step1. train custom weights on macOS using origin darknet source code(alexAB's version tried, the same as this);
step2. test some images using custom weights on macOS, it's OK;
step3. compile darknet-nnpack on pi4;
step4. on pi4, using yolov3-tiny.weights is ok;
step5. on pi4, using custom weights , I cannot get any output except some messages
pi@raspberrypi:~/pt3t/Captcha/darknet-nnpack $ ./darknet detector test tiny_captcha/yolo.data tiny_captcha/yolov3-tiny.cfg tiny_captcha/backup/yolov3-tiny_900.weights tiny_captcha/samples/00000.png
layer filters size input output
0 conv 16 3 x 3 / 1 384 x 384 x 3 -> 384 x 384 x 16 0.127 BFLOPs
1 max 2 x 2 / 2 384 x 384 x 16 -> 192 x 192 x 16
2 conv 32 3 x 3 / 1 192 x 192 x 16 -> 192 x 192 x 32 0.340 BFLOPs
3 max 2 x 2 / 2 192 x 192 x 32 -> 96 x 96 x 32
4 conv 64 3 x 3 / 1 96 x 96 x 32 -> 96 x 96 x 64 0.340 BFLOPs
5 max 2 x 2 / 2 96 x 96 x 64 -> 48 x 48 x 64
6 conv 128 3 x 3 / 1 48 x 48 x 64 -> 48 x 48 x 128 0.340 BFLOPs
7 max 2 x 2 / 2 48 x 48 x 128 -> 24 x 24 x 128
8 conv 256 3 x 3 / 1 24 x 24 x 128 -> 24 x 24 x 256 0.340 BFLOPs
9 max 2 x 2 / 2 24 x 24 x 256 -> 12 x 12 x 256
10 conv 512 3 x 3 / 1 12 x 12 x 256 -> 12 x 12 x 512 0.340 BFLOPs
11 max 2 x 2 / 1 12 x 12 x 512 -> 12 x 12 x 512
12 conv 1024 3 x 3 / 1 12 x 12 x 512 -> 12 x 12 x1024 1.359 BFLOPs
13 conv 256 1 x 1 / 1 12 x 12 x1024 -> 12 x 12 x 256 0.075 BFLOPs
14 conv 512 3 x 3 / 1 12 x 12 x 256 -> 12 x 12 x 512 0.340 BFLOPs
15 conv 123 1 x 1 / 1 12 x 12 x 512 -> 12 x 12 x 123 0.018 BFLOPs
16 yolo
17 route 13
18 conv 128 1 x 1 / 1 12 x 12 x 256 -> 12 x 12 x 128 0.009 BFLOPs
19 upsample 2x 12 x 12 x 128 -> 24 x 24 x 128
20 route 19 8
21 conv 256 3 x 3 / 1 24 x 24 x 384 -> 24 x 24 x 256 1.019 BFLOPs
22 conv 123 1 x 1 / 1 24 x 24 x 256 -> 24 x 24 x 123 0.036 BFLOPs
23 yolo
Loading weights from tiny_captcha/backup/yolov3-tiny_900.weights...Done!
tiny_captcha/samples/00000.png: Predicted in 0.554214 seconds.
step1. train custom weights on macOS using origin darknet source code(alexAB's version tried, the same as this); step2. test some images using custom weights on macOS, it's OK; step3. compile darknet-nnpack on pi4; step4. on pi4, using yolov3-tiny.weights is ok; step5. on pi4, using custom weights , I cannot get any output except some messages