darknet-convolution层

forward

void forward_convolutional_layer(convolutional_layer l, network_state state)
{
    int out_h = convolutional_out_height(l);
    int out_w = convolutional_out_width(l);
    int i;

    fill_cpu(l.outputs*l.batch, 0, l.output, 1); //填0

    if(l.xnor){
        binarize_weights(l.weights, l.n, l.c*l.size*l.size, l.binary_weights);
        swap_binary(&l);
        binarize_cpu(state.input, l.c*l.h*l.w*l.batch, l.binary_input);
        state.input = l.binary_input;
    }

    int m = l.n;
    int k = l.size*l.size*l.c;
    int n = out_h*out_w;


    float *a = l.weights;
    float *b = state.workspace;
    float *c = l.output;

    for(i = 0; i < l.batch; ++i){
        im2col_cpu(state.input, l.c, l.h, l.w, 
                l.size, l.stride, l.pad, b); //用当前层(c,h,w)和前一层输出做卷积,结果是2维向量(此处被展开成1维向量)
        gemm(0,0,m,n,k,1,a,k,b,n,1,c,n); //计算当前层输出初始值(没有加偏置和激励)
        c += n*m; //当前层输出一个样本对应的特征尺寸 n*m
        state.input += l.c*l.h*l.w;
    }

    if(l.batch_normalize){
        forward_batchnorm_layer(l, state);
    }
    add_bias(l.output, l.biases, l.batch, l.n, out_h*out_w);

    activate_array(l.output, m*n*l.batch, l.activation);
    if(l.binary || l.xnor) swap_binary(&l);
}

backward

最终 l l 层的输出 Oli O i l 对偏置 blk b k l 求导就是激励函数的导数 ∂g(x) ∂ g ( x ) ;对
权重 Wli W i l 求导结果是 ∂g(x)×Ol−1i ∂ g ( x ) × O i l − 1

void backward_convolutional_layer(convolutional_layer l, network_state state)
{
    int i;
    int m = l.n; //当前层filter个数,从cfg中读取
    int n = l.size*l.size*l.c; //当前层使用的卷积核尺寸,注意c是前一层输出通道数
    int k = convolutional_out_height(l)*
    convolutional_out_width(l); //当前层,一个通道下特征图尺寸  

    //l.delta中保存激励函数的导数
    gradient_array(l.output, m*k*l.batch, l.activation, l.delta); 
    //l.bias_updates中保存每个通道的偏置的导数
    backward_bias(l.bias_updates, l.delta, l.batch, l.n, k);

    if(l.batch_normalize){
        backward_batchnorm_layer(l, state);
    }

    for(i = 0; i < l.batch; ++i){
        float *a = l.delta + i*m*k; 
        float *b = state.workspace; //临时内存
        float *c = l.weight_updates; //权重更新(什么时候做了初始化?)

        float *im = state.input+i*l.c*l.h*l.w; //前一层的输出

        im2col_cpu(im, l.c, l.h, l.w, 
                l.size, l.stride, l.pad, b);
        gemm(0,1,m,n,k,1,a,k,b,k,1,c,n); //计算权重的导数

        if(state.delta){ //前一层有输出误差,把当前层的梯度向前传播
            a = l.weights;
            b = l.delta + i*m*k;
            c = state.workspace;

            gemm(1,0,n,k,m,1,a,n,b,k,0,c,k);

            col2im_cpu(state.workspace, l.c,  l.h,  l.w,  l.size,  l.stride, l.pad, state.delta+i*l.c*l.h*l.w);
        }
    }
}

gradient_array(l.output, m*k*l.batch, l.activation, l.delta); 

完成 df(x)dx d f ( x ) d x 的计算,结果保存在l.delta中, 尺度等于当前层输入(n,h,w)

 backward_bias(l.bias_updates, l.delta, l.batch, l.n, k);

计算每个输出通道的偏置,结果保存在l.bias_updates中,尺寸等于输出通道数n