Skip to main content

Training with Existing Materials

This guide explains how to train the system with existing materials, focusing on dataset import capabilities and model training.

Dataset Import Options​

Our system supports multiple methods for importing existing material data:

1. Hugging Face Datasets​

The system can import datasets directly from Hugging Face:

// Example dataset: Material in Context (MINC-2500)
// https://huggingface.co/datasets/mcimpoi/minc-2500_split_1

Import Process:

  1. Navigate to the Admin Panel → Datasets section
  2. Select "Import External Dataset"
  3. Enter the Hugging Face dataset ID: mcimpoi/minc-2500_split_1
  4. Configure import options:
    • Map dataset categories to system material types
    • Set maximum images per class (recommended: 250-500)
    • Enable/disable metadata import
    • Select specific classes to import (optional)
  5. Start the import process

The system will:

  • Download dataset samples from Hugging Face
  • Extract category and image information
  • Map external dataset fields to internal metadata fields
  • Create properly categorized material samples

2. CSV/Structured Dataset Import​

For CSV or structured datasets with mapped fields:

  1. Prepare a CSV with columns mapping to system fields
  2. Include material type, properties, and image file paths
  3. Upload through the Admin Panel → Datasets → Import CSV

3. Local Directory Import​

For datasets stored in local directories:

  1. Organize materials by category in subdirectories
  2. Specify the root directory path during import
  3. The system will analyze the structure and suggest mappings

Field Mapping System​

The system includes a flexible field mapping capability:

  • Maps external dataset fields to internal metadata fields
  • Provides predefined mappings for common material datasets
  • Allows custom mapping configuration
  • Handles automatic property extraction

Training Models with Imported Materials​

After importing materials, you can train recognition models:

  1. Navigate to Admin Panel → Models → Training
  2. Select the imported dataset from available datasets
  3. Configure training parameters:
    • Base model (ResNet, MobileNet, etc.)
    • Batch size and learning rate
    • Number of epochs
    • Transfer learning settings
  4. Start training
  5. Monitor progress in real-time
  6. Evaluate model performance with validation metrics

Detailed Training Process​

The training process involves several sophisticated steps:

  1. Model Initialization

    • Base models are loaded dynamically from ML framework libraries (not stored in our repo)
    • TensorFlow or PyTorch pre-trained architectures (MobileNetV2, ResNet, EfficientNet)
    • Classification layers are replaced with custom layers for material recognition

  2. Transfer Learning Optimization

    • Initial layers of base models are frozen to preserve general features
    • Only the top classification layers are trained initially
    • Sparse categorical cross-entropy loss is used for classification tasks
    • Adam optimizer with carefully tuned learning rate (typically 0.0001)
    • Later training phases gradually unfreeze more layers for fine-tuning

  3. Training Enhancement Techniques

    • Early stopping with validation loss monitoring
    • Learning rate reduction on plateau
    • Data augmentation specific to material properties
    • Regularization to prevent overfitting (dropout, L2)

  4. Model Storage and Versioning

    • Trained models are saved with metadata in the output directory structure:
    /models/
      ├── /
      │   ├── model.h5 (or .pt for PyTorch)
      │   ├── metadata.json
      │   ├── training_history.json
      │   └── hyperparameters.json
    • Complete training history is preserved for analysis
    • Models are versioned for tracking improvements

  5. Vector Database Integration

    • The trained models generate embeddings for all materials
    • These embeddings (not the models themselves) are stored in the vector database
    • FAISS indexing enables efficient similarity search
    • Each material's embedding links to knowledge base entries through material IDs

Example: Training with MINC-2500​

For the MINC-2500 dataset, which contains material images across 10 categories:

  1. Import the dataset using the Hugging Face importer:

    • Use ID: mcimpoi/minc-2500_split_1
    • The system will automatically map categories like 'wood', 'metal', 'fabric', etc.
    • System will assign appropriate material types based on content

  2. Train a material recognition model:

    • Use transfer learning on a pre-trained image model
    • Configure 10-20 epochs for good results
    • Enable data augmentation for improved generalization
    • Set learning rate to ~0.0001 for stable training
    • Apply sparse categorical cross-entropy loss for classification

  3. Evaluate results:

    • The system will display accuracy per material category
    • Review performance metrics to identify areas for improvement
    • Test with sample images to verify recognition quality
    • Analyze confusion matrix to understand misclassifications
    • Review embedding quality metrics for similarity search applications

  4. Model Deployment:

    • The trained model is automatically versioned and stored
    • Embeddings are generated for all materials in the dataset
    • Vector database is updated with new embeddings
    • Recognition system starts using the new model immediately

Implementation Notes​

  • The dataset importer supports automatic detection of dataset structure
  • Field mapping can be customized for specific material types
  • Metadata extracted from datasets is stored for future reference
  • Training parameters are automatically optimized based on dataset characteristics

For more advanced training techniques, see the ML Training Documentation.