Regarding the compiled benchmark model for arm64-v8a

[elenazy]

It is mentioned that the latency measurements of tflite model on 4x ARM Cortex-A55 and 4x ARM Cortex-A75 in your paper. I would like to realize it. if you kindly provide the corresponding code. More importantly, I hope you provide a compiled benchmark model to facilitate the delay latency measurements on arm64-v8a-based Android phones. You can email me to [email protected]!

[tair1]

Regarding the TFLite models, you can find them in their respective folders: KWT1 KWT2 KWT3

For the latency measurement,s we used the benchmark tools provided by tensorflow. You can find more information on how to run it on this page: https://www.tensorflow.org/lite/performance/measurement

You can either install the benchmark apk or run it as a native binary (our recommended approach). Download the binary to a folder inside your phone, e.g. /data/local/tmp, send the keyword transformer .tflite model to your phone as well and then run benchmark through shell. You can either use a shell client on your phone or just use adb through a computer. I provide an example of a run command below using adb.

./adb shell /data/local/tmp/android_aarch64_benchmark_model --graph=/data/local/tmp/kwt1.tflite --warmup_runs=20 --num_runs=1000 --min_secs=0 --warmup_min_secs=0 --num_threads=1

You can see further documentation on how to run the adb commands here: https://github.com/ARM-software/keyword-transformer/blob/master/kws_streaming/experiments/kws_experiments_paper_12_labels.md#compile-tflite-benchmarking-tools (keep in mind that you can just download the latest benchmark binary and don't need to build it).