apply_dl_model — Apply a deep-learning-based network on a set of images for inference.
apply_dl_model( : : DLModelHandle, DLSampleBatch, Outputs : DLResultBatch)
apply_dl_model applies the deep-learning-based network given
by DLModelHandle on the batch of input images
handed over through the tuple of dictionaries DLSampleBatch.
The operator returns DLResultBatch, a tuple with a result
dictionary DLResult for every input image.
Please see the chapter Deep Learning / Model for more information on the concept and the dictionaries of the deep learning model in HALCON.
In order to apply the network on images, you have to hand them over
through a tuple of dictionaries DLSampleBatch, where a dictionary
refers to a single image. You can create such a dictionary conveniently
using the procedure gen_dl_samples_from_images. The tuple
DLSampleBatch can contain an arbitrary number of
dictionaries. The operator apply_dl_model always processes
a batch with up to 'batch_size' images simultaneously.
In case the tuple contains more images, apply_dl_model iterates
over the necessary number of batches internally. For a
DLSampleBatch with less than 'batch_size' images,
the tuple is padded to a full batch which means that the time required
to process a DLSampleBatch is independent of whether the
batch is filled up or just consists of a single image. This also means
that if fewer images than 'batch_size' are processed in one
operator call, the network still requires the same amount of memory as
for a full batch. The current value of 'batch_size' can be
retrieved using get_dl_model_param.
Note that the images might have to be preprocessed before feeding them into
the operator apply_dl_model in order to fulfill the network
requirements. You can retrieve the current requirements of your network,
such as e.g., the image dimensions, using get_dl_model_param.
The procedure preprocess_dl_dataset provides guidance on how to
implement such a preprocessing stage.
The results are returned in DLResultBatch, a tuple with a
dictionary DLResult for every input image.
Please see the chapter Deep Learning / Model
for more information to the output dictionaries in DLResultBatch
and their keys.
In Outputs you can specify, which output data is returned in
DLResult.
Outputs can be a single string, a tuple of strings, or
an empty tuple with which you retrieve all possible outputs.
If apply_dl_model is used with an AI
2-interface, it might be
required to set 'is_inference_output' = 'true' for all
requested layers in Outputs before the model is optimized for the
AI
2-interface, see optimize_dl_model_for_inference and
set_dl_model_layer_param for further details.
The values for Outputs depend on the model type of your network:
Outputs='[]': DLResult containing:
'gripping_map': Binary image, indicating for each pixel of
the scene whether the model predicted a gripping point
(pixel value = 1.0) or not (0.0).
'gripping_confidence': Image, containing raw, uncalibrated
confidence values for every point in the scene.
Outputs='[]': DLResult
contains an image where each pixel has the score of the according input
image pixel. Additionally it contains a score for the entire image.
This model type cannot be run with the operator apply_dl_model.
For each value of Outputs, DLResult
contains an image where each pixel has the score of the according input
image pixel.
Additionally it contains a score for the entire image.
Outputs='[]':
The scores of each input image pixel are calculated as a combination of
all available networks.
Outputs='anomaly_image_local':
The scores of each input image pixel are calculated from the
'local' network only. If the 'local' network is not
available, an error is raised.
Outputs='anomaly_image_global':
The scores of each input image pixel are calculated from the
'global' network only. If the 'global' network is not
available, an error is raised.
Outputs='anomaly_image_combined':
The scores of each input image pixel are calculated by combining the
'global' and the 'local' networks.
If one or both of the networks are not available, an error is raised.
Outputs='[]': DLResult contains a
tuple with confidence values in descending order and tuples with the
class names and class IDs sorted accordingly.
Outputs='[]': DLResult contains a
tuple with the selected class names, class IDs and the corresponding
confidence values according to the model parameter 'min_confidence'.
Additionally, it contains tuples with all class names, class IDs and
corresponding confidence values.
Outputs='[]': DLResult contains
the bounding box coordinates as well as the inferred classes and their
confidence values resulting from all levels.
Outputs=
'[bboxhead + level + _prediction, classhead + level + _prediction]',
where 'level' stands for the selected level which lies
between 'min_level' and 'max_level':
DLResult contains the bounding box coordinates as well
as the inferred classes and their confidence values resulting from
specific levels.
Outputs='[]': DLResult
contains the recognized word.
Additionally it contains candidates for each character of the
word and their confidences.
Outputs='[]': DLResult
contains the bounding boxes coordinates of localized words.
Outputs='segmentation_image':
DLResult contains an image where each pixel has a value
corresponding to the class
its corresponding pixel has been assigned to.
Outputs='segmentation_confidence':
DLResult contains an image where each pixel has the
confidence value out of the classification of the according pixel.
Outputs='[]':
DLResult contains all output values.
System requirements:
To run this operator on GPU by setting 'device' to 'gpu'
(see get_dl_model_param), cuDNN and cuBLAS are required.
For further details, please refer to the “Installation Guide”,
paragraph “Requirements for Deep Learning and Deep-Learning-Based Methods”.
This operator supports canceling timeouts and interrupts.
This operator supports breaking timeouts and interrupts.
DLModelHandle (input_control) dl_model → (handle)
Handle of the deep learning model.
DLSampleBatch (input_control) dict-array → (handle)
Input data.
Outputs (input_control) string-array → (string)
Requested outputs.
Default: []
List of values: [], 'bboxhead2_prediction', 'classhead2_prediction', 'segmentation_confidence', 'segmentation_image'
DLResultBatch (output_control) dict-array → (handle)
Result data.
If the parameters are valid, the operator apply_dl_model
returns the value 2 (
H_MSG_TRUE)
. If necessary, an exception is raised.
read_dl_model,
train_dl_model_batch,
train_dl_model_anomaly_dataset,
set_dl_model_param
Foundation. This operator uses dynamic licensing (see the ``Installation Guide''). Which of the following modules is required depends on the specific usage of the operator:
3D Metrology, OCR/OCV, Deep Learning Inference