So, I’m working through the VTA tutorials, and it seems like it might be a good fit for what I need to do.
However, it seems to be a very manual process, and I’m wondering: is there an automated path from a TensorFlow design (presumably, imported via the relay.frontend.from_tensorflow() function) to a VTA-based implementation, which uses the configuration given in vta_config.json to do the transformation?
We are working on more automation, I believe there is room for automated graph compilation. For now we need to manually specify in the quantization and graph packing transformations what operators are targeted to the accelerator.