Limitations of Torch Mobile//Torchscript (Object Detection + Segmentation)


I was wondering if its possible to script YOLO, MobileSSD, or FasterRCNN as Object Detection models to run in iOS or android via Libtorch on Pytorch mobile. Are all of these models currently unavailable? Is there a custom approach to running FasterRCNN via Torch/TorchScript on mobile?

Same question but in the case of Semantic Segmentation models such as PSPNet, DeepLabv3, and MaskRCNN.

I believe someone mentioned that TorchVision is not supported but I don’t understand the difference between that and the Resnet and Inception models that are already currently supported/scriptable for mobile. Are resnets/inceptions not considered Torchvision?

Thanks in advance,

Basic torchvision models are supported, but specialized ops for detection are currently not included. We are looking into it.


Is there an ETA for obj detection in PyTorch mobile? In the meantime is there a tutorial on how to use Caffe2 mobile for obj detection?

1 Like