch Lightning-based implementation of the AlexNet architecture for CIFAR-10 and CIFAR-100 datasets. The project is designed for experimental and diagnostic purposes, not production use or performance benchmarking.

---

Project Structure

alexnet/
├── configs/             # Hydra configuration files
│   ├── checkpoint/
│   ├── dataset/
│   ├── model/
│   ├── optimizer/
│   └── trainer/
├── data/                # Data loading logic
├── models/              # AlexNet architecture
├── training/            # Lightning training wrapper

Environment Setup

This project uses Python 3.10. Development was done using conda for environment management and pip for dependency resolution.

1. Clone the Repository

git clone https://github.com/https://github.com/cmsolson75/AlexNet.git
cd AlexNet

2. Create and Activate Environment

conda create -n alexnet python=3.10
conda activate alexnet

3. Install Dependencies

pip install -r requirements.txt

---

Dependencies

All runtime dependencies are specified in requirements.txt. Key libraries:

• torch
• torchvision
• lightning
• wandb
• omegaconf
• hydra-core
• torchmetrics
• tqdm

---

Configuration System

Hydra is used to compose configuration files located under alexnet/configs/.

Primary config: alexnet/configs/config.yaml

Modular components:

• dataset/ — Dataset-specific parameters (e.g. batch size, location)
• model/ — AlexNet architecture variants
• optimizer/ — SGD and LR scheduler
• trainer/ — Lightning trainer settings
• checkpoint/ — Checkpointing frequency and naming

Override configs at runtime as needed:

python train.py model.pool.type=average optimizer.lr=0.01

---

Training

Before starting, log into Weights & Biases (only required once per machine):

wandb login

Then launch training

python train.py

This will:

• Load config from alexnet/configs/config.yaml
• Create experiment-specific output directory: outputs/<project>/<run_name>/
• Log training to Weights & Biases
• Save checkpoints via Lightning

---

Evaluation

python evaluate.py \
    --model-path path/to/model.pth

This uses the test loader for the dataset defined in the active config.

---

Unwrap Trained Model

Use this to strip the Lightning wrapper and extract the raw AlexNet model:

python unwrap_model.py \
    --ckpt-path path/to/lightning.ckpt \
    --output-path path/to/alexnet.pth

---

Notes

• Run names include a deterministic hash of the config and a timestamp.
• Checkpoints and logs are stored in machine-independent paths.
• All configurations are reproducible with fixed seeds.

---