get_dl_model_param — Return the parameters of a deep learning model.
get_dl_model_param( : : DLModelHandle, GenParamName : GenParamValue)
get_dl_model_param returns the parameter values
GenParamValue of GenParamName for the deep learning
model DLModelHandle.
For a deep learning model, parameters GenParamName can be set using
set_dl_model_param or create_dl_model_detection, depending
on the parameter and the model type.
With this operator, get_dl_model_param, you can retrieve
the parameter values GenParamValue.
Below we give an overview of the different parameters and an explanation,
except of those you can only set. For latter ones, please see the
documentation of corresponding operator.
GenParamName |
Ano. Det. | Class. | Obj. Det. | Sem. Seg. |
|---|---|---|---|---|
| 'batch_size' | ||||
| 'batch_size_multiplier' | ||||
| 'class_ids' | ||||
| 'class_weights' | ||||
| 'enable_resizing' | ||||
| 'fuse_bn_relu' | ||||
| 'fuse_conv_relu' | ||||
| 'gpu' | ||||
| 'image_dimensions' | ||||
| 'image_height' | ||||
| 'image_width' | ||||
| 'image_num_channels' | ||||
| 'image_range_max' | ||||
| 'image_range_min' | ||||
| 'learning_rate' | ||||
| 'momentum' | ||||
'num_classes' (NumClasses) |
||||
| 'optimize_for_inference' | ||||
| 'runtime' | ||||
| 'runtime_init' | ||||
| 'summary' | ||||
| 'type' | ||||
| 'weight_prior' | ||||
GenParamName |
Ano. Det. | Class. | Obj. Det. | Sem. Seg. |
|---|---|---|---|---|
| 'complexity' | ||||
| 'standard_deviation_factor' | ||||
| 'class_names' | ||||
| 'extract_feature_maps' | ||||
| 'anchor_angles' | ||||
| 'anchor_aspect_ratios' | ||||
| 'anchor_num_subscales' | ||||
'backbone' (Backbone) |
||||
| 'backbone_docking_layers' | ||||
| 'bbox_heads_weight' | ||||
| 'capacity' | ||||
| 'class_heads_weight' | ||||
| 'class_ids_no_orientation' | ||||
| 'freeze_backbone_level' | ||||
| 'ignore_direction' | ||||
| 'instance_type' | ||||
| 'max_level' | ||||
| 'min_level' | ||||
| 'max_num_detections' | ||||
| 'max_overlap' | ||||
| 'max_overlap_class_agnostic' | ||||
| 'min_confidence' | ||||
| 'ignore_class_ids' | ||||
Thereby, the symbols denote the following :
's': The parameter can be set using set_dl_model_param.
'g': The parameter can be retrieved using get_dl_model_param.
'c': The parameter can be set using
create_dl_model_detection.
'x': The parameter has no effect for this type of model.
'-': The parameter is not applicable for this type of model.
Certain parameters are set as non-optional parameters, the corresponding notation is given in brackets.
In the following we list and explain the parameters GenParamName
for which you can retrieve their value using this operator,
get_dl_model_param.
They are sorted according to the model type.
Note, for models of 'type'='segmentation' the default
values depend on the specific network and therefore have to be retrieved.
Number of input images (and corresponding labels) in a batch that is transferred to device memory.
For a trainig using train_dl_model_batch, the batch of images
which are processed simultaneously in a single training iteration
contains a number of images which is equal to 'batch_size'
times 'batch_size_multiplier'. Please refer to
train_dl_model_batch for further details.
For inference, the 'batch_size' can be generally set
independently from the number of input images.
See apply_dl_model for details on how to set this parameter
for greater efficiency.
The parameter 'batch_size' is stored in the pretrained
classifier. Per default, the 'batch_size' is set such
that a training of the pretrained classifier with up to
100 classes can be easily performed on a device with
8 gigabyte of memory.
| pretrained classifier | default value of 'batch_size' |
|---|---|
| 'pretrained_dl_classifier_alexnet.hdl' | 230 |
| 'pretrained_dl_classifier_compact.hdl' | 160 |
| 'pretrained_dl_classifier_enhanced.hdl' | 96 |
| 'pretrained_dl_classifier_resnet50.hdl' | 23 |
Multiplier for 'batch_size' to enable training with larger
numbers of images in one step which would otherwise not be possible
due to GPU memory limitations. This model parameter does only affect
train_dl_model_batch and thus has no
impact during evaluation and inference.
For detailed information see train_dl_model_batch.
The parameter 'batch_size_multiplier' has no effect.
Default: 'batch_size_multiplier' = 1
Unique IDs of the classes the model shall distinguish. The tuple is of length 'num_classes'.
We stress out the slightly different meanings and restrictions depending on the model type:
'class_ids' is not supported.
The IDs are unique identifiers, which are automatically assigned to each class. The ID of a class corresponds to the index within the tuple 'class_names'.
Only the classes of the objects to be detected are included and therewith no background class. Thereby, you can set any integer within the interval as class id value.
Note that the values of 'class_ids_no_orientation' depend on 'class_ids'. Thus if 'class_ids' is changed after the creation of the model, 'class_ids_no_orientation' is reset to an empty tuple.
Default: 'class_ids' = '[0,...,num_classes-1]'
Every class used for training has to be included and therewith also the class ID of the 'background' class. Therefore, for such a model the tuple has a minimal length of 2. Thereby, you can set any integer within the interval as class id value.
The parameter 'class_weights' is a tuple of class specific weighting factors for the loss. Giving the unique classes a different weight, it is possible to force the network to learn the classes with different importance. This is useful in cases where a class dominates the dataset. The weighting factors have to be within the interval . Thereby a class gets a stronger impact during the training the larger its weight is. The weights in the tuple 'class_weights' are sorted the same way as the classes in the tuple 'class_ids'. We stress out the slightly different meanings and restrictions depending on the model type:
'class_weights' is not supported.
Default: 'class_weights' = 1.0 for each class.
Default: 'class_weights' = 0.25 for each class.
'class_weights' is not supported.
Identifier of the GPU where the training and inference operators
(train_dl_model_batch and apply_dl_model) are executed.
Per default, the first available GPU is used.
get_system with 'cuda_devices' can be used to
retrieve a list of available GPUs. Pass the index in this list to
'gpu'.
Default: 'gpu' = 0
Tuple containing the input image dimensions 'image_width', 'image_height', and number of channels 'image_num_channels'.
The respective default values and possible value ranges depend on the model and model type. Please see the individual dimension parameter description for more details.
Height and width of the input images, respectively, that the network will process.
This parameter can attain different values depending on the model type:
The default values depend on the specific pretrained network,
see read_dl_model.
The network architectures allow changes of the image dimensions,
which can be done using set_dl_model_param. Please also
refer to read_dl_model for restrictions each of the
delivered networks has on the input image size.
Note that these parameters have to be set before training the model.
Setting them on an already trained model would render this model
useless. When trying to do so, an error is returned but the model
itself is unchanged.
The default values depend on the specific pretrained classifier,
see read_dl_model.
The network architectures allow changes of the image dimensions,
which can be done using set_dl_model_param.
But for networks with at least one fully connected layer such a
change makes a retraining necessary.
Networks without fully connected layers are directly applicable to
different image sizes. However, images with a size differing from
the size with which the classifier has been trained are likely to
show a reduced classification accuracy.
The network architectures allow changes of the image dimensions.
But the image lengths are halved for every level, thats why the
dimensions 'image_width' and 'image_height'
need to be an integer multiple of
.
depends on the 'backbone'
and the parameter 'max_level', see
create_dl_model_detection for further information.
Default: 'image_height' = 640, 'image_width' = 640
The network architectures allow changes of the image dimensions.
The default and minimal values are given by the network, see
read_dl_model.
Number of channels of the input images the network will process.
The default value is given by the network, see
read_dl_model and create_dl_model_detection.
For models of 'type'='anomaly_detection', only the values 1 and 3 are supported. In addition, this parameter can only be set on a model of 'type'='anomaly_detection' before the model is trained. Setting 'image_num_channels' on an already trained model would render this model useless. When trying to do so, an error is returned but the model itself is unchanged.
For other models, any number of input image channels is possible.
If number of channels is changed to a value >1, the weights of the first layers after the input image layer will be initialized with random values. Note, in this case more data for the retraining is needed. If the number of channels is changed to 1, the weights of the concerned layers are fused.
Default: 'image_num_channels' = 3
Default: 'image_num_channels' = 3
Maximum and minimum gray value of the input images, respectively, the network will process.
The default values are given by the network, see
read_dl_model and create_dl_model_detection.
Value of the factor determining the gradient influence during
training using train_dl_model_batch.
Please refer to train_dl_model_batch for further details.
The default values depend on the model.
The parameter 'learning_rate' has no effect.
When updating the weights of the network,
the hyperparameter 'momentum' specifies to which extent
previous updating vectors will be added to the current updating
vector.
Please refer to train_dl_model_batch for further details.
The default value is given by the model.
The parameter 'momentum' has no effect.
Number of distinct classes that the model is able to distinguish for its predictions.
This parameter differs between the model types:
'num_classes' is not supported.
'num_classes' is determined implicitly by the length of 'class_names' and therefore not supported.
This parameter is set as NumClasses over
create_dl_model_detection
and 'class_ids' always needs to have a number of entries
equal to 'num_classes'.
A model of 'type'='segmentation' does predict background and therefore in this case the 'background' class is included in 'num_classes'. For these models, 'num_classes' is determined implicitly by the length of 'class_ids'.
Defines whether the model is optimized and only applicable for inference.
Setting this parameter to 'true' frees model memory not needed
for inference (e.g., memory of gradients).
This can significantly reduce the amount of memory needed by the model.
As a consequence, models with this characteristic have no gradients
accessible (needed e.g., for training or calculations of heatmaps with
mode 'grad_cam').
Operators using values from freed memory (e.g.,
train_dl_model_batch) will automatically reset this parameter
value to 'false'.
In case the value is reset to 'false' (both, manually or automatically), memory needed by the model for training is reallocated. This implies, a following training behaves as if 'momentum' is temporarily set to 0 (as possible updating vectors have to be accumulated again).
'optimize_for_inference' is not supported.
Default: 'false'
Defines the device on which the operators will be executed.
Default: 'runtime' = 'gpu'
The training and inference operator will be executed on CPU.
Note, training is only supported for specific platform types,
please see the HALCON “Installation Guide”.
In case the GPU has been used before, CPU memory is initialized, and if necessary values stored on the GPU memory are moved to the CPU memory.
For parallelization:
The runtime is highly dependent on the number of threads set.
The use of all available threads does not necessarily create
a faster performance.
How many threads are currently set can be queried with the operator
get_system.
The implemented CPU parallelization is dependent on the architecture:
Intel or AMD architecture: OpenMP.
By default all available threads of the OpenMP runtime environment
are used.
The number of threads used can be specified with the thread specific
parameter 'tsp_thread_num' of the operator set_system.
Arm architectures: Global Thread Pool.
The number of threads can be set with the global parameter 'thread_num'
of the operator set_system.
However, it is not possible to specify a thread-specific number of threads
(via the parameter ValRef{'tsp_thread_num'} of the operator set_system).
The GPU memory is initialized.
The operators apply_dl_model, train_dl_model_batch, and
train_dl_model_anomaly_dataset will be executed on the GPU.
For the specific requirements please refer to the HALCON
“Installation Guide”.
This parameter returns information on the layers of the model.
More precisely, it returns a tuple with a string for every layer.
The string is as follows:
ID; NAME; TYPE; OUTPUT_SHAPE; CONNECTED_NODES
ID: Index of the layer in the CNN graph.
NAME: Human-readable identifier (optional).
TYPE: Human-readable identifier
representing the type of the layer (e.g., input or
convolution).
OUTPUT_SHAPE: Size of the output, given in the form
(Width, Height, Depth, 'batch_size').
This means, the layer has feature maps of size Width times Height
and therefrom Depth many.
Together they form an iconic object with a channel for every
feature map.
The parameter 'batch_size' determines, how many objects
are returned together.
CONNECTED_NODES: Comma separated list with IDs
of the layers using the output of the current layer as input
E.g., '3; conv1; convolution; (112, 112, 64, 160); 4'.
Note, for some networks distributed with HALCON, the network
architecture is confidential. In this case
get_dl_model_param returns an empty tuple with 'summary'.
This parameter returns the model type. The following types are distinguished:
'anomaly_detection'
'classification'
'detection'
'segmentation'
'generic' - only for certain read in models, see
set_dl_model_param.
Regularization parameter used for
the regularization of the loss function.
For a detailed description of the regularization term we refer to
train_dl_model_batch.
Simply put: Regularization favors simpler models that are less likely
to learn noise in the data and generalize better.
In case the classifier overfits the data, it is strongly
recommended to try different values for the parameter
'weight_prior' to improve the generalization properties of
the neural network. Choosing its value is a trade-off between the
models ability to generalize, overfitting, and underfitting.
If is too small, the model might overfit, if its too
large the model might loose its ability to fit the data, because all
weights are effectively zero. For finding an ideal value for
, we recommend a cross-validation, i.e. to perform the
training for a range of values and choose the value that results in
the best validation error. For typical applications, we recommend
testing the values for 'weight_prior' on a logarithmic scale
between . If the training
takes a very long time, one might consider performing the
hyperparameter optimization on a reduced amount of data.
The parameter 'weight_prior' has no effect.
Default: 'weight_prior' = 0.0
(with exception of the pretrained classifiers
pretrained_dl_classifier_resnet50:
'weight_prior' = 0.0001,
pretrained_dl_classifier_alexnet:
'weight_prior' = 0.0005)
This parameter controls the capacity of the model to deal with more complex applications. A higher value allows for the model to represent images showing more complexity. Increasing the parameter leads to higher runtimes during training and inference. Please note that this parameter can only be set before the model is trained. Setting 'complexity' on an already trained model would render this model useless. When trying to do so, an error is returned but the model itself is unchanged.
Default: 'complexity' = 15
The anomaly score is calculated internally as the mean of certain internal scores s plus lambda times their standard deviation.
Where s denotes a pixel value of the internal
anomaly_image,
the mean value of s and
the standard deviation of s.
The parameter 'standard_deviation_factor' sets the value
and thus controls how import the standard deviation
is in comparison to the mean.
Default: 'standard_deviation_factor' = 3.0
The parameter 'backbone_docking_layers' specifies which layers of the backbone are to be used as docking layers by the feature pyramid. Thereby the layers are referenced by their names.
The docking layers can be specified for every classifier, also without using them as backbone. The specification is only considered for object detection backbones. When selecting the docking layers, consider that the feature map lengths have to be halved from one docking layer to the other. Rule of thumb: Use the deepest layers for every (lateral) resolution in the backbone (corresponding to one of the required levels for your object detection task).
Information about the names and sizes of the layers in a model can be enquired using 'summary'.
Default: For the pretrained backbones delivered by HALCON the defaults depend on the classifier. Other classifiers do not have any docking layers set by default and therefore need to have this parameter set before they can be used as backbone.
Unique names of the classes the model shall distinguish. The order of the class names remains unchanged after the setting. The tuple is of length 'num_classes'.
With this parameter value you can extract feature maps of the
specified model layer for an inferred image.
The selected layer must be part of the existing network. An overview of
all existing layers of the model can be returned by the operator
get_dl_model_param with the corresponding parameter
'summary'.
Note, using this option modifies the network architecture:
The network is truncated after a selected layer.
This modification can not be reversed.
If the original network architecture should be used again
it must be read in again with the operator read_dl_model.
The parameter 'anchor_angles' determines the orientation angle of the anchors for a model of 'instance_type' = 'rectangle2'.
Thereby, the orientation is given in arc measure and indicates the angle between the horizontal axis and 'Length1' (mathematically positive). See the chapter Deep Learning / Object Detection for more explanations to anchors.
You can set a tuple of values. A higher number of angles increases the number of anchors which might lead to a better localization but also increases the runtime and memory-consumption.
Assertion: 'anchor_angles' for 'ignore_direction' = 'false', 'anchor_angles' for 'ignore_direction' = 'true'
Default: 'anchor_angles' = '[0.0]'
The parameter 'anchor_aspect_ratios' determines the aspect ratio of the anchors. Thereby, the definition of the ratio depends on the 'instance_type':
'rectangle1': height-to-width ratio
'rectangle2': ration length1 to length2
E.g., for instance type 'rectangle1' the ratio 2 gives a narrow and 0.5 a broad anchor. The size of the anchor is affected by the parameter 'anchor_num_subscales' and with its explanation we give the formula for the sizes and lengths of the generated anchors. See the chapter Deep Learning / Object Detection for more explanations to anchors.
You can set a tuple of values. A higher number of aspect ratios increases the number of anchors which might lead to a better localization but also increases the runtime and memory-consumption.
For reasons of backward compatibility, the parameter name 'aspect_ratios' can be used instead of 'anchor_aspect_ratios'.
Default: 'anchor_aspect_ratios' = '[1.0, 2.0, 0.5]'
This parameter determines the number of different sizes with which the anchors are generated at the different levels used.
In HALCON for every anchor point, thus every pixel of every feature map of the feature pyramid, a set of anchors is proposed. See the chapter Deep Learning / Object Detection for more explanations to anchors. Thereby the parameter 'anchor_num_subscales' affects the size of the anchors. An example is shown in the figure below.
An anchor of level has by default a edge lengths of in the input image, whereby the parameter has the value . With the parameter 'anchor_num_subscales' additional anchors can be generated, which converge in size to the smallest anchor of the level . More precisely, these anchors of level have in the input image the edge lengths where . For subscale , this results on level in an anchor of height and width equal where is the ratio of this anchor (see 'anchor_aspect_ratios').
A larger number of subscales increases the number of anchors and will therefore increase the runtime and memory-consumption.
For reasons of backward compatibility, the parameter name 'num_subscales' can be used instead of 'anchor_num_subscales'.
Default: 'anchor_num_subscales' = 3
The parameter 'backbone' is the name (together with the path)
of the backbone network which is used to create the model. A list of
the delivered backbone networks can be found under
create_dl_model_detection.
The parameters 'bbox_heads_weight' and 'class_heads_weight' are weighting factors for the calculation of the total loss. This means, when the losses of the individual networks are summed up, the contributions from the bounding box regression heads are weighted by a factor 'bbox_heads_weight' and the contributions from the classification heads are weighted by a factor 'class_heads_weight'.
Default: 'bbox_heads_weight' = 1.0, 'class_heads_weight' = 1.0
This parameter roughly determines the number of parameters (or filter weights) in the deeper sections of the object detection network (after the backbone). Its possible values are 'high', 'medium', and 'low'.
It can be used to trade-off between detection performance and speed. For simpler object detection tasks, the 'low' or 'medium' settings may be sufficient to achieve the same detection performance as with 'high'.
Default: 'capacity' = 'high'
With this parameter you can declare classes, for which the orientation will not be considered, e.g., round or other point symmetrical objects. For each class, whose class ID is present in 'class_ids_no_orientation', the network returns axis-aligned bounding boxes.
Note, this parameter only affects networks of 'instance_type' = 'rectangle2'.
Note that the values of 'class_ids_no_orientation' depend on 'class_ids'. Thus if 'class_ids' is changed after the creation of the model, 'class_ids_no_orientation' is reset to an empty tuple.
Default: 'class_ids_no_orientation' = []
This parameter determines the backbone levels whose weights are kept (meaning not updated and thus frozen) during training. Thereby the given number signifies the highest level whose layers are frozen in the backbone. Setting 'freeze_backbone_level' to 0, for no level the weights are frozen and as a consequence the weights of all layers are updated. It is recommended to set this in case the weights have been randomly initialized (e.g., after certain changes of the number of image channels) or the in case the backbone is not pretrained.
Default: 'freeze_backbone_level' = 2
The parameter 'instance_type' determines, which instance type is used for the object model. The current implementations differ regarding the allowed orientations of the bounding boxes. See the chapter Deep Learning / Object Detection for more explanations to the different types and their bounding boxes.
Possible values: 'rectangle1', 'rectangle2'
Default: 'instance_type' = 'rectangle1'
These parameters determine on which levels the additional networks are attached on the feature pyramid. We refer to the chapter Deep Learning / Object Detection for further explanations to the feature pyramid and the attached networks.
From these ('max_level' - 'min_level' + 1) networks all predictions with a minimum confidence value are kept as long they do not strongly overlap (see 'min_confidence' and 'max_overlap').
The level declares how often the size of the feature map already has been scaled down. Thus, level 0 corresponds to the feature maps with size of the input image, level 1 to feature maps subscaled once, and so on. As a consequence, smaller objects are detected in the lower levels, whereas larger objects are detected in higher levels.
The value for 'min_level' needs to be at least 2.
If 'max_level' is larger than the number of levels the backbone can provide, the backbone is extended with additional (randomly initialized) convolutional layers in order to generate deeper levels. Further, 'max_level' may have an influence on the minimal input image size.
Note, for small input image dimensions, high levels might not be meaningful, as the feature maps could already be too small to contain meaningful information.
A higher number of used levels might increase the runtime and memory-consumption, whereby especially lower levels carry weight.
Default: 'max_level' = 6, 'min_level' = 2
This parameter determines the maximum number of detections (bounding boxes) per image proposed from the network.
Default: 'max_num_detections' = 100
The maximum allowed intersection over union (IoU) for two predicted bounding boxes of the same class. Or, vice-versa, when two bounding boxes are classified into the same class and have an IoU higher than 'max_overlap', the one with lower confidence value gets suppressed. We refer to the chapter Deep Learning / Object Detection for further explanations to the IoU.
Default: 'max_overlap' = 0.5
The maximum allowed intersection over union (IoU) for two predicted bounding boxes independently of their predicted classes. Or, vice-versa, when two bounding boxes have an IoU higher than 'max_overlap_class_agnostic', the one with lower confidence value gets suppressed. As default, 'max_overlap_class_agnostic' is set to 1.0, hence class agnostic bounding box suppression has no influence.
Default: 'max_overlap_class_agnostic' = 1.0
This parameter determines the minimum confidence, when the image part
within the bounding box is classified in order to keep the proposed
bounding box.
This means, when apply_dl_model is called, all output bounding
boxes with a confidence value smaller than 'min_confidence'
are suppressed.
Default: 'min_confidence' = 0.5
This parameter determines whether for the oriented bounding box also the direction of the object within the bounding box is considered or not. In case the direction within the bounding box is not to be considered you can set 'ignore_direction' to 'true'. In order to determine the bounding box unambiguously, in this case (but only in this case) the following conventions apply:
'phi'
'bbox_length1' > 'bbox_length2'
This is consistent to smallest_rectangle2.
Note, this parameter only affects networks of 'instance_type' = 'rectangle2'.
Possible values: 'true', 'false'
Default: 'ignore_direction' = 'false'
With this parameter you can declare one or multiple classes as 'ignore' classes, see the chapter Deep Learning / Semantic Segmentation for further information. These classes are declared over their ID (integers).
Note, you can not set a class ID in 'ignore_class_ids' and 'class_ids' simultaneously.
DLModelHandle (input_control) dl_model → (handle)
Handle of the deep learning model.
GenParamName (input_control) attribute.name → (string)
Name of the generic parameter.
Default value: 'batch_size'
List of values: 'anchor_angles', 'anchor_aspect_ratios', 'anchor_num_subscales', 'backbone', 'backbone_docking_layers', 'batch_size', 'batch_size_multiplier', 'bbox_heads_weight', 'capacity', 'class_heads_weight', 'class_ids', 'class_ids_no_orientation', 'class_names', 'class_weights', 'complexity', 'extract_feature_maps', 'gpu', 'ignore_class_ids', 'ignore_direction', 'image_dimensions', 'image_height', 'image_num_channels', 'image_range_max', 'image_range_min', 'image_width', 'instance_type', 'learning_rate', 'max_level', 'max_num_detections', 'max_overlap', 'max_overlap_class_agnostic', 'min_confidence', 'min_level', 'momentum', 'num_classes', 'optimize_for_inference', 'runtime', 'standard_deviation_factor', 'summary', 'type', 'weight_prior'
GenParamValue (output_control) attribute.name(-array) → (integer / string / real)
Value of the generic parameter.
If the parameters are valid, the operator get_dl_model_param
returns the value 2 (H_MSG_TRUE). If necessary, an exception is raised.
read_dl_model,
set_dl_model_param
set_dl_model_param,
apply_dl_model,
train_dl_model_batch,
train_dl_model_anomaly_dataset
Deep Learning Inference