wargoninnovation/wargon-clothing-classifier

Wargon Clothing Classifier

A Vision Transformer (ViT) based model for clothing classification, trained on secondhand clothing images. This model can classify 27 different types of clothing items with 73% accuracy.

Model Details

Model Description

This is a Vision Transformer model fine-tuned for clothing classification. It was developed to solve real-world clothing categorization challenges in secondhand fashion applications.

• Developed by: Wargon Innovation
• Model type: Image Classification
• Language(s): N/A (Vision model)
• License: Apache 2.0
• Finetuned from model: google/vit-base-patch16-224

Model Sources

• Repository: Wargon Innovation Clothing Dataset
• Base Model: google/vit-base-patch16-224

Uses

Direct Use

This model can be used for:

• Automatic clothing categorization in e-commerce
• Fashion inventory management
• Secondhand clothing marketplaces
• Fashion recommendation systems

Downstream Use

The model can be fine-tuned for:

• Specific clothing brand recognition
• Size estimation from images
• Style classification
• Multi-label clothing attribute detection

How to Get Started with the Model

from transformers import AutoImageProcessor, AutoModelForImageClassification
from PIL import Image
import torch

# Load model and processor
processor = AutoImageProcessor.from_pretrained("wargoninnovation/wargon-clothing-classifier")
model = AutoModelForImageClassification.from_pretrained("wargoninnovation/wargon-clothing-classifier")

# Load and preprocess image
image = Image.open("path_to_clothing_image.jpg")
inputs = processor(image, return_tensors="pt")

# Make prediction
with torch.no_grad():
    outputs = model(**inputs)
    predictions = torch.nn.functional.softmax(outputs.logits, dim=-1)

# Get top prediction
predicted_class_id = predictions.argmax().item()

Training Details

Training Data

The model was trained on the wargoninnovation/clothingdatasetsecondhand dataset, which contains over 30,000 images of secondhand clothing items across 34+ categories.

Data Preprocessing:

• Filtered classes with fewer than 10 samples to ensure robust train/validation splits
• Final dataset contains 27 clothing categories
• Images resized to 224x224 pixels
• Stratified train/validation split (80/20)

Training Procedure

Preprocessing

• Image Size: 224x224 pixels
• Normalization: ImageNet statistics
• Data Augmentation: Standard transformations applied

Training Hyperparameters

• Training regime: Mixed precision (fp16)
• Learning Rate: 2e-5
• Batch Size: 16
• Epochs: 6
• Optimizer: AdamW
• Weight Decay: 0.01
• Warmup Steps: 500
• Label Smoothing: 0.1

Hardware

• GPU: NVIDIA RTX 3060 (12GB VRAM)
• Training Time: ~1.5 hours

Evaluation

Testing Data, Factors & Metrics

The model was evaluated on a stratified validation set (20% of the filtered dataset).

Metrics

• Validation Accuracy: 73.0%
• F1 Score: 72.7%
• Precision: 72.8%
• Recall: 73.0%

Results

The model achieves balanced performance across major clothing categories, with particular strength in:

• Common items (T-shirts, Jeans, Dresses)
• Well-represented categories in the training data
• Clean product photography (as in the training dataset)

Clothing Categories

The model can classify the following 27 clothing types:

1. Blazer
2. Blouse
3. Cardigan
4. Dress
5. Hoodie
6. Jacket
7. Jeans
8. Nightgown
9. Outerwear
10. Pajamas
11. Rain jacket
12. Rain trousers
13. Robe
14. Shirt
15. Shorts
16. Skirt
17. Sweater
18. T-shirt
19. Tank top
20. Tights
21. Top
22. Training top
23. Trousers
24. Tunic
25. Vest
26. Winter jacket
27. Winter trousers

Limitations and Bias

Limitations

• Image Quality: Best performance on clean, well-lit product photos similar to training data
• Background: Optimized for images with minimal background distractions
• Viewpoint: Trained primarily on front-facing clothing images
• Categories: Limited to the 27 categories present in training data

Bias

• Data Source: Trained on secondhand clothing, may not generalize well to new/luxury items
• Cultural Bias: Dataset may reflect specific regional fashion preferences
• Class Imbalance: Some categories had limited representation even after filtering

Environmental Impact

• Hardware Type: NVIDIA RTX 3060
• Hours Used: ~1.5 hours training time
• Cloud Provider: N/A (Local training)
• Compute Region: Local

Technical Specifications

Model Architecture

• Base: Vision Transformer (ViT-Base/16)
• Parameters: ~86M parameters
• Input Size: 224x224x3
• Patch Size: 16x16
• Number of Classes: 27

Software

• Framework: PyTorch
• Libraries: HuggingFace Transformers, Datasets
• Training Libraries: Weights & Biases (W&B)

Citation

@misc{wargon_clothing_classifier_2024,
  title={Wargon Clothing Classifier: A Vision Transformer for Secondhand Fashion Classification},
  author={Wargon Innovation},
  year={2024},
  publisher={Hugging Face},
  howpublished={\url{https://huggingface.co/wargoninnovation/wargon-clothing-classifier}},
}

Model Card Authors

Wargon Innovation Team

Model Card Contact

For questions and feedback, please open an issue in the model repository or contact the Wargon Innovation team.