def generate_submodel(self, submodel):
"""
Generates multiple PMML object for the regression and classification submodel of RetinaNet for each connected pyramid layers
Parameters
----------
submodel :
The Regression or the Classification submodel
Returns
-------
List of Nyoka's NetworkLayer object for all the submodels
"""
net_layers_group = list()
for idx, name in enumerate(self._pyramid_layers):
nyoka_pmml_reg_mod = kerasAPI.KerasToPmml(submodel)
del nyoka_pmml_reg_mod.DeepNetwork[0].NetworkLayer[0]
nyoka_pmml_reg_mod.DeepNetwork[0].NetworkLayer[
0].connectionLayerId = name
for idx_, lay in enumerate(
nyoka_pmml_reg_mod.DeepNetwork[0].NetworkLayer):
lay.layerId = lay.layerId + "_" + name
if idx_ != 0:
lay.connectionLayerId = lay.connectionLayerId + "_" + name
net_layers_group.extend(
nyoka_pmml_reg_mod.DeepNetwork[0].NetworkLayer)
return net_layers_group
def generate_beckbone_anchors(self, model, input_format, trained_classes):
"""
Generates PMML object for the backbone + anchors
Parameters
----------
model :
RetinaNet model object
input_format : string
Input format to be used during inference with the PMML. Valid values are -
"image" : Original image in png format
"encoded" : Base64 encoded string of the image
trained_classes : List
List of class names for which the model was trained
Returns
-------
Nyoka's PMML object
"""
from keras.models import Sequential
mod = Sequential()
for l in model.layers[1:]:
if l.__class__.__name__ == "Model":
break
mod.add(l)
if trained_classes == None:
warnings.warn(
f"trained_classes are not provided. Maximum 80 classes will be considered."
)
trained_classes = [
"Category_" + str(i + 1).zfill(2) for i in range(80)
]
group1_pmml = kerasAPI.KerasToPmml(mod,
model_name=self.model_name,
dataSet=input_format,
description=self.description,
predictedClasses=trained_classes,
script_args=self.script_args)
return group1_pmml