yolov4 prediction crashes on mac
bogdad opened this issue · 4 comments
hey,
thank you for opencl version, this seems very cool to have!
i am getting this crash on mac big sour, trying to run yolo4, regular one or tiny.
(lldb) target create "./darknet"
Current executable set to '/Users/vladimir/workspace/darknet/darknet' (x86_64).
(lldb) settings set -- target.run-args "-i" "1" "yolo4" "test" "cfg/yolov4-tiny.cfg" "yolov4-tiny.weights" "1.png"
(lldb) run
Process 61001 launched: '/Users/vladimir/workspace/darknet/darknet' (x86_64)
Device IDs: 2
Device ID: 1
Device name: AMD Radeon Pro 560X Compute Engine
Device vendor: AMD
Device opencl availability: OpenCL 1.2
Device opencl used: 1.2 (May 8 2021 06:19:33)
Device double precision: YES
Device max group size: 256
Device address bits: 32
layer filters size input output
0 conv 32 3 x 3 / 2 416 x 416 x 3 -> 208 x 208 x 32 0.075 BFLOPs
1 conv 64 3 x 3 / 2 208 x 208 x 32 -> 104 x 104 x 64 0.399 BFLOPs
2 conv 64 3 x 3 / 1 104 x 104 x 64 -> 104 x 104 x 64 0.797 BFLOPs
...
[yolo4] params: iou loss: ciou (4), iou_norm: 0.07, obj_norm: 1.00, cls_norm: 1.00, delta_norm: 1.00, scale_x_y: 1.05
nms_kind: greedynms (1), beta = 0.600000
Loading weights from yolov4-tiny.weights...done!
darknet was compiled with optimization - stepping may behave oddly; variables may not be available.
Process 61001 stopped
* thread #1, queue = 'com.apple.main-thread', stop reason = EXC_BAD_ACCESS (code=1, address=0x0)
frame #0: 0x000000010003f89c darknet`avg_predictions(net=0x00007ffeefbf89f0, nboxes=0x00007ffeefbf91cc) at demo.c:75:45 [opt]
72 int count = 0;
73 fill_cpu(demo_total, 0, avg, 1);
74 for(j = 0; j < demo_frame; ++j){
-> 75 axpy_cpu(demo_total, 1./demo_frame, predictions[j], 1, avg, 1);
76 }
77 for(i = 0; i < net->n; ++i){
78 layer l = net->layers[i];
Target 0: (darknet) stopped.
(lldb) bt
* thread #1, queue = 'com.apple.main-thread', stop reason = EXC_BAD_ACCESS (code=1, address=0x0)
* frame #0: 0x000000010003f8cc darknet`avg_predictions(net=0x00007ffeefbf89f0, nboxes=0x00007ffeefbf91cc) at demo.c:75:45 [opt]
frame #1: 0x00000001000b370b darknet`test_yolo4(cfgfile=<unavailable>, weightfile=<unavailable>, filename="1.png", thresh=0.200000003) at yolo.c:647:16 [opt]
frame #2: 0x00000001000b3b8c darknet`run_yolo4(argc=<unavailable>, argv=<unavailable>) at yolo.c:687:36 [opt] [artificial]
frame #3: 0x00000001000d075f darknet`main(argc=8, argv=0x00007ffeefbfeca0) at darknet.c:452:9 [opt]
frame #4: 0x00007fff2065bf5d libdyld.dylib`start + 1
(
it seems predictions is 0x000000000...
tried debugging a bit and seems in pull_network_output(netp); in the
1295 void pull_network_output(network *net)
1296 {
1297 layer l = get_network_output_layer(net);
-> 1298 opencl_pull_array(l.output_gpu, l.output, l.outputs*l.batch);
1299 }
l.output and l.output_gpu are not zero. but also no pointers to what may be wrong.
can continue debugging, will appreciate hints though!
It looks like a hardware issue to me... you may change to opencl_pull_array_map (same parameters) function in line 1298, but I doubt it helps. I have just checked. 
it works!
@bogdad the command is:
./darknet detect cfg/coco.data cfg/yolov4-tiny.cfg ../weights/yolov4-tiny.weights data/dog.jpg -i 1
and in the cfg file, you should replace [yolo] with [yolo4] once I support [yolo] tag as ver. 3, ok..?
but it is an unknown issue that version v4 of Yolo works as I showed you but tiny does... not... yet!
let me add my 2 cents... error is in the clBLAS and/or in CLBlast for GEMM computing... maybe I should find some other project that multiplies matrixes? I am not sure... anyway, appreciate your try to find the issues ;-). thanks!
you are trying to use yolo4-tiny that did not work.. till now :).
