can extract be reused？

[cscyangj]

<1> below code get correct result;
for(i = 0 ; i < I; i++)
{
ncnn::Extractor Net_ex_ = Net_.create_extractor();
Net_ex_.set_light_mode(true);

Net_ex_.input("data", input);
Net_ex_.extract("conv3",output);

}

<2> after i change the code below, can not get correct result, so can bet extract be re-used?In my opinion, net should be independent to data input to net.

ncnn::Extractor Net_ex_ = Net_.create_extractor();
Net_ex_.set_light_mode(true);
for(i = 0 ; i < I; i++)
{
Net_ex_.input("data", input);
Net_ex_.extract("conv3",output);
}

[nihui]

Net can be reused, as you've shown in <1>
Extractor is the context of each inference, which should not be reused.

[cscyangj]

@nihui ,thanks.
in such case, if Extractor can be re-used, may get better performance.

[nihui]

Creating an Extractor is nearly instantly like allocating an array.
you will get little performance if reused, compared with the inference itself.

[GarrickLin]

In mtcnn, there are a few pyramids, can I create a number of pnets to avoid reallocate memory again and again ?

[Paseam]

Creating an Extractor is nearly instantly like allocating an array.
you will get little performance if reused, compared with the inference itself.

but i think every time extractor will allocate new memory for intermediate layer output and release it later, this will be bad results for meachine