MaxPoolGrad ================= 描述 MaxPool 的反向传播(梯度)计算。该算子将上游梯度(dy)只回传到前向最大池化过程中被选为最大值的位置;其它位置的梯度为 0。 数学定义: .. math:: \text{output}_{b,\ h_i,\ w_i,\ c} = \begin{cases} \text{dy}_{b,\ h_o,\ w_o,\ c}, & \text{if } (h_i,\ w_i) = \displaystyle \arg\max_{(h,w)\in\mathcal{W}(h_o,w_o)} \text{input}_{b,\ h,\ w,\ c}, \\ 0, & \text{otherwise}. \end{cases} 其中,:math:`\mathcal{W}(h_o, w_o)` 表示输出位置 :math:`(h_o, w_o)` 对应的池化窗口区域。窗口像素位置 :math:`(h, w)` 可表示为: .. math:: h = h_o \cdot \text{stride}_h - \text{pad}_u + \Delta h .. math:: w = w_o \cdot \text{stride}_w - \text{pad}_l + \Delta w .. math:: \Delta h \in [0,\ \text{win}_h - 1], \qquad \Delta w \in [0,\ \text{win}_w - 1] 并且仅当采样点落在输入有效范围内时会被考虑: .. math:: 0 \le h < \text{in}_h, \qquad 0 \le w < \text{in}_w. 实现细节说明: - 前向池化使用窗口 :math:`\text{win}_h \times \text{win}_w`,步长为 :math:`\text{stride}_h`, :math:`\text{stride}_w`,并且在边界处使用 pad(pad\_u, pad\_l)。 - 反向传播时,输出梯度 tensor(即需要写入的输入梯度)在每个 batch 开始前先被初始化为 0(代码中有一次整体清零)。 - 对于每个输出像素 :math:`(h_o,w_o)` 以及每个通道 c: - 在对应的输入窗口中找到前向最大值的位置 :math:`(h^*,w^*)`; - 将上游梯度 :math:`\text{dy}_{b,h_o,w_o,c}` 累加到该位置::math:`\text{output}_{b,h^*,w^*,c} \mathrel{+}= \text{dy}_{b,h_o,w_o,c}`。 - 其他位置梯度保持 0。 输入: - **input** - 输入张量指针,采用 **NHWC 格式**,形状为 :math:`[batch,\ in\_h,\ in\_w,\ channel]` - **dy** - 上游梯度张量指针,采用 **NHWC 格式**,形状为 :math:`[batch,\ output\_h,\ output\_w,\ channel]` - **params** - 参数数组,包含所有输入参数,顺序如下: - **in_w** - 输入张量的宽度 (W) - **in_h** - 输入张量的高度 (H) - **win_w** - 池化窗口的宽度,即窗口在 W 方向的大小 - **win_h** - 池化窗口的高度,即窗口在 H 方向的大小 - **output_w** - 输出特征图的宽度 - **output_h** - 输出特征图的高度 - **batch** - 批次大小,即输入中的 batch 数 - **channel** - 通道数 C ,每个池化位置都分别对 C 个通道独立执行最大池化与裁剪 - **stride_w** - 池化窗口在 W 方向的步长 - **stride_h** - 池化窗口在 H 方向的步长 - **pad_l** - 输入特征图左侧的填充大小 - **pad_u** - 输入特征图上侧的填充大小 - **minf** - 输出结果的下界值,传指针 - **maxf** - 输出结果的上界值,传指针 - **core_mask** - 核心掩码,指定使用的计算核心 输出: - **output** - 输出张量指针,采用 **NHWC 格式**,形状为 :math:`[batch,\ in\_h,\ in\_w,\ channel]`。 支持平台: ``FT78NE`` ``MT7004`` .. note:: - FT78NE 支持fp32, fp64 - MT7004 支持fp16, fp32 - 调用时将除 core_mask 外的参数打包通过 long long params 数组传入,顺序为: input, dy, output, in_w, in_h, win_w, win_h, output_w, output_h, batch, channel, stride_w, stride_h, pad_l, pad_u, minf, maxf **共享存储版本:** .. c:function:: void fp_maxpool_grad_s(float *input_ptr, float *dy_ptr, float *output_ptr, long long *params, int core_mask); .. c:function:: void hp_maxpool_grad_s(float16 *input_ptr, float16 *dy_ptr, float16 *output_ptr, long long *params, int core_mask); **C调用示例:** .. code-block:: c :linenos: :emphasize-lines: 57 //FT78NE示例 #include int main(int argc, char* argv[]) { float *input_ptr = (float *)0x81000000; float *dy_ptr = (float *)0x82000000; float *output_ptr = (float *)0x83000000; float *check_ptr = (float *)0x84000000; int in_w = gin_w; int in_h = gin_h; int win_w = 6; int win_h = 6; int output_batch = gbatch; //batch数 int channel = 1; int stride_w = 4; int stride_h = 4; int pad_l = 1; int pad_u = 1; float minf = 0; float maxf = 50; //计算output_w和output_h int dividor = in_w + pad_l*2 - win_w; int output_w = (dividor + stride_w - 1) / stride_w + 1; int dividor2 = in_h + pad_u*2 - win_h; int output_h = (dividor2 + stride_h - 1) / stride_h + 1; long long params[17]; params[0] = (long long)in_w; params[1] = (long long)in_h; params[2] = (long long)win_w; params[3] = (long long)win_h; params[4] = (long long)output_w; params[5] = (long long)output_h; params[6] = (long long)output_batch; params[7] = (long long)channel; params[8] = (long long)stride_w; params[9] = (long long)stride_h; params[10] = (long long)pad_l; params[11] = (long long)pad_u; params[12] = (long long)&minf; //注意这里传指针,不能直接强制转换成long long params[13] = (long long)&maxf; srand(time(NULL)); //初始化output_ptr int input_size = output_batch * channel * in_w * in_h; int dy_size = output_batch * channel * output_w * output_h; int i; for (i = 0; i < input_size; i++) { input_ptr[i] = (float)(rand() % 100); } for (i = 0; i < dy_size; i++) { dy_ptr[i] = (float)(rand() % 100); } int core_mask = 0b1111; fp_maxpool_grad_s(input_ptr, dy_ptr, output_ptr, params, core_mask); return 0; } **私有存储版本:** .. c:function:: void fp_maxpool_grad_p(float *input_ptr, float *dy_ptr, float *output_ptr, long long *params); .. c:function:: void hp_maxpool_grad_p(float16 *input_ptr, float16 *dy_ptr, float16 *output_ptr, long long *params); **C调用示例:** .. code-block:: c :linenos: :emphasize-lines: 57 //FT04示例 #include int main(int argc, char* argv[]) { float *input_ptr = (float *)0x10010000; float *dy_ptr = (float *)0x10020000; float *output_ptr = (float *)0x10030000; float *check_ptr = (float *)0x10040000; int in_w = gin_w; int in_h = gin_h; int win_w = 6; int win_h = 6; int output_batch = gbatch; //batch数 int channel = 1; int stride_w = 4; int stride_h = 4; int pad_l = 1; int pad_u = 1; float minf = 0; float maxf = 50; //计算output_w和output_h int dividor = in_w + pad_l*2 - win_w; int output_w = (dividor + stride_w - 1) / stride_w + 1; int dividor2 = in_h + pad_u*2 - win_h; int output_h = (dividor2 + stride_h - 1) / stride_h + 1; long long params[17]; params[0] = (long long)in_w; params[1] = (long long)in_h; params[2] = (long long)win_w; params[3] = (long long)win_h; params[4] = (long long)output_w; params[5] = (long long)output_h; params[6] = (long long)output_batch; params[7] = (long long)channel; params[8] = (long long)stride_w; params[9] = (long long)stride_h; params[10] = (long long)pad_l; params[11] = (long long)pad_u; params[12] = (long long)&minf; //注意这里传指针,不能直接强制转换成long long params[13] = (long long)&maxf; srand(time(NULL)); //初始化output_ptr int input_size = output_batch * channel * in_w * in_h; int dy_size = output_batch * channel * output_w * output_h; int i; for (i = 0; i < input_size; i++) { input_ptr[i] = (float)(rand() % 100); } for (i = 0; i < dy_size; i++) { dy_ptr[i] = (float)(rand() % 100); } fp_maxpool_grad_p(input_ptr, dy_ptr, output_ptr, params); return 0; }