Advanced settings

Advanced settings are available for each model type and are hidden by default.

When you enable advanced settings, the following sections appear:

• Image classification: Base model and model hyperparameter.
• Object detection: Base model and model hyperparameter.
• Action detection: Model hyperparameter

Base model

You must select a base model when training for image classification with GoogLeNet. You can optionally choose a base model when training for object detection with Faster R-CNN, Detectron2, High resolution, and YOLO v3.

When you specify a base model, IBM® Maximo® Visual Inspection uses the information in the base model to train the new model. Selecting a base model can result in faster convergence. Therefore, fewer training iterations are required, especially if the base model shares characteristics with the data set that you are using to train the model.

You can choose a model that is included with IBM Maximo Visual Inspection, or you can choose your own model that you previously trained or imported. For models that were trained in IBM PowerAI Vision versions before 1.1.2, the list of associated objects or categories is not shown in the user interface. However, those models are still usable.

The new model must have the model type that the base model's network was trained on. For example, to train a new object detection model with Faster R-CNN, then the base model must be built on Faster R-CNN. Only viable models are listed in the Base model table.

Note: Starting in Maximo Visual Inspection 8.7, custom models are not supported. Custom models are still supported in Maximo Visual Inspection 8.6 and earlier versions.

Base models are not available for SSD, anomaly optimzed, action detection, Tiny YOLO v3 and custom models that are used for object detection, or custom models that are used for image classification.

IBM Maximo Visual Inspection comes with several common models, such as flowers and food, that you can use to help classify your data. If you do not select a base model when training with GoogLeNet, General is used. For more information, see Base models included with IBM Maximo Visual Inspection.

Model hyperparameters

For advanced users, model hyperparameters are available to help fine-tune the training. The user interface shows ranges for several of the hyperparameters, with lower and upper bounds marked by either a parenthesis, ( ), or a bracket, [ ]. A parenthesis indicates a noninclusive bound, and a bracket indicates an inclusive bound. For example, (0-0.5] indicates that the value must be greater than 0 and can be up to and including 0.5.

Epochs

This value is the number of times the entire data set is passed through the training algorithm. Large data sets are divided into smaller parts to fit the GPU memory and are processed as batches. One batch is passed through the algorithm during each iteration. Therefore, each epoch is made up of many iterations.

Specifying many epochs can increase the training time substantially, especially for larger data sets.

This hyperparameter is available only when you select Action detection.

Max iteration

This value is the maximum number of times the data is passed through the training algorithm, up to 1,000,000 iterations. In general, the more iterations the model is trained, the more accurate the model is. However, in many cases the test accuracy plateaus at some point beyond which further iterations do not result in a significant improvement in the accuracy of the model.

The optimal number of iterations might be higher or lower than the default. This variation depends on training factors, such as the number of data set samples, number of classes, inter-similarity and intra-similarity between classes and objects, and target accuracy. One strategy for determining the optimal number of iterations is to choose a target accuracy. Then, train the model until the target accuracy is reached or the loss rate no longer is improving.

This hyperparameter is available only when you select Image classification or Object detection.

Momentum

This value increases the step size that is used when you try to find the minimum value of the error curve. A larger step size can keep the algorithm from stopping at a local minimum instead of finding the global minimum.

This hyperparameter is available only when you select Object detection.

Ratio

IBM Maximo Visual Inspection automatically “splits” the data set for internal validation of the model’s performance during training. The default Ratio value of 80:20 results in 80% of the images in the data set, at random, being used for training, and 20% being used for measurement or validation.

Image classification models do not allow a ratio of 100% because some images are required for validation.

Test iteration

This value is the number of times data is passed through the training algorithm before possible completion. For example, if this value is 100, and Test interval is 50, the model is run through the algorithm at least 100 times, being tested every 50 times.

This hyperparameter is available only when you select Image classification.

Test interval

This value is the number of times the model is passed through the algorithm before testing. For example, if this value is 50, the model is tested every 50 iterations. Each of these tests becomes a data point on the metrics graphs.

This hyperparameter is available only when you select Image classification.

Learning rate

This option determines how much the weights in the network are adjusted regarding the loss gradient. A correctly tuned value can result in a shorter training time. Change this value only if you are an advanced user. A learning rate that is too large can result in significant oscillations in the loss function, and in the worst case, "exploding gradients" result in failure to train the model.

If training fails for Detectron2, Faster R-CNN, High resolution, or Anomaly optimized models, you can slow the training learning rate, and then run the training again. Training for these models might fail due to the speed of the learning rate. For example, a training configuration that contains the same model and data set might succeed and fail in two different training tasks that use the same learning rate.

Weight decay

This value specifies regularization in the network. It protects against over-fitting and is used to multiply the weights when training.