gperdrizet/CIFAR10

CIFAR-10 Image Classification Models

This repository contains a collection of PyTorch models trained on the CIFAR-10 dataset, demonstrating progressive improvements in image classification through different architectures and training techniques.

• Package: image-classification-tools on PyPI
• Documentation: Full Tutorial
• Repository: GitHub

Model Overview

Model	Architecture	Test Accuracy	Parameters
DNN	Fully-connected	~45%	1.4M
CNN	Basic CNN	~65%	122K
RGB CNN	CNN with color	~70%	370K
Optimized CNN	Optuna-tuned architecture	~75%	Variable
Augmented CNN	Optimized + augmentation	~78%	Variable
ResNet50	Transfer learning	~85%+	23.5M

Note: Exact accuracies may vary slightly between training runs

Models

1. DNN (Deep Neural Network)

File: dnn.pth

A baseline fully-connected neural network operating on grayscale CIFAR-10 images.

Architecture:

• Input: 1024 (32×32 grayscale)
• 4 fully-connected layers: 1024 → 1024 → 256 → 64 → 10
• ReLU activations
• Dropout (0.5) for regularization

Training:

• Optimizer: Adam (lr=1e-3)
• Early stopping with patience=15
• Grayscale preprocessing

2. CNN (Convolutional Neural Network)

File: cnn.pth

Basic convolutional architecture to exploit spatial relationships in images.

Architecture:

• 3 convolutional blocks with max pooling
• Batch normalization
• Fully-connected classifier

Training:

• Optimizer: Adam
• Grayscale preprocessing
• Early stopping

3. RGB CNN

File: rgb_cnn.pth

CNN modified to process full RGB color information.

Architecture:

• Similar to CNN but accepts 3-channel RGB input
• Additional filters to handle color information

Training:

• RGB preprocessing (no grayscale conversion)
• Same optimization strategy as CNN

4. Optimized CNN

File: architecture_optimized_cnn.pth

Architecture optimized using Optuna hyperparameter search.

Optimized Parameters:

• Number of convolutional blocks
• Filters per layer
• Dropout rates
• Kernel sizes

Training:

• Optuna study with 100+ trials
• Best architecture selected based on validation accuracy

5. Training-Optimized CNN

File: training_optimized_cnn.pth

Further optimization of training hyperparameters.

Optimized Parameters:

• Optimizer type (Adam/SGD/AdamW)
• Learning rate
• Weight decay
• Batch size

6. Augmented CNN

File: augmented_cnn.pth

Combines optimized architecture with data augmentation.

Data Augmentation:

• Random horizontal flips
• Random crops with padding
• Color jittering
• Random rotations

Training:

• Extended training on augmented dataset
• Best-performing model overall (non-transfer learning)

7. ResNet50

File: resnet50.pth

Transfer learning using pre-trained ResNet50.

Architecture:

• ResNet50 backbone pre-trained on ImageNet
• Custom classifier head for CIFAR-10
• Images upscaled to 224×224

Training:

• Fine-tuned on CIFAR-10
• Lower learning rate for pre-trained layers
• Achieves highest accuracy

Dataset

CIFAR-10 consists of 60,000 32×32 color images in 10 classes:

• airplane
• automobile
• bird
• cat
• deer
• dog
• frog
• horse
• ship
• truck

Split:

• Training: 40,000 images
• Validation: 10,000 images
• Test: 10,000 images

Usage

Loading Models

import torch

# Load a model
model = torch.load('dnn.pth', map_location='cpu')
model.eval()

# Or download from Hugging Face Hub
from huggingface_hub import hf_hub_download

model_path = hf_hub_download(
    repo_id='YOUR_USERNAME/CIFAR10',
    filename='dnn.pth'
)
model = torch.load(model_path, map_location='cpu')

Making Predictions

from torchvision import transforms
from PIL import Image

# Preprocessing for grayscale models (DNN, CNN)
transform = transforms.Compose([
    transforms.Grayscale(num_output_channels=1),
    transforms.ToTensor(),
    transforms.Normalize((0.5,), (0.5,))
])

# Preprocessing for RGB models
transform = transforms.Compose([
    transforms.ToTensor(),
    transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))
])

# For ResNet50 (224x224 input)
transform = transforms.Compose([
    transforms.Resize(256),
    transforms.CenterCrop(224),
    transforms.ToTensor(),
    transforms.Normalize(mean=[0.485, 0.456, 0.406], 
                        std=[0.229, 0.224, 0.225])
])

# Load and preprocess image
image = Image.open('path/to/image.jpg')
input_tensor = transform(image).unsqueeze(0)

# Make prediction
with torch.no_grad():
    output = model(input_tensor)
    probabilities = torch.softmax(output, dim=1)
    predicted_class = output.argmax(dim=1).item()

classes = ['airplane', 'automobile', 'bird', 'cat', 'deer', 
           'dog', 'frog', 'horse', 'ship', 'truck']
print(f'Predicted: {classes[predicted_class]}')

Training Details

All models were trained with:

• Framework: PyTorch 2.0+
• Loss: CrossEntropyLoss
• Validation: 20% of training data
• Early Stopping: Prevents overfitting
• Hardware: CUDA-enabled GPU (when available)

Hyperparameter Optimization

Models 4-6 used Optuna for automated hyperparameter tuning:

• Search Algorithm: TPE (Tree-structured Parzen Estimator)
• Trials: 100+ per optimization study
• Objective: Validation accuracy
• Pruning: MedianPruner for early termination of unpromising trials

Performance Metrics

Each model includes:

• Test accuracy
• Per-class accuracy
• Confusion matrix
• ROC curves
• Precision-Recall curves
• Class probability distributions

See the companion notebooks for detailed performance analysis.

Repository Structure

CIFAR10/
├── notebooks/          # Jupyter notebooks for each model
├── models/            # Saved model files (.pth)
├── src/               # image-classification-tools package
└── data/              # CIFAR-10 dataset (auto-downloaded)

Installation

Install the companion tools package:

pip install image-classification-tools

Or clone the repository for full notebooks:

git clone https://github.com/YOUR_USERNAME/CIFAR10.git
cd CIFAR10

Citation

If you use these models or code, please cite:

@software{cifar10_tutorial,
  author = {Your Name},
  title = {CIFAR-10 Image Classification Tutorial},
  year = {2026},
  url = {https://github.com/YOUR_USERNAME/CIFAR10}
}

License

This project is licensed under the GNU General Public License v3.0 - see the LICENSE file for details.

---

Last updated: February 2026