Hi,
I read SNPE-SDK doc and found this sentence:
[ --bitwidth=<val> ] Use the --bitwidth option to select the bitwidth to use when quantizing the weights/activation/bias, either 8 (default) or 16. Can't mix with --weights_bitwidth or --act_bitwidth or --bias_bitdwith
But when I used the quantization command below, I reported an error
command : snpe-dlc-quant --input_dlc=xxx.dlc --input_list=xxx.txt --output_dlc=xxx.dlc --bitwidth=16
ERROR INFO : IrQuantizer: Unsupported weight bitwidth: 16
BR.
Dear developer,
Your snpe version is not supported bitwidth=8.
BR.
Yunxiang
Dear Yunxiang
I use --bitwidth=16 because the 8-bit quantization model has insufficient accuracy.
uesd snpe version : snpe-2.10.0.4541
BR.
Dear developer,
You can use the commands '--skip_quantization -enabel_htp' to improve the accuracy. And it's for fp16 quantization.
Br.
Yunxiang
Dear Yunxiang
I quantified the model according to your method, but failed to load the quantified model on the Snapdragon 888 chip. How can I solve this?
SNPE version: snpe-2.10.0.4541
BR.
Dear developer,
Would you mind providing the model you used? You can upload the model to git and I will run it on Snapdragon 888 .
Br.
Yunxiang
Dear Yunxiang
sorry, i can't providing the model.
The type of model is object detection.
BR.