Advanced model settings
Open a draft model version, select Overview, and expand Advanced Settings. A setting is editable only while the version is a draft.
Sections
| Section | What it controls |
|---|---|
| Model initialization | Random, resumed, or transferred weights |
| Hyperparameters | Architecture, size, epochs, optimizer, learning rate, and regularization |
| Thresholds | Threshold policy for non-regression outputs |
| Loss function | Type-specific optimization objective |
| Training data generation | Synthetic mixing or ground-truth batches for Unmixing and Target Detection |
| Preprocessing | Saved spectral transforms |
| Inference | Inference optimizer |
| Early stopping | Metric, minimum change, patience, and start epoch |
| Data Augmentation | Optional randomized spectral transforms |
Only sections and choices applicable to the selected model type are shown.
Initial defaults
A newly created draft currently starts with:
| Setting | Default |
|---|---|
| Model architecture | Base CNN |
| Model size | Light |
| Batch size | 512 |
| Max epochs | 10 |
| Optimizer | Adam |
| Learning rate decay type | Fixed |
| Learning rate | 0.001 |
| Dropout rate | 0 |
| L1 and L2 regularizers | 0 |
| Preprocessing | None |
| Data augmentation | Disabled |
| Inference optimizer | TensorRT |
The values displayed in the draft are authoritative if they differ from this page.
Hyperparameters
The architecture choices are Base CNN and Adaptive CNN, with Light, Medium, and Large model sizes. Batch-size presets are 32, 64, 128, 256, and 512.
Optimizer choices include AdaDelta, AdaGrad, Adam, FTRL, SGD, and RMSProp. Learning-rate schedules include Fixed, Exponential, Polynomial, Cosine, Cosine Restarts, and Inverse Time.
Max epochs applies to training from scratch and transfer initialization. A resumed checkpoint uses Additional epochs under Model initialization instead. Validation interval (epochs) controls how often validation is run.
Larger models and batches can require more memory. More epochs can overfit as well as improve a model; use validation evidence when changing these values. Evaluate the selected, locked configuration on untouched test data afterward.
Type-specific settings
Classification offers cross-entropy and focal loss. Regression offers MSE, MAE, Huber, and log-cosh. Unmixing and Target Detection expose their own abundance, focal, false-positive, and spectral-loss controls.
Threshold policies include type-appropriate choices such as maximum F1, recall at minimum precision, a false-positive budget, or a manual threshold. Argmax is a class-decoding rule rather than a scalar threshold.
For Classification, current test predictions and deployed class maps use argmax. Non-argmax Classification policies affect validation threshold and curve metrics, but they do not change those class decisions. Treat this as a current product limitation rather than an operational threshold setting. For Unmixing and Target Detection, select evaluation thresholds according to the cost of false positives and missed detections; a selected threshold does not guarantee performance on new imagery.
Prefer policies that choose thresholds from Validation data. A policy that optimizes on Test data makes that Test bucket part of model selection, so its reported performance is no longer an independent final estimate.
Unmixing and Target Detection can use ground-truth labels or synthetic mixing batches. Mixing models and nonlinear parameters appear only when applicable.