georgymironov / single_shot_detection Goto Github PK

Hello, firstly thanks for open-sourcing this wonderful and clean code for embedding different object detectors.
I am trying to export a trained m2det model to onnx format and then do the inference by wrapping it in a tensorflow serving module. However, in the exported model only the prediction outputs of network are included by excluding the anchor box priors. These priors are needed to post-process and derive the final object boxes. I tried looking into your process of generating the prior boxes, but it seems to be tightly interlinked with the model architecture itself and this makes it necessary to always keep the source code around. Is there a way I could disregard the source code for architecture and generate the anchor boxes as a standalone script so that I could finally use only the frozen model, priors and post-processing functions to generate predictions? Thanks in advance.

hello, how to input the mobilenetv2 pretrained model.