Are there any best practices for using torchscript with pytorch-mobile? Im currently trying to implement a conv-net with pytorch-mobile on android. The model is very slow in comparison to an identical tensorflow trained tflite model. Perhaps this might have something to do with the torchscript tracing.
I tries freeze the graph using torch.utils.mobile_optimizer.optimize_for_mobile which is new for pytorch 1.6.0. It seems to improve performance a little but not significantly.
Did you find anything new? We are experiencing a very similar pattern in terms of performance while comparing a TorchScript model against its TFLite counterpart