checkpoint

Joseph Redmon

2014-11-18 d407bffde934ea4c1ee392f24cdf26d9a987199b

 src/connected_layer.c

@@ -7,58 +7,69 @@
#include <stdlib.h>
#include <string.h>

connected_layer *make_connected_layer(int batch, int inputs, int outputs, float dropout, ACTIVATION activation)
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));

    layer->learning_rate = learning_rate;
    layer->momentum = momentum;
    layer->decay = decay;

    layer->inputs = inputs;
    layer->outputs = outputs;
    layer->batch=batch;
    layer->dropout = dropout;

    layer->output = calloc(batch*outputs, sizeof(float*));
    layer->delta = calloc(batch*outputs, sizeof(float*));

    layer->weight_updates = calloc(inputs*outputs, sizeof(float));
    layer->weight_adapt = calloc(inputs*outputs, sizeof(float));
    layer->weight_momentum = 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*(rand_uniform());
        layer->weights[i] = scale*2*(rand_uniform()-.5);

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

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

    layer->weight_updates_cl = cl_make_array(layer->weight_updates, inputs*outputs);
    layer->bias_updates_cl = cl_make_array(layer->bias_updates, outputs);

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

void update_connected_layer(connected_layer layer, float step, float momentum, float decay)
void update_connected_layer(connected_layer layer)
{
    int i;
    for(i = 0; i < layer.outputs; ++i){
        layer.bias_momentum[i] = step*(layer.bias_updates[i]) + momentum*layer.bias_momentum[i];
        layer.biases[i] += layer.bias_momentum[i];
    }
    for(i = 0; i < layer.outputs*layer.inputs; ++i){
        layer.weight_momentum[i] = step*(layer.weight_updates[i] - decay*layer.weights[i]) + momentum*layer.weight_momentum[i];
        layer.weights[i] += layer.weight_momentum[i];
    }
    memset(layer.bias_updates, 0, layer.outputs*sizeof(float));
    memset(layer.weight_updates, 0, layer.outputs*layer.inputs*sizeof(float));
    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.learning_rate, layer.weight_updates, 1, layer.weights, 1);
    scal_cpu(layer.inputs*layer.outputs, layer.momentum, layer.weight_updates, 1);
}

void forward_connected_layer(connected_layer layer, float *input, int train)
void forward_connected_layer(connected_layer layer, float *input)
{
    if(!train) layer.dropout = 0;
    memcpy(layer.output, layer.biases, layer.outputs*sizeof(float));
    int i;
    for(i = 0; i < layer.batch; ++i){
        copy_cpu(layer.outputs, layer.biases, 1, layer.output + i*layer.outputs, 1);
    }
    int m = layer.batch;
    int k = layer.inputs;
    int n = layer.outputs;
@@ -66,15 +77,15 @@
    float *b = layer.weights;
    float *c = layer.output;
    gemm(0,0,m,n,k,1,a,k,b,n,1,c,n);
    activate_array(layer.output, layer.outputs*layer.batch, layer.activation, layer.dropout);
    activate_array(layer.output, layer.outputs*layer.batch, layer.activation);
}

void backward_connected_layer(connected_layer layer, float *input, float *delta)
{
    int i;
    for(i = 0; i < layer.outputs*layer.batch; ++i){
        layer.delta[i] *= gradient(layer.output[i], layer.activation);
        layer.bias_updates[i%layer.batch] += layer.delta[i];
    gradient_array(layer.output, layer.outputs*layer.batch, layer.activation, layer.delta);
    for(i = 0; i < layer.batch; ++i){
        axpy_cpu(layer.outputs, 1, layer.delta + i*layer.outputs, 1, layer.bias_updates, 1);
    }
    int m = layer.inputs;
    int k = layer.batch;
@@ -82,16 +93,83 @@
    float *a = input;
    float *b = layer.delta;
    float *c = layer.weight_updates;
    gemm(0,0,m,n,k,1,a,k,b,n,1,c,n);
    gemm(1,0,m,n,k,1,a,m,b,n,1,c,n);

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

    a = layer.weights;
    b = layer.delta;
    a = layer.delta;
    b = layer.weights;
    c = delta;

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

#ifdef GPU

void pull_connected_layer(connected_layer layer)
{
    cl_read_array(layer.weights_cl, layer.weights, layer.inputs*layer.outputs);
    cl_read_array(layer.biases_cl, layer.biases, 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);
}

void update_connected_layer_gpu(connected_layer layer)
{
    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.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);
}

void forward_connected_layer_gpu(connected_layer layer, cl_mem input)
{
    int i;
    for(i = 0; i < layer.batch; ++i){
        copy_ongpu_offset(layer.outputs, layer.biases_cl, 0, 1, layer.output_cl, i*layer.outputs, 1);
    }
    int m = layer.batch;
    int k = layer.inputs;
    int n = layer.outputs;
    cl_mem a = input;
    cl_mem b = layer.weights_cl;
    cl_mem c = layer.output_cl;
    gemm_ongpu(0,0,m,n,k,1,a,k,b,n,1,c,n);
    activate_array_ongpu(layer.output_cl, layer.outputs*layer.batch, layer.activation);
}

void backward_connected_layer_gpu(connected_layer layer, cl_mem input, cl_mem delta)
{
    int i;
    gradient_array_ongpu(layer.output_cl, layer.outputs*layer.batch, layer.activation, layer.delta_cl);
    for(i = 0; i < layer.batch; ++i){
        axpy_ongpu_offset(layer.outputs, 1, layer.delta_cl, i*layer.outputs, 1, layer.bias_updates_cl, 0, 1);
    }
    int m = layer.inputs;
    int k = layer.batch;
    int n = layer.outputs;
    cl_mem a = input;
    cl_mem b = layer.delta_cl;
    cl_mem c = layer.weight_updates_cl;
    gemm_ongpu(1,0,m,n,k,1,a,m,b,n,1,c,n);

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

    a = layer.delta_cl;
    b = layer.weights_cl;
    c = delta;

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