Darknet函数分析

Darknet中函数分析

随机打乱数据

代码在data.c源文件中

void randomize_data(data d)
{
    int i;
    for(i = d.X.rows-1; i > 0; --i){ // 从最后一个元素位置开始，当前位置为i
        int index = rand()%i;// 从i（不包括）之前的所有位置随机选择一个位置index
        float *swap = d.X.vals[index]; // 交换index和i处的指针
        d.X.vals[index] = d.X.vals[i];
        d.X.vals[i] = swap;
    swap <span class="token operator">=</span> d<span class="token punctuation">.</span>y<span class="token punctuation">.</span>vals<span class="token punctuation">[</span>index<span class="token punctuation">]</span><span class="token punctuation">;</span>
    d<span class="token punctuation">.</span>y<span class="token punctuation">.</span>vals<span class="token punctuation">[</span>index<span class="token punctuation">]</span> <span class="token operator">=</span> d<span class="token punctuation">.</span>y<span class="token punctuation">.</span>vals<span class="token punctuation">[</span>i<span class="token punctuation">]</span><span class="token punctuation">;</span>
    d<span class="token punctuation">.</span>y<span class="token punctuation">.</span>vals<span class="token punctuation">[</span>i<span class="token punctuation">]</span> <span class="token operator">=</span> swap<span class="token punctuation">;</span>
<span class="token punctuation">}</span>

}

• 1
• 2
• 3
• 4
• 5
• 6
• 7
• 8
• 9
• 10
• 11
• 12
• 13
• 14

可以看到整个打乱过程只有指针指向在不断发生变化，数据在内存中的位置不变。

---

im2col实现

代码在im2col.c源文件中

float im2col_get_pixel(float *im, int height, int width, int channels,
                        int row, int col, int channel, int pad)
{
    row -= pad;
    col -= pad;
<span class="token keyword">if</span> <span class="token punctuation">(</span>row <span class="token operator">&lt;</span> <span class="token number">0</span> <span class="token operator">||</span> col <span class="token operator">&lt;</span> <span class="token number">0</span> <span class="token operator">||</span>
    row <span class="token operator">&gt;=</span> height <span class="token operator">||</span> col <span class="token operator">&gt;=</span> width<span class="token punctuation">)</span> <span class="token keyword">return</span> <span class="token number">0</span><span class="token punctuation">;</span>
<span class="token keyword">return</span> im<span class="token punctuation">[</span>col <span class="token operator">+</span> width<span class="token operator">*</span><span class="token punctuation">(</span>row <span class="token operator">+</span> height<span class="token operator">*</span>channel<span class="token punctuation">)</span><span class="token punctuation">]</span><span class="token punctuation">;</span>

}

//From Berkeley Vision’s Caffe!
//https://github.com/BVLC/caffe/blob/master/LICENSE
void im2col_cpu(float data_im, // 输入的图像数据，内存中按行排列成一维
int channels, // 通道数
int height, int width, // 图像的高和宽
int ksize, // 卷积核的高和宽，这里默认卷积核高和宽大小一样
int stride, // 卷积时的步长，这里默认高和宽两个方向上步长一样
int pad, // 图像的填充，这里默认高和宽上填充一样
float data_col // 最终输出的数据
) {
int c,h,w;
int height_col = (height + 2pad - ksize) / stride + 1; // 卷积后的图像尺寸，可以想象，其中每个点对应图像的一个卷积区域，卷积区域大小是channels * ksize * ksize，即下面的channels_col。
int width_col = (width + 2pad - ksize) / stride + 1;

<span class="token keyword">int</span> channels_col <span class="token operator">=</span> channels <span class="token operator">*</span> ksize <span class="token operator">*</span> ksize<span class="token punctuation">;</span> <span class="token comment">// 图像上每个卷积区域展成一列后的大小，比如                                                          // ksize=3，图像通道数为3，那么channels_col为27</span>
<span class="token keyword">for</span> <span class="token punctuation">(</span>c <span class="token operator">=</span> <span class="token number">0</span><span class="token punctuation">;</span> c <span class="token operator">&lt;</span> channels_col<span class="token punctuation">;</span> <span class="token operator">++</span>c<span class="token punctuation">)</span> <span class="token punctuation">{</span>
    <span class="token keyword">int</span> w_offset <span class="token operator">=</span> c <span class="token operator">%</span> ksize<span class="token punctuation">;</span> <span class="token comment">// 0, 1, 2, 0, 1, 2, 0, 1, 2, ... 宽的相对偏移</span>
    <span class="token keyword">int</span> h_offset <span class="token operator">=</span> <span class="token punctuation">(</span>c <span class="token operator">/</span> ksize<span class="token punctuation">)</span> <span class="token operator">%</span> ksize<span class="token punctuation">;</span> <span class="token comment">// 0, 0, 0, 1, 1, 1, 2, 2, 2, 0, 0, 0, 1, 1, 1, ...高的相对偏移</span>
    <span class="token keyword">int</span> c_im <span class="token operator">=</span> c <span class="token operator">/</span> ksize <span class="token operator">/</span> ksize<span class="token punctuation">;</span> <span class="token comment">// 9个0， 9个1， 9个2 通道序号，（c_im, h_offset, w_offset）是卷积区域的相对座标</span>
    <span class="token keyword">for</span> <span class="token punctuation">(</span>h <span class="token operator">=</span> <span class="token number">0</span><span class="token punctuation">;</span> h <span class="token operator">&lt;</span> height_col<span class="token punctuation">;</span> <span class="token operator">++</span>h<span class="token punctuation">)</span> <span class="token punctuation">{</span> <span class="token comment">// 有height_col * width_col个卷积区域，遍历每一个卷积区域，对于每个区域，计算它的c位置处的图像像素值</span>
        <span class="token keyword">for</span> <span class="token punctuation">(</span>w <span class="token operator">=</span> <span class="token number">0</span><span class="token punctuation">;</span> w <span class="token operator">&lt;</span> width_col<span class="token punctuation">;</span> <span class="token operator">++</span>w<span class="token punctuation">)</span> <span class="token punctuation">{</span>
            <span class="token keyword">int</span> im_row <span class="token operator">=</span> h_offset <span class="token operator">+</span> h <span class="token operator">*</span> stride<span class="token punctuation">;</span> <span class="token comment">// 相对于图像的行座标</span>
            <span class="token keyword">int</span> im_col <span class="token operator">=</span> w_offset <span class="token operator">+</span> w <span class="token operator">*</span> stride<span class="token punctuation">;</span> <span class="token comment">// 相对于图像的列座标</span>
            <span class="token keyword">int</span> col_index <span class="token operator">=</span> <span class="token punctuation">(</span>c <span class="token operator">*</span> height_col <span class="token operator">+</span> h<span class="token punctuation">)</span> <span class="token operator">*</span> width_col <span class="token operator">+</span> w<span class="token punctuation">;</span> <span class="token comment">// </span>
            data_col<span class="token punctuation">[</span>col_index<span class="token punctuation">]</span> <span class="token operator">=</span> <span class="token function">im2col_get_pixel</span><span class="token punctuation">(</span>data_im<span class="token punctuation">,</span> height<span class="token punctuation">,</span> width<span class="token punctuation">,</span> channels<span class="token punctuation">,</span>
                    im_row<span class="token punctuation">,</span> im_col<span class="token punctuation">,</span> c_im<span class="token punctuation">,</span> pad<span class="token punctuation">)</span><span class="token punctuation">;</span>
        <span class="token punctuation">}</span>
    <span class="token punctuation">}</span>
<span class="token punctuation">}</span>

}

• 1
• 2
• 3
• 4
• 5
• 6
• 7
• 8
• 9
• 10
• 11
• 12
• 13
• 14
• 15
• 16
• 17
• 18
• 19
• 20
• 21
• 22
• 23
• 24
• 25
• 26
• 27
• 28
• 29
• 30
• 31
• 32
• 33
• 34
• 35
• 36
• 37
• 38
• 39
• 40
• 41

这是caffe中卷积操作之前的关键步骤，对三维图像进行变形，将传统卷积操作变成矩阵形式的卷积操作。

简便起见，设channels = 1，即图像是个单通道图。data_im为
$05101520161116212712172238131823491419240123456789101112131415161718192021222324 \begin{matrix}0 & 1 & 2 & 3 & 4 \\5 & 6 & 7 & 8 & 9 \\10 & 11 & 12 & 13 & 14 \\15 & 16 & 17 & 18 & 19 \\20 & 21 & 22 & 23 & 24 \\\end{matrix}$3×3×3=27列。此时，再将kernel拉伸成一列，卷积过程变成了矩阵乘法：
$⎡⎣⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢0125671011121236781112132347891213145671011121516176781112131617187891213141718191011121516172021221112131617182122231213141718$