YOLOv2源码分析（六）

文章全部YOLOv2源码分析

我们再次回到了parse_network_cfg函数

        //parse_network_cfg
        else if(lt == ACTIVE){
            l = parse_activation(options, params);
        }

接着看后面这个parse_activation函数

0x01 parse_activation

layer parse_activation(list *options, size_params params)
{
    char *activation_s = option_find_str(options, "activation", "linear");
    ACTIVATION activation = get_activation(activation_s);

    layer l = make_activation_layer(params.batch, params.inputs, activation);

    l.out_h = params.h;
    l.out_w = params.w;
    l.out_c = params.c;
    l.h = params.h;
    l.w = params.w;
    l.c = params.c;

    return l;
}

上面的一些参数我在之前的文章中已经说过了，这里就不再说明了。直接看关键函数make_activation_layer

layer make_activation_layer(int batch, int inputs, ACTIVATION activation)
{
...
    l.forward = forward_activation_layer;
    l.backward = backward_activation_layer;
...
    return l;
}

前面的参数信息我这里也不再提了，直接看关键的两个函数，先看第一个forward_activation_layer

0x0101 forward_activation_layer

void forward_activation_layer(layer l, network net)
{
    copy_cpu(l.outputs*l.batch, net.input, 1, l.output, 1);
    activate_array(l.output, l.outputs*l.batch, l.activation);
}

貌似这里没什么好说的b（￣▽￣）d　

0x0102 backward_activation_layer

void backward_activation_layer(layer l, network net)
{
    gradient_array(l.output, l.outputs*l.batch, l.activation, l.delta);
    copy_cpu(l.outputs*l.batch, l.delta, 1, net.delta, 1);
}

貌似这里也没什么好说的d( •̀ ω •́ )y　

回到parse_network_cfg函数

        //parse_network_cfg
        else if(lt == RNN){
            l = parse_rnn(options, params);
        }

0x02 parse_rnn

layer parse_rnn(list *options, size_params params)
{
    int output = option_find_int(options, "output",1);
    int hidden = option_find_int(options, "hidden",1);
    char *activation_s = option_find_str(options, "activation", "logistic");
    ACTIVATION activation = get_activation(activation_s);
    int batch_normalize = option_find_int_quiet(options, "batch_normalize", 0);
    int logistic = option_find_int_quiet(options, "logistic", 0);

    layer l = make_rnn_layer(params.batch, params.inputs, hidden, output, params.time_steps, activation, batch_normalize, logistic);

    l.shortcut = option_find_int_quiet(options, "shortcut", 0);

    return l;
}

我先说说这里的几个参数的含义，因为我之前有的没有讲过。

• hidden:RNN隐藏层的元素个数
• time_steps:RNN的步长
• logistic:Logistic激活函数

接着我们来看关键函数make_rnn_layer

0x02 make_rnn_layer

作者这里使用的是vanilla RNN结构，有三个全连接层组成。

layer make_rnn_layer(int batch, int inputs, int outputs, int steps, ACTIVATION activation, int batch_normalize, int adam)
{
...
    l.input_layer = malloc(sizeof(layer));//隐藏层1
    fprintf(stderr, "\t\t");
    *(l.input_layer) = make_connected_layer(batch*steps, inputs, outputs, activation, batch_normalize, adam);
    l.input_layer->batch = batch;

    l.self_layer = malloc(sizeof(layer));//隐藏层2
    fprintf(stderr, "\t\t");
    *(l.self_layer) = make_connected_layer(batch*steps, outputs, outputs, activation, batch_normalize, adam);
    l.self_layer->batch = batch;

    l.output_layer = malloc(sizeof(layer));//隐藏层3
    fprintf(stderr, "\t\t");
    *(l.output_layer) = make_connected_layer(batch*steps, outputs, outputs, activation, batch_normalize, adam);
    l.output_layer->batch = batch;

    l.outputs = outputs;
    l.output = l.output_layer->output;
    l.delta = l.output_layer->delta;

    l.forward = forward_rnn_layer;
    l.backward = backward_rnn_layer;
    l.update = update_rnn_layer;
...

我们看这里的make_connected_layer函数

0x0201 make_connected_layer

layer make_connected_layer(int batch, int inputs, int outputs, ACTIVATION activation, int batch_normalize, int adam)
{
...

    l.forward = forward_connected_layer;
    l.backward = backward_connected_layer;
    l.update = update_connected_layer;   

...
}

这里的参数信息也没什么好说的，直接看函数吧

0x020101 forward_connected_layer

void forward_connected_layer(layer l, network net)
{
    fill_cpu(l.outputs*l.batch, 0, l.output, 1);
    int m = l.batch;
    int k = l.inputs;
    int n = l.outputs;
    float *a = net.input;
    float *b = l.weights;
    float *c = l.output;
    gemm(0,1,m,n,k,1,a,k,b,k,1,c,n);
    if(l.batch_normalize){
        forward_batchnorm_layer(l, net);
    } else {
        add_bias(l.output, l.biases, l.batch, l.outputs, 1);
    }
    activate_array(l.output, l.outputs*l.batch, l.activation);
}

这个函数其实没什么好说的，要注意的地方就是这里的b是转置的。还有一个地方要注意的是，这里没有了groups和batch，这也非常好理解。

0x020102 backward_connected_layer

void backward_connected_layer(layer l, network net)
{
    gradient_array(l.output, l.outputs*l.batch, l.activation, l.delta);

    if(l.batch_normalize){
        backward_batchnorm_layer(l, net);
    } else {
        backward_bias(l.bias_updates, l.delta, l.batch, l.outputs, 1);
    }

    int m = l.outputs;
    int k = l.batch;
    int n = l.inputs;
    float *a = l.delta;
    float *b = net.input;
    float *c = l.weight_updates;
    gemm(1,0,m,n,k,1,a,m,b,n,1,c,n);

    m = l.batch;
    k = l.outputs;
    n = l.inputs;

    a = l.delta;
    b = l.weights;
    c = net.delta;

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

没什么好说的

0x020103 update_connected_layer

void update_connected_layer(layer l, update_args a)
{
    float learning_rate = a.learning_rate*l.learning_rate_scale;
    float momentum = a.momentum;
    float decay = a.decay;
    int batch = a.batch;
    axpy_cpu(l.outputs, learning_rate/batch, l.bias_updates, 1, l.biases, 1);
    scal_cpu(l.outputs, momentum, l.bias_updates, 1);

    if(l.batch_normalize){
        axpy_cpu(l.outputs, learning_rate/batch, l.scale_updates, 1, l.scales, 1);
        scal_cpu(l.outputs, momentum, l.scale_updates, 1);
    }

    axpy_cpu(l.inputs*l.outputs, -decay*batch, l.weights, 1, l.weight_updates, 1);
    axpy_cpu(l.inputs*l.outputs, learning_rate/batch, l.weight_updates, 1, l.weights, 1);
    scal_cpu(l.inputs*l.outputs, momentum, l.weight_updates, 1);
}

没什么说的(￣▽￣)”

0x0202 forward_rnn_layer

void forward_rnn_layer(layer l, network net)
{
    network s = net;
    s.train = net.train;
    int i;
    layer input_layer = *(l.input_layer);
    layer self_layer = *(l.self_layer);
    layer output_layer = *(l.output_layer);

    fill_cpu(l.outputs * l.batch * l.steps, 0, output_layer.delta, 1);
    fill_cpu(l.outputs * l.batch * l.steps, 0, self_layer.delta, 1);
    fill_cpu(l.outputs * l.batch * l.steps, 0, input_layer.delta, 1);
    if(net.train) fill_cpu(l.outputs * l.batch, 0, l.state, 1);//如果网络是训练状态的话，将state设置为0

    for (i = 0; i < l.steps; ++i) {
        s.input = net.input;
        forward_connected_layer(input_layer, s);

        s.input = l.state;
        forward_connected_layer(self_layer, s);

        float *old_state = l.state;//存储当前状态
        if(net.train) l.state += l.outputs*l.batch;
        if(l.shortcut){//是否添加shortcut连接
            copy_cpu(l.outputs * l.batch, old_state, 1, l.state, 1);
        }else{
            fill_cpu(l.outputs * l.batch, 0, l.state, 1);
        }
        axpy_cpu(l.outputs * l.batch, 1, input_layer.output, 1, l.state, 1);
        axpy_cpu(l.outputs * l.batch, 1, self_layer.output, 1, l.state, 1);

        s.input = l.state;
        forward_connected_layer(output_layer, s);
        //结束后，三个层都向前移动一步
        net.input += l.inputs*l.batch;
        increment_layer(&input_layer, 1);
        increment_layer(&self_layer, 1);
        increment_layer(&output_layer, 1);
    }
}

先说说这个函数吧increment_layer

static void increment_layer(layer *l, int steps)
{
    int num = l->outputs*l->batch*steps;
    l->output += num;
    l->delta += num;
    l->x += num;
    l->x_norm += num;

#ifdef GPU
    l->output_gpu += num;
    l->delta_gpu += num;
    l->x_gpu += num;
    l->x_norm_gpu += num;
#endif
}

这个函数的作用就是rnn向前的过程中，调整参数的初始位置。

这个RNN网络主要前向传播time_steps次。网络总共分为三层，第一层将输入数据编码（one-hot），是一个256的向量。第二层主要是传入上一次的状态和前一层的输出。第二层主要是传入上一次的状态和前一层的输出，并且输出结果转化为一个256维的向量，并且进行归一化处理。

0x0203 backward_rnn_layer

void backward_rnn_layer(layer l, network net)
{
    network s = net;
    s.train = net.train;
    int i;
    layer input_layer = *(l.input_layer);
    layer self_layer = *(l.self_layer);
    layer output_layer = *(l.output_layer);

    increment_layer(&input_layer, l.steps-1);
    increment_layer(&self_layer, l.steps-1);
    increment_layer(&output_layer, l.steps-1);

    l.state += l.outputs*l.batch*l.steps;
    for (i = l.steps-1; i >= 0; --i) {
        copy_cpu(l.outputs * l.batch, input_layer.output, 1, l.state, 1);
        axpy_cpu(l.outputs * l.batch, 1, self_layer.output, 1, l.state, 1);

        s.input = l.state;
        s.delta = self_layer.delta;
        backward_connected_layer(output_layer, s);

        l.state -= l.outputs*l.batch;
        /*
           if(i > 0){
           copy_cpu(l.outputs * l.batch, input_layer.output - l.outputs*l.batch, 1, l.state, 1);
           axpy_cpu(l.outputs * l.batch, 1, self_layer.output - l.outputs*l.batch, 1, l.state, 1);
           }else{
           fill_cpu(l.outputs * l.batch, 0, l.state, 1);
           }
         */

        s.input = l.state;
        s.delta = self_layer.delta - l.outputs*l.batch;
        if (i == 0) s.delta = 0;
        backward_connected_layer(self_layer, s);

        copy_cpu(l.outputs*l.batch, self_layer.delta, 1, input_layer.delta, 1);
        if (i > 0 && l.shortcut) axpy_cpu(l.outputs*l.batch, 1, self_layer.delta, 1, self_layer.delta - l.outputs*l.batch, 1);
        s.input = net.input + i*l.inputs*l.batch;
        if(net.delta) s.delta = net.delta + i*l.inputs*l.batch;
        else s.delta = 0;
        backward_connected_layer(input_layer, s);
        //误差传回后，后退一步
        increment_layer(&input_layer, -1);
        increment_layer(&self_layer, -1);
        increment_layer(&output_layer, -1);
    }
}

没啥好说的(￣▽￣)”

0x0204 update_rnn_layer

void update_rnn_layer(layer l, update_args a)
{
    update_connected_layer(*(l.input_layer),  a);
    update_connected_layer(*(l.self_layer),   a);
    update_connected_layer(*(l.output_layer), a);
}

更新参数信息，调用三次全连接层的更新函数

回到parse_network_cfg函数

        else if(lt == GRU){
            l = parse_gru(options, params);
        }

0x03 parse_gru

这一部分内容我本来不打算写了，因为这和YOLOv2没有一点关系，但是抱着一种学习的态度，我还是将这部分代码做一些简要地分析，如果有不对的地方，希望大家指出，互相学习。参考这篇论文

layer parse_gru(list *options, size_params params)
{
    int output = option_find_int(options, "output",1);
    int batch_normalize = option_find_int_quiet(options, "batch_normalize", 0);

    layer l = make_gru_layer(params.batch, params.inputs, output, params.time_steps, batch_normalize, params.net->adam);
    l.tanh = option_find_int_quiet(options, "tanh", 0);//tanh激活函数

    return l;
}

我们现在看看这个make_gru_layer函数

layer make_gru_layer(int batch, int inputs, int outputs, int steps, int batch_normalize, int adam)
{
...

    l.uz = malloc(sizeof(layer));
    fprintf(stderr, "\t\t");
    *(l.uz) = make_connected_layer(batch*steps, inputs, outputs, LINEAR, batch_normalize, adam);
    l.uz->batch = batch;

    l.wz = malloc(sizeof(layer));
    fprintf(stderr, "\t\t");
    *(l.wz) = make_connected_layer(batch*steps, outputs, outputs, LINEAR, batch_normalize, adam);
    l.wz->batch = batch;

    l.ur = malloc(sizeof(layer));
    fprintf(stderr, "\t\t");
    *(l.ur) = make_connected_layer(batch*steps, inputs, outputs, LINEAR, batch_normalize, adam);
    l.ur->batch = batch;

    l.wr = malloc(sizeof(layer));
    fprintf(stderr, "\t\t");
    *(l.wr) = make_connected_layer(batch*steps, outputs, outputs, LINEAR, batch_normalize, adam);
    l.wr->batch = batch;

    l.uh = malloc(sizeof(layer));
    fprintf(stderr, "\t\t");
    *(l.uh) = make_connected_layer(batch*steps, inputs, outputs, LINEAR, batch_normalize, adam);
    l.uh->batch = batch;

    l.wh = malloc(sizeof(layer));
    fprintf(stderr, "\t\t");
    *(l.wh) = make_connected_layer(batch*steps, outputs, outputs, LINEAR, batch_normalize, adam);
    l.wh->batch = batch;

    l.batch_normalize = batch_normalize;


    l.outputs = outputs;
    l.output = calloc(outputs*batch*steps, sizeof(float));
    l.delta = calloc(outputs*batch*steps, sizeof(float));
    l.state = calloc(outputs*batch, sizeof(float));
    l.prev_state = calloc(outputs*batch, sizeof(float));
    l.forgot_state = calloc(outputs*batch, sizeof(float));
    l.forgot_delta = calloc(outputs*batch, sizeof(float));

    l.r_cpu = calloc(outputs*batch, sizeof(float));
    l.z_cpu = calloc(outputs*batch, sizeof(float));
    l.h_cpu = calloc(outputs*batch, sizeof(float));

    l.forward = forward_gru_layer;
    l.backward = backward_gru_layer;
    l.update = update_gru_layer;
...
}

0x0301 forward_gru_layer

void forward_gru_layer(layer l, network net)
{
    network s = net;
    s.train = net.train;
    int i;
    layer uz = *(l.uz);
    layer ur = *(l.ur);
    layer uh = *(l.uh);

    layer wz = *(l.wz);
    layer wr = *(l.wr);
    layer wh = *(l.wh);

    fill_cpu(l.outputs * l.batch * l.steps, 0, uz.delta, 1);
    fill_cpu(l.outputs * l.batch * l.steps, 0, ur.delta, 1);
    fill_cpu(l.outputs * l.batch * l.steps, 0, uh.delta, 1);

    fill_cpu(l.outputs * l.batch * l.steps, 0, wz.delta, 1);
    fill_cpu(l.outputs * l.batch * l.steps, 0, wr.delta, 1);
    fill_cpu(l.outputs * l.batch * l.steps, 0, wh.delta, 1);
    if(net.train) {
        fill_cpu(l.outputs * l.batch * l.steps, 0, l.delta, 1);
        copy_cpu(l.outputs*l.batch, l.state, 1, l.prev_state, 1);
    }

    for (i = 0; i < l.steps; ++i) {
        s.input = l.state;
        forward_connected_layer(wz, s);
        forward_connected_layer(wr, s);

        s.input = net.input;
        forward_connected_layer(uz, s);
        forward_connected_layer(ur, s);
        forward_connected_layer(uh, s);


        copy_cpu(l.outputs*l.batch, uz.output, 1, l.z_cpu, 1);
        axpy_cpu(l.outputs*l.batch, 1, wz.output, 1, l.z_cpu, 1);

        copy_cpu(l.outputs*l.batch, ur.output, 1, l.r_cpu, 1);
        axpy_cpu(l.outputs*l.batch, 1, wr.output, 1, l.r_cpu, 1);

        activate_array(l.z_cpu, l.outputs*l.batch, LOGISTIC);
        activate_array(l.r_cpu, l.outputs*l.batch, LOGISTIC);

        copy_cpu(l.outputs*l.batch, l.state, 1, l.forgot_state, 1);
        mul_cpu(l.outputs*l.batch, l.r_cpu, 1, l.forgot_state, 1);

        s.input = l.forgot_state;
        forward_connected_layer(wh, s);

        copy_cpu(l.outputs*l.batch, uh.output, 1, l.h_cpu, 1);
        axpy_cpu(l.outputs*l.batch, 1, wh.output, 1, l.h_cpu, 1);

        if(l.tanh){
            activate_array(l.h_cpu, l.outputs*l.batch, TANH);
        } else {
            activate_array(l.h_cpu, l.outputs*l.batch, LOGISTIC);
        }

        weighted_sum_cpu(l.state, l.h_cpu, l.z_cpu, l.outputs*l.batch, l.output);

        copy_cpu(l.outputs*l.batch, l.output, 1, l.state, 1);

        net.input += l.inputs*l.batch;
        l.output += l.outputs*l.batch;
        increment_layer(&uz, 1);
        increment_layer(&ur, 1);
        increment_layer(&uh, 1);

        increment_layer(&wz, 1);
        increment_layer(&wr, 1);
        increment_layer(&wh, 1);
    }
}

• axpy_cpu(l.outputs*l.batch, 1, wz.output, 1, l.z_cpu, 1);

  对应 iz(t)=xtUz+st−1Wz

• axpy_cpu(l.outputs*l.batch, 1, wr.output, 1, l.r_cpu, 1);

  对应 iz(t)=xtUz+st−1Wz

• mul_cpu(l.outputs*l.batch, l.r_cpu, 1, l.forgot_state, 1);

  对应 st−1⊙r(t)

• axpy_cpu(l.outputs*l.batch, 1, wh.output, 1, l.h_cpu, 1);

  对应 ih(t)=xtUh+(st−1⊙r(t))Wh

看一下weighted_sum_cpu函数的作用

void weighted_sum_cpu(float *a, float *b, float *s, int n, float *c)
{
    int i;
    for(i = 0; i < n; ++i){
        c[i] = s[i]*a[i] + (1-s[i])*(b ? b[i] : 0);
    }
}

• weighted_sum_cpu(l.state, l.h_cpu, l.z_cpu, l.outputs*l.batch, l.output);

  对应 (1−z)⊙h+z⊙st−1

0x0302 update_gru_layer

void update_gru_layer(layer l, update_args a)
{
    update_connected_layer(*(l.ur), a);
    update_connected_layer(*(l.uz), a);
    update_connected_layer(*(l.uh), a);
    update_connected_layer(*(l.wr), a);
    update_connected_layer(*(l.wz), a);
    update_connected_layer(*(l.wh), a);
}

我现在看到的这个源码，作者还没有实现backward_gru_layer

由于本人水平有限，文中有不对之处，希望大家指出，谢谢^_^!