EasySegmentUnsupervisedTrain

Support
Required licenses	EasySegment
Recommended images	n/a
Location	Deep Learning Inspection/EasySegmentUnsupervisedTrain

Purpose

This sample program demonstrates how to:

• Train an EasySegment Unsupervised tool in console mode.

Code highlights

By default, the sample program trains a tool for the dataset Fabric in the Deep Learning Additional Resources.

Note

In the code, you need to replace DEEP_LEARNING_ADDITIONAL_RESOURCES by the path of the Deep Learning Additional Resources.

1. Select the first available GPU for the training.

   // initialize a segmenter that will be trained
   EUnsupervisedSegmenter newUnsupervisedSegmenter;
   
   // make sure that we have a GPU to use (note: GPU's can only be detected in x64 configuration)
   std::vector<EDeepLearningDevice> devices = newClassifier.GetAvailableDevices();
   
   bool foundGPU = false;
   for (size_t i = 0; i < devices.size(); i++)
   {
     if (devices[i].GetDeviceType() == EDeepLearningDeviceType_GPU)
     {
       newClassifier.SetActiveDevice(devices[i]);
       foundGPU = true;
       break;
     }
   }
   

2. Load a dataset from a Deep Learning Studio project.

   const EClassificationDataset* dataset;
   EDeepLearningProject project;
   project.Load(PROJECT_PATH);
   dataset = &project.GetDataset();
   

3. Alternatively, load a dataset directly using a .edldataset file.

   const EClassificationDataset* dataset;
   EClassificationDataset datasetHolder;
   datasetHolder.Load(DATASET_PATH);
   dataset = &datasetHolder;
   

4. Split the dataset in a training part and a validation part.

Note

Be careful that the method for splitting a dataset is different for each type of tool. EasySegment Unsupervised uses the same method as EasyClassify to create a split.

```c++
// Put 80% of images into the training dataset and 20% in the validation dataset
EClassificationDataset trainingDataset;
EClassificationDataset validationDataset;
dataset->SplitDataset(trainingDataset, validationDataset, 0.8f);
```

5. Configure and train the EasySegment Unsupervised tool and wait for the training to end.

   newUnsupervisedSegmenter.SetCapacity(EUnsupervisedSegmenterCapacity_Small);
   newUnsupervisedSegmenter.SetForceGrayscale(true);
   newUnsupervisedSegmenter.SetGoodLabel("Good");
   newUnsupervisedSegmenter.Train(trainingDataset, validationDataset, 50);
   newUnsupervisedSegmenter.WaitForTrainingCompletion();
   

6. Check the validation accuracy of the tool we just trained.

   float newUnsupervisedSegmenterAccuracy = newUnsupervisedSegmenter.GetValidationMetrics(newUnsupervisedSegmenter.GetBestIteration()).GetBestAccuracy();
   std::cout << "New segmenter Accuracy:" << newUnsupervisedSegmenterAccuracy << std::endl;