Operator Reference
create_dl_model (Operator)
create_dl_model — Create a deep learning model.
Signature
create_dl_model( : : OutputLayers : DLModelHandle)
Description
The operator create_dl_model creates a deep learning model from a
graph and returns its handle in DLModelHandle.
A deep learning model in HALCON mainly consists of a directed acyclic
graph that defines the networks architecture.
Further components of a deep learning model in HALCON are parameters as
'class_names', 'class_ids', and many others,
or hyperparameters that are needed to train a model, as for example the
'learning_rate'.
While parameters and hyperparameters can be set after creation of the model
using set_dl_model_param, the model itself can only be created
using create_dl_model if its network architecture is given in form of
a graph.
To build a graph that defines the models network architecture, one needs to
put together the networks layers. In general, a graph starts with an input
layer. A subsequent layer that follows after the input layer uses the
input layer as feeding layer, and the new layer itself might be used as
a feeding layer for the next layer, and so on. This is repeated until
the graphs output layers (e.g., softmax or loss layers) are
appended to the graph.
To create a layer, use its specified creation operator, e.g., an input
layer is created using create_dl_layer_input, a convolution layer
is created using create_dl_layer_convolution, and so on.
When the graph is defined, a model can be created using
create_dl_model by passing over the graphs output layer handles in
OutputLayers.
Note that the output layer handles save all other layers that directly or
indirectly serve as feeding input layers for the output layers during
their creation. This means that the output layer handles keep the
whole network architecture necessary for the creation of the model
using create_dl_model.
The type of the created model, hence the task the model is designed for
(classification, object detection, segmentation), is only given by the
networks architecture.
However, if the networks architecture allows it, the type of the model,
'type', can be set using set_dl_model_param.
A specified model type allows a more user friendly usage in the HALCON
deep learning workflow.
Supported types are:
- 'generic':
This is the default model type. The task the model's neuronal network can solve is defined by its architecture. When
apply_dl_modelis applied for inference, the operator returns the activations of the output layers. To train the model usingtrain_dl_model_batch, the underlying graph requires loss layers.- 'classification':
-
The model is specified for classification and all layers required for training the model are adapted to the model. When
apply_dl_modelis applied for inference, the output is adapted according to the type, seeapply_dl_modelfor more details. See Deep Learning / Classification for further information.In addition, the operator
gen_dl_model_heatmapcan be used to display the models heatmap. - 'detection':
The model is specified for object detection and instance segmentation and all layers and anchors required for training the model are adapted to the model. When
apply_dl_modelis applied for inference, the output is adapted according to the type, seeapply_dl_modelfor more details. See Deep Learning / Object Detection and Instance Segmentation for further information.- 'multi_label_classification':
The model is specified for multi-label classification and all layers required for training the model are adapted to the model. When
apply_dl_modelis applied for inference, the output is adapted according to the type, seeapply_dl_modelfor more details. See Deep Learning / Multi-Label Classification for further information.- 'segmentation':
The model is specified for semantic segmentation or edge extraction respectively and all layers required for training the model are adapted to the model. When
apply_dl_modelis applied for inference, the output is adapted according to the type, seeapply_dl_modelfor more details. See Deep Learning / Semantic Segmentation and Edge Extraction for further information.
Furthermore, many deep learning procedures provide more functionality for the
model if its type is set. As an example, dev_display_dl_data
can be used to display the inferred results more nicely.
Note that setting a model type requires that the graph fulfills certain structure conditions. We recommend to follow the architecture of our delivered neuronal networks if the model type should be set to one of these types.
Execution Information
- Multithreading type: reentrant (runs in parallel with non-exclusive operators).
- Multithreading scope: global (may be called from any thread).
- Processed without parallelization.
This operator returns a handle. Note that the state of an instance of this handle type may be changed by specific operators even though the handle is used as an input parameter by those operators.
Parameters
OutputLayers (input_control) dl_layer(-array) → (handle)
Output layers of the graph.
DLModelHandle (output_control) dl_model → (handle)
Handle of the deep learning model.
Result
If the parameters are valid, the operator create_dl_model
returns the value 2 (
H_MSG_TRUE)
. If necessary, an exception is raised.
Possible Predecessors
create_dl_layer_softmax,
create_dl_layer_loss_cross_entropy,
create_dl_layer_loss_focal,
create_dl_layer_loss_huber
Possible Successors
Module
Deep Learning Professional