Model Configuration Reference

This page is a detailed reference for the Models Configuration dialog.

If you want a shorter walkthrough, see Model Configuration.

How to use this reference

• Use the section that matches the tab you are editing
• Fields only appear when they are relevant for the current model type

Configuration tab

General

Setting	What it means	When to change it
Name	The model name shown throughout the project.	Use a clear name that matches the job of the model.
Primary	Makes this the main model used as the project entry point.	Turn this on for the main model users should start from.
Input width	Width of the images fed into the model.	Change only if you know you need a different input size.
Input height	Height of the images fed into the model.	Change only if you know you need a different input size.
Description	Short explanation of what the model is meant to do.	Use this to keep several models easy to understand.

Base

Setting	What it means	Notes
Purpose	Chooses whether the model is for locating, classifying, or identifying.	This choice affects which other tabs and options appear.
Accuracy	Simple speed-versus-quality control.	Higher settings usually favor a slower, more accurate setup.
Bounding boxes	Lets the model return object locations.	Not available for Identify models.
Key points	Lets the model return landmarks.	Not available for Identify models.
Multiple categories	Lets the model work with more than one class.	Only available for Classify models.
Based on	Starting point for training.	COCO 2017 can preload categories and key points.

Purpose options

Purpose	Best used for
Locate	Finding where objects are in an image
Classify	Assigning classes, optionally with boxes or key points
Identify	Matching previously found objects to identities

Inference tab

Use this tab to control how the model behaves when it is run on images.

Accuracy Thresholds

Setting	What it means	Practical guidance
Base Accuracy (%)	The model's default accuracy point.	This is the center point used before the global find-accuracy slider shifts the threshold.
Lower Limit (%)	Lowest accuracy the model can use when the global find-accuracy slider moves downward.	Keep this lower than the base value.
Upper Limit (%)	Highest accuracy the model can use when the global find-accuracy slider moves upward.	Keep this higher than the base value.

These three values work together with the global find-accuracy slider in the Engine tab of the menu. From a user perspective, the model config sets the allowed range, and the global slider moves within that range at runtime.

Limits

Setting	What it means	Notes
Maximum Detections	Caps how many detections are returned.	Only useful when the model can return multiple categories or multiple detections.

Inference and Non-Max Suppression

These settings are relevant for models that use bounding boxes.

NMS stands for Non-Maximum Suppression. It removes duplicate overlapping detections by keeping the strongest result and suppressing weaker boxes that likely refer to the same object.

Setting	What it means	Practical guidance
Use NMS	Turns duplicate cleanup on or off. When enabled, overlapping detections are compared and weaker duplicates are removed.	Turn this on when the model often returns several boxes for the same object.
Pre NMS Top K	Number of candidate detections kept before NMS runs.	Raise it when crowded images need more candidates to survive until cleanup.
NMS IoU Threshold	Controls how much overlap is allowed before one detection suppresses another. IoU means Intersection over Union, a measure of box overlap.	Lower values remove overlaps more aggressively, higher values keep more overlapping boxes.
Class-Agnostic NMS	Lets detections suppress each other across different categories instead of only within the same category.	Use it when different classes are close alternatives and should compete with each other.
NMS Overlap Groups	Lets you define specific groups of categories that should suppress each other when class-agnostic NMS is off.	Useful when only certain categories should compete, such as similar object types.

NMS Overlap Groups

Setting	What it means
Group name	Name used when assigning categories to a group
IoU Threshold	Overlap threshold used for that group

Categories tab

Use this tab to define the classes the model can detect.

Per-category fields

Setting	What it means	Notes
Value	Numeric id for the category.	Keep values stable once you have labeled data.
Name	Display name of the category.	Use short, consistent names.
Show ROI	Controls whether the category’s boxes or key points are shown.	Only relevant when the model uses boxes or key points.
Per-category accuracy	Lets this category use its own accuracy range instead of the model-wide default.	Useful when one class needs stricter or looser filtering with the global find-accuracy slider.
Base / Lower / Upper	Category-specific accuracy values.	These appear after enabling per-category accuracy.
NMS IoU Threshold	Per-category overlap threshold.	Available when NMS is in use.
NMS Groups	Assigns the category to one or more overlap groups.	Available when overlap groups exist.
Weight	Per-category training weight for CenterNet heatmap loss.	Available when per-class weighting is enabled in training.
Observation count	Shows how many labeled observations exist for the category.	Useful when checking class balance.

Compute button for class weights

When per-class weighting is enabled in the Training tab, the Categories tab also shows:

Setting	What it means
Method	Chooses how the app computes class weights from the current category frequencies
Compute	Fills the category weights automatically

Use this when some categories have much less training data than others.

Key points tab

This tab appears when key points are enabled.

Categories that detect key points

Only shown for Classify models.

Setting	What it means
Category list	Chooses which categories produce key points

List of key points

Setting	What it means	Notes
Key point labels	Ordered list of the landmarks the model predicts	The order matters and should stay stable once data is labeled.

Meshes

Meshes define visual or structural connections between key points.

Setting	What it means
Path	Treats the selected points as a connected path
Triangles	Treats the selected points as a triangle-based mesh
Key points in mesh	Chooses which key points belong to the mesh

Attributes tab

Attributes let you compute extra information from the model output.

Shared vs per-category attributes

Section	What it means
Shared	Attributes used across the whole model
Per-category	Attributes used only for specific categories

Common fields

Setting	What it means
Attribute type	Chooses how the attribute value is produced
Name	Label for the attribute
Skip frames	Cadence control: sets how many frames to skip between calculations. 0 means every frame. 4 means calculated every 5th frame. Use this to reduce the cost of attributes that do not need to update every frame.
Requires categories	Multi-select that restricts when this attribute runs. When one or more categories are selected, the attribute is only calculated when at least one of those categories is also detected in the same image or frame. Leave empty to run the attribute unconditionally. Only available when the model has Multiple categories enabled and at least one category defined. For category-specific attributes, the owning category is excluded from the list.

Attribute types

Type	What it does	Extra fields
Color	Extracts one or more dominant colors	Number of colors, Max colors
Another model	Uses another model to calculate the value	model selection, optional clip/alignment
Identification	Uses an identification model to assign identity	model selection, optional clip/alignment
Border width calculator	Measures the pixel width of a circular bright border by sampling 360 radial profiles. Outputs a sections array; each section contains minWidthPx, maxWidthPx, validRatio, startDegrees, endDegrees, minDegrees, maxDegrees, startWidthPx, and endWidthPx. Result is omitted when the border cannot be measured reliably.	Min valid profile ratio, optional clip/alignment
Angle	Calculates face or object angle from 4 key points	key point selection
Bearing	Calculates direction from 5 key points	key point selection
Adjacent angle	Calculates the angle between two lines defined by key points	key point selection

Clip and alignment options

These are available for model-based attributes, Border width calculator, and for training-time crops.

Setting	What it means	Requirements
Clip	Crops to the bounding box before running the next step	The source model must provide boxes
Align with key points	Rotates or aligns the crop using selected key points	The source model must provide key points
Align points	Chooses the points used for alignment	Only relevant when alignment is enabled
Margin	Adds extra space around the crop	Only relevant when clipping is enabled

Training tab

The Training tab is split into several sections.

Training tab: Data preparation

Setting	What it means	Practical guidance
Prepare data	Rebuilds the train/validation split	Use this when you want a fresh split.
Percentage training data	Percentage used for training rather than validation	Keep enough validation data to judge training quality.
Level categories	Tries to balance categories during data preparation	Useful when some classes have much less data.

Training tab: Clip settings

Setting	What it means
Clip	Crop training samples to the object box
Align with key points	Rotate training samples using key points
Align points	Which key points to use for alignment
Margin	Extra crop padding around the box

Training tab: Data augmentation

Setting	What it means
Rotate	Random rotation augmentation
Rotation	Maximum rotation amount
Scale	Random resizing augmentation
Range for scale	Maximum scale change
Flip Horizontal	Horizontal mirroring
Flip Vertical	Vertical mirroring
Translate X	Horizontal movement augmentation
Translate Y	Vertical movement augmentation
Brightness	Brightness variation
Contrast	Contrast variation
Saturation	Saturation variation
Hue	Hue variation
Nr of combinations	Number of augmentations combined on a single image

Training tab: Core training settings

Setting	What it means	Practical guidance
Batch size	Number of samples per training step	Larger values are faster if your hardware can handle them.
Epochs	Minimum number of passes through the training data	Start moderate, then review results.
Max epochs	Hard upper limit for training	Mostly relevant when patience is greater than 0.
Optimizer	Training optimizer	Adam is a common default.
Momentum	Extra momentum used by SGD	Only relevant for SGD.
Patience	Stops training after enough non-improving evaluations	Lower values stop sooner.
Weight decay	Regularization strength for AdamW	Only relevant for AdamW.
Clip norm	Gradient clipping limit	Leave empty unless you are dealing with unstable training.
Fine tune	Trains more of the model instead of only the final layer	Usually best left on when boxes or key points are involved.
Validation metric	Metric used for early stopping and best checkpoint selection. Single-label ResNet models always use accuracy. Multi-label models support accuracy, exactmatch, precision, recall, and f1.	Classify models with Multiple categories only. Defaults to accuracy for single-label and f1 for multi-label.
Multi-label threshold	Confidence threshold applied to sigmoid scores when computing multi-label training and validation metrics. Does not affect the loss function, only the reported numbers.	Classify models with Multiple categories only. Default 0.5.

Training tab: Learning rate

Setting	What it means	Notes
Learning rate decay	Overall schedule used during training	Cosine is the simplest default.
Initial learning rate	Starting learning rate	Disabled when differential learning rates are enabled.
Decay steps	How long the cosine schedule runs before reaching its minimum	Mainly relevant for cosine schedules.
Differential learning rates	Uses different learning rates for the head and backbone	Advanced tuning option.
Head learning rate	Learning rate for the prediction head	Usually higher than the backbone rate.
Backbone learning rate	Learning rate for the backbone	Usually lower than the head rate.
Warmup epochs	Gradually ramps the learning rate at the start	Useful for smoother starts.
Restart mode	How the first cosine restart cycle is defined	Only appears for cosine restart schedules.
First decay steps	Absolute first-cycle length	Cosine restart mode only.
Restart every epochs	Defines restart timing in epochs	Cosine restart mode only.
Warmup each restart	Adds warmup at each restart	Cosine restart mode only.
Cycle multiplier	Changes cycle length over time	Cosine restart mode only.
Decay	Changes how much each restart peak drops	Cosine restart mode only.
Minimum	Minimum learning-rate fraction	Cosine restart mode only.

Training tab: Advanced training parameters

These are the advanced fields shown in the Training parameters card.

Backbone fine-tuning

Setting	What it means
Freeze backbone epochs	Keeps the backbone frozen for the first part of training
Unfreeze stages/interval	How many backbone stages are opened each time
Unfreeze interval	Number of epochs between unfreeze steps
Stop unfreezing at stage	Limits how far unfreezing goes
Train backbone BatchNorm	Allows backbone batch norm layers to update
BN unfreeze epoch	Delays when backbone batch norm starts updating

Detection and key point loss settings

These fields appear only when the model uses boxes or key points.

Setting	What it means
Object center loss w	Weight for the object-center heatmap loss
Size loss weight	Weight for bounding-box size regression
Box loss weight	Weight for box offset regression
Keypoint loss weight	Weight for the key point heatmap loss
Offset loss weight	Weight for per-keypoint sub-cell offset regression
Center-to-joint loss w	Weight for the center-to-joint regression term. Only active when Keypoint head mode is center_to_joint.
Direction pair vector loss w	Weight for the direction-pair vector regression term. Only active when Keypoint head mode is direction_pairs.
Center accuracy threshold	Threshold used when checking center prediction quality during training metrics
Focal alpha	Focal-loss alpha parameter
Focal beta	Focal-loss beta parameter
Heatmap bias	Initial bias for heatmap output
Gaussian radius	Fixed radius used for box heatmap targets when size-based Gaussians is off
Keypoint Gaussian radius	Fixed radius used for key point heatmap targets
Gaussian min overlap	Minimum IoU overlap used by the size-based Gaussian radius heuristic. Only active when Size-based Gaussians is on.
CenterNet debug images	Saves debug images during training
Size-based Gaussians	Adapts Gaussian size to object size using a CenterNet IoU heuristic instead of a fixed radius
Box upsample filters	Advanced CenterNet FPN decoder channel configuration
Box head mode	Selects the box prediction contract: center is the standard mode, center_radius_xy predicts half-extents, extreme_points predicts top/bottom/left/right point heatmaps, center_rotated predicts an angle so tilted objects get tight, rotated boxes.
Keypoint head mode	Selects how key points are predicted: offset is the direct per-keypoint path, center_to_joint conditions each joint on the decoded object center, direction_pairs predicts pair-center vectors. Only meaningful when key points are enabled.
Keypoint decode radius	Local search radius used when refining center_to_joint predictions from the heatmap. Only active when Keypoint head mode is center_to_joint.

Keypoint augmentation

These settings appear when the model uses key points.

Setting	What it means
Horizontal flip pairs	Pairs of key point indices swapped during horizontal flip augmentation. Configure this whenever left/right semantics matter, such as body joints or facial landmarks with side-specific labels.
Vertical flip pairs	Pairs of key point indices swapped during vertical flip augmentation. Configure this whenever top/bottom semantics matter, such as ordered endpoints or vertically paired landmarks.

Synthetic boxes

These settings appear when the model uses key points and no box labels are supplied. Synthetic boxes are generated from the visible keypoint bounds so the box head receives supervision even when no explicit boxes exist in the training data.

Setting	What it means
Synthetic boxes	Enables generating box targets from visible keypoint bounds when a sample has keypoints but no box label.
Box padding ratio	Relative padding added around the keypoint envelope. 0.05 is the default.
Box padding pixels	Minimum absolute padding in pixels added around the envelope.
Box min size pixels	Minimum generated box height and width before padding is applied.

Regularization

Setting	What it means
Dropout	Amount of dropout used during training
Label smoothing	Softens target labels to reduce overconfidence

Label smoothing is only available for basic ResNet classification without multiple categories, boxes, or key points.

Training tab: CenterNet per-class weighting

This section appears for models that use boxes or key points.

Setting	What it means
Enable per-class weights	Turns on class-based weighting for the heatmap loss
Normalize to mean 1.0	Keeps the average weight centered around 1
Clip weights	Limits extreme weights
Warmup epochs	Gradually applies class weights over the first epochs
Clip min	Lowest allowed class weight when clipping is enabled
Clip max	Highest allowed class weight when clipping is enabled

Training tab: Genetic optimization

Setting	What it means
Use genetic optimization	Enables automatic search over selected training values
Max search time (s)	Maximum time spent searching
Consecutive worse limit	Stops the search after enough non-improving attempts
Field	Training field to optimize
Min	Lowest value to try
Max	Highest value to try
Step	Increment used while searching

Parameter dependencies

Some settings only take effect when another setting has a specific value. The most important dependencies are listed below.

Inference

Setting	Requires
Base Accuracy, Lower Limit, Upper Limit, Accuracy Delta, per-category accuracy ranges	Threshold must not be set. When Threshold is present all accuracy-range fields are inactive.

Training

Setting	Requires
Multi-label threshold, Validation metric (exactmatch / precision / recall / f1)	Multiple categories = on
Label smoothing	Multiple categories = off — rejected when multiple categories is enabled
Head learning rate, Backbone learning rate	Differential learning rates = on
Weight decay	Optimizer = AdamW — silently unused for Adam and SGD
Momentum	Optimizer = SGD — silently unused for Adam and AdamW
Warmup each restart	Learning rate decay = cosine_restarts and Warmup epochs > 0
First decay steps, Restart every epochs, Warmup each restart, Cycle multiplier, Decay, Minimum	Learning rate decay = cosine_restarts
Freeze backbone epochs, Unfreeze stages/interval, Unfreeze interval, Stop unfreezing at stage	Fine tune = on
BN unfreeze epoch	Train backbone BatchNorm = off — when BatchNorm is always trainable an unfreeze epoch has no effect
Keypoint head mode, Keypoint Gaussian radius, Keypoint decode radius, Center-to-joint loss w, Direction pair vector loss w, Offset loss weight, keypoint flip pairs, synthetic boxes	Key points must be enabled (nrKeyPoints > 0)
Center-to-joint loss w, Keypoint decode radius	Keypoint head mode = center_to_joint
Direction pair vector loss w	Keypoint head mode = direction_pairs
Gaussian min overlap	Size-based Gaussians = on
Box padding ratio, Box padding pixels, Box min size pixels	Synthetic boxes = on
Class weights, Normalize to mean 1.0, Clip weights, Warmup epochs (class weights), Clip min, Clip max	Enable per-class weights = on
Clip min, Clip max	Clip weights = on and Enable per-class weights = on

Attributes

Setting	Requires
Requires categories	Multiple categories = on and at least one category defined

Mutually exclusive pairs

Settings	Note
Threshold vs accuracy range fields	Presence of Threshold makes all accuracy-range fields inactive in both CenterNet and ResNet Classify
Label smoothing vs Multiple categories	Label smoothing is rejected when Multiple categories is enabled
First decay steps vs Restart every epochs	Both control the first cosine-restart cycle length; First decay steps wins when both are set

Related

• Model Configuration
• Getting started