l = parse_connected(options, params);

l = parse_connected(options, params);函数主体

layer parse_connected(list *options, size_params params)
{
	//设置网络参数
    int output = option_find_int(options, "output",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);

    layer l = make_connected_layer(params.batch, params.inputs, output, activation, batch_normalize, params.net->adam);//构建全连接层
    return l;
}

make_connected_layer(params.batch, params.inputs, output, activation, batch_normalize, params.net->adam);

layer make_connected_layer(int batch, int inputs, int outputs, ACTIVATION activation, int batch_normalize, int adam)
{
	//前面与其他层一样，先设置网络参数
    int i;
    layer l = {0};
    l.learning_rate_scale = 1;
    l.type = CONNECTED;

    l.inputs = inputs;
    l.outputs = outputs;//这里的输出是人为设计的
    l.batch=batch;
    l.batch_normalize = batch_normalize;
    l.h = 1;
    l.w = 1;
    l.c = inputs;
    l.out_h = 1;
    l.out_w = 1;
    l.out_c = outputs;

    l.output = calloc(batch*outputs, sizeof(float));//为min_batch的所有输出分配空间
    l.delta = calloc(batch*outputs, sizeof(float));

    l.weight_updates = calloc(inputs*outputs, sizeof(float));//为更新的权重分配空间
    l.bias_updates = calloc(outputs, sizeof(float));//为更新的偏置分配空间

    l.weights = calloc(outputs*inputs, sizeof(float));//为权重分配空间
    l.biases = calloc(outputs, sizeof(float));//为偏置分配空间
	//为函数指针设置指向函数地址
    l.forward = forward_connected_layer;
    l.backward = backward_connected_layer;
    l.update = update_connected_layer;

    //float scale = 1./sqrt(inputs);
    // 初始化权重：缩放因子*标准均匀分布随机数（-1,1）之间，缩放因子等于sqrt(2./(size*size*c))，
    float scale = sqrt(2./inputs);
    for(i = 0; i < outputs*inputs; ++i){
        l.weights[i] = scale*rand_uniform(-1, 1);
    }
	// 初始化偏置，都设置为0
    for(i = 0; i < outputs; ++i){
        l.biases[i] = 0;
    }
    
	//为优化参数分配空间
    if(adam){
        l.m = calloc(l.inputs*l.outputs, sizeof(float));
        l.v = calloc(l.inputs*l.outputs, sizeof(float));
        l.bias_m = calloc(l.outputs, sizeof(float));
        l.scale_m = calloc(l.outputs, sizeof(float));
        l.bias_v = calloc(l.outputs, sizeof(float));
        l.scale_v = calloc(l.outputs, sizeof(float));
    }
    //为批标准化参数分配空间
    if(batch_normalize){
        l.scales = calloc(outputs, sizeof(float));
        l.scale_updates = calloc(outputs, sizeof(float));
        for(i = 0; i < outputs; ++i){
            l.scales[i] = 1;
        }

        l.mean = calloc(outputs, sizeof(float));
        l.mean_delta = calloc(outputs, sizeof(float));
        l.variance = calloc(outputs, sizeof(float));
        l.variance_delta = calloc(outputs, sizeof(float));

        l.rolling_mean = calloc(outputs, sizeof(float));
        l.rolling_variance = calloc(outputs, sizeof(float));

        l.x = calloc(batch*outputs, sizeof(float));
        l.x_norm = calloc(batch*outputs, sizeof(float));
    }
//若是定义了GPU则运行下面代码，与上面功能相同，效率更高
#ifdef GPU
    l.forward_gpu = forward_connected_layer_gpu;
    l.backward_gpu = backward_connected_layer_gpu;
    l.update_gpu = update_connected_layer_gpu;

    l.weights_gpu = cuda_make_array(l.weights, outputs*inputs);
    l.biases_gpu = cuda_make_array(l.biases, outputs);

    l.weight_updates_gpu = cuda_make_array(l.weight_updates, outputs*inputs);
    l.bias_updates_gpu = cuda_make_array(l.bias_updates, outputs);

    l.output_gpu = cuda_make_array(l.output, outputs*batch);
    l.delta_gpu = cuda_make_array(l.delta, outputs*batch);
    if (adam) {
        l.m_gpu =       cuda_make_array(0, inputs*outputs);
        l.v_gpu =       cuda_make_array(0, inputs*outputs);
        l.bias_m_gpu =  cuda_make_array(0, outputs);
        l.bias_v_gpu =  cuda_make_array(0, outputs);
        l.scale_m_gpu = cuda_make_array(0, outputs);
        l.scale_v_gpu = cuda_make_array(0, outputs);
    }

    if(batch_normalize){
        l.mean_gpu = cuda_make_array(l.mean, outputs);
        l.variance_gpu = cuda_make_array(l.variance, outputs);

        l.rolling_mean_gpu = cuda_make_array(l.mean, outputs);
        l.rolling_variance_gpu = cuda_make_array(l.variance, outputs);

        l.mean_delta_gpu = cuda_make_array(l.mean, outputs);
        l.variance_delta_gpu = cuda_make_array(l.variance, outputs);

        l.scales_gpu = cuda_make_array(l.scales, outputs);
        l.scale_updates_gpu = cuda_make_array(l.scale_updates, outputs);

        l.x_gpu = cuda_make_array(l.output, l.batch*outputs);
        l.x_norm_gpu = cuda_make_array(l.output, l.batch*outputs);
//若是定义了CUDNN则运行下面代码，由于GPU训练的加速
#ifdef CUDNN
        cudnnCreateTensorDescriptor(&l.normTensorDesc);
        cudnnCreateTensorDescriptor(&l.dstTensorDesc);
        cudnnSetTensor4dDescriptor(l.dstTensorDesc, CUDNN_TENSOR_NCHW, CUDNN_DATA_FLOAT, l.batch, l.out_c, l.out_h, l.out_w); 
        cudnnSetTensor4dDescriptor(l.normTensorDesc, CUDNN_TENSOR_NCHW, CUDNN_DATA_FLOAT, 1, l.out_c, 1, 1); 
#endif
    }
#endif
    l.activation = activation;//设置激活函数
    fprintf(stderr, "connected                            %4d  ->  %4d\n", inputs, outputs);
    return l;
}