Graph for training models

As IBM® Maximo® Visual Inspection trains the model, a graph shows the relative performance of the model over time.

This graph is displayed when you are Training a model.

The goal is for the model to converge at the end of the training with low error and high accuracy.

In the figure, you can see the Loss CLS line and the Loss Bbox lines start to plateau. In the training graph, the lower the value for the loss line, the better. Therefore, you can stop the training process when the loss value stops decreasing. When the loss value stops decreasing, enough iterations are completed by the training model, and you can continue to the next step.

Figure 1. Model training graph
The image shows a loss on the vertical axis and iterations on the horizontal axis. The more iterations that occur the line for loss converges to a flat line.
Important: If the training graph converges quickly and has 100% accuracy, the data set does not have enough information. The same is true if the accuracy of the training graph fails to rise or the errors in the graph do not decrease at the end of the training process. For example, a model with high accuracy might be able to discover all instances of different race cars. However, the same model might be unable to differentiate between specific race cars or cars that have different colors. In this situation, add more images, video frames, or videos to the data set. Then, label those objects and try the training again.

The Loss versus Iteration graph is different for the SSD model than for other object detection models. The difference is because the SSD model combines the Train Loss Bbox and Train Loss CLS into a single statistic that is presented in the training graph. For SSD models, the Train Loss Bbox value is not valid. The graph shows a constant value of 0, while the Train Loss CLS tracks a combination of Bbox and CLS loss.

Note: Starting in Maximo Visual Inspection 9.1, Single Shot Detector (SSD) models are no longer supported for model training. However, you can continue to import and deploy SSD models in Maximo Visual Inspection and Maximo Visual Inspection Edge.

Click the icons to take the following actions:

  • View a tabular representation of the graph data.
  • View the graph in full-screen mode.
  • Export the graph data to a comma-separated values (CSV) file.
  • Save the graph locally to a JPG or PNG image file.

For High resolution models, a different version of the graph is displayed. When you train a model, the data set is split into training and validation subsets. In this version of the graph, which is displayed in Figure 2, the Training Loss and Validation Loss loss lines show how accurately the model classifies images and identifies objects in the training and validation subsets.

Figure 2. High resolution model training graph
The image shows two loss lines on the vertical axis and iterations on the horizontal axis. As more iterations occur the lines for loss converge to flat lines.

The training loss line combines the Train Loss Bbox and Train Loss CLS measurements into a single measurement. Training loss data is captured when training begins and appears on the graph after every 20 training iterations. Validation loss data is captured later and appears less often. To determine how often validation data loss is captured and displayed, divide the number of images in your data set by 4. For example, if your data set contains 1000 images, validation loss data appears after every 250 iterations.

After the loss lines plateau, if the validation loss line value is lower than the training loss line, let the training continue. If the validation loss line or value is equal to or is higher on the graph than the training loss line, such as the validation loss line that is shown in Figure 3, you can stop the training. When you train a High resolution model, select the Enable auto early stop option to automatically stop the training when the most accurate model is available.

Figure 3. Stopping point for training a High resolution model
The image shows two loss lines on the vertical axis and iterations on the horizontal axis. As more iterations occur the lines for loss converge to flat lines.

For Anomaly optimized models, only the Training Loss line is displayed. This loss line combines the Train Loss Bbox and Train Loss CLS measurements into a single measurement in the same way that these measurements are combined for High resolution models.

Figure 4. Anomaly optimized model training graph
The image shows one loss line on the vertical axis and iterations on the horizontal axis. As more iterations occur the line for loss converges to a flat line.