jostmey / controls

Training examples of different foundational models used as controls for my experiments

Deep Learning Models for Experimental Controls

Overview

This repository hosts deep learning models that function as experimental controls. The models included are pivotal for research experiments aimed at publication. The code is crafted to ensure that each model can be trained from scratch, critical for executing the controls on identical hardware as experimental approaches, providing a consistent baseline for comparison.

Models

Currently, the repository includes two primary models:

• A Fully Connected Neural Network (FCNN) trained on the MNIST dataset.
• A Residual Neural Network (ResNet) model trained on the CIFAR10 dataset.

Requirements

• PyTorch 2.1
• Python 3.10
• TorchMetrics
• TorchVision
• TensorBoard

Execution

Execute models using PyTorch with the following commands:

# ResNet on MNIST
python3 models/run_fcnn_cntrl.py --epochs 256 --batch 1024 --basepath results/fcnn_cntrl

# ResNet on CIFAR10
python3 models/run_resnet_cntrl.py --step 0.0001 --epochs 256 --device cuda --batch 512 --basepath results/resnet_cntrl

# Visualize results
tensorboard --logdir=results

Performance

• MNIST (FCNN): ~98.5% accuracy.
• CIFAR10 (ResNet): ~92.5% accuracy.

Summary

This repository provides deep learning models for consistent experimental controls, requiring specific software and libraries for operation, and includes detailed instructions for model training and expected performance outcomes.