(LVI) CIFAR-100 Image Classification

This project implements various image classification models for the CIFAR-100 dataset, including ResNet, ResNeXt, ViT, Swin Transformer, PyramidNet, and EfficientNet. It was developed as part of the Learning Vision Intelligence (LVI) course project.

Table of Contents

• Overview
• Getting Started
  • Prerequisites
  • Installation
  • Random Seed
• Usage
• Dataset
• Model Architectures
• Training
• Results
• Team Members
• License

Overview

The CIFAR-100 dataset consists of 60,000 32x32 color images in 100 classes. This project, developed as part of the Learning Vision Intelligence (LVI) course, aims to develop and compare high-performance image classification models using various state-of-the-art architectures.

Getting Started

Prerequisites

• Python 3.11.7
• PyTorch 2.1.2+cu121
• CUDA-capable GPU (recommended)
• For other required libraries, please refer to the requirements file

Installation

1. Clone the repository:

   git clone https://github.com/bigeco/lvi-cifar100-classifier-pytorch.git
   cd lvi-cifar100-classifier-pytorch

2. Install the required packages:

   pip install -r requirements.txt

Random Seed

To ensure reproducibility, we use a fixed random seed. The default random seed is set to 42. To change the random seed, use the --seed argument when running the training script:

python3 src/train.py --seed 123 

We use the following code to set the random seed:

import os
import torch
import numpy as np
import random

def seed_everything(seed):
    random.seed(seed)
    os.environ['PYTHONHASHSEED'] = str(seed)
    np.random.seed(seed)
    torch.manual_seed(seed)
    torch.cuda.manual_seed(seed)
    torch.cuda.manual_seed_all(seed)
    torch.backends.cudnn.deterministic = True
    torch.backends.cudnn.benchmark = False

def main(args):
    seed_everything(args.seed)  # Seed 고정

This function is called at the beginning of our training script to ensure consistent results across runs.

Usage

To train the model:

1. ResNet9

python3 src/train.py --model_name "resnet9" --epochs 240 --batch_size 128 --optimizer_name "Adam" --lr 0.005 --scheduler_name "OneCycleLR" --select_transform 'RandomCrop,RandomHorizontalFlip,ColorJitter' --mixup True

2. ResNet18

python3 src/train.py --model_name "resnet18" --epochs 100 --batch_size 64 --optimizer_name "AdamW" --lr 0.008 --criterion_name "LabelSmoothingLoss" --scheduler_name "OneCycleLR" --select_transform 'RandomCrop,RandomHorizontalFlip,ColorJitter' --mixup True --split True --train_ratio 0.8

3. ResNeXt50

python3 src/train.py --model_name "resnext50" --epochs 100 --batch_size 64 --optimizer_name "Adam" --lr 0.001 --select_transform 'RandomCrop,RandomHorizontalFlip,ColorJitter' --mixup True --split True --train_ratio 0.8

4. Densenet201

python3 src/train.py --model_name "densenet201" --epochs 100 --batch_size 64 --optimizer_name "Adam" --lr 0.001 --select_transform '' --split True --train_ratio 0.8

5. WideResNet28_10

python3 src/train.py --model_name "wide_resnet28_10" --epochs 100 --batch_size 64 --optimizer_name "Adam" --lr 0.001 --select_transform 'RandomCrop,RandomHorizontalFlip,ColorJitter' --mixup True --split True --train_ratio 0.8

6. ViT

python3 src/train.py --model_name "vit" --epochs 100 --batch_size 64 --optimizer_name "Adam" --lr 0.001 --select_transform '' --split True --train_ratio 0.8

7. Swin

python3 src/train.py --model_name "swin6" --epochs 100 --batch_size 64 --optimizer_name "AdamW" --lr 0.001 --weight_decay 0.05 --scheduler_name "CosineAnnealingLR" --select_transform 'RandomCrop,RandomHorizontalFlip' --split True --train_ratio 0.8

8. PyramidNet110 with Shake-Drop (Best Score)

python3 src/train.py 

Dataset

The CIFAR-100 dataset is automatically downloaded by the PyTorch torchvision library. It includes:

• 50,000 training images
• 10,000 testing images
• 100 classes
• 32x32 pixel resolution

Model Architecture

ResNet18 Architecture

ResNeXt29 Architecture

Training for Best Score

We implemented PyramidNet with ShakeDrop regularization and achieved the highest top-1 accuracy of 83.42% among all models on the CIFAR-100 dataset.

• Optimizer: SGD
• Learning Rate: 0.1 with MultiStepLR and ReduceLROnPlateau
• Batch Size: 128
• Epochs: 200
• Data Augmentation: Random crop, Random horizontal flip, AutoAugment, Cutout
• loss function: LabelSmoothingLoss

Result

PyramidNet with Shake-Drop achieved Best Score.

Model	Seed	Loss	Top-1 Accuracy	Top-5 Accuracy	Super Top-1 Accuracy
PyramidNet	42	1.26	83.42%	97.42%	91.14%

Team Members

This project was developed collaboratively by the following team members:

• Lee Songeun

  • Role:

    implement

    • utils(e.g. top-k accuracy function, train/valid utils) and CifarDataset class
    • models(e.g. ResNet9, ResNet56, ResNet110, Swin, PyramidNet with Shake Drop, Wide-ResNet)
    • lr scheduler (OneCycleLR, LambdaLR, StepLR, CosineAnnealingLR, CosineAnnealingWarmRestarts)

  • GitHub: @bigeco

• Park Jihye

  • Role:

    implement

    • models(e.g. EfficientNet, ResNet18, ResNeXt50, DenseNet201)
    • augmentation(rotation, flipping)
    • Loss functions(Focal Loss and Label Smoothing Loss functions)

  • GitHub: @park-ji-hye

• Song Daeun

  • Role:

    implement

    • models(e.g. ResNet, ResNeXt, ViT)
    • augmentation(Mixup, Color Jittering, Cutout, AutoAugment)
    • Modifying Model(Resnet, ResNeXt baseline)

  • GitHub: @Song-Daeun

License

This project is licensed under the MIT License - see the LICENSE file for details.