Darknet: Memory Leak in detect/classify functions in Python API.

[nivangio]

It seems that detect is not correctly freeing memory in Python. I ran detection on multiple images (basically on a video) and found out that memory leaks slowly. After isolating the different components I found out it was an issue of detect function. Is it possible that free_imgs() and free_ptrs() are not working correctly.

from cv2 import VideoCapture, imwrite
import sys
from tempfile import NamedTemporaryFile

cap = VideoCapture(0)

sys.path.append("darknet/python")

from darknet import detect, load_net, load_meta

yolo_net = load_net("darknet/cfg/yolo.cfg",
                    "darknet/weights/yolo.weights",0)
yolo_meta = load_meta("darknet/cfg/coco.data")

def detect_objects_darknet(img, net, meta):

    tempf = NamedTemporaryFile(suffix=".jpg")
    imwrite(tempf.name,img)

    detect(net=net, meta=meta, image=tempf.name)

    return 1


while (cap.isOpened()):
    aaa, bbb = cap.read()
    if aaa == False:
        cap.release()

    else:
        detect_objects_darknet(img = bbb,net = yolo_net, meta = yolo_meta)

If I comment out the detect line, the memory usage stays stable, tht's why I am assuming the issue relies there and not anywhere else. Please let me know if you see sth different

[johnsilvester]

+1

[johnsilvester]

I'm having the same issue. Any updates?

[nivangio]

I managed to solve it by adding some memory freeing lines in Python's detect and classify functions

def classify(net, meta, im):
    out = predict_image(net, im)
    res = []
    for i in range(meta.classes):
        res.append((meta.names[i], out[i]))
    res = sorted(res, key=lambda x: -x[1])
    free_image(im)
    return res

def detect(net, meta, image, thresh=.5, hier_thresh=.5, nms=.45):
    im = load_image(image, 0, 0)
    boxes = make_boxes(net)
    probs = make_probs(net)
    num =   num_boxes(net)
    network_detect(net, im, thresh, hier_thresh, nms, boxes, probs)
    res = []
    for j in range(num):
        for i in range(meta.classes):
            if probs[j][i] > 0:
                res.append((meta.names[i], probs[j][i], (boxes[j].x, boxes[j].y, boxes[j].w, boxes[j].h)))
    res = sorted(res, key=lambda x: -x[1])
    free_image(im)
    lib.free(boxes)
    free_ptrs(cast(probs, POINTER(c_void_p)), num)
    return res

[AndyYangMao]

I solve this issue by modifying the original darknet.py by add:

free_network = lib.free_network
reset_rnn.argtypes = [c_void_p]

and

def freeNetwork(net):
free_network(net)
return 0

this will create a new interface to free the memory usage after loading the net and finish the detection

example code will be something like this
import darknet as dn
dn.set_gpu(0)
net = dn.load_net(bytes(cfg_name, encoding='utf-8'), bytes(weight_name, encoding='utf-8'), 0)
meta = dn.load_meta(bytes(meta_name,encoding='utf-8'))
outs = dn.detect(net, meta, bytes(test_img_name,encoding='utf-8'), thresh, hier_thresh, nms)
dn.freeNetwork(net)