Trying some stuff w/ dropout

Joseph Redmon

2014-08-28 76ee68f96d864a27312c9aa09856ddda559a5cd9

 src/network.c

@@ -4,11 +4,13 @@
#include "data.h"
#include "utils.h"

#include "crop_layer.h"
#include "connected_layer.h"
#include "convolutional_layer.h"
#include "maxpool_layer.h"
#include "normalization_layer.h"
#include "softmax_layer.h"
#include "dropout_layer.h"

network make_network(int n, int batch)
{
@@ -25,114 +27,17 @@
    return net;
}

void print_convolutional_cfg(FILE *fp, convolutional_layer *l, int first)
{
    int i;
    fprintf(fp, "[convolutional]\n");
    if(first) fprintf(fp,   "batch=%d\n"
                            "height=%d\n"
                            "width=%d\n"
                            "channels=%d\n",
                            l->batch,l->h, l->w, l->c);
    fprintf(fp, "filters=%d\n"
                "size=%d\n"
                "stride=%d\n"
                "activation=%s\n",
                l->n, l->size, l->stride,
                get_activation_string(l->activation));
    fprintf(fp, "data=");
    for(i = 0; i < l->n; ++i) fprintf(fp, "%g,", l->biases[i]);
    for(i = 0; i < l->n*l->c*l->size*l->size; ++i) fprintf(fp, "%g,", l->filters[i]);
    fprintf(fp, "\n\n");
}
void print_connected_cfg(FILE *fp, connected_layer *l, int first)
{
    int i;
    fprintf(fp, "[connected]\n");
    if(first) fprintf(fp, "batch=%d\ninput=%d\n", l->batch, l->inputs);
    fprintf(fp, "output=%d\n"
            "activation=%s\n",
            l->outputs,
            get_activation_string(l->activation));
    fprintf(fp, "data=");
    for(i = 0; i < l->outputs; ++i) fprintf(fp, "%g,", l->biases[i]);
    for(i = 0; i < l->inputs*l->outputs; ++i) fprintf(fp, "%g,", l->weights[i]);
    fprintf(fp, "\n\n");
}

void print_maxpool_cfg(FILE *fp, maxpool_layer *l, int first)
{
    fprintf(fp, "[maxpool]\n");
    if(first) fprintf(fp,   "batch=%d\n"
            "height=%d\n"
            "width=%d\n"
            "channels=%d\n",
            l->batch,l->h, l->w, l->c);
    fprintf(fp, "stride=%d\n\n", l->stride);
}

void print_normalization_cfg(FILE *fp, normalization_layer *l, int first)
{
    fprintf(fp, "[localresponsenormalization]\n");
    if(first) fprintf(fp,   "batch=%d\n"
            "height=%d\n"
            "width=%d\n"
            "channels=%d\n",
            l->batch,l->h, l->w, l->c);
    fprintf(fp, "size=%d\n"
                "alpha=%g\n"
                "beta=%g\n"
                "kappa=%g\n\n", l->size, l->alpha, l->beta, l->kappa);
}

void print_softmax_cfg(FILE *fp, softmax_layer *l, int first)
{
    fprintf(fp, "[softmax]\n");
    if(first) fprintf(fp, "batch=%d\ninput=%d\n", l->batch, l->inputs);
    fprintf(fp, "\n");
}

void save_network(network net, char *filename)
{
    FILE *fp = fopen(filename, "w");
    if(!fp) file_error(filename);
    int i;
    for(i = 0; i < net.n; ++i)
    {
        if(net.types[i] == CONVOLUTIONAL)
            print_convolutional_cfg(fp, (convolutional_layer *)net.layers[i], i==0);
        else if(net.types[i] == CONNECTED)
            print_connected_cfg(fp, (connected_layer *)net.layers[i], i==0);
        else if(net.types[i] == MAXPOOL)
            print_maxpool_cfg(fp, (maxpool_layer *)net.layers[i], i==0);
        else if(net.types[i] == NORMALIZATION)
            print_normalization_cfg(fp, (normalization_layer *)net.layers[i], i==0);
        else if(net.types[i] == SOFTMAX)
            print_softmax_cfg(fp, (softmax_layer *)net.layers[i], i==0);
    }
    fclose(fp);
}

#ifdef GPU
void forward_network(network net, float *input, int train)
void forward_network_gpu(network net, cl_mem input_cl, int train)
{
    cl_setup();
    size_t size = get_network_input_size(net);
    if(!net.input_cl){
        net.input_cl = clCreateBuffer(cl.context,
            CL_MEM_READ_WRITE, size*sizeof(float), 0, &cl.error);
        check_error(cl);
    }
    cl_write_array(net.input_cl, input, size);
    cl_mem input_cl = net.input_cl;
    int i;
    for(i = 0; i < net.n; ++i){
        if(net.types[i] == CONVOLUTIONAL){
            convolutional_layer layer = *(convolutional_layer *)net.layers[i];
            forward_convolutional_layer_gpu(layer, input_cl);
            input_cl = layer.output_cl;
            input = layer.output;
        }
        /*
        else if(net.types[i] == CONNECTED){
            connected_layer layer = *(connected_layer *)net.layers[i];
            forward_connected_layer(layer, input, train);
@@ -143,6 +48,11 @@
            forward_softmax_layer(layer, input);
            input = layer.output;
        }
        else if(net.types[i] == CROP){
            crop_layer layer = *(crop_layer *)net.layers[i];
            forward_crop_layer(layer, input);
            input = layer.output;
        }
        else if(net.types[i] == MAXPOOL){
            maxpool_layer layer = *(maxpool_layer *)net.layers[i];
            forward_maxpool_layer(layer, input);
@@ -153,10 +63,11 @@
            forward_normalization_layer(layer, input);
            input = layer.output;
        }
        */
    }
}

#else
#endif

void forward_network(network net, float *input, int train)
{
@@ -169,7 +80,12 @@
        }
        else if(net.types[i] == CONNECTED){
            connected_layer layer = *(connected_layer *)net.layers[i];
            forward_connected_layer(layer, input, train);
            forward_connected_layer(layer, input);
            input = layer.output;
        }
        else if(net.types[i] == CROP){
            crop_layer layer = *(crop_layer *)net.layers[i];
            forward_crop_layer(layer, input);
            input = layer.output;
        }
        else if(net.types[i] == SOFTMAX){
@@ -187,17 +103,21 @@
            forward_normalization_layer(layer, input);
            input = layer.output;
        }
        else if(net.types[i] == DROPOUT){
            if(!train) continue;
            dropout_layer layer = *(dropout_layer *)net.layers[i];
            forward_dropout_layer(layer, input);
        }
    }
}
#endif

void update_network(network net, float step, float momentum, float decay)
void update_network(network net)
{
    int i;
    for(i = 0; i < net.n; ++i){
        if(net.types[i] == CONVOLUTIONAL){
            convolutional_layer layer = *(convolutional_layer *)net.layers[i];
            update_convolutional_layer(layer, step, momentum, decay);
            update_convolutional_layer(layer);
        }
        else if(net.types[i] == MAXPOOL){
            //maxpool_layer layer = *(maxpool_layer *)net.layers[i];
@@ -210,7 +130,7 @@
        }
        else if(net.types[i] == CONNECTED){
            connected_layer layer = *(connected_layer *)net.layers[i];
            update_connected_layer(layer, step, momentum, decay);
            update_connected_layer(layer);
        }
    }
}
@@ -226,6 +146,8 @@
    } else if(net.types[i] == SOFTMAX){
        softmax_layer layer = *(softmax_layer *)net.layers[i];
        return layer.output;
    } else if(net.types[i] == DROPOUT){
        return get_network_output_layer(net, i-1);
    } else if(net.types[i] == CONNECTED){
        connected_layer layer = *(connected_layer *)net.layers[i];
        return layer.output;
@@ -251,6 +173,8 @@
    } else if(net.types[i] == SOFTMAX){
        softmax_layer layer = *(softmax_layer *)net.layers[i];
        return layer.delta;
    } else if(net.types[i] == DROPOUT){
        return get_network_delta_layer(net, i-1);
    } else if(net.types[i] == CONNECTED){
        connected_layer layer = *(connected_layer *)net.layers[i];
        return layer.delta;
@@ -326,63 +250,34 @@
    return error;
}

float train_network_datum(network net, float *x, float *y, float step, float momentum, float decay)
float train_network_datum(network net, float *x, float *y)
{
    forward_network(net, x, 1);
    //int class = get_predicted_class_network(net);
    float error = backward_network(net, x, y);
    update_network(net, step, momentum, decay);
    update_network(net);
    //return (y[class]?1:0);
    return error;
}

float train_network_sgd(network net, data d, int n, float step, float momentum,float decay)
float train_network_sgd(network net, data d, int n)
{
    int batch = net.batch;
    float *X = calloc(batch*d.X.cols, sizeof(float));
    float *y = calloc(batch*d.y.cols, sizeof(float));

    int i,j;
    int i;
    float sum = 0;
    for(i = 0; i < n; ++i){
        for(j = 0; j < batch; ++j){
            int index = rand()%d.X.rows;
            memcpy(X+j*d.X.cols, d.X.vals[index], d.X.cols*sizeof(float));
            memcpy(y+j*d.y.cols, d.y.vals[index], d.y.cols*sizeof(float));
        }
        float err = train_network_datum(net, X, y, step, momentum, decay);
        get_batch(d, batch, X, y);
        float err = train_network_datum(net, X, y);
        sum += err;
        //train_network_datum(net, X, y, step, momentum, decay);
        /*
        float *y = d.y.vals[index];
        int class = get_predicted_class_network(net);
        correct += (y[class]?1:0);
        */

/*
        for(j = 0; j < d.y.cols*batch; ++j){
            printf("%6.3f ", y[j]);
        }
        printf("\n");
        for(j = 0; j < d.y.cols*batch; ++j){
            printf("%6.3f ", get_network_output(net)[j]);
        }
        printf("\n");
        printf("\n");
        */


        //printf("%d %f %f\n", i,net.output[0], d.y.vals[index][0]);
        //if((i+1)%10 == 0){
        //    printf("%d: %f\n", (i+1), (float)correct/(i+1));
        //}
    }
    //printf("Accuracy: %f\n",(float) correct/n);
    free(X);
    free(y);
    return (float)sum/(n*batch);
}
float train_network_batch(network net, data d, int n, float step, float momentum,float decay)
float train_network_batch(network net, data d, int n)
{
    int i,j;
    float sum = 0;
@@ -395,18 +290,18 @@
            forward_network(net, x, 1);
            sum += backward_network(net, x, y);
        }
        update_network(net, step, momentum, decay);
        update_network(net);
    }
    return (float)sum/(n*batch);
}


void train_network(network net, data d, float step, float momentum, float decay)
void train_network(network net, data d)
{
    int i;
    int correct = 0;
    for(i = 0; i < d.X.rows; ++i){
        correct += train_network_datum(net, d.X.vals[i], d.y.vals[i], step, momentum, decay);
        correct += train_network_datum(net, d.X.vals[i], d.y.vals[i]);
        if(i%100 == 0){
            visualize_network(net);
            cvWaitKey(10);
@@ -430,6 +325,9 @@
    else if(net.types[i] == CONNECTED){
        connected_layer layer = *(connected_layer *)net.layers[i];
        return layer.inputs;
    } else if(net.types[i] == DROPOUT){
        dropout_layer layer = *(dropout_layer *) net.layers[i];
        return layer.inputs;
    }
    else if(net.types[i] == SOFTMAX){
        softmax_layer layer = *(softmax_layer *)net.layers[i];
@@ -453,6 +351,9 @@
    else if(net.types[i] == CONNECTED){
        connected_layer layer = *(connected_layer *)net.layers[i];
        return layer.outputs;
    } else if(net.types[i] == DROPOUT){
        dropout_layer layer = *(dropout_layer *) net.layers[i];
        return layer.inputs;
    }
    else if(net.types[i] == SOFTMAX){
        softmax_layer layer = *(softmax_layer *)net.layers[i];
@@ -518,6 +419,10 @@
        normalization_layer layer = *(normalization_layer *)net.layers[i];
        return get_normalization_image(layer);
    }
    else if(net.types[i] == CROP){
        crop_layer layer = *(crop_layer *)net.layers[i];
        return get_crop_image(layer);
    }
    return make_empty_image(0,0,0);
}

@@ -536,6 +441,7 @@
    image *prev = 0;
    int i;
    char buff[256];
    show_image(get_network_image_layer(net, 0), "Crop");
    for(i = 0; i < net.n; ++i){
        sprintf(buff, "Layer %d", i);
        if(net.types[i] == CONVOLUTIONAL){
@@ -556,6 +462,31 @@
    return out;
}

matrix network_predict_data_multi(network net, data test, int n)
{
    int i,j,b,m;
    int k = get_network_output_size(net);
    matrix pred = make_matrix(test.X.rows, k);
    float *X = calloc(net.batch*test.X.rows, sizeof(float));
    for(i = 0; i < test.X.rows; i += net.batch){
        for(b = 0; b < net.batch; ++b){
            if(i+b == test.X.rows) break;
            memcpy(X+b*test.X.cols, test.X.vals[i+b], test.X.cols*sizeof(float));
        }
        for(m = 0; m < n; ++m){
            float *out = network_predict(net, X);
            for(b = 0; b < net.batch; ++b){
                if(i+b == test.X.rows) break;
                for(j = 0; j < k; ++j){
                    pred.vals[i+b][j] += out[j+b*k]/n;
                }
            }
        }
    }
    free(X);
    return pred;   
}

matrix network_predict_data(network net, data test)
{
    int i,j,b;
@@ -597,6 +528,12 @@
            image m = get_maxpool_image(layer);
            n = m.h*m.w*m.c;
        }
        else if(net.types[i] == CROP){
            crop_layer layer = *(crop_layer *)net.layers[i];
            output = layer.output;
            image m = get_crop_image(layer);
            n = m.h*m.w*m.c;
        }
        else if(net.types[i] == CONNECTED){
            connected_layer layer = *(connected_layer *)net.layers[i];
            output = layer.output;
@@ -625,4 +562,12 @@
    return acc;
}

float network_accuracy_multi(network net, data d, int n)
{
    matrix guess = network_predict_data_multi(net, d, n);
    float acc = matrix_accuracy(d.y, guess);
    free_matrix(guess);
    return acc;
}