Faster

Joseph Redmon

2015-01-13 aa5996d58e68edfbefe51061856aecd549dd09c4

 src/connected_layer.c

@@ -9,7 +9,6 @@

connected_layer *make_connected_layer(int batch, int inputs, int outputs, ACTIVATION activation, float learning_rate, float momentum, float decay)
{
    fprintf(stderr, "Connected Layer: %d inputs, %d outputs\n", inputs, outputs);
    int i;
    connected_layer *layer = calloc(1, sizeof(connected_layer));

@@ -25,22 +24,26 @@
    layer->delta = calloc(batch*outputs, sizeof(float*));

    layer->weight_updates = calloc(inputs*outputs, sizeof(float));
    //layer->weight_adapt = calloc(inputs*outputs, sizeof(float));
    layer->weights = calloc(inputs*outputs, sizeof(float));
    float scale = 1./inputs;
    scale = .01;
    for(i = 0; i < inputs*outputs; ++i)
        layer->weights[i] = scale*2*(rand_uniform()-.5);

    layer->bias_updates = calloc(outputs, sizeof(float));
    //layer->bias_adapt = calloc(outputs, sizeof(float));

    layer->weight_prev = calloc(inputs*outputs, sizeof(float));
    layer->bias_prev = calloc(outputs, sizeof(float));

    layer->weights = calloc(inputs*outputs, sizeof(float));
    layer->biases = calloc(outputs, sizeof(float));
    for(i = 0; i < outputs; ++i){
        //layer->biases[i] = rand_normal()*scale + scale;
        layer->biases[i] = 1;


    float scale = 1./sqrt(inputs);
    //scale = .01;
    for(i = 0; i < inputs*outputs; ++i){
        layer->weights[i] = scale*rand_normal();
    }

    #ifdef GPU
    for(i = 0; i < outputs; ++i){
        layer->biases[i] = scale;
    }

#ifdef GPU
    layer->weights_cl = cl_make_array(layer->weights, inputs*outputs);
    layer->biases_cl = cl_make_array(layer->biases, outputs);

@@ -49,17 +52,44 @@

    layer->output_cl = cl_make_array(layer->output, outputs*batch);
    layer->delta_cl = cl_make_array(layer->delta, outputs*batch);
    #endif
#endif
    layer->activation = activation;
    fprintf(stderr, "Connected Layer: %d inputs, %d outputs\n", inputs, outputs);
    return layer;
}

void secret_update_connected_layer(connected_layer *layer)
{
    int n = layer->outputs*layer->inputs;
    float dot = dot_cpu(n, layer->weight_updates, 1, layer->weight_prev, 1);
    float mag = sqrt(dot_cpu(n, layer->weight_updates, 1, layer->weight_updates, 1))
                * sqrt(dot_cpu(n, layer->weight_prev, 1, layer->weight_prev, 1));
    float cos = dot/mag;
    if(cos > .3) layer->learning_rate *= 1.1;
    else if (cos < -.3) layer-> learning_rate /= 1.1;

    scal_cpu(n, layer->momentum, layer->weight_prev, 1);
    axpy_cpu(n, 1, layer->weight_updates, 1, layer->weight_prev, 1);
    scal_cpu(n, 0, layer->weight_updates, 1);

    scal_cpu(layer->outputs, layer->momentum, layer->bias_prev, 1);
    axpy_cpu(layer->outputs, 1, layer->bias_updates, 1, layer->bias_prev, 1);
    scal_cpu(layer->outputs, 0, layer->bias_updates, 1);

    //printf("rate:   %f\n", layer->learning_rate);

    axpy_cpu(layer->outputs, layer->learning_rate, layer->bias_prev, 1, layer->biases, 1);

    axpy_cpu(layer->inputs*layer->outputs, -layer->decay, layer->weights, 1, layer->weight_prev, 1);
    axpy_cpu(layer->inputs*layer->outputs, layer->learning_rate, layer->weight_prev, 1, layer->weights, 1);
}

void update_connected_layer(connected_layer layer)
{
    axpy_cpu(layer.outputs, layer.learning_rate, layer.bias_updates, 1, layer.biases, 1);
    scal_cpu(layer.outputs, layer.momentum, layer.bias_updates, 1);

    scal_cpu(layer.inputs*layer.outputs, 1.-layer.learning_rate*layer.decay, layer.weights, 1);
    axpy_cpu(layer.inputs*layer.outputs, -layer.decay, layer.weights, 1, layer.weight_updates, 1);
    axpy_cpu(layer.inputs*layer.outputs, layer.learning_rate, layer.weight_updates, 1, layer.weights, 1);
    scal_cpu(layer.inputs*layer.outputs, layer.momentum, layer.weight_updates, 1);
}
@@ -112,12 +142,16 @@
{
    cl_read_array(layer.weights_cl, layer.weights, layer.inputs*layer.outputs);
    cl_read_array(layer.biases_cl, layer.biases, layer.outputs);
    cl_read_array(layer.weight_updates_cl, layer.weight_updates, layer.inputs*layer.outputs);
    cl_read_array(layer.bias_updates_cl, layer.bias_updates, layer.outputs);
}

void push_connected_layer(connected_layer layer)
{
    cl_write_array(layer.weights_cl, layer.weights, layer.inputs*layer.outputs);
    cl_write_array(layer.biases_cl, layer.biases, layer.outputs);
    cl_write_array(layer.weight_updates_cl, layer.weight_updates, layer.inputs*layer.outputs);
    cl_write_array(layer.bias_updates_cl, layer.bias_updates, layer.outputs);
}

void update_connected_layer_gpu(connected_layer layer)
@@ -125,10 +159,10 @@
    axpy_ongpu(layer.outputs, layer.learning_rate, layer.bias_updates_cl, 1, layer.biases_cl, 1);
    scal_ongpu(layer.outputs, layer.momentum, layer.bias_updates_cl, 1);

    scal_ongpu(layer.inputs*layer.outputs, 1.-layer.learning_rate*layer.decay, layer.weights_cl, 1);
    axpy_ongpu(layer.inputs*layer.outputs, -layer.decay, layer.weights_cl, 1, layer.weight_updates_cl, 1);
    axpy_ongpu(layer.inputs*layer.outputs, layer.learning_rate, layer.weight_updates_cl, 1, layer.weights_cl, 1);
    scal_ongpu(layer.inputs*layer.outputs, layer.momentum, layer.weight_updates_cl, 1);
    pull_connected_layer(layer);
    //pull_connected_layer(layer);
}

void forward_connected_layer_gpu(connected_layer layer, cl_mem input)
@@ -172,4 +206,4 @@

    if(c) gemm_ongpu(0,1,m,n,k,1,a,k,b,k,0,c,n);
}
    #endif
#endif