共计 6963 个字符,预计需要花费 18 分钟才能阅读完成。
欢送关注我的公众号 [极智视界],回复 001 获取 Google 编程标准
O_o
>_<
o_O
O_o
~_~
o_O
本文剖析下 darknet load_weights 接口,这个接口次要做模型权重的加载。
1、darknet 数据加载流程
之前的文章曾经介绍了一下 darknet 指标检测的数据加载流程,并介绍了.data、.names 和 .cfg 的加载实现。
接下来这里 load_weights 接口次要做 .weights 模型权重的加载。
2、load_weights 接口
先来看一下接口调用:
load_weights(&net, weightfile);
其中 net 为 network 构造体的实例,weightfile 为权重的文件门路,看一下 load_weights 的实现:
/// parser.c
void load_weights(network *net, char *filename)
{load_weights_upto(net, filename, net->n);
}
次要调用了 load_weights_upto 函数:
/// parser.c
void load_weights_upto(network *net, char *filename, int cutoff)
{
#ifdef GPU
if(net->gpu_index >= 0){cuda_set_device(net->gpu_index); // 设置 gpu_index
}
#endif
fprintf(stderr, "Loading weights from %s...", filename);
fflush(stdout); // 强制马上输入
FILE *fp = fopen(filename, "rb");
if(!fp) file_error(filename);
int major;
int minor;
int revision;
fread(&major, sizeof(int), 1, fp); // 一些标记位的加载
fread(&minor, sizeof(int), 1, fp);
fread(&revision, sizeof(int), 1, fp);
if ((major * 10 + minor) >= 2) {printf("\n seen 64");
uint64_t iseen = 0;
fread(&iseen, sizeof(uint64_t), 1, fp);
*net->seen = iseen;
}
else {printf("\n seen 32");
uint32_t iseen = 0;
fread(&iseen, sizeof(uint32_t), 1, fp);
*net->seen = iseen;
}
*net->cur_iteration = get_current_batch(*net);
printf(", trained: %.0f K-images (%.0f Kilo-batches_64) \n", (float)(*net->seen / 1000), (float)(*net->seen / 64000));
int transpose = (major > 1000) || (minor > 1000);
int i;
for(i = 0; i < net->n && i < cutoff; ++i){ // 辨认不同算子进行权重加载
layer l = net->layers[i];
if (l.dontload) continue;
if(l.type == CONVOLUTIONAL && l.share_layer == NULL){load_convolutional_weights(l, fp);
}
if (l.type == SHORTCUT && l.nweights > 0) {load_shortcut_weights(l, fp);
}
if (l.type == IMPLICIT) {load_implicit_weights(l, fp);
}
if(l.type == CONNECTED){load_connected_weights(l, fp, transpose);
}
if(l.type == BATCHNORM){load_batchnorm_weights(l, fp);
}
if(l.type == CRNN){load_convolutional_weights(*(l.input_layer), fp);
load_convolutional_weights(*(l.self_layer), fp);
load_convolutional_weights(*(l.output_layer), fp);
}
if(l.type == RNN){load_connected_weights(*(l.input_layer), fp, transpose);
load_connected_weights(*(l.self_layer), fp, transpose);
load_connected_weights(*(l.output_layer), fp, transpose);
}
if(l.type == GRU){load_connected_weights(*(l.input_z_layer), fp, transpose);
load_connected_weights(*(l.input_r_layer), fp, transpose);
load_connected_weights(*(l.input_h_layer), fp, transpose);
load_connected_weights(*(l.state_z_layer), fp, transpose);
load_connected_weights(*(l.state_r_layer), fp, transpose);
load_connected_weights(*(l.state_h_layer), fp, transpose);
}
if(l.type == LSTM){load_connected_weights(*(l.wf), fp, transpose);
load_connected_weights(*(l.wi), fp, transpose);
load_connected_weights(*(l.wg), fp, transpose);
load_connected_weights(*(l.wo), fp, transpose);
load_connected_weights(*(l.uf), fp, transpose);
load_connected_weights(*(l.ui), fp, transpose);
load_connected_weights(*(l.ug), fp, transpose);
load_connected_weights(*(l.uo), fp, transpose);
}
if (l.type == CONV_LSTM) {if (l.peephole) {load_convolutional_weights(*(l.vf), fp);
load_convolutional_weights(*(l.vi), fp);
load_convolutional_weights(*(l.vo), fp);
}
load_convolutional_weights(*(l.wf), fp);
if (!l.bottleneck) {load_convolutional_weights(*(l.wi), fp);
load_convolutional_weights(*(l.wg), fp);
load_convolutional_weights(*(l.wo), fp);
}
load_convolutional_weights(*(l.uf), fp);
load_convolutional_weights(*(l.ui), fp);
load_convolutional_weights(*(l.ug), fp);
load_convolutional_weights(*(l.uo), fp);
}
if(l.type == LOCAL){
int locations = l.out_w*l.out_h;
int size = l.size*l.size*l.c*l.n*locations;
fread(l.biases, sizeof(float), l.outputs, fp);
fread(l.weights, sizeof(float), size, fp);
#ifdef GPU
if(gpu_index >= 0){push_local_layer(l);
}
#endif
}
if (feof(fp)) break;
}
fprintf(stderr, "Done! Loaded %d layers from weights-file \n", i);
fclose(fp);
}
以上有几个点不容易看懂,如以下这段:
int major;
int minor;
int revision;
fread(&major, sizeof(int), 1, fp);
fread(&minor, sizeof(int), 1, fp);
fread(&revision, sizeof(int), 1, fp);
这个最好联合保留权重的接口一起来看,load_weights 是 save_weights 的解码过程,来看一下 save_weights_upto 的后面局部:
void save_weights_upto(network net, char *filename, int cutoff, int save_ema)
{
#ifdef GPU
if(net.gpu_index >= 0){cuda_set_device(net.gpu_index);
}
#endif
fprintf(stderr, "Saving weights to %s\n", filename);
FILE *fp = fopen(filename, "wb");
if(!fp) file_error(filename);
int major = MAJOR_VERSION;
int minor = MINOR_VERSION;
int revision = PATCH_VERSION;
fwrite(&major, sizeof(int), 1, fp); // 先打上 major
fwrite(&minor, sizeof(int), 1, fp); // 再打上 minor
fwrite(&revision, sizeof(int), 1, fp); // 再打上 revision
(*net.seen) = get_current_iteration(net) * net.batch * net.subdivisions; // remove this line, when you will save to weights-file both: seen & cur_iteration
fwrite(net.seen, sizeof(uint64_t), 1, fp); // 最初打上 net.seen
......
}
从下面的 save_weights 接口能够看出 darknet 的权重在后面会先打上几个标记:major、minor、revision、net.seen,而后再间断存储各层的权重数据,这样就不难理解 load_weights 的时候做这个解码了,以下是这几个参数的宏定义:
/// version.h
#define MAJOR_VERSION 0
#define MINOR_VERSION 2
#define PATCH_VERSION 5
再回到 load_weights,在加载这些标记后是加载各层的权重,以卷积权重加载来说,外面的逻辑分两个:
(1) 单 conv,应用 fread 依据特定大小顺次加载 biases 和 weights;
(2) conv + bn 交融,应用 fread 依据特定大小顺次加载 biases、scales、rolling_mean、rolling_variance、weights。
来看实现:
/// parser.c
void load_convolutional_weights(layer l, FILE *fp)
{if(l.binary){//load_convolutional_weights_binary(l, fp);
//return;
}
int num = l.nweights;
int read_bytes;
read_bytes = fread(l.biases, sizeof(float), l.n, fp); // load biases
if (read_bytes > 0 && read_bytes < l.n) printf("\n Warning: Unexpected end of wights-file! l.biases - l.index = %d \n", l.index);
//fread(l.weights, sizeof(float), num, fp); // as in connected layer
if (l.batch_normalize && (!l.dontloadscales)){read_bytes = fread(l.scales, sizeof(float), l.n, fp); // load scales
if (read_bytes > 0 && read_bytes < l.n) printf("\n Warning: Unexpected end of wights-file! l.scales - l.index = %d \n", l.index);
read_bytes = fread(l.rolling_mean, sizeof(float), l.n, fp); // load rolling_mean
if (read_bytes > 0 && read_bytes < l.n) printf("\n Warning: Unexpected end of wights-file! l.rolling_mean - l.index = %d \n", l.index);
read_bytes = fread(l.rolling_variance, sizeof(float), l.n, fp); // load rolling_variance
if (read_bytes > 0 && read_bytes < l.n) printf("\n Warning: Unexpected end of wights-file! l.rolling_variance - l.index = %d \n", l.index);
if(0){
int i;
for(i = 0; i < l.n; ++i){printf("%g,", l.rolling_mean[i]);
}
printf("\n");
for(i = 0; i < l.n; ++i){printf("%g,", l.rolling_variance[i]);
}
printf("\n");
}
if(0){fill_cpu(l.n, 0, l.rolling_mean, 1);
fill_cpu(l.n, 0, l.rolling_variance, 1);
}
}
read_bytes = fread(l.weights, sizeof(float), num, fp); // load weights
if (read_bytes > 0 && read_bytes < l.n) printf("\n Warning: Unexpected end of wights-file! l.weights - l.index = %d \n", l.index);
//if(l.adam){// fread(l.m, sizeof(float), num, fp);
// fread(l.v, sizeof(float), num, fp);
//}
//if(l.c == 3) scal_cpu(num, 1./256, l.weights, 1);
if (l.flipped) {transpose_matrix(l.weights, (l.c/l.groups)*l.size*l.size, l.n);
}
//if (l.binary) binarize_weights(l.weights, l.n, (l.c/l.groups)*l.size*l.size, l.weights);
#ifdef GPU
if(gpu_index >= 0){push_convolutional_layer(l);
}
#endif
}
再说一下 fread,这个函数在框架源码中数据读取方面会用的比拟多,来看一下这个 C 语言的函数:
size_t fread(void *ptr, size_t size, size_t nmemb, FILE *stream)
参数阐明:
- ptr:指向带有最小尺寸
size * nmemb
字节的内存块的指针; - size:要读取的每个元素的大小,以字节为单位;
- nmemb:元素的个数,每个元素的大小为 size 字节;
- stream:指向 FILE 对象的指针,指定了一个输出流;
返回值:胜利读取的元素个数以 size_t 对象返回,返回值与 nmenb 参数一样,若不一样,则可能产生了读谬误或达到了文件尾。
好了,以上剖析了 darknet 的 load_weights 接口及 weights 数据结构,再联合之前的文章就曾经集齐了 darkent 指标检测数据加载局部的解读,心愿我的分享对你的学习能有一点帮忙。
【公众号传送】
《【编程艺术】分析 darknet load_weights 接口》