extend_dl_continual_learning — Extend a Continual Learning model with with new data.
extend_dl_continual_learning( : : DLModelHandle, DLDataset, GenParam : )
extend_dl_continual_learning extends a Continual Learning model
DLModelHandle of 'type' = 'classification'
with new data provided by DLDataset. The new data can either
contain new classes that the model should learn to predict, or new samples
of an existing class to improve the model for that particular class.
DLModelHandle learns from the provided data in a manner specific to
Continual Learning. After calling extend_dl_continual_learning, the model
will either be able to predict more classes or the internal knowledge of existing
classes will have been extended, depending on the dataset provided.
See Deep Learning / Continual Learning for further information.
DLDataset must only contain new data which the model should
learn. It is not necessary to provide data samples from previous Continual Learning
training steps, because the model preserves previously learned classes.
Besides extend_dl_continual_learning can be called
with a DLModelHandle optimized for an AI
2-interface. As a result,
you can extend your Continual Learning model directly on an edge device.
The operator expects a dataset that is preprocessed by the supplied
procedure preprocess_dl_dataset. Changes made within the procedure
to the path of the samples are not supported and lead to errors.
extend_dl_continual_learning will change the ClassIDs of the provided
DLDataset internally to match the ClassIDs of the provided model. New unknown
classes will receive a new and unused ClassID. The ClassID
of known classes within DLDataset will be changed to the ClassID given by the
model. These changes are applied both to the provided DLDataset dictionary
and to all preprocessed samples on disk. This is required for correct evaluation.
The performance of the Continual Learning model can be
evaluated using the procedure evaluate_dl_model. You can provide one or
more datasets, containing all classes the model has been trained on or only a subset
of those. It is strongly recommended to evaluate the model with test data
corresponding to all classes and Continual Learning steps, so that the evaluation
result accurately reflects the capabilities of the model on all relevant data.
The operator also recalculates the Out-of-Distribution threshold. Any
threshold value previously set by the user will be overwritten.
Confidence calibration is done only once (either in init_dl_continual_learning, if
a dataset was provided, or in the first call to extend_dl_continual_learning)
and not repeated. As more classes are added,
the absolute confidence values will generally decrease.
GenParam is a dictionary for setting generic parameters.
Currently no generic parameters are supported.
The operator makes changes to the samples of the provided dataset. If an error occurs during the changes, the dataset is incompletely adjusted and corrupted. It must then be re-read and pre-processed again. This is indicated by the operator's error message in the event of an error.
Note that the operator does not behave additively. Adding images of the same class in several calls does not result in an identical model compared to adding the same, combined images in a single call.
This operator modifies the state of the following input parameters:
During execution of this operator, access to the values of these parameters must be synchronized if it is used across multiple threads.
DLModelHandle (input_control, state is modified) dl_model → (handle)
Handle of a deep learning classification model initialized for Continual Learning.
DLDataset (input_control, state is modified) dict → (handle)
Dataset for new Continual Learning training step.
GenParam (input_control) dict → (handle)
Dictionary for generic parameters.
Default: []
Deep Learning Professional