Detection is back, baby\!

Joseph Redmon

2015-02-11 0f645836f193e75c4c3b718369e6fab15b5d19c5

 src/network.c

@@ -8,6 +8,7 @@
#include "crop_layer.h"
#include "connected_layer.h"
#include "convolutional_layer.h"
#include "deconvolutional_layer.h"
#include "maxpool_layer.h"
#include "cost_layer.h"
#include "normalization_layer.h"
@@ -15,6 +16,35 @@
#include "softmax_layer.h"
#include "dropout_layer.h"

char *get_layer_string(LAYER_TYPE a)
{
    switch(a){
        case CONVOLUTIONAL:
            return "convolutional";
        case DECONVOLUTIONAL:
            return "deconvolutional";
        case CONNECTED:
            return "connected";
        case MAXPOOL:
            return "maxpool";
        case SOFTMAX:
            return "softmax";
        case NORMALIZATION:
            return "normalization";
        case DROPOUT:
            return "dropout";
        case FREEWEIGHT:
            return "freeweight";
        case CROP:
            return "crop";
        case COST:
            return "cost";
        default:
            break;
    }
    return "none";
}

network make_network(int n, int batch)
{
    network net;
@@ -24,14 +54,14 @@
    net.types = calloc(net.n, sizeof(LAYER_TYPE));
    net.outputs = 0;
    net.output = 0;
    net.seen = 0;
    #ifdef GPU
    net.input_cl = calloc(1, sizeof(cl_mem));
    net.truth_cl = calloc(1, sizeof(cl_mem));
    net.input_gpu = calloc(1, sizeof(float *));
    net.truth_gpu = calloc(1, sizeof(float *));
    #endif
    return net;
}


void forward_network(network net, float *input, float *truth, int train)
{
    int i;
@@ -41,6 +71,11 @@
            forward_convolutional_layer(layer, input);
            input = layer.output;
        }
        else if(net.types[i] == DECONVOLUTIONAL){
            deconvolutional_layer layer = *(deconvolutional_layer *)net.layers[i];
            forward_deconvolutional_layer(layer, input);
            input = layer.output;
        }
        else if(net.types[i] == CONNECTED){
            connected_layer layer = *(connected_layer *)net.layers[i];
            forward_connected_layer(layer, input);
@@ -48,7 +83,7 @@
        }
        else if(net.types[i] == CROP){
            crop_layer layer = *(crop_layer *)net.layers[i];
            forward_crop_layer(layer, input);
            forward_crop_layer(layer, train, input);
            input = layer.output;
        }
        else if(net.types[i] == COST){
@@ -74,12 +109,16 @@
            if(!train) continue;
            dropout_layer layer = *(dropout_layer *)net.layers[i];
            forward_dropout_layer(layer, input);
            input = layer.output;
        }
        else if(net.types[i] == FREEWEIGHT){
            if(!train) continue;
            freeweight_layer layer = *(freeweight_layer *)net.layers[i];
            forward_freeweight_layer(layer, input);
            //freeweight_layer layer = *(freeweight_layer *)net.layers[i];
            //forward_freeweight_layer(layer, input);
        }
        //char buff[256];
        //sprintf(buff, "layer %d", i);
        //cuda_compare(get_network_output_gpu_layer(net, i), input, get_network_output_size_layer(net, i)*net.batch, buff);
    }
}

@@ -91,14 +130,9 @@
            convolutional_layer layer = *(convolutional_layer *)net.layers[i];
            update_convolutional_layer(layer);
        }
        else if(net.types[i] == MAXPOOL){
            //maxpool_layer layer = *(maxpool_layer *)net.layers[i];
        }
        else if(net.types[i] == SOFTMAX){
            //maxpool_layer layer = *(maxpool_layer *)net.layers[i];
        }
        else if(net.types[i] == NORMALIZATION){
            //maxpool_layer layer = *(maxpool_layer *)net.layers[i];
        else if(net.types[i] == DECONVOLUTIONAL){
            deconvolutional_layer layer = *(deconvolutional_layer *)net.layers[i];
            update_deconvolutional_layer(layer);
        }
        else if(net.types[i] == CONNECTED){
            connected_layer layer = *(connected_layer *)net.layers[i];
@@ -112,6 +146,9 @@
    if(net.types[i] == CONVOLUTIONAL){
        convolutional_layer layer = *(convolutional_layer *)net.layers[i];
        return layer.output;
    } else if(net.types[i] == DECONVOLUTIONAL){
        deconvolutional_layer layer = *(deconvolutional_layer *)net.layers[i];
        return layer.output;
    } else if(net.types[i] == MAXPOOL){
        maxpool_layer layer = *(maxpool_layer *)net.layers[i];
        return layer.output;
@@ -119,12 +156,16 @@
        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);
        dropout_layer layer = *(dropout_layer *)net.layers[i];
        return layer.output;
    } else if(net.types[i] == FREEWEIGHT){
        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;
    } else if(net.types[i] == CROP){
        crop_layer layer = *(crop_layer *)net.layers[i];
        return layer.output;
    } else if(net.types[i] == NORMALIZATION){
        normalization_layer layer = *(normalization_layer *)net.layers[i];
        return layer.output;
@@ -143,6 +184,9 @@
    if(net.types[i] == CONVOLUTIONAL){
        convolutional_layer layer = *(convolutional_layer *)net.layers[i];
        return layer.delta;
    } else if(net.types[i] == DECONVOLUTIONAL){
        deconvolutional_layer layer = *(deconvolutional_layer *)net.layers[i];
        return layer.delta;
    } else if(net.types[i] == MAXPOOL){
        maxpool_layer layer = *(maxpool_layer *)net.layers[i];
        return layer.delta;
@@ -150,6 +194,7 @@
        softmax_layer layer = *(softmax_layer *)net.layers[i];
        return layer.delta;
    } else if(net.types[i] == DROPOUT){
        if(i == 0) return 0;
        return get_network_delta_layer(net, i-1);
    } else if(net.types[i] == FREEWEIGHT){
        return get_network_delta_layer(net, i-1);
@@ -211,9 +256,13 @@
            prev_input = get_network_output_layer(net, i-1);
            prev_delta = get_network_delta_layer(net, i-1);
        }

        if(net.types[i] == CONVOLUTIONAL){
            convolutional_layer layer = *(convolutional_layer *)net.layers[i];
            backward_convolutional_layer(layer, prev_input, prev_delta);
        } else if(net.types[i] == DECONVOLUTIONAL){
            deconvolutional_layer layer = *(deconvolutional_layer *)net.layers[i];
            backward_deconvolutional_layer(layer, prev_input, prev_delta);
        }
        else if(net.types[i] == MAXPOOL){
            maxpool_layer layer = *(maxpool_layer *)net.layers[i];
@@ -242,17 +291,15 @@
    }
}




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
    forward_network(net, x, y, 1);
    //int class = get_predicted_class_network(net);
    backward_network(net, x);
    float error = get_network_cost(net);
    update_network(net);
    //return (y[class]?1:0);
    return error;
}

@@ -274,6 +321,25 @@
    return (float)sum/(n*batch);
}

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));

    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);
}

float train_network_batch(network net, data d, int n)
{
    int i,j;
@@ -293,40 +359,6 @@
    return (float)sum/(n*batch);
}

float train_network_data_cpu(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;
    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);
}

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]);
        if(i%100 == 0){
            visualize_network(net);
            cvWaitKey(10);
        }
    }
    visualize_network(net);
    cvWaitKey(100);
    fprintf(stderr, "Accuracy: %f\n", (float)correct/d.X.rows);
}

void set_learning_network(network *net, float rate, float momentum, float decay)
{
    int i;
@@ -358,6 +390,9 @@
        if(net->types[i] == CONVOLUTIONAL){
            convolutional_layer *layer = (convolutional_layer *)net->layers[i];
            layer->batch = b;
        }else if(net->types[i] == DECONVOLUTIONAL){
            deconvolutional_layer *layer = (deconvolutional_layer *)net->layers[i];
            layer->batch = b;
        }
        else if(net->types[i] == MAXPOOL){
            maxpool_layer *layer = (maxpool_layer *)net->layers[i];
@@ -382,6 +417,10 @@
            cost_layer *layer = (cost_layer *)net->layers[i];
            layer->batch = b;
        }
        else if(net->types[i] == CROP){
            crop_layer *layer = (crop_layer *)net->layers[i];
            layer->batch = b;
        }
    }
}

@@ -392,6 +431,10 @@
        convolutional_layer layer = *(convolutional_layer *)net.layers[i];
        return layer.h*layer.w*layer.c;
    }
    if(net.types[i] == DECONVOLUTIONAL){
        deconvolutional_layer layer = *(deconvolutional_layer *)net.layers[i];
        return layer.h*layer.w*layer.c;
    }
    else if(net.types[i] == MAXPOOL){
        maxpool_layer layer = *(maxpool_layer *)net.layers[i];
        return layer.h*layer.w*layer.c;
@@ -402,6 +445,9 @@
    } else if(net.types[i] == DROPOUT){
        dropout_layer layer = *(dropout_layer *) net.layers[i];
        return layer.inputs;
    } else if(net.types[i] == CROP){
        crop_layer layer = *(crop_layer *) net.layers[i];
        return layer.c*layer.h*layer.w;
    }
    else if(net.types[i] == FREEWEIGHT){
        freeweight_layer layer = *(freeweight_layer *) net.layers[i];
@@ -411,6 +457,7 @@
        softmax_layer layer = *(softmax_layer *)net.layers[i];
        return layer.inputs;
    }
    printf("Can't find input size\n");
    return 0;
}

@@ -421,11 +468,20 @@
        image output = get_convolutional_image(layer);
        return output.h*output.w*output.c;
    }
    else if(net.types[i] == DECONVOLUTIONAL){
        deconvolutional_layer layer = *(deconvolutional_layer *)net.layers[i];
        image output = get_deconvolutional_image(layer);
        return output.h*output.w*output.c;
    }
    else if(net.types[i] == MAXPOOL){
        maxpool_layer layer = *(maxpool_layer *)net.layers[i];
        image output = get_maxpool_image(layer);
        return output.h*output.w*output.c;
    }
     else if(net.types[i] == CROP){
        crop_layer layer = *(crop_layer *) net.layers[i];
        return layer.c*layer.crop_height*layer.crop_width;
    }
    else if(net.types[i] == CONNECTED){
        connected_layer layer = *(connected_layer *)net.layers[i];
        return layer.outputs;
@@ -442,6 +498,7 @@
        softmax_layer layer = *(softmax_layer *)net.layers[i];
        return layer.inputs;
    }
    printf("Can't find output size\n");
    return 0;
}

@@ -451,21 +508,31 @@
    for (i = 0; i < net.n; ++i){
        if(net.types[i] == CONVOLUTIONAL){
            convolutional_layer *layer = (convolutional_layer *)net.layers[i];
            resize_convolutional_layer(layer, h, w, c);
            resize_convolutional_layer(layer, h, w);
            image output = get_convolutional_image(*layer);
            h = output.h;
            w = output.w;
            c = output.c;
        } else if(net.types[i] == DECONVOLUTIONAL){
            deconvolutional_layer *layer = (deconvolutional_layer *)net.layers[i];
            resize_deconvolutional_layer(layer, h, w);
            image output = get_deconvolutional_image(*layer);
            h = output.h;
            w = output.w;
            c = output.c;
        }else if(net.types[i] == MAXPOOL){
            maxpool_layer *layer = (maxpool_layer *)net.layers[i];
            resize_maxpool_layer(layer, h, w, c);
            resize_maxpool_layer(layer, h, w);
            image output = get_maxpool_image(*layer);
            h = output.h;
            w = output.w;
            c = output.c;
        }else if(net.types[i] == DROPOUT){
            dropout_layer *layer = (dropout_layer *)net.layers[i];
            resize_dropout_layer(layer, h*w*c);
        }else if(net.types[i] == NORMALIZATION){
            normalization_layer *layer = (normalization_layer *)net.layers[i];
            resize_normalization_layer(layer, h, w, c);
            resize_normalization_layer(layer, h, w);
            image output = get_normalization_image(*layer);
            h = output.h;
            w = output.w;
@@ -495,6 +562,10 @@
        convolutional_layer layer = *(convolutional_layer *)net.layers[i];
        return get_convolutional_image(layer);
    }
    else if(net.types[i] == DECONVOLUTIONAL){
        deconvolutional_layer layer = *(deconvolutional_layer *)net.layers[i];
        return get_deconvolutional_image(layer);
    }
    else if(net.types[i] == MAXPOOL){
        maxpool_layer layer = *(maxpool_layer *)net.layers[i];
        return get_maxpool_image(layer);
@@ -503,6 +574,9 @@
        normalization_layer layer = *(normalization_layer *)net.layers[i];
        return get_normalization_image(layer);
    }
    else if(net.types[i] == DROPOUT){
        return get_network_image_layer(net, i-1);
    }
    else if(net.types[i] == CROP){
        crop_layer layer = *(crop_layer *)net.layers[i];
        return get_crop_image(layer);
@@ -549,6 +623,10 @@

float *network_predict(network net, float *input)
{
    #ifdef GPU
    if(gpu_index >= 0)  return network_predict_gpu(net, input);
    #endif

    forward_network(net, input, 0, 0);
    float *out = get_network_output(net);
    return out;
@@ -646,6 +724,31 @@
    }
}

void compare_networks(network n1, network n2, data test)
{
    matrix g1 = network_predict_data(n1, test);
    matrix g2 = network_predict_data(n2, test);
    int i;
    int a,b,c,d;
    a = b = c = d = 0;
    for(i = 0; i < g1.rows; ++i){
        int truth = max_index(test.y.vals[i], test.y.cols);
        int p1 = max_index(g1.vals[i], g1.cols);
        int p2 = max_index(g2.vals[i], g2.cols);
        if(p1 == truth){
            if(p2 == truth) ++d;
            else ++c;
        }else{
            if(p2 == truth) ++b;
            else ++a;
        }
    }
    printf("%5d %5d\n%5d %5d\n", a, b, c, d);
    float num = pow((abs(b - c) - 1.), 2.);
    float den = b + c;
    printf("%f\n", num/den); 
}

float network_accuracy(network net, data d)
{
    matrix guess = network_predict_data(net, d);