does TTSAndroid support int8 infer?

[elliotzheng]

General Question

I Quantify speedyspeech and mb_melban models using paddle-lite with the following command.
./opt_linux_x86 --model_file=speedyspeech_csmsc.pdmodel --param_file=speedyspeech_csmsc.pdiparams --valid_targets=arm --optimize_out_type=naive_buffer --optimize_out=./speedyspeech_csmsc_int8.nb --quant_model=true --quant_type=QUANT_INT8
and copy the two models to TTSAndroid/app/src/main/assets/models/cpu/, then change model names in mainactivity.java, finally change input tensor type from float to long.
But no decrease of cpu usage or infer time is seen on my armv8 cpu phone.
So I want to ask whether int8 infer is supported or not?
Any help is appreciated.

[zxcd]

Maybe you can try convert model to quantized model first , such as:

PaddleSpeech/examples/csmsc/tts3/run.sh

Lines 77 to 83 in 50ef94b

	# PTQ_dynamic
	if [ ${stage} -le 9 ] && [ ${stop_stage} -ge 9 ]; then
	./local/PTQ_dynamic.sh ${train_output_path} fastspeech2_csmsc 8
	# ./local/PTQ_dynamic.sh ${train_output_path} pwgan_csmsc 8
	# ./local/PTQ_dynamic.sh ${train_output_path} mb_melgan_csmsc 8
	# ./local/PTQ_dynamic.sh ${train_output_path} hifigan_csmsc 8
	fi