GenSynth Documentation

Model Builder Packages

Model Builder packages provide scripts to automatically build custom, ready-to-go models based on your custom specifications (e.g., number of classes, input resolution, etc.) The Model Builder packages also include full instructions for how to easily use the custom-built models with GenSynth to create high-performance, highly efficient models for your operational targets. When using the built models, make sure to turn Improve with GenSynth Learn on in GenSynth.

Note

As of GenSynth 1.18 you can build all model builders using the New Model Wizard through the GenSynth UI.

Available Model Builder packages include:

Package Name

Description

Applies to:

autoencoder_fc_builder

Fully-connected autoencoder model builder. An auto-encoder for anomaly detection.

1.17+

avod_builder

Provides a script to create an Aggregate View Object Detection (AVOD) model to detect 3D objects.

1.12+

centernet_builder

Provides a script to create a centernet_resnet50 object-detection model according to parameters of image size, number of classes, and number of boxes.

1.13+

dncnn_builder

Provides a script to build custom dncnn models.

1.14+

depthnet_builder

Provides a script to create a depth estimation model called DepthNet.

1.12+

edge_speechnet_residual15_builder

Provides a script to create an untrained EdgeSpeechNet model for classifying speech utterances. You may specify the number of classes.

1.11+

frcnn_resnet50_builder

Provides a script to construct an untrained Faster-R-CNN (ResNet50) object detection network. You may specify the input image size and number of classes.

1.12+

inceptionv1_builder

Provides a script to create an InceptionV1 model for image classification.

1.15+

maskrcnn_builder

Provides a script to create a MaskRCNN model for image segmentation.

1.13+

refinenet_builder

Provides a script to construct a RefineNet image semantic segmentation model using a ResNet50 backbone, with untrained front-end weights. You may specify a configuration file for mapping the label-image color palette to pixel classes.

Includes instructions for use with the CamVid dataset.

1.11+

resnet_builder

Provides a script to create an untrained ResNet image classification model for a specified number of classes.

1.11+

ssd_fpn_resnet50_builder

Provides a script to construct an untrained SSD-FPN (ResNet50) object detection network. You may specify the input image size and number of classes.

1.11+

ssdlite_mobilenet_builder

Provides a script to build an SSDLite MobileNet network with custom parameters.

1.14+

yolov3_builder

Provides a script to create a YOLOv3 model to detect 2D objects. You may specify the input image size and number of classes.

1.12+