Prune and transform an existing TF2 model


Is there any way in TAO to load a TensorFlow saved_model.pb model in order to directly apply pruning and transformation to it? It is an SSD Mobilenet_v2 (640x640) object detection model that I have already trained, and I want to see if I can improve the performance and throughput without the need of training it again from scratch in TAO.

I tried transforming it to Keras .h5 or .ONNX but none of the scripts that I found worked.

I am trying this because I need an SSD model with 640x640 input images, and the ones in the Nvidia NGC are 300x300.

Thanks in advance.

There is no update from you for a period, assuming this is not an issue anymore. Hence we are closing this topic. If need further support, please open a new one. Thanks

Officially, BYOM with TF1 (Classification and UNet) in TAO can convert any open-source ONNX model to a TAO-compatible model. The TAO BYOM Converter provides a CLI to import an ONNX model and convert it to Keras. The converted model is stored in .tltb format. Refer to Bring Your Own Model (BYOM) - NVIDIA Docs.

For your mentioned 3rd-party TensorFlow saved_model.pb model, it should be possible to save to .h5 format file.

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.