Yolov3之darknet下训练与预测

在做一些实时性要求比较高的目标检测时候，经常会选择Yolov3。本文介绍其训练和预测过程：
官网：https://pjreddie.com/darknet/
github地址：https://github.com/pjreddie/darknet

一、制作数据集

数据生成直接参见另一篇博文 Yolov3之生成训练数据

二、修改配置文件

本文以人手检测为例配置（只有一个label：hand）
1、添加或修改data/hand.names文件，此文件记录label，每行一个label(注意：对应于训练数据的labels)
2、添加或修改cfg/hand.data文件

classes= 1  # 自己数据集的类别数(不包含背景类)
train  = /home/xxx/darknet/train.txt  # train文件的路径
valid  = /home/xxx/darknet/test.txt   # test文件的路径
names = /home/xxx/darknet/data/hand.names
backup = /home/xxx/darknet/backup   # 生成权重存放文件夹，如果不存在，需提前创建

3、修改cfg/yolov3.cfg文件

(1)基本修改

(2)修改3处，直接根据“yolo”关键字查询定位，然后根据注释修改

三、训练

下载预训练模型放在models文件夹下

darknet53.conv.74下载链接：https://pjreddie.com/media/files/darknet53.conv.74

终端访问dartnet目录，输入一下命令：

# 从头开始训练
./darknet detector train cfg/hand.data cfg/yolov3_hand.cfg models/darknet53.conv.74

# 从某个权重快照继续训练
./darknet detector train cfg/hand.data cfg/yolov3_hand.cfg models/yolov3_hand_150000.weights

# 测试单张照片，会在当前目录生成一个predictions.jpg的测试图片
./darknet detector test cfg/hand.data cfg/yolov3_hand.cfg models/yolov3_hand_150000.weights data/tmp_hand.jpg

四、训练日志可视化

主要根据日志文件生成loss和iou曲线，当然日志需要训练时候从定向来生成日志文件

详情见另一博文：https://blog.csdn.net/oTengYue/article/details/81365185

五、预测过程

由于github上的master版本的python/darknet.py针对yolov3版本目前不太完善（截止2018-08-02），如果darknet采用master分支编译的话，建议把yolov3分支的python/darknet.py替换掉master版本的该文件。同时需要更改该脚本中的libdarknet.so路径为本机编译后的路径，不修改可能引起报错。

此处记录备份一下当前版本的darknet.py文件

from ctypes import *
import math
import random

def sample(probs):
    s = sum(probs)
    probs = [a/s for a in probs]
    r = random.uniform(0, 1)
    for i in range(len(probs)):
        r = r - probs[i]
        if r <= 0:
            return i
    return len(probs)-1

def c_array(ctype, values):
    new_values = values.ctypes.data_as(POINTER(ctype))
    return new_values

def array_to_image(arr):
    import numpy as np
    # need to return old values to avoid python freeing memory
    arr = arr.transpose(2,0,1)
    c = arr.shape[0]
    h = arr.shape[1]
    w = arr.shape[2]
    arr = np.ascontiguousarray(arr.flat, dtype=np.float32) / 255.0
    data = arr.ctypes.data_as(POINTER(c_float))
    im = IMAGE(w,h,c,data)
    return im, arr


class BOX(Structure):
    _fields_ = [("x", c_float),
                ("y", c_float),
                ("w", c_float),
                ("h", c_float)]

class DETECTION(Structure):
    _fields_ = [("bbox", BOX),
                ("classes", c_int),
                ("prob", POINTER(c_float)),
                ("mask", POINTER(c_float)),
                ("objectness", c_float),
                ("sort_class", c_int)]


class IMAGE(Structure):
    _fields_ = [("w", c_int),
                ("h", c_int),
                ("c", c_int),
                ("data", POINTER(c_float))]

class METADATA(Structure):
    _fields_ = [("classes", c_int),
                ("names", POINTER(c_char_p))]


#lib = CDLL("/home/pjreddie/documents/darknet/libdarknet.so", RTLD_GLOBAL)
lib = CDLL("libdarknet.so", RTLD_GLOBAL)
lib.network_width.argtypes = [c_void_p]
lib.network_width.restype = c_int
lib.network_height.argtypes = [c_void_p]
lib.network_height.restype = c_int

predict = lib.network_predict
predict.argtypes = [c_void_p, POINTER(c_float)]
predict.restype = POINTER(c_float)

set_gpu = lib.cuda_set_device
set_gpu.argtypes = [c_int]

make_image = lib.make_image
make_image.argtypes = [c_int, c_int, c_int]
make_image.restype = IMAGE

get_network_boxes = lib.get_network_boxes
get_network_boxes.argtypes = [c_void_p, c_int, c_int, c_float, c_float, POINTER(c_int), c_int, POINTER(c_int)]
get_network_boxes.restype = POINTER(DETECTION)

make_network_boxes = lib.make_network_boxes
make_network_boxes.argtypes = [c_void_p]
make_network_boxes.restype = POINTER(DETECTION)

free_detections = lib.free_detections
free_detections.argtypes = [POINTER(DETECTION), c_int]

free_ptrs = lib.free_ptrs
free_ptrs.argtypes = [POINTER(c_void_p), c_int]

network_predict = lib.network_predict
network_predict.argtypes = [c_void_p, POINTER(c_float)]

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

load_net = lib.load_network
load_net.argtypes = [c_char_p, c_char_p, c_int]
load_net.restype = c_void_p

do_nms_obj = lib.do_nms_obj
do_nms_obj.argtypes = [POINTER(DETECTION), c_int, c_int, c_float]

do_nms_sort = lib.do_nms_sort
do_nms_sort.argtypes = [POINTER(DETECTION), c_int, c_int, c_float]

free_image = lib.free_image
free_image.argtypes = [IMAGE]

letterbox_image = lib.letterbox_image
letterbox_image.argtypes = [IMAGE, c_int, c_int]
letterbox_image.restype = IMAGE

load_meta = lib.get_metadata
lib.get_metadata.argtypes = [c_char_p]
lib.get_metadata.restype = METADATA

load_image = lib.load_image_color
load_image.argtypes = [c_char_p, c_int, c_int]
load_image.restype = IMAGE

rgbgr_image = lib.rgbgr_image
rgbgr_image.argtypes = [IMAGE]

predict_image = lib.network_predict_image
predict_image.argtypes = [c_void_p, IMAGE]
predict_image.restype = POINTER(c_float)

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])
    return res

def detect(net, meta, image, thresh=.5, hier_thresh=.5, nms=.45):
    im = load_image(image, 0, 0)
    num = c_int(0)
    pnum = pointer(num)
    predict_image(net, im)
    dets = get_network_boxes(net, im.w, im.h, thresh, hier_thresh, None, 0, pnum)
    num = pnum[0]
    if (nms): do_nms_obj(dets, num, meta.classes, nms);

    res = []
    for j in range(num):
        for i in range(meta.classes):
            if dets[j].prob[i] > 0:
                b = dets[j].bbox
                res.append((meta.names[i], dets[j].prob[i], (b.x, b.y, b.w, b.h)))
    res = sorted(res, key=lambda x: -x[1])
    free_image(im)
    free_detections(dets, num)
    return res

def detect_numpy(net, meta, image, thresh=.5, hier_thresh=.5, nms=.45):
    im, arr = array_to_image(image)
    num = c_int(0)
    pnum = pointer(num)
    predict_image(net, im)
    dets = get_network_boxes(net, im.w, im.h, thresh, hier_thresh, None, 0, pnum)
    num = pnum[0]
    if (nms): do_nms_obj(dets, num, meta.classes, nms);

    res = []
    for j in range(num):
        for i in range(meta.classes):
            if dets[j].prob[i] > 0:
                b = dets[j].bbox
                res.append((meta.names[i], dets[j].prob[i], (b.x, b.y, b.w, b.h)))
    res = sorted(res, key=lambda x: -x[1])
    free_detections(dets, num)
    return res


if __name__ == "__main__":
    #net = load_net("cfg/densenet201.cfg", "/home/pjreddie/trained/densenet201.weights", 0)
    #im = load_image("data/wolf.jpg", 0, 0)
    #meta = load_meta("cfg/imagenet1k.data")
    #r = classify(net, meta, im)
    #print r[:10]
    net = load_net("cfg/tiny-yolo.cfg", "tiny-yolo.weights", 0)
    meta = load_meta("cfg/coco.data")
    import scipy.misc
    import time
    ''' t_start = time.time() for ii in range(100): r = detect(net, meta, 'data/dog.jpg') print(time.time() - t_start) print(r) image = scipy.misc.imread('data/dog.jpg') for ii in range(100): scipy.misc.imsave('/tmp/image.jpg', image) r = detect(net, meta, '/tmp/image.jpg') print(time.time() - t_start) print(r) '''

    image = scipy.misc.imread('data/dog.jpg')
    t_start = time.time()
    for ii in range(100):
        r = detect_numpy(net, meta, image)
    print(time.time() - t_start)
    print(r)

人手检测预测代码：examples/yolov3_detector_hand.py

#coding=utf-8
import cv2
import sys, os
sys.path.append('/export/songhongwei/code/darknet/python/')
import scipy.misc
import darknet as dn
from PIL import Image

class Yolov3HandDetector:
    hand_cfg_path = "/export/songhongwei/code/darknet/cfg/yolov3_hand.cfg"
    hand_weights_path = "/export/songhongwei/code/darknet/backup/yolov3_hand_150000.weights"
    hand_data_path = "/export/songhongwei/code/darknet/cfg/hand.data"

    def __init__(self):
        # Darknet
        self.net = dn.load_net(self.hand_cfg_path, self.hand_weights_path, 0)
        self.meta = dn.load_meta(self.hand_data_path)

    def img_cv2pil(self, cv_im):
        pil_im = Image.fromarray(cv2.cvtColor(cv_im, cv2.COLOR_BGR2RGB))
        return pil_im

    def detect_hand(self,cv_im):
        im = self.img_cv2pil(cv_im)
        im = scipy.misc.fromimage(im)
        res_infos = dn.detect_numpy(self.net, self.meta, im)
        bbox_map = {}
        for label, probability, bbox in res_infos:
            if label not in bbox_map:
                bbox_map[label] = []
            bbox_map[label].append([int(bbox[0]-bbox[2]/2),int(bbox[1]-bbox[3]/2),int(bbox[0]+bbox[2]/2),int(bbox[1]+bbox[3]/2)])
        return bbox_map

if __name__ == '__main__':
    handDetector = Yolov3HandDetector()
    pic_path = '/export/songhongwei/code/darknet/data/tmp_hand.jpg'
    cv_im = cv2.imread(pic_path)
    bbox_map = handDetector.detect_hand(cv_im)
    print(bbox_map)

输出：

{'hand': [[155, 180, 218, 265], [239, 241, 302, 291]]}

注：每个中括号内数字代表格式[top_left_x, top_left_y, bottom_right_x, bottom_right_y]