Skip to main content

Advanced Settings

Hyperparameters​

The Hyperparameters panel defines how a model learns during training — including its architecture, optimization behavior, and validation schedule. These parameters can be adjusted before starting a training job.

Model Architecture​

Select the network type and size that best match your dataset and task.

Options​

Base CNN (Recommended default) – Standard convolutional neural network with configurable parameters.

Adaptive CNN – Automatically determines model size, regularization, and dropout based on dataset size (number of samples × bands).

Base CNN​

When using the standard CNN architecture, the following can be configured manually:

Model size: Light, Medium, or Large

  • Light – Lightweight model with fewer parameters; faster training and lower memory usage. Equivalent to legacy Baseline_Light model. Ideal for smaller datasets or quick experiments.
  • Medium – Balanced model with moderate capacity; good trade-off between performance and training time.
  • Large – Full-capacity model with more layers and parameters; best for complex datasets but requires more GPU memory and training time. Equivalent to legacy Baseline model.

Adaptive CNN​

Automatically selects model size, dropout, and regularization to balance model capacity with dataset scale.

  • You can still override these values if you need fine control

Regularization and Dropout​

Dropout rate: Adds robustness by randomly disabling neurons each step.

L2 Regularizer: Reduces overfitting by penalizing large weights (common default: 0.0001).

Training Parameters​

Batch size​

Number of samples per training step. (Default: 512)

  • Larger values train faster but require more GPU memory

Max epochs​

Number of complete passes through the dataset.

  • Higher values may improve accuracy but can increase overfitting risk

Validation interval (epochs)​

Defines how often validation runs during training. (Default: 1)

  • For long runs (≥20 epochs), set this to 2+ to reduce computation time

Spatial bins​

Groups nearby pixels for architectures that support spatial context. (Default: 1)

Optimization Settings​

Optimizer​

Controls how the model updates weights during training.

Available options:

  • Adam (Adaptive Moment Estimation) – Recommended default; adapts learning rates for each parameter
  • SGD (Stochastic Gradient Descent) – Classical optimizer; useful for stable convergence in large datasets
  • AdaGrad (Adaptive Gradient) – Suitable for sparse data; decreases learning rate over time
  • AdaDelta (Adaptive Delta) – Robust to noisy gradients; maintains adaptive step sizes
  • Adamax (Variant of Adam) – More stable when gradients are large
  • RMSProp (Root Mean Square Propagation) – Works well for non-stationary problems
  • FTRL (Follow the Regularized Leader) – Used for strongly regularized models or online learning

Learning rate​

Recommended default: 0.0001

Defines the magnitude of updates applied to model weights.

Learning rate decay type​

Determines how the learning rate decreases during training.

Available options:

  • Fixed – Constant learning rate throughout training
  • Exponential – Gradually decays at a fixed exponential rate
  • Polynomial – Decays polynomially from initial to final value
  • Cosine – Cyclic cosine-shaped decay curve
  • Cosine Restarts (Recommended default) – Cosine schedule that resets periodically to encourage convergence
  • Inverse Time – Decays inversely with iteration count (slower than exponential)

Threshold Optimization​

Defines how the decision threshold is determined after training (used for classification and target detection).

Available options:

  • ROC (on validation) – Recommended default – Optimizes threshold based on validation ROC curve for best balance of true/false positives
  • ROC (on test) – Uses test dataset ROC to compute threshold; useful for evaluating final model performance

Best Practices​

For optimal results with most hyperspectral datasets, we recommend the following configuration:

  • Model Architecture: Use Base CNN with Light model size for efficient training and good generalization
  • Training Duration: Train for 20 epochs to ensure the model has sufficient iterations to learn patterns without overfitting
  • Optimizer: Use Adam optimizer (default) for adaptive learning rate adjustment and stable convergence
  • Learning Rate: Set to 0.0001 for balanced training speed and stability
  • Learning Rate Decay: Use Cosine decay type to help the model escape local minima and improve convergence by gradually reducing the learning rate in a cyclic pattern
  • Regularization:
    • Dropout: Set to 0.1 (10%) to prevent overfitting by randomly disabling neurons during training
    • L2 Regularizer: Set to 0.0001 to penalize large weights and improve model generalization
  • Batch Size: Keep default (512) or adjust based on available GPU memory
  • Monitoring: Watch validation metrics closely to detect overfitting early

General Tips​

  • Start with recommended settings above, then fine-tune based on your specific dataset characteristics
  • Increase epochs gradually if validation metrics show continued improvement
  • Monitor GPU memory usage and reduce batch size if you encounter out-of-memory errors
  • Compare model versions using the Compare tab to track performance improvements across configurations

Need Help?​

If you're unsure about which advanced settings to modify or need guidance on optimizing your model for a specific use case, our team is here to help!

Contact us for guidance:

  • 📧 Email: support@metaspectral.com
  • 💬 Chat: Use the Clarity AI assistant in Spectral Explorer
  • 📋 Consultation: Schedule a technical consultation for complex projects