artifact-efficientnet

Artifact Classification Model v2 - Best Model Usage Guide

This directory contains the improved v2 artifact classification model with state-of-the-art performance for classifying museum artifacts by both object type and material.

Hosted Model

The best model is available on Hugging Face at: SpyC0der77/artifact-efficientnet

You can use the model directly from Hugging Face without downloading it locally.

Model Overview

The v2 model is an advanced multi-output neural network that predicts two attributes simultaneously:

• Object Name: The type/category of the artifact (e.g., "vase", "statue", "pottery")
• Material: The material composition (e.g., "ceramic", "bronze", "stone")

Key Improvements Over v1

• EfficientNet Backbone: Uses EfficientNet-B0 instead of ResNet-50 for better feature extraction
• Attention Mechanism: Includes an attention layer to focus on relevant features
• Advanced Training: Incorporates CutMix augmentation, Focal Loss, and mixed precision training
• Better Regularization: Uses dropout and batch normalization for improved generalization

Architecture & Usage

The v2 model uses an EfficientNet-B0 backbone with an attention mechanism for multi-output classification. It processes RGB images of artifacts and outputs predictions for both object type and material composition.

Input

• Format: RGB images (224×224 pixels after preprocessing)
• Preprocessing: Resize to 256×256, center crop to 224×224, normalize with ImageNet statistics

Output

• Object Classification: Predicts artifact type (e.g., "vase", "statue", "pottery")
• Material Classification: Predicts material composition (e.g., "ceramic", "bronze", "stone")
• Confidence Scores: Probability scores for each prediction
• Format: Dictionary with 'object_name' and 'material' logits

Model Architecture

ImprovedMultiOutputModel(
    backbone: EfficientNet-B0 (pretrained)
    attention: Linear(1280 → 512 → 1280) with Sigmoid
    object_classifier: Linear(1280 → 1024 → 512 → num_object_classes)
    material_classifier: Linear(1280 → 1024 → 512 → num_material_classes)
)

Input Requirements

• Image Size: 224×224 pixels (automatically resized and cropped)
• Format: RGB images
• Normalization: ImageNet normalization (mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])

Output Format

Returns a dictionary with:

• 'object_name': Logits for object classification
• 'material': Logits for material classification

Training Details

The model was trained with the following configuration:

• Dataset: ArtifactClfDurham/OrientalMuseum-white
• Training Split: 85% of data
• Validation Split: 15% of data
• Batch Size: 32
• Epochs: 20
• Optimizer: AdamW with differential learning rates
  • Backbone: 2e-4 (0.1× base LR)
  • Heads: 2e-3 (base LR)
• Augmentation: Advanced (CutMix, rotation, color jitter, Gaussian blur)
• Loss Function: Cross-Entropy (or Focal Loss if enabled)
• Scheduler: Cosine annealing with warmup

Advanced Training Features

• CutMix Augmentation: Randomly mixes image patches between samples
• Focal Loss: Addresses class imbalance (optional)
• Mixed Precision: Automatic mixed precision training for speed
• Gradient Scaling: Prevents gradient underflow
• Early Stopping: Saves best model based on validation accuracy

Troubleshooting

Common Issues

1. CUDA Out of Memory

   • Reduce batch size: --batch_size 8
   • Use CPU: Set device to "cpu"

2. Import Errors

   • Ensure all dependencies are installed
   • Check Python path includes project root

3. Model Loading Errors

   • Verify the model file path is correct
   • Ensure PyTorch version compatibility

4. Low Confidence Scores

   • Model may not be trained on similar artifacts
   • Check image preprocessing matches training setup

Performance Tips

• Use GPU for faster inference
• Process images in batches for efficiency
• Consider model quantization for deployment

Model Limitations

• Trained specifically on Oriental Museum artifacts
• May not generalize well to artifacts from other cultures/regions
• Performance depends on image quality and lighting
• Multi-output nature may have trade-offs between object and material accuracy