ANE support?
Opened this issue · 3 comments
s2nnc seems a bit slower compared to coreml. any plans to support ANE?
There is no official path for ANE support (see also tinygrad / PyTorch), but it seems plausible through MPSGraph with OptimizationLevel1 (that seems to use ANE).
Having said that, ANE support through that seems require op authors to maintain a list which can be successfully translated, which will not, and that is a very labor-intensive task involves a lot of trials-and-errors with different devices (the ANE support on different devices also varies).
hmm, how is apple doing it for TF and pytorch. any idea guys?
Maybe try building interface for non apple NEs on the market
I am not sure about TF. PyTorch doesn't support ANE. The way Apple doing is to introduce coremltools which can trace TF / PyTorch execution and build computation graph internally through CoreML to support computing capabilities. That's why it doesn't support full set of PyTorch ops.