[CONTRIB] TFLite Runtime

[ZihengJiang]

A runtime that can load and execute TFLite models via TVM calling convention, while compiled tflite library linked with tvm runtime. This enables us to do streamline baseline benchmarks of tflite over RPC:

# code snippet shows how to do inference via remote tvm tflite runtime
server = rpc.Server(remote_address)
remote = rpc.connect(server.host, server.port)
ctx = remote.cpu(0)

# create remote tflite runtime
tflite_model_bytes = open('model.tflite', 'rb').open()
runtime = tflite_runtime.create(tflite_model_bytes, ctx)
runtime.allocate_tensors()
runtime.set_input(0, tvm.nd.array(input_data, ctx)))
runtime.invoke()
out = runtime.get_output(0)

@tmoreau89 @tqchen

[tqchen]

Please modify the description to add a bit about motivated usecase. e.g. streamline baseline benchmarks of tflite over RPC

[tmoreau89]

LGTM

[FrozenGene]

cc @jackwish maybe you are interested

[tqchen]

Thanks @ZihengJiang @tmoreau89 !