Sketch Image Classification

Overview

2024.09.10 ~ 2024.09.26

This project focuses on classifying ImageNet-Sketch images as part of a private competition organized by Naver Connect Foundation and Upstage.

Contributors

김기수	문채원	안주형	은의찬	이재훈	장지우

Experimental Techniques

We experimented with various techniques to improve performance:

• Augmentation: Aspect ratio, CLAHE, grayscale, flip, rotation, mixup, cutmix, Text-to-Image
• Optimization: Pseudo labeling, label smoothing, Error-Driven Learning, Progressive resizing, Model Soups, TTA
• Visualization: Confusion matrix, error-class visualization
• Ensemble: Soft voting, Hard voting, Ensemble of Ensembles, stacking, snapshot

Folder Structure

level1-cv-10
├── data/
│   ├── train/
│   ├── test/
│   ├── train.csv
│   └── test.csv
├── src/
│   ├── args.py
│   ├── dataset.py
│   ├── transforms.py
│   ├── model.py
│   ├── loss.py
│   ├── trainer.py
├── train.py
├── train.sh
├── inference.py
├── inference.sh
├── ensemble.py
├── requirements.txt
├── .gitignore
└── README.md

• data/train/: 15,021 training images across 500 classes (29-31 images per class)
• data/test/: 10,014 test images

Getting Started

Requirements

Install the required packages:

pip install -r requirements.txt

Hyperparameters

The following arguments can be adjusted via command line to experiment with different settings:

argument option	default	Description
traindata_dir	./data/train	Path to training data dir
traindata_info_file	./data/train.csv	Name of the training info file
testdata_dir	./data/test	Path to test data dir
testdata_info_file	./data/test.csv	Name of the test info file
val_ratio	0.2	Validation data ratio
model_type	timm	Model type
model_name	swin_s3_base_224	Model name
pretrained	True	Use pretrained model
num_classes	500	Number of classes
device	cuda	Device to use(gpu or cpu)
batch_size	64	Batch size
num_workers	4	Number of workers
epochs	10	Number of epochs
lr	0.001	Learning rate
weight_decay	0.0001	Weight decay coefficient
label_smoothing	0.0	Label smoothing factor
model_dir	./train_result	Path to save training results (.pt)
output_dir	./output	Path to save prediction results (.csv)

Training Step

To train a single model, run following command.

python train.py --model {model_name} --batch_size {batch_size} --lr {learning rate} --epochs {epochs} --label_smoothing {label smoothing}

To train multiple models, run following shell script file.

bash train.sh

Inference Step

To test model, run following command.

python inference.py --model_name {model_name}

Ensemble

To perform ensemble prediction:

python ensemble.py