provide pytorch_android_lite version？

[theta666]

Since Pytorch team is updating pytorch_android to pytorch_android_lite, I found the newly built version can only use models created by "model._save_for_lite_interpreter",instead of ".save", so do we have such version model file?

[snakers4]

Are you referring to this tutorial - https://pytorch.org/tutorials/prototype/lite_interpreter.html or this - https://pytorch.org/tutorials/recipes/mobile_interpreter.html?

@adamnsandle can you try exporting using this method for a test?

[theta666]

yes, another model of my program is saved in this "lite" way, and loaded by manually built pytorch-android.aar(inside jni/arm64-v8: alibc++shared.so ibfbjni.so libpytorch_jni_lite.so), then I found my vad codes(using model_micro_mobile.jit) not working

[theta666]

I happened to find that I can convert the pytorch_android to pytorch_android_lite by this:

model = torch.jit.load('./old.jit', map_location=torch.device('cpu'))
model._save_for_lite_interpreter('lite.jit')

and load it on android

import org.pytorch.LiteModuleLoader;
Module vadmodule = LiteModuleLoader.load(modulePath);

results are slightly different, but okay to use
inference time is longer, 2.7ms(pytorch_android 1.8) to 3.3ms(pytorch_android_lite 1.10?) per 4000/16000 second

[snakers4]

per 4000/16000 second

I believe that the recommended chunk sizes for 16 kHz are [512, 1024, 1536] samples

https://github.com/snakers4/silero-vad/blob/master/utils_vad.py#L196

[theta666]

thank you, I realize the result is computed using first version model

[snakers4]

please use the newest models