The advanced settings menu gives access to additional parameters to configure MIPAR’s Deep Learning model training. These settings can help tune the trainer for model accuracy, model robustness, and training performance. To access the Advanced Settings menu, select the training settings icon (wrench) and select the Advanced Settings button (1).
Advanced Settings Window
 |
 |
 |
 |
2. Model Parameters
Model Type
Select the type of deep learning model to train:
- SegNet / U-Net: Semantic segmentation models that classify each pixel in the image.
- YOLO: Object detection model that identifies objects with bounding boxes. Boxes can be passed to Spotlight as prompts to extract perform semantic segmentation of the detected objects.
Switching from SegNet to U-Net will reduce the selected Initial Learning Rate by a factor of 10, while switching from U-Net to SegNet will increase the selected Initial Learning Rate by a factor of 10.
Model Depth (SegNet/U-Net only)
Number of network layers (building blocks). More layers can help the model learn larger, more complex features, but risk overfitting, or memorizing the data set instead of general learning, and may lose the fine details within the features.
Filter Size (SegNet/U-Net only)
Window for how many surrounding pixels the model considers when analyzing a pixel. A larger window will result in more awareness of the neighboring environment but may lose the fine details within the features.
Model Complexity (YOLO only)
When YOLO is selected as the Model Type, Model Complexity replaces Model Depth and Filter Size.
| Basic | Advanced | |
|---|---|---|
| Training Time | Faster | Slower |
| Application Speed | Faster | Slower |
| Memory Usage | Lower | Higher |
| Best For | Most applications, simpler objects | Complex objects, overlapping instances, maximum accuracy |
Recommendation: Start with Basic – it works for 80% of applications. Only upgrade to Advanced if detection accuracy is insufficient after proper training.
3. Data Augmentation: Geometry
All data augmentation options apply to both segmentation and YOLO models. For YOLO, bounding box coordinates are automatically transformed along with the images.
Rotation
Rotate the images a random number of degrees clockwise or counterclockwise up to the specified value to simulate variation in feature orientation.
Shift
Translate the images a random number of pixels in the x and y directions up to the specified value to simulate variation in feature location.
Min and Max Scale Factors
Resize the image in a range between the minimum and the maximum to simulate variation in feature size or resolution.
Reflection
Randomly flip the image over the specified axes to simulate variation in feature reflected duplication.
4. Data Augmentation: Color and Grayscale Intensity
Color Intensity
Randomly shift the color channel intensities to simulate variation in feature color. Channel intensity shift can be defined individually, using the R, G and B entries.
Grayscale Intensity
Randomly shift the overall image brightness to simulate variation in feature exposure.
5. Training Options: General Settings
Batch Size
Define number of images (or tiles, if tiling is used) sent through the network trainer at once. A lower number can result in suboptimal detection accuracy and GPU resource utilization, a higher number ran result in overfitting (data memorization) and GPU resource overflow.
Image Order Shuffle
Mitigates image dropping if the batch size is not divisible into the number of images (or tiles, if tiling is used). This setting also helps reduce overfitting.
Tiles X and Y
Define the number of tiles the images should be divided into before being passed into the trainer. Defined by the number of tiles in the X and Y axis. Helps increase the training data set and can be used to reduce GPU memory usage.
The default is 1×1 for segmentation models (SegNet/U-Net) and 2×2 for YOLO models.
Size Factor
Resizes the images before passing them into the model trainer to reduce GPU memory usage.
Processor
Select the training hardware. A dedicated Nvidia GPU is strongly recommended. See Deep Learning hardware requirements.
Epochs
Number of times to iterate through all images in the training set. If the value is too low, the model will not converge to anything useful. If the epochs are too high and early stop is not turned on, the training can run longer than needed or risk overfitting.
6. Training Options: Learning Rate
Initial Learning Rate
Defines how much a model changes each time it is presented with training data. Values that are too high may result in failure to converge.
Progressively Drop Learning Rate
Recommended for the model training to reach maximum accuracy, but can result in longer training times.
Learning Rate Drop Period
Defines how often the learning rate is adjusted, expressed in number of epochs.
Learning Rate Drop Factor
Defines the rate at which the Learning Rate is reduced.
7. Training Options: Validation
Validation
Enables checking of model accuracy on all images every specified number epochs, at the cost of increased training time. Model accuracy is reported in the Session Processor once the training has started. (9)
Validation Frequency
Defines how often the model accuracy is checked, expressed in number of epochs.
Output Model
Best: returns the model that resulted in the highest accuracy score.
Last: returns the model that was returned at the end of the training.
Early Stop Patience
Will stop the model training early if the model training accuracy has reached an asymptote, defined by a number of validations to be considered when concluding a maximum accuracy.
8. Training Options: Preprocessing Recipe
Load Recipe
Select a recipe to apply to the images before training begins. For example, to reduce noise or increase contrast of the features of interest. Recipes cannot have Layers or Measurements. It is not recommended for the recipe to have a crop or resize step.
Model Accuracy
What the Accuracy Metric Measures
The Accuracy metric during Deep Learning training provides a measure of how well your model is performing. The meaning varies by model type:
- SegNet/U-Net: Calculates the percentage of labeled pixels that your model correctly classifies.
- YOLO: Displays mAP50 (mean Average Precision at 50% IoU threshold), which measures how accurately the model detects and locates objects.
How Accuracy Is Calculated
For Segmentation Models (SegNet/U-Net)
With validation enabled, after each specified number of epochs:
- Your model processes each pixel in your dataset and produces a set of scores (one for each possible class)
- For each pixel, the class with the highest score is selected as the predicted class
- For pixels which have been labelled for training, the system compares this predicted class with the actual class label
- The Accuracy Metric counts how many predictions match the actual labels
- This count is divided by the total number of labelled pixels and multiplied by 100 to give you a percentage
For example, if your model correctly classifies 900 out of 1,000 labelled pixels, the Accuracy Metric would report 90%.
For YOLO Models
The mAP50 metric evaluates object detection by measuring how well predicted bounding boxes match the ground truth boxes. A prediction is considered correct if its overlap (IoU) with the ground truth is at least 50%. The metric combines precision and recall across all classes.
Interpreting the Accuracy Percentage
The accuracy percentage shown after each specified number of epochs represents how well your model recognizes patterns in the data it has seen. A higher percentage indicates better performance, with 100% meaning perfect classification.
However, this accuracy is calculated on the same data used for training. This means the metric shows how well your model has learned the training data, but doesn’t necessarily indicate how well it will perform on new, unseen data. The ability to validate on data excluded from the training set is a planned feature for a future release.
Need more help with this?
Chat with an expert now ››
