Darknet 训练（train）源码分析

1.查看Yolo中train函数

Paste_Image.png

这里面确定训练数据路径和中间备份路径、准备随机batch数据、确定网络结构，just as always do。这里主要查找其使怎样训练的（关于定位和识别）
查找loss函数

float loss = train_network(net, train);
if (avg_loss < 0) avg_loss = loss;
        avg_loss = avg_loss*.9 + loss*.1;

这里比较有意思avg_loss may be <0，还不明白为啥会小于0
然后做一个加权
发现其调用network.c中的train_network函数

float train_network_datum(network net, float *x, float *y)
//train_network()调用这个函数，把batch中所有x，y摊平
{
#ifdef GPU
    if(gpu_index >= 0) return train_network_datum_gpu(net, x, y);
#endif
    network_state state;
    *net.seen += net.batch;
    state.index = 0;
    state.net = net;
    state.input = x;
    state.delta = 0;
    state.truth = y;
    state.train = 1;
    forward_network(net, state);
    backward_network(net, state);
    float error = get_network_cost(net);//here cost
    if(((*net.seen)/net.batch)%net.subdivisions == 0) update_network(net);
    return error;
}

查看 get_network_cost(net)

float get_network_cost(network net)
{
    int i;
    float sum = 0;
    int count = 0;
    for(i = 0; i < net.n; ++i){
        if(net.layers[i].cost){//here important：maybe two cost layers
            sum += net.layers[i].cost[0];
            ++count;
        }
    }
    return sum/count;
}

typedef struct network{
    float *workspace;
    int n;
    int batch;
    int *seen;
    float epoch;
    int subdivisions;
    float momentum;
    float decay;
    layer *layers;
    int outputs;
    float *output;
    learning_rate_policy policy;

    float learning_rate;
    float gamma;
    float scale;
    float power;
    int time_steps;
    int step;
    int max_batches;
    float *scales;
    int   *steps;
    int num_steps;
    int burn_in;

    int inputs;
    int h, w, c;
    int max_crop;
    int min_crop;
    float angle;
    float aspect;
    float exposure;
    float saturation;
    float hue;

    int gpu_index;

    #ifdef GPU
    float **input_gpu;
    float **truth_gpu;
    #endif
} network;

然后看layer，amazing data structure

struct layer{
    LAYER_TYPE type;
    ACTIVATION activation;
    COST_TYPE cost_type;
    int batch_normalize;
    int shortcut;
    int batch;
    int forced;
    int flipped;
    int inputs;
    int outputs;
    int truths;
    int h,w,c;
    int out_h, out_w, out_c;
    int n;
    int max_boxes;
    int groups;
    int size;
    int side;
    int stride;
    int pad;
    int sqrt;
    int flip;
    int index;
    int binary;
    int xnor;
    int steps;
    int hidden;
    float dot;
    float angle;
    float jitter;
    float saturation;
    float exposure;
    float shift;
    float ratio;
    int softmax;
    int classes;
    int coords;
    int background;
    int rescore;
    int objectness;
    int does_cost;
    int joint;
    int noadjust;
    int reorg;
    int log;

    float alpha;
    float beta;
    float kappa;

    float coord_scale;
    float object_scale;
    float noobject_scale;
    float class_scale;
    int random;

    int dontload;
    int dontloadscales;

    float temperature;
    float probability;
    float scale;

    int *indexes;
    float *rand;
    float *cost;
    char  *cweights;
    float *state;
    float *prev_state;
    float *forgot_state;
    float *forgot_delta;
    float *state_delta;

    float *concat;
    float *concat_delta;

    float *binary_weights;

    float *biases;
    float *bias_updates;

    float *scales;
    float *scale_updates;

    float *weights;
    float *weight_updates;

    float *col_image;
    int   * input_layers;
    int   * input_sizes;
    float * delta;
    float * output;
    float * squared;
    float * norms;

    float * spatial_mean;
    float * mean;
    float * variance;

    float * mean_delta;
    float * variance_delta;

    float * rolling_mean;
    float * rolling_variance;

    float * x;
    float * x_norm;

    struct layer *input_layer;
    struct layer *self_layer;
    struct layer *output_layer;

    struct layer *input_gate_layer;
    struct layer *state_gate_layer;
    struct layer *input_save_layer;
    struct layer *state_save_layer;
    struct layer *input_state_layer;
    struct layer *state_state_layer;

    struct layer *input_z_layer;
    struct layer *state_z_layer;

    struct layer *input_r_layer;
    struct layer *state_r_layer;

    struct layer *input_h_layer;
    struct layer *state_h_layer;

    float *z_cpu;
    float *r_cpu;
    float *h_cpu;

    float *binary_input;

    size_t workspace_size;

    #ifdef GPU
    float *z_gpu;
    float *r_gpu;
    float *h_gpu;

    int *indexes_gpu;
    float * prev_state_gpu;
    float * forgot_state_gpu;
    float * forgot_delta_gpu;
    float * state_gpu;
    float * state_delta_gpu;
    float * gate_gpu;
    float * gate_delta_gpu;
    float * save_gpu;
    float * save_delta_gpu;
    float * concat_gpu;
    float * concat_delta_gpu;

    float *binary_input_gpu;
    float *binary_weights_gpu;

    float * mean_gpu;
    float * variance_gpu;

    float * rolling_mean_gpu;
    float * rolling_variance_gpu;

    float * variance_delta_gpu;
    float * mean_delta_gpu;

    float * col_image_gpu;

    float * x_gpu;
    float * x_norm_gpu;
    float * weights_gpu;
    float * weight_updates_gpu;

    float * biases_gpu;
    float * bias_updates_gpu;

    float * scales_gpu;
    float * scale_updates_gpu;

    float * output_gpu;
    float * delta_gpu;
    float * rand_gpu;
    float * squared_gpu;
    float * norms_gpu;
    #ifdef CUDNN
    cudnnTensorDescriptor_t srcTensorDesc, dstTensorDesc;
    cudnnTensorDescriptor_t dsrcTensorDesc, ddstTensorDesc;
    cudnnFilterDescriptor_t weightDesc;
    cudnnFilterDescriptor_t dweightDesc;
    cudnnConvolutionDescriptor_t convDesc;
    cudnnConvolutionFwdAlgo_t fw_algo;
    cudnnConvolutionBwdDataAlgo_t bd_algo;
    cudnnConvolutionBwdFilterAlgo_t bf_algo;
    #endif
    #endif
};

所以是在forward propogation时就计算好了cost

void forward_network(network net, network_state state)
{
    state.workspace = net.workspace;
    int i;
    for(i = 0; i < net.n; ++i){
        state.index = i;
        layer l = net.layers[i];
        if(l.delta){
            scal_cpu(l.outputs * l.batch, 0, l.delta, 1);
        }
        if(l.type == CONVOLUTIONAL){
            forward_convolutional_layer(l, state);
        } else if(l.type == DECONVOLUTIONAL){
            forward_deconvolutional_layer(l, state);
        } else if(l.type == ACTIVE){
            forward_activation_layer(l, state);
        } else if(l.type == LOCAL){
            forward_local_layer(l, state);
        } else if(l.type == NORMALIZATION){
            forward_normalization_layer(l, state);
        } else if(l.type == BATCHNORM){
            forward_batchnorm_layer(l, state);
        } else if(l.type == DETECTION){
            forward_detection_layer(l, state);
        } else if(l.type == REGION){
            forward_region_layer(l, state);
        } else if(l.type == CONNECTED){
            forward_connected_layer(l, state);
        } else if(l.type == RNN){
            forward_rnn_layer(l, state);
        } else if(l.type == GRU){
            forward_gru_layer(l, state);
        } else if(l.type == CRNN){
            forward_crnn_layer(l, state);
        } else if(l.type == CROP){
            forward_crop_layer(l, state);
        } else if(l.type == COST){
            forward_cost_layer(l, state);
        } else if(l.type == SOFTMAX){
            forward_softmax_layer(l, state);
        } else if(l.type == MAXPOOL){
            forward_maxpool_layer(l, state);
        } else if(l.type == REORG){
            forward_reorg_layer(l, state);
        } else if(l.type == AVGPOOL){
            forward_avgpool_layer(l, state);
        } else if(l.type == DROPOUT){
            forward_dropout_layer(l, state);
        } else if(l.type == ROUTE){
            forward_route_layer(l, net);
        } else if(l.type == SHORTCUT){
            forward_shortcut_layer(l, state);
        }
        state.input = l.output;
    }
}

下面我们看到他用smooth_l1 l2进行计算代价函数

 if(l.cost_type == SMOOTH){
        smooth_l1_cpu(l.batch*l.inputs, state.input, state.truth, l.delta, l.output);
    } else {
        l2_cpu(l.batch*l.inputs, state.input, state.truth, l.delta, l.output);
    }
    l.cost[0] = sum_array(l.output, l.batch*l.inputs);
void smooth_l1_cpu(int n, float *pred, float *truth, float *delta, float *error)
{
    int i;
    for(i = 0; i < n; ++i){
        float diff = truth[i] - pred[i];
        float abs_val = fabs(diff);
        if(abs_val < 1) {
            error[i] = diff * diff;
            delta[i] = diff;
        }
        else {
            error[i] = 2*abs_val - 1;
            delta[i] = (diff < 0) ? -1 : 1;
        }
    }
}

void l2_cpu(int n, float *pred, float *truth, float *delta, float *error)
{
    int i;
    for(i = 0; i < n; ++i){
        float diff = truth[i] - pred[i];
        error[i] = diff * diff;
        delta[i] = diff;
    }
}

主要用到状态state里的truth和看n的大小l.inputs

typedef struct network_state {
    float *truth;
    float *input;
    float *delta;
    float *workspace;
    int train;
    int index;
    network net;
} network_state;

float train_network_datum(network net, float *x, float *y)
{
#ifdef GPU
    if(gpu_index >= 0) return train_network_datum_gpu(net, x, y);
#endif
    network_state state;
    *net.seen += net.batch;
    state.index = 0;
    state.net = net;
    state.input = x;
    state.delta = 0;
    state.truth = y;//HERE
    state.train = 1;
    forward_network(net, state);
    backward_network(net, state);
    float error = get_network_cost(net);
    if(((*net.seen)/net.batch)%net.subdivisions == 0) update_network(net);
    return error;
}

float train_network(network net, data d)
{
    int batch = net.batch;
    int n = d.X.rows / batch;
    float *X = calloc(batch*d.X.cols, sizeof(float));
    float *y = calloc(batch*d.y.cols, sizeof(float));//HERE

    int i;
    float sum = 0;
    for(i = 0; i < n; ++i){
        get_next_batch(d, batch, i*batch, X, y);
        float err = train_network_datum(net, X, y);
        sum += err;
    }
    free(X);
    free(y);
    return (float)sum/(n*batch);
}

typedef struct{
    int w, h;
    matrix X;
    matrix y;//HERE
    int *indexes;
    int shallow;
    int *num_boxes;
    box **boxes;
} data;

可以看到只与y有关而与box无关