What's new in Maximo Visual Inspection 8.8

Learn more about what's new and changed in Maximo Visual Inspection 8.8.

Edge management dashboard

You can now manage multiple edge devices on the Edge management dashboard, including devices in Maximo Visual Inspection Edge. From the dashboard, you can track where each device is installed and check a device status.

For more information about the Edge management dashboard, see Edge management.

TensorRT support on Maximo Visual Inspection training platform

The following models are converted to NVIDIA TensorRT models at deployment time:

  • GoogLeNet
  • SSD
  • YOLOv3
  • Tiny YOLOv3

The TensorRT model conversion results in models that run more efficiently and use less GPU memory.

For more information about TensorRT support, see TensorRT models.

Maximo Application Suite as a service

AppPoint metrics for MAS-Image-Inferred

When a model is deployed, the MAS-Images-Inferred metric tracks and counts the total number of successful image inferences that occurred within an hour.

Dimensionality reduction in anomaly detection models

When you train an object detection model with anomaly optimization, dimensionality reduction is automatically enabled.

Enabling dimensionality reduction increases anomaly detection model performance as follows:

  • Reduces the number of feature embeddings that are used in the anomaly detection phase of the anomaly model.
  • Improves memory usage without substantially reducing model accuracy.
Note: When dimensionality reduction is enabled, the model runs at a 25% slower speed. It can also cause a loss of model accuracy. Therefore, deselect dimensionality reduction when you are deploying the model by itself on a GPU or when you are deploying 1 or 2 other models with 16 GB memory or more.

Updates to Grafana dashboard metrics

The Grafana dashboard is now installed automatically and the import of .zip files is no longer required.

The following metrics are now supported on the Grafana dashboard:

  • JVM: Metrics include GC time, GC count, threads used, threads state, JVM memory area (Heap), and JVM memory pool.
  • GPU: Metrics include GPU usage and memory usage.

Detectron2 model optimization

Detectron2 model optimization has the Enable auto early stop and Enable online augmentation training options. Training metrics include training and validation loss.

The network backbone is modified resulting in improvements to accuracy, lower latency, and less GPU memory requirements at deployment time.

Issues fixed

Anomaly model: out of memory during training of large data sets

When the number of images and labeled objects in a data set reaches the thousands, an out of memory condition can occur causing training to fail.

To resolve the issue, the number of images that are used to train the anomaly detector in the second stage of training is limited to 1000. The full number of images in the data set is still used to train the object detector in the first stage of training.

Out of memory error response code

When a model inference causes an out of memory condition, an out of memory error response code 507 is returned to the caller instead of 200.

This condition is most likely to occur with Anomaly optimized, High resolution, and Detectron2 models.