Training a Deep Learning Tool

In Deep Learning Studio

1. Create a new tool.
2. Configure the tool settings.
3. Select the dataset split to use for this tool.
4. Configure the training settings and click on Train.

In the API

In the API, to train a deep learning tool, call the method EDeepLearningTool::Train(trainingDataset, validationDataset, numberOfIterations).

The training settings

● The Number of iterations. An Iteration corresponds to going through all the images in the training set once.
□ The training process requires a large number of iterations to obtain good results.
□ The larger the number of iterations, the longer the training is and the better the results you obtain.
□ The default number of iterations is 50.

We recommend to use more iterations than the default value:

□ For EasyClassify without pretraining, EasyLocate and EasySegment.
□ For smaller dataset because the training is harder (for example, 100 iterations for a dataset with 100 images, 200 iterations for a dataset with 50 images, 400 iterations for a dataset with 25 images...).
● The Batch size corresponds to the number of image patches that are processed together.
□ The training is influenced by the batch size.
□ A large batch size increases the processing speed of a single iteration on a GPU but requires more memory.
□ The training process is not able to learn a good model with too small batch sizes.
□ By default, the batch size is determined automatically during the training to optimize the training speed with respect to the available memory.

- Use EDeepLearningTool::SetOptimizeBatchSize(false) to disable this behavior.

- Use EDeepLearningTool::SetBatchSize to change the size of your batch.

□ EDeepLearningTool::GetBatchSizeForMaximumInferenceSpeed gets the batch size that maximizes the batch classification speed on a GPU according to the available memory.
□ It is common to choose powers of 2 as the batch size for performance reasons.
● The Data augmentation.
● Whether to use Deterministic training or not.
□ The deterministic training allows to reproduce the exact same results when all the settings are the same (tool settings, dataset split and training settings).
□ The deterministic training fixes random seeds used in the training algorithm and uses deterministic algorithms.
□ The deterministic training is usually slower than a non-deterministic training.
□ In Deep Learning Studio, the option to use deterministic training and the random seed are available in the advanced parameters.

□ In the API, use EDeepLearningTool::EnableDeterministicTraining. and EDeepLearningTool::DeterministicTrainingRandomSeed.

Continue the training

You can continue to train a tool that is already trained.

In Deep Learning Studio, the dataset split associated with a trained tool is locked.

□ You can only continue training a tool with the same dataset split.
□ You can still add new training or validation images to the split by moving test images to the training set or the validation set of that split.

Asynchronous training

The training process is asynchronous and performed in the background.

● In Deep Learning Studio:
□ The training processes are queued.
□ They are automatically executed one after the other.
□ You can manually reorder the training in the processing queue.

● In the API:
□ EDeepLearningTool::Train launches a new thread that does the training in the background.
□ EDeepLearningTool::WaitForTrainingCompletion suspends the program until the whole training is completed.
□ EDeepLearningTool::WaitForIterationCompletion suspends the program until the current iteration is completed.
□ During the training, EDeepLearningTool::GetCurrentTrainingProgression shows the progression of the training.