Run code on the mobile about mobile-segmentation

[sercant]

Hi @InstantWindy, thank you for your interest. There is no plan to implement the code in pytorch at the moment but a re-write using the new tensorflow API might happen. Since tensorflow is easy to be exported for mobile targets such as smartphones using tf-lite, it was the choice of framework. I am not sure if pytorch runs on Android or iOS.

To run on the network on a mobile platform one should export the trained model to tf-lite using the export_tflite.py. After that, you can see a simple Android application that I have implemented at this repository: sercant/android-segmentation-app.

[InstantWindy]

Thanks for your reply. How do you implement DPC(dense predictioncell ) ?

[sercant]

I'm not quite sure what did you mean by how do I implement DPC but it is a part of the network so the exported model already contains the architecture. The architecture of the module is described in the paper and also here in the repository you can see the core/dense_prediction_cell_branch5_top1_cityscapes.json file and model.py for the implementation.

[InstantWindy]

I mean how you get the DPC module . How do you get the parameters of file "dense_prediction_cell_branch5_top1_cityscapes.json" . Thanks!

[sercant]

The parameters of the DPC module was found in Searching for Efficient Multi-Scale Architectures for Dense Image Prediction by LC Chen. The code that parses the JSON file is under the core folder. Let me know if this answers your question and don't hesitate to ask more if you have any.