Train and Apply AI Models
Once you have painted enough labels, click Train to train an AI model for segmentation. he progress of the training is shown in a separate dialog.
|
Note! The loss reported for the Unet models during training refers to the training loss because there is currently no separation between training and test data. All available data is used for training. |
After completing the training, select the Preview Overlay to see what a segmentation using the trained model would look like. In the example below, the Boosted Tree model was used. Therefore, only a few labels were needed for this gray value image of a Berea sandstone.
You can improve a previously trained and currently loaded model by clicking Continue Training. A model can be loaded by training it in the current Image Processing session or by clicking Load Model to load a previously trained model. For example, Continue Training can improve the model when you have added more labels. Note that not only are the new labels considered, but also the labels used for the loaded model, so that all available information is used for training. Additionally, a model can be improved by loading and labeling another similar image, as explained here, to provide more training data.
|
Note! If you click Train while a trained model is already loaded, the current model will be discarded and a new one will be trained from scratch. This can be helpful when a model does not improve with Continue Training. |
Click Save Model to save a trained machine learning model. You can load this model again by clicking Load Model whenever you need to segment similar scans.
The file formats for the different AI models are as follows:
- *.XGBM for Boosted Tree
- *.FOREST for Random Forest
- *.UNET2D for Unet 2D
- *.UNET3D for Unet 3D
The final step is to click Create Segmentation to apply the trained AI model to the entire image dataset.
