Save model for inference. How to generate *.prototxt?

[IgorKasianenko]

Hello, my goal is to run Maks R-CNN inference independently without depending on Detectron code base.

First approach was to convert model to *.pb, but at the moment FPN is not supported even that caffe2 team is working on it.

Meanwhile I'd like to run the model in C++, according to nVidia guide. Their sample uses FasterRCNN, from this source py-faster-rcnn that has been deprecated and leads to Detectron repo. Nevertheless, py-faster-rcnn has very useful file test.prototxt, with all layers description. That file is needed to write cpp plugins for custom layers, that can't be translated.

My question is how can I generate or find maskrcnn.prototxt file for Detectron, for two solutions, VGG16 and ResNet101?

I assume in case of ResNet101 it would be much bigger, as Detectron is not only adding masking to FasterRCNN, but also adds Feature Pyramid Networks.

[rbgirshick]

The only viable route is to use the convert_pkl_to_pb.py script, but as you noted it does not yet support the full set of models that Detectron can train.