Darknet源码剖析第一讲：src/activations.h 激活函数分析

这个头文件里总共有以下几个函数

ACTIVATION get_activation(char *s);

char *get_activation_string(ACTIVATION a);
float activate(float x, ACTIVATION a);
float gradient(float x, ACTIVATION a);
void gradient_array(const float *x, const int n, const ACTIVATION a, float *delta);
void activate_array(float *x, const int n, const ACTIVATION a);
#ifdef GPU
void activate_array_gpu(float *x, int n, ACTIVATION a);
void gradient_array_gpu(float *x, int n, ACTIVATION a, float *delta);
#endif

其中ACTIVATION是枚举类型，定义了各种激活函数

typedef enum{
    LOGISTIC, RELU, RELIE, LINEAR, RAMP, TANH, PLSE, LEAKY, ELU, LOGGY, STAIR, HARDTAN, LHTAN, SELU
} ACTIVATION;

接下来通过inline快速实现一些简单的激活函数计算

static inline float linear_activate(float x){return x;}
static inline float logistic_activate(float x){return 1./(1. + exp(-x));}
static inline float loggy_activate(float x){return 2./(1. + exp(-x)) - 1;}
static inline float relu_activate(float x){return x*(x>0);}
static inline float elu_activate(float x){return (x >= 0)*x + (x < 0)*(exp(x)-1);}
static inline float selu_activate(float x){return (x >= 0)*1.0507*x + (x < 0)*1.0507*1.6732*(exp(x)-1);}
static inline float relie_activate(float x){return (x>0) ? x : .01*x;}
static inline float ramp_activate(float x){return x*(x>0)+.1*x;}
static inline float leaky_activate(float x){return (x>0) ? x : .1*x;}
static inline float tanh_activate(float x){return (exp(2*x)-1)/(exp(2*x)+1);}

然后实现统一的激活函数计算

float activate(float x, ACTIVATION a)
{
    switch(a){
        case LINEAR:
            return linear_activate(x);
        case LOGISTIC:
            return logistic_activate(x);
        case LOGGY:
            return loggy_activate(x);
        case RELU:
            return relu_activate(x);
        case ELU:
            return elu_activate(x);
        case SELU:
            return selu_activate(x);
        case RELIE:
            return relie_activate(x);
        case RAMP:
            return ramp_activate(x);
        case LEAKY:
            return leaky_activate(x);
        case TANH:
            return tanh_activate(x);
        case PLSE:
            return plse_activate(x);
        case STAIR:
            return stair_activate(x);
        case HARDTAN:
            return hardtan_activate(x);
        case LHTAN:
            return lhtan_activate(x);
    }
    return 0;
}

void activate_array(float *x, const int n, const ACTIVATION a)
{
    int i;
    for(i = 0; i < n; ++i){
        x[i] = activate(x[i], a);
    }
}

然后梯度的实现也是相同的

float gradient(float x, ACTIVATION a)
{
    switch(a){
        case LINEAR:
            return linear_gradient(x);
        case LOGISTIC:
            return logistic_gradient(x);
        case LOGGY:
            return loggy_gradient(x);
        case RELU:
            return relu_gradient(x);
        case ELU:
            return elu_gradient(x);
        case SELU:
            return selu_gradient(x);
        case RELIE:
            return relie_gradient(x);
        case RAMP:
            return ramp_gradient(x);
        case LEAKY:
            return leaky_gradient(x);
        case TANH:
            return tanh_gradient(x);
        case PLSE:
            return plse_gradient(x);
        case STAIR:
            return stair_gradient(x);
        case HARDTAN:
            return hardtan_gradient(x);
        case LHTAN:
            return lhtan_gradient(x);
    }
    return 0;
}

然后去看activations_layer的实现

layer make_activation_layer(int batch, int inputs, ACTIVATION activation);

void forward_activation_layer(layer l, network net);
void backward_activation_layer(layer l, network net);

#ifdef GPU
void forward_activation_layer_gpu(layer l, network net);
void backward_activation_layer_gpu(layer l, network net);
#endif

layer和net是两个结构体

layer这个结构体东西相当多，net也是结构体，东西很多。

layer make_activation_layer(int batch, int inputs, ACTIVATION activation)
{
    layer l = {0};
    l.type = ACTIVE;

    l.inputs = inputs;
    l.outputs = inputs;
    l.batch=batch;

    l.output = calloc(batch*inputs, sizeof(float*));
    l.delta = calloc(batch*inputs, sizeof(float*));

    l.forward = forward_activation_layer;
    l.backward = backward_activation_layer;
#ifdef GPU
    l.forward_gpu = forward_activation_layer_gpu;
    l.backward_gpu = backward_activation_layer_gpu;

    l.output_gpu = cuda_make_array(l.output, inputs*batch);
    l.delta_gpu = cuda_make_array(l.delta, inputs*batch);
#endif
    l.activation = activation;
    fprintf(stderr, "Activation Layer: %d inputs\n", inputs);
    return l;
}

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);
}

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);
}

#ifdef GPU

void forward_activation_layer_gpu(layer l, network net)
{
    copy_gpu(l.outputs*l.batch, net.input_gpu, 1, l.output_gpu, 1);
    activate_array_gpu(l.output_gpu, l.outputs*l.batch, l.activation);
}

void backward_activation_layer_gpu(layer l, network net)
{
    gradient_array_gpu(l.output_gpu, l.outputs*l.batch, l.activation, l.delta_gpu);
    copy_gpu(l.outputs*l.batch, l.delta_gpu, 1, net.delta_gpu, 1);
}
#endif

重点看forward和backward

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);
}

 

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);
}

其中

void copy_cpu(int N, float *X, int INCX, float *Y, int INCY)
{
    int i;
    for(i = 0; i < N; ++i) Y[i*INCY] = X[i*INCX];
}