dropout probably ok

Joseph Redmon

2015-03-28 cf0300ea55538d4ca139d68cd24b0ee452cce015

 src/network.c

@@ -4,105 +4,121 @@
#include "image.h"
#include "data.h"
#include "utils.h"
#include "params.h"

#include "crop_layer.h"
#include "connected_layer.h"
#include "convolutional_layer.h"
#include "deconvolutional_layer.h"
#include "detection_layer.h"
#include "maxpool_layer.h"
#include "cost_layer.h"
#include "normalization_layer.h"
#include "freeweight_layer.h"
#include "softmax_layer.h"
#include "dropout_layer.h"

network make_network(int n, int batch)
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 DETECTION:
            return "detection";
        case NORMALIZATION:
            return "normalization";
        case DROPOUT:
            return "dropout";
        case CROP:
            return "crop";
        case COST:
            return "cost";
        default:
            break;
    }
    return "none";
}

network make_network(int n)
{
    network net;
    net.n = n;
    net.batch = batch;
    net.layers = calloc(net.n, sizeof(void *));
    net.types = calloc(net.n, sizeof(LAYER_TYPE));
    net.outputs = 0;
    net.output = 0;
    net.seen = 0;
    net.batch = 0;
    net.inputs = 0;
    net.h = net.w = net.c = 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)
void forward_network(network net, network_state state)
{
    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(layer, input);
            input = layer.output;
            forward_convolutional_layer(*(convolutional_layer *)net.layers[i], state);
        }
        else if(net.types[i] == DECONVOLUTIONAL){
            forward_deconvolutional_layer(*(deconvolutional_layer *)net.layers[i], state);
        }
        else if(net.types[i] == DETECTION){
            forward_detection_layer(*(detection_layer *)net.layers[i], state);
        }
        else if(net.types[i] == CONNECTED){
            connected_layer layer = *(connected_layer *)net.layers[i];
            forward_connected_layer(layer, input);
            input = layer.output;
            forward_connected_layer(*(connected_layer *)net.layers[i], state);
        }
        else if(net.types[i] == CROP){
            crop_layer layer = *(crop_layer *)net.layers[i];
            forward_crop_layer(layer, input);
            input = layer.output;
            forward_crop_layer(*(crop_layer *)net.layers[i], state);
        }
        else if(net.types[i] == COST){
            cost_layer layer = *(cost_layer *)net.layers[i];
            forward_cost_layer(layer, input, truth);
            forward_cost_layer(*(cost_layer *)net.layers[i], state);
        }
        else if(net.types[i] == SOFTMAX){
            softmax_layer layer = *(softmax_layer *)net.layers[i];
            forward_softmax_layer(layer, input);
            input = layer.output;
            forward_softmax_layer(*(softmax_layer *)net.layers[i], state);
        }
        else if(net.types[i] == MAXPOOL){
            maxpool_layer layer = *(maxpool_layer *)net.layers[i];
            forward_maxpool_layer(layer, input);
            input = layer.output;
            forward_maxpool_layer(*(maxpool_layer *)net.layers[i], state);
        }
        else if(net.types[i] == NORMALIZATION){
            normalization_layer layer = *(normalization_layer *)net.layers[i];
            forward_normalization_layer(layer, input);
            input = layer.output;
            forward_normalization_layer(*(normalization_layer *)net.layers[i], state);
        }
        else if(net.types[i] == DROPOUT){
            if(!train) continue;
            dropout_layer layer = *(dropout_layer *)net.layers[i];
            forward_dropout_layer(layer, input);
            forward_dropout_layer(*(dropout_layer *)net.layers[i], state);
        }
        else if(net.types[i] == FREEWEIGHT){
            if(!train) continue;
            freeweight_layer layer = *(freeweight_layer *)net.layers[i];
            forward_freeweight_layer(layer, input);
        }
        state.input = get_network_output_layer(net, i);
    }
}

void update_network(network net)
{
    int i;
    int update_batch = net.batch*net.subdivisions;
    for(i = 0; i < net.n; ++i){
        if(net.types[i] == CONVOLUTIONAL){
            convolutional_layer layer = *(convolutional_layer *)net.layers[i];
            update_convolutional_layer(layer);
            update_convolutional_layer(layer, update_batch, net.learning_rate, net.momentum, net.decay);
        }
        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, net.learning_rate, net.momentum, net.decay);
        }
        else if(net.types[i] == CONNECTED){
            connected_layer layer = *(connected_layer *)net.layers[i];
            update_connected_layer(layer);
            update_connected_layer(layer, update_batch, net.learning_rate, net.momentum, net.decay);
        }
    }
}
@@ -110,30 +126,27 @@
float *get_network_output_layer(network net, int i)
{
    if(net.types[i] == CONVOLUTIONAL){
        convolutional_layer layer = *(convolutional_layer *)net.layers[i];
        return layer.output;
        return ((convolutional_layer *)net.layers[i]) -> output;
    } else if(net.types[i] == DECONVOLUTIONAL){
        return ((deconvolutional_layer *)net.layers[i]) -> output;
    } else if(net.types[i] == MAXPOOL){
        maxpool_layer layer = *(maxpool_layer *)net.layers[i];
        return layer.output;
        return ((maxpool_layer *)net.layers[i]) -> output;
    } else if(net.types[i] == DETECTION){
        return ((detection_layer *)net.layers[i]) -> output;
    } else if(net.types[i] == SOFTMAX){
        softmax_layer layer = *(softmax_layer *)net.layers[i];
        return layer.output;
        return ((softmax_layer *)net.layers[i]) -> output;
    } else if(net.types[i] == DROPOUT){
        return get_network_output_layer(net, i-1);
    } 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;
        return ((connected_layer *)net.layers[i]) -> output;
    } else if(net.types[i] == CROP){
        crop_layer layer = *(crop_layer *)net.layers[i];
        return layer.output;
        return ((crop_layer *)net.layers[i]) -> output;
    } else if(net.types[i] == NORMALIZATION){
        normalization_layer layer = *(normalization_layer *)net.layers[i];
        return layer.output;
        return ((normalization_layer *)net.layers[i]) -> output;
    }
    return 0;
}

float *get_network_output(network net)
{
    int i;
@@ -146,15 +159,20 @@
    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;
    } else if(net.types[i] == SOFTMAX){
        softmax_layer layer = *(softmax_layer *)net.layers[i];
        return layer.delta;
    } else if(net.types[i] == DETECTION){
        detection_layer layer = *(detection_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] == FREEWEIGHT){
        if(i == 0) return 0;
        return get_network_delta_layer(net, i-1);
    } else if(net.types[i] == CONNECTED){
        connected_layer layer = *(connected_layer *)net.layers[i];
@@ -176,24 +194,6 @@
    return get_network_delta_layer(net, net.n-1);
}

float calculate_error_network(network net, float *truth)
{
    float sum = 0;
    float *delta = get_network_delta(net);
    float *out = get_network_output(net);
    int i;
    for(i = 0; i < get_network_output_size(net)*net.batch; ++i){
        //if(i %get_network_output_size(net) == 0) printf("\n");
        //printf("%5.2f %5.2f, ", out[i], truth[i]);
        //if(i == get_network_output_size(net)) printf("\n");
        delta[i] = truth[i] - out[i];
        //printf("%.10f, ", out[i]);
        sum += delta[i]*delta[i];
    }
    //printf("\n");
    return sum;
}

int get_predicted_class_network(network net)
{
    float *out = get_network_output(net);
@@ -201,46 +201,53 @@
    return max_index(out, k);
}

void backward_network(network net, float *input)
void backward_network(network net, network_state state)
{
    int i;
    float *prev_input;
    float *prev_delta;
    float *original_input = state.input;
    for(i = net.n-1; i >= 0; --i){
        if(i == 0){
            prev_input = input;
            prev_delta = 0;
            state.input = original_input;
            state.delta = 0;
        }else{
            prev_input = get_network_output_layer(net, i-1);
            prev_delta = get_network_delta_layer(net, i-1);
            state.input = get_network_output_layer(net, i-1);
            state.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);
            backward_convolutional_layer(layer, state);
        } else if(net.types[i] == DECONVOLUTIONAL){
            deconvolutional_layer layer = *(deconvolutional_layer *)net.layers[i];
            backward_deconvolutional_layer(layer, state);
        }
        else if(net.types[i] == MAXPOOL){
            maxpool_layer layer = *(maxpool_layer *)net.layers[i];
            if(i != 0) backward_maxpool_layer(layer, prev_delta);
            if(i != 0) backward_maxpool_layer(layer, state);
        }
        else if(net.types[i] == DROPOUT){
            dropout_layer layer = *(dropout_layer *)net.layers[i];
            backward_dropout_layer(layer, prev_delta);
            backward_dropout_layer(layer, state);
        }
        else if(net.types[i] == DETECTION){
            detection_layer layer = *(detection_layer *)net.layers[i];
            backward_detection_layer(layer, state);
        }
        else if(net.types[i] == NORMALIZATION){
            normalization_layer layer = *(normalization_layer *)net.layers[i];
            if(i != 0) backward_normalization_layer(layer, prev_input, prev_delta);
            if(i != 0) backward_normalization_layer(layer, state);
        }
        else if(net.types[i] == SOFTMAX){
            softmax_layer layer = *(softmax_layer *)net.layers[i];
            if(i != 0) backward_softmax_layer(layer, prev_delta);
            if(i != 0) backward_softmax_layer(layer, state);
        }
        else if(net.types[i] == CONNECTED){
            connected_layer layer = *(connected_layer *)net.layers[i];
            backward_connected_layer(layer, prev_input, prev_delta);
            backward_connected_layer(layer, state);
        }
        else if(net.types[i] == COST){
            cost_layer layer = *(cost_layer *)net.layers[i];
            backward_cost_layer(layer, prev_input, prev_delta);
            backward_cost_layer(layer, state);
        }
    }
}
@@ -250,10 +257,14 @@
    #ifdef GPU
    if(gpu_index >= 0) return train_network_datum_gpu(net, x, y);
    #endif
    forward_network(net, x, y, 1);
    backward_network(net, x);
    network_state state;
    state.input = x;
    state.truth = y;
    state.train = 1;
    forward_network(net, state);
    backward_network(net, state);
    float error = get_network_cost(net);
    update_network(net);
    if((net.seen/net.batch)%net.subdivisions == 0) update_network(net);
    return error;
}

@@ -266,6 +277,7 @@
    int i;
    float sum = 0;
    for(i = 0; i < n; ++i){
        net.seen += batch;
        get_random_batch(d, batch, X, y);
        float err = train_network_datum(net, X, y);
        sum += err;
@@ -286,6 +298,7 @@
    float sum = 0;
    for(i = 0; i < n; ++i){
        get_next_batch(d, batch, i*batch, X, y);
        net.seen += batch;
        float err = train_network_datum(net, X, y);
        sum += err;
    }
@@ -297,15 +310,17 @@
float train_network_batch(network net, data d, int n)
{
    int i,j;
    network_state state;
    state.train = 1;
    float sum = 0;
    int batch = 2;
    for(i = 0; i < n; ++i){
        for(j = 0; j < batch; ++j){
            int index = rand()%d.X.rows;
            float *x = d.X.vals[index];
            float *y = d.y.vals[index];
            forward_network(net, x, y, 1);
            backward_network(net, x);
            state.input = d.X.vals[index];
            state.truth = d.y.vals[index];
            forward_network(net, state);
            backward_network(net, state);
            sum += get_network_cost(net);
        }
        update_network(net);
@@ -313,29 +328,6 @@
    return (float)sum/(n*batch);
}

void set_learning_network(network *net, float rate, float momentum, float decay)
{
    int i;
    net->learning_rate=rate;
    net->momentum = momentum;
    net->decay = decay;
    for(i = 0; i < net->n; ++i){
        if(net->types[i] == CONVOLUTIONAL){
            convolutional_layer *layer = (convolutional_layer *)net->layers[i];
            layer->learning_rate=rate;
            layer->momentum = momentum;
            layer->decay = decay;
        }
        else if(net->types[i] == CONNECTED){
            connected_layer *layer = (connected_layer *)net->layers[i];
            layer->learning_rate=rate;
            layer->momentum = momentum;
            layer->decay = decay;
        }
    }
}


void set_batch_network(network *net, int b)
{
    net->batch = b;
@@ -344,6 +336,9 @@
        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];
@@ -355,9 +350,8 @@
        } else if(net->types[i] == DROPOUT){
            dropout_layer *layer = (dropout_layer *) net->layers[i];
            layer->batch = b;
        }
        else if(net->types[i] == FREEWEIGHT){
            freeweight_layer *layer = (freeweight_layer *) net->layers[i];
        } else if(net->types[i] == DETECTION){
            detection_layer *layer = (detection_layer *) net->layers[i];
            layer->batch = b;
        }
        else if(net->types[i] == SOFTMAX){
@@ -368,6 +362,10 @@
            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;
        }
    }
}

@@ -378,6 +376,10 @@
        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;
@@ -388,19 +390,18 @@
    } else if(net.types[i] == DROPOUT){
        dropout_layer layer = *(dropout_layer *) net.layers[i];
        return layer.inputs;
    } else if(net.types[i] == DETECTION){
        detection_layer layer = *(detection_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];
        return layer.inputs;
    }
    else if(net.types[i] == SOFTMAX){
        softmax_layer layer = *(softmax_layer *)net.layers[i];
        return layer.inputs;
    }
    printf("Can't find input size\n");
    fprintf(stderr, "Can't find input size\n");
    return 0;
}

@@ -411,12 +412,21 @@
        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] == DETECTION){
        detection_layer layer = *(detection_layer *)net.layers[i];
        return get_detection_layer_output_size(layer);
    }
    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){
    else if(net.types[i] == CROP){
        crop_layer layer = *(crop_layer *) net.layers[i];
        return layer.c*layer.crop_height*layer.crop_width;
    }
@@ -428,15 +438,11 @@
        dropout_layer layer = *(dropout_layer *) net.layers[i];
        return layer.inputs;
    }
    else if(net.types[i] == FREEWEIGHT){
        freeweight_layer layer = *(freeweight_layer *) net.layers[i];
        return layer.inputs;
    }
    else if(net.types[i] == SOFTMAX){
        softmax_layer layer = *(softmax_layer *)net.layers[i];
        return layer.inputs;
    }
    printf("Can't find output size\n");
    fprintf(stderr, "Can't find output size\n");
    return 0;
}

@@ -446,21 +452,31 @@
    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;
@@ -490,6 +506,10 @@
        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);
@@ -498,6 +518,9 @@
        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);
@@ -544,11 +567,16 @@

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

    forward_network(net, input, 0, 0);
    network_state state;
    state.input = input;
    state.truth = 0;
    state.train = 0;
    state.delta = 0;
    forward_network(net, state);
    float *out = get_network_output(net);
    return out;
}
@@ -645,6 +673,31 @@
    }
}

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);