Operator Reference
get_dl_model_layer_weights (Operator)
get_dl_model_layer_weights — Get the weights (or values) of a Deep Learning model layer.
Signature
get_dl_model_layer_weights( : Weights : DLModelHandle, LayerName, WeightsType : )
Description
The operator get_dl_model_layer_weights returns in Weights
the values of a LayerName of the model DLModelHandle.
The parameter WeightsType determines which type of layer values are
retrieved.
The following values are supported for WeightsType:
-
'batchnorm_mean': Batch-wise calculated mean values to normalize the inputs. For further information, please refer to
create_dl_layer_batch_normalization.Restriction: This value is only supported if the layer is of type 'batchnorm'.
-
'batchnorm_mean_avg': Average of the batch-wise calculated mean values to normalize the inputs. For further information, please refer to
create_dl_layer_batch_normalization.Restriction: This value is only supported if the layer is of type 'batchnorm'.
-
'batchnorm_variance': Batch-wise calculated variance values to normalize the inputs. For further information, please refer to
create_dl_layer_batch_normalization.Restriction: This value is only supported if the layer is of type 'batchnorm'.
-
'batchnorm_variance_avg': Average of the batch-wise calculated variance values to normalize the inputs. For further information, please refer to
create_dl_layer_batch_normalization.Restriction: This value is only supported if the layer is of type 'batchnorm'.
-
'bias': Biases of the layer.
-
'bias_gradient': Gradients of the biases of the layer.
-
'bias_gradient_norm_l2': Gradients of the biases of the layer in terms of L2 norm.
-
'bias_norm_l2': Biases of the layer in terms of L2 norm.
-
'bias_update': Update of the biases of the layer. This is used in e.g., a solver which uses the last update.
-
'bias_update_norm_l2': Update of the biases of the layer in terms of L2 norm. This is used in a solver which uses the last update.
-
'weights': Weights of the layer.
-
'weights_gradient': Gradients of the weights of the layer.
-
'weights_gradient_norm_l2': Gradients of the weights of the layer in terms of L2 norm.
-
'weights_norm_l2': Weights of the layer in terms of L2 norm.
-
'weights_update': Update of the weights of the layer. This is used in a solver which uses the last update.
-
'weights_update_norm_l2': Update of the weights of the layer in terms of L2 norm. This is used in a solver which uses the last update.
The following tables give an overview, which parameters for WeightsType
can be set using set_dl_model_layer_weights and which ones can be retrieved
using get_dl_model_layer_weights.
| Layer Parameters | set |
get |
|---|---|---|
| 'batchnorm_mean' | x |
x
|
| 'batchnorm_mean_avg' | x |
x
|
| 'batchnorm_variance' | x |
x
|
| 'batchnorm_variance_avg' | x |
x
|
| 'bias' | x |
x
|
| 'bias_gradient' | x
|
|
| 'bias_gradient_norm_l2' | x
|
|
| 'bias_norm_l2' | x
|
|
| 'bias_update' | x
|
|
| 'bias_update_norm_l2' | x
|
|
| 'weights' | x |
x
|
| 'weights_gradient' | x
|
|
| 'weights_gradient_norm_l2' | x
|
|
| 'weights_norm_l2' | x
|
|
| 'weights_update' | x
|
|
| 'weights_update_norm_l2' | x
|
Attention
The operator get_dl_model_layer_weights is only applicable to self-created
networks. For networks delivered by HALCON, the operator returns an
empty tuple.
Execution Information
- Multithreading type: reentrant (runs in parallel with non-exclusive operators).
- Multithreading scope: global (may be called from any thread).
- Processed without parallelization.
Parameters
Weights (output_object) image(-array) → object (real)
Output weights.
DLModelHandle (input_control) dl_model → (handle)
Handle of the deep learning model.
LayerName (input_control) string → (string)
Name of the layer to be queried.
WeightsType (input_control) string → (string)
Selected type of layer values to be returned.
Default: 'weights'
List of values: 'batchnorm_mean', 'batchnorm_mean_avg', 'batchnorm_variance', 'batchnorm_variance_avg', 'bias', 'bias_gradient', 'bias_gradient_norm_l2', 'bias_norm_l2', 'bias_update', 'bias_update_norm_l2', 'weights', 'weights_gradient', 'weights_gradient_norm_l2', 'weights_norm_l2', 'weights_update', 'weights_update_norm_l2'
Example (HDevelop)
set_system ('seed_rand', 42)
* Create a small model network.
create_dl_layer_input ('input', [InputImageSize[0],InputImageSize[1],1], [],\
[], DLGraphNodeInput)
create_dl_layer_convolution (DLGraphNodeInput, 'conv', 3, 1, 1, 2, 1, 'none',\
'none', [], [], DLGraphNodeConvolution)
create_dl_layer_activation (DLGraphNodeConvolution, 'relu', 'relu', [], [],\
DLGraphNodeActivation)
create_dl_layer_dense (DLGraphNodeActivation, 'dense', 3, [], [],\
DLGraphNodeDense)
create_dl_layer_softmax (DLGraphNodeDense, 'softmax', [], [],\
DLGraphNodeSoftMax)
create_dl_model (DLGraphNodeSoftMax, DLModelHandle)
*
set_dl_model_param (DLModelHandle, 'type', 'classification')
set_dl_model_param (DLModelHandle, 'batch_size', 1)
set_dl_model_param (DLModelHandle, 'runtime', 'gpu')
set_dl_model_param (DLModelHandle, 'runtime_init', 'immediately')
*
* Train for 5 iterations.
for TrainIterations := 1 to NumTrainIterations by 1
train_dl_model_batch (DLModelHandle, DLSample, DLTrainResult)
endfor
*
* Get the gradients, weights, and activations.
get_dl_model_layer_gradients (GradientsSoftmax, DLModelHandle, 'softmax')
get_dl_model_layer_gradients (GradientsDense, DLModelHandle, 'dense')
get_dl_model_layer_gradients (GradientsConv, DLModelHandle, 'conv')
*
get_dl_model_layer_weights (WeightsDense, DLModelHandle, 'dense',\
'weights_gradient')
get_dl_model_layer_weights (WeightsConv, DLModelHandle, 'conv',\
'weights_gradient')
*
get_dl_model_layer_activations (ActivationsDense, DLModelHandle, 'dense')
get_dl_model_layer_activations (ActivationsConv, DLModelHandle, 'conv')
Possible Predecessors
create_dl_model,
train_dl_classifier_batch,
set_dl_model_layer_weights
Possible Successors
Alternatives
get_dl_model_layer_activations,
get_dl_model_layer_gradients
Module
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:
Deep Learning Professional