YOLO配置文件理解

[net]
batch=64                           每batch个样本更新一次参数。
subdivisions=8                     如果内存不够大，将batch分割为subdivisions个子batch，每个子batch的大小为batch/subdivisions。
                                   在darknet代码中，会将batch/subdivisions命名为batch。
height=416                         input图像的高
width=416                          Input图像的宽
channels=3                         Input图像的通道数
momentum=0.9                       动量
decay=0.0005                       权重衰减正则项，防止过拟合
angle=0                            通过旋转角度来生成更多训练样本
saturation = 1.5                   通过调整饱和度来生成更多训练样本
exposure = 1.5                     通过调整曝光量来生成更多训练样本
hue=.1                             通过调整色调来生成更多训练样本
learning_rate=0.0001               初始学习率
max_batches = 45000                训练达到max_batches后停止学习
policy=steps                       调整学习率的policy，有如下policy：CONSTANT, STEP, EXP, POLY, STEPS, SIG, RANDOM
steps=100,25000,35000              根据batch_num调整学习率
scales=10,.1,.1                    学习率变化的比例，累计相乘
[convolutional]
batch_normalize=1                  是否做BN
filters=32                         输出多少个特征图
size=3                             卷积核的尺寸
stride=1                           做卷积运算的步长
pad=1                              如果pad为0,padding由 padding参数指定。如果pad为1，padding大小为size/2
activation=leaky                   激活函数：
                                   logistic，loggy，relu，elu，relie，plse，hardtan，lhtan，linear，ramp，leaky，tanh，stair
[maxpool]
size=2                             池化层尺寸
stride=2                           池化步进
[convolutional]
batch_normalize=1
filters=64
size=3
stride=1
pad=1
activation=leaky
[maxpool]
size=2
stride=2
......
......
#######
[convolutional]
batch_normalize=1
size=3
stride=1
pad=1
filters=1024
activation=leaky
[convolutional]
batch_normalize=1
size=3
stride=1
pad=1
filters=1024
activation=leaky
[route]                            the route layer is to bring finer grained features in from earlier in the network
layers=-9
[reorg]                            the reorg layer is to make these features match the feature map size at the later layer.
                                   The end feature map is 13x13, the feature map from earlier is 26x26x512.
                                   The reorg layer maps the 26x26x512 feature map onto a 13x13x2048 feature map
                                   so that it can be concatenated with the feature maps at 13x13 resolution.
stride=2
[route]
layers=-1,-3
[convolutional]
batch_normalize=1
size=3
stride=1
pad=1
filters=1024
activation=leaky
[convolutional]
size=1
stride=1
pad=1
filters=125                        region前最后一个卷积层的filters数是特定的，计算公式为filter=num*(classes+5)
                                   5的意义是5个坐标，论文中的tx,ty,tw,th,to
activation=linear
[region]
anchors = 1.08,1.19,  3.42,4.41,  6.63,11.38,  9.42,5.11,  16.62,10.52          预选框，可以手工挑选，
                                                                                也可以通过k means 从训练样本中学出
bias_match=1
classes=20                         网络需要识别的物体种类数
coords=4                           每个box的4个坐标tx,ty,tw,th
num=5                              每个grid cell预测几个box
softmax=1                          使用softmax做激活函数
jitter=.2                          通过抖动增加噪声来抑制过拟合
rescore=1                          暂理解为一个开关，非0时通过重打分来调整l.delta（预测值与真实值的差）
object_scale=5                     暂理解为计算损失时预测框中有物体时的权重
noobject_scale=1                   暂理解为计算损失时预测框中无物体时的权重
class_scale=1                      暂理解为计算类别损失时的权重
coord_scale=1                      暂理解为计算损失时坐标偏差的权重
absolute=1
thresh = .6
random=0                           是否随机确定最后一个预测框

darknet对应代码

找到cfg文件解析的代码，选择detector demo 作为入口

darknet.c文件 main 函数开始

    } else if (0 == strcmp(argv[1], "detector")){
    run_detector(argc, argv);

Detector.c文件 run_detector函数

char *prefix = find_char_arg(argc, argv, "-prefix", 0);
float thresh = find_float_arg(argc, argv, "-thresh", .24);
float hier_thresh = find_float_arg(argc, argv, "-hier", .5);
int cam_index = find_int_arg(argc, argv, "-c", 0);
int frame_skip = find_int_arg(argc, argv, "-s", 0);
if(argc < 4){
    fprintf(stderr, "usage: %s %s [train/test/valid] [cfg] [weights (optional)]\n", argv[0], argv[1]);
    return;
}
char *gpu_list = find_char_arg(argc, argv, "-gpus", 0);
char *outfile = find_char_arg(argc, argv, "-out", 0);
......
......
else if(0==strcmp(argv[2], "demo")) {
    list *options = read_data_cfg(datacfg);
    int classes = option_find_int(options, "classes", 20);
    char *name_list = option_find_str(options, "names", "data/names.list");
    char **names = get_labels(name_list);
    demo(cfg, weights, thresh, cam_index, filename, names, classes, frame_skip, prefix, hier_thresh);
}

read_data_cfg函数解析配置文件，保存到options指针。

class

int classes = option_find_int(options, "classes", 20);

classes为YOLO可识别的种类数

batch、learning_rate、momentum、decay和 subdivisions

demo.c文件demo函数

net = parse_network_cfg(cfgfile);

Parser.c文件 parse_network_cfg函数

list *sections = read_cfg(filename);
node *n = sections->front;
if(!n) error("Config file has no sections");
network net = make_network(sections->size - 1);
net.gpu_index = gpu_index;
size_params params;
section *s = (section *)n->val;
list *options = s->options;
if(!is_network(s)) error("First section must be [net] or [network]");
parse_net_options(options, &net);

parse_net_options函数

net->batch = option_find_int(options, "batch",1);
net->learning_rate = option_find_float(options, "learning_rate", .001);
net->momentum = option_find_float(options, "momentum", .9);
net->decay = option_find_float(options, "decay", .0001);
int subdivs = option_find_int(options, "subdivisions",1);
net->time_steps = option_find_int_quiet(options, "time_steps",1);
net->batch /= subdivs;
net->batch *= net->time_steps;
net->subdivisions = subdivs;

learning_rate为初始学习率，训练时的真正学习率和学习率的策略及初始学习率有关。

momentum为动量，在训练时加入动量可以帮助走出local minima 以及saddle point。

decay是权重衰减正则项，用来防止过拟合。

batch的值等于cfg文件中的batch/subdivisions 再乘以time_steps。
time_steps在yolo默认的cfg中是没有配置的，所以是默认值1。
因此batch可以认为就是cfg文件中的batch/subdivisions。

前面有提到batch的意义是每batch个样本更新一次参数。

而subdivisions的意义在于降低对GPU memory的要求。
darknet将batch分割为subdivisions个子batch，每个子batch的大小为batch/subdivisions，并将子batch命名为batch。

我们看下训练时和batch有关的代码

Detector.c文件的train_detector函数

#ifdef GPU
    if(ngpus == 1){
        loss = train_network(net, train);
    } else {
        loss = train_networks(nets, ngpus, train, 4);
    }
#else
    loss = train_network(net, train);
#endif

Network.c文件的train_network函数

int batch = net.batch;
int n = d.X.rows / batch;
float *X = calloc(batch*d.X.cols, sizeof(float));
float *y = calloc(batch*d.y.cols, sizeof(float));
int i;
float sum = 0;
for(i = 0; i < n; ++i){
    get_next_batch(d, batch, i*batch, X, y);
    float err = train_network_datum(net, X, y);
    sum += err;
}

train_network_datum函数

*net.seen += net.batch;
......
......
forward_network(net, state);
backward_network(net, state);
float error = get_network_cost(net);
if(((*net.seen)/net.batch)%net.subdivisions == 0) update_network(net);

我们看到，只有((*net.seen)/net.batch)%net.subdivisions == 0时才会更新网络参数。
*net.seen是已经训练过的子batch数，((*net.seen)/net.batch)%net.subdivisions的意义正是已经训练过了多少个真正的batch。

policy、steps和scales

Parser.c文件 parse_network_cfg函数

char *policy_s = option_find_str(options, "policy", "constant");
net->policy = get_policy(policy_s);
net->burn_in = option_find_int_quiet(options, "burn_in", 0);
if(net->policy == STEP){
    net->step = option_find_int(options, "step", 1);
    net->scale = option_find_float(options, "scale", 1);
} else if (net->policy == STEPS){
    char *l = option_find(options, "steps");
    char *p = option_find(options, "scales");
    if(!l || !p) error("STEPS policy must have steps and scales in cfg file");
    int len = strlen(l);
    int n = 1;
    int i;
    for(i = 0; i < len; ++i){
        if (l[i] == ',') ++n;
    }
    int *steps = calloc(n, sizeof(int));
    float *scales = calloc(n, sizeof(float));
    for(i = 0; i < n; ++i){
        int step    = atoi(l);
        float scale = atof(p);
        l = strchr(l, ',')+1;
        p = strchr(p, ',')+1;
        steps[i] = step;
        scales[i] = scale;
    }
    net->scales = scales;
    net->steps = steps;
    net->num_steps = n;
} else if (net->policy == EXP){
    net->gamma = option_find_float(options, "gamma", 1);
} else if (net->policy == SIG){
    net->gamma = option_find_float(options, "gamma", 1);
    net->step = option_find_int(options, "step", 1);
} else if (net->policy == POLY || net->policy == RANDOM){
    net->power = option_find_float(options, "power", 1);
}

get_policy函数

if (strcmp(s, "random")==0) return RANDOM;
if (strcmp(s, "poly")==0) return POLY;
if (strcmp(s, "constant")==0) return CONSTANT;
if (strcmp(s, "step")==0) return STEP;
if (strcmp(s, "exp")==0) return EXP;
if (strcmp(s, "sigmoid")==0) return SIG;
if (strcmp(s, "steps")==0) return STEPS;
fprintf(stderr, "Couldn't find policy %s, going with constant\n", s);
return CONSTANT;

学习率动态调整的策略有多种，YOLO默认使用的是steps。

yolo-voc.cfg文件：

steps=100,25000,35000

scales=10,.1,.1

Network.c文件get_current_rate函数

int batch_num = get_current_batch(net);
int i;
float rate;
switch (net.policy) {
    case CONSTANT:
        return net.learning_rate;
    case STEP:
        return net.learning_rate * pow(net.scale, batch_num/net.step);
    case STEPS:
        rate = net.learning_rate;
        for(i = 0; i < net.num_steps; ++i){
            if(net.steps[i] > batch_num) return rate;
            rate *= net.scales[i];
            //if(net.steps[i] > batch_num - 1 && net.scales[i] > 1) reset_momentum(net);
        }
        return rate;

get_current_batch获取的是(*net.seen)/(net.batch*net.subdivisions)，即真正的batch。

steps的每个阶段是根据batch_num划分的，根据配置文件，学习率会在batch_num达到100、25000、35000时发生改变。

当前的学习率是初始学习率与当前阶段及之前所有阶段对应的scale的总乘积。

convolutional超参数加载

Parser.c文件parse_network_cfg函数

LAYER_TYPE lt = string_to_layer_type(s->type);
        if(lt == CONVOLUTIONAL){
            l = parse_convolutional(options, params);

parse_convolutional函数

int n = option_find_int(options, "filters",1);
int size = option_find_int(options, "size",1);
int stride = option_find_int(options, "stride",1);
int pad = option_find_int_quiet(options, "pad",0);
int padding = option_find_int_quiet(options, "padding",0);
if(pad) padding = size/2;
char *activation_s = option_find_str(options, "activation", "logistic");
ACTIVATION activation = get_activation(activation_s);
int batch,h,w,c;
h = params.h;
w = params.w;
c = params.c;
batch=params.batch;
if(!(h && w && c)) error("Layer before convolutional layer must output image.");
int batch_normalize = option_find_int_quiet(options, "batch_normalize", 0);

需要注意的是如果enable了pad，cfg文件中的padding不会生效，实际的padding值为size/2。