Darknet中函數分析
隨機打亂數據
代碼在data.c源文件中
void randomize_data(data d) { int i; for(i = d.X.rows-1; i > 0; --i){ // 從最後一個元素位置開始,當前位置爲i int index = rand()%i;// 從i(不包括)之前的所有位置隨機選擇一個位置index float *swap = d.X.vals[index]; // 交換index和i處的指針 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
可以看到整個打亂過程只有指針指向在不斷髮生變化,數據在內存中的位置不變。
im2col實現
代碼在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"><</span> <span class="token number">0</span> <span class="token operator">||</span> col <span class="token operator"><</span> <span class="token number">0</span> <span class="token operator">||</span> row <span class="token operator">>=</span> height <span class="token operator">||</span> col <span class="token operator">>=</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, // 輸入的圖像數據,內存中按行排列成一維
int channels, // 通道數
int height, int width, // 圖像的高和寬
int ksize, // 卷積核的高和寬,這裏默認卷積核高和寬大小一樣
int stride, // 卷積時的步長,這裏默認高和寬兩個方向上步長一樣
int pad, // 圖像的填充,這裏默認高和寬上填充一樣
float data_col // 最終輸出的數據
) {
int c,h,w;
int height_col = (height + 2pad - ksize) / stride + 1; // 卷積後的圖像尺寸,可以想象,其中每個點對應圖像的一個卷積區域,卷積區域大小是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">// 圖像上每個卷積區域展成一列後的大小,比如 // ksize=3,圖像通道數爲3,那麼channels_col爲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"><</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, ... 寬的相對偏移</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, ...高的相對偏移</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個0, 9個1, 9個2 通道序號,(c_im, h_offset, w_offset)是卷積區域的相對座標</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"><</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個卷積區域,遍歷每一個卷積區域,對於每個區域,計算它的c位置處的圖像像素值</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"><</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">// 相對於圖像的行座標</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">// 相對於圖像的列座標</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
這是caffe中卷積操作之前的關鍵步驟,對三維圖像進行變形,將傳統卷積操作變成矩陣形式的卷積操作。
簡便起見,設channels = 1,即圖像是個單通道圖。data_im爲
3×3×3=27列。此時,再將kernel拉伸成一列,卷積過程變成了矩陣乘法: