create_dl_layer_class_id_conversion — Create a class ID conversion layer.
create_dl_layer_class_id_conversion( : : DLLayerInput, LayerName, ConversionMode, GenParamName, GenParamValue : DLLayerClassIdConversion)
The operator create_dl_layer_class_id_conversion creates a class ID
conversion layer whose handle is returned in DLLayerClassIdConversion.
The layer converts between the IDs used internally by the network and the
target / output class IDs.
The network internally uses consecutive integer values starting from 0
as IDs (the number of values depends on the model type). In case
the target / output class IDs differ from the internal IDs, this layer
can be used to convert between them. The target / output class IDs are
stored in the model parameter 'class_ids' (see
get_dl_model_param for more information on this parameter).
If no 'class_ids' are set, this layer copies the input to the
output.
The parameter ConversionMode specifies the conversion direction
and accepts the following values:
'from_class_id': Convert class IDs into internal IDs. This mode is typically used after a target input layer.
'to_class_id': Convert internal IDs into class IDs. This mode is typically used after an inference output layer.
The parameter DLLayerInput determines the feeding input layer.
The parameter LayerName sets an individual layer name.
Note that if creating a model using create_dl_model each layer of
the created network must have an unique name.
The following generic parameters GenParamName and the corresponding
values GenParamValue are supported:
Determines whether apply_dl_model will include the output of this
layer in the dictionary DLResultBatch even without specifying this
layer in Outputs ('true') or not ('false').
Default: 'false'
Number of trainable parameters (weights and biases) of the layer.
Certain parameters of layers created using this operator
create_dl_layer_class_id_conversion can be set and retrieved using
further operators.
The following tables give an overview, which parameters can be set
using set_dl_model_layer_param and which ones can be retrieved
using get_dl_model_layer_param or get_dl_layer_param. Note, the
operators set_dl_model_layer_param and get_dl_model_layer_param
require a model created by create_dl_model.
| Layer Parameters | set | get |
|---|---|---|
'input_layer' (DLLayerInput) |
||
'name' (LayerName) |
||
'output_layer' (DLLayerClassIdConversion) |
||
| 'shape' | ||
'to_class_id' (ConversionMode) |
||
| 'type' |
| Generic Layer Parameters | set | get |
|---|---|---|
| 'is_inference_output' | ||
| 'num_trainable_params' |
DLLayerInput (input_control) dl_layer → (handle)
Feeding layer.
LayerName (input_control) string → (string)
Name of the layer.
ConversionMode (input_control) string → (string)
Direction of the class ID conversion.
Default value: 'from_class_id'
List of values: 'from_class_id', 'to_class_id'
GenParamName (input_control) attribute.name(-array) → (string)
Generic input parameter names.
Default value: []
List of values: 'is_inference_output'
GenParamValue (input_control) attribute.value(-array) → (string / integer / real)
Generic input parameter values.
Default value: []
Suggested values: 'true', 'false'
DLLayerClassIdConversion (output_control) dl_layer → (handle)
Class IDs conversion layer.
* Example demonstrating the usage of
* create_dl_layer_class_id_conversion.
*
dev_update_off ()
set_system ('seed_rand', 42)
*
* Create simple segmentation model.
NumClasses := 3
InputShape := [32, 32, 3]
*
* Input feeding layers.
create_dl_layer_input ('image', InputShape, [], [], DLLayerInput)
create_dl_layer_input ('target', [InputShape[0],InputShape[1],1], [], [], \
DLLayerTarget)
create_dl_layer_class_id_conversion (DLLayerTarget, 'target_internal', \
'from_class_id', [], [], \
DLLayerTargetInternal)
* Feature extraction layers.
create_dl_layer_convolution (DLLayerInput, 'conv1', 3, 1, 1, 32, 1, \
'half_kernel_size', 'relu', [], [], \
DLLayerConv1)
create_dl_layer_convolution (DLLayerConv1, 'conv2', 3, 1, 1, 32, 1, \
'half_kernel_size', 'relu', [], [], \
DLLayerConv2)
* Output generation layers.
create_dl_layer_convolution (DLLayerConv2, 'conv_final', 1, 1, 1, \
NumClasses, 1, 'none', 'none', [], [], \
DLLayerConvFinal)
create_dl_layer_softmax (DLLayerConvFinal, 'softmax', [], [], \
DLLayerSoftMax)
create_dl_layer_depth_max (DLLayerSoftMax, 'output_internal', \
'argmax', [], [], DLLayerOutputInternal, _)
create_dl_layer_class_id_conversion (DLLayerOutputInternal, 'output', \
'to_class_id', [], [], DLLayerOutput)
* Loss layer.
create_dl_layer_loss_cross_entropy (DLLayerSoftMax, DLLayerTargetInternal, \
[], 'loss', 1.0, [], [], DLLayerLoss)
*
* Create the model.
create_dl_model ([DLLayerOutput, DLLayerLoss], DLModelHandle)
set_dl_model_param (DLModelHandle, 'type', 'segmentation')
set_dl_model_param (DLModelHandle, 'runtime', 'cpu')
*
* Test model on dummy example data.
read_image (Image, 'claudia')
zoom_image_size (Image, Image, InputShape[0], InputShape[1], 'constant')
convert_image_type (Image, Image, 'real')
*
* Fill target image with specific target class IDs.
ClassIDs := [42, 17, 5]
gen_image_const (Target, 'real', InputShape[0], InputShape[1])
paint_region (Target, Target, Target, ClassIDs[0], 'fill')
gen_rectangle1 (RectClass1, 1, 3, 16, 27)
paint_region (RectClass1, Target, Target, ClassIDs[1], 'fill')
gen_rectangle1 (RectClass2, 19, 1, 30, 30)
paint_region (RectClass2, Target, Target, ClassIDs[2], 'fill')
*
* Set class IDs in the model.
set_dl_model_param (DLModelHandle, 'class_ids', ClassIDs)
*
* Create test sample.
create_dict (DLSample)
set_dict_object (Image, DLSample, 'image')
set_dict_object (Target, DLSample, 'target')
*
* Train model for a few iterations. Note that training would not
* work without the first class ID conversion layer 'target_internal'.
for Idx := 1 to 100 by 1
train_dl_model_batch (DLModelHandle, DLSample, DLTrainResult)
endfor
*
* Apply model on test image. With the second class ID conversion
* layer 'output', the image now contains values according to the
* target IDs in segmentation_image.
apply_dl_model (DLModelHandle, DLSample, [], DLApplyResult)
get_dict_object (SegmentationImage, DLApplyResult, 'output')
dev_display (SegmentationImage)
Deep Learning Training