Can tflite models be compiled for ARM cpu?
Yes. Earlier the problem was that ARM schedule did not have NHWC schedule.
Currently, Relay’s Legalize pass inserts layout transform before and after each conv to go to NCHW, to use the ARM NCHW schedule.
I am working on a PR that optimizes this little bit further, reducing the number of transposes. Ideally, there will be only 2 transposes in whole network
ok, I tried to compile
mobilenet_v2_0.5_224.tflite for ARMv8.
It works. Output result is correct. Inference time is 2x faster than in TFLite. Great!
Wohoo! Awesome. My current PR should hopefully get little more.