How to use batch inference?

[zengjie617789]

I want to test the images with batch inference.
I construct the data which shape is (target_size, target_size, batch_size, 3), and the key code is like below:

for (int i = 0; i < batch_size; ++i) {
            dfd_ex.input(0, in_batch);
            ncnn::Mat out;
            dfd_ex.extract(499, out);
            float pred = out.row(0)[0];
            pred = (1.f / (1.f + exp(-pred)));
            preds.push_back(pred);
        }

I found the pred results is not correct and alway got the same outputs;
Futhermore, I found the official doc like this:

int max_batch_size = vkdev->info.compute_queue_count();

ncnn::Mat inputs[1000];
ncnn::Mat outputs[1000];

#pragma omp parallel for num_threads(max_batch_size)
for (int i=0; i<1000; i++)
{
    ncnn::Extractor ex = net1.create_extractor();
    ex.input("data", inputs[i]);
    ex.extract("prob", outputs[i]);
}

I canot do this like because I create a data which make batch_size as the depth dims.
Anyone who can help me will be appreciated !