Hacks to get nightmare to not break gridsizing

Joseph Redmon

2015-08-11 d0b9326a352ed2fbc3ae66fdef40b4533a2f211d

 src/network.c

@@ -4,18 +4,20 @@
#include "image.h"
#include "data.h"
#include "utils.h"
#include "params.h"
#include "blas.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 "maxpool_layer.h"
#include "avgpool_layer.h"
#include "cost_layer.h"
#include "softmax_layer.h"
#include "dropout_layer.h"
#include "route_layer.h"

char *get_layer_string(LAYER_TYPE a)
{
@@ -28,18 +30,22 @@
            return "connected";
        case MAXPOOL:
            return "maxpool";
        case AVGPOOL:
            return "avgpool";
        case SOFTMAX:
            return "softmax";
        case DETECTION:
            return "detection";
        case NORMALIZATION:
            return "normalization";
        case DROPOUT:
            return "dropout";
        case CROP:
            return "crop";
        case COST:
            return "cost";
        case ROUTE:
            return "route";
        case NORMALIZATION:
            return "normalization";
        default:
            break;
    }
@@ -48,16 +54,9 @@

network make_network(int n)
{
    network net;
    network net = {0};
    net.n = n;
    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;
    net.layers = calloc(net.n, sizeof(layer));
    #ifdef GPU
    net.input_gpu = calloc(1, sizeof(float *));
    net.truth_gpu = calloc(1, sizeof(float *));
@@ -69,37 +68,36 @@
{
    int i;
    for(i = 0; i < net.n; ++i){
        if(net.types[i] == CONVOLUTIONAL){
            forward_convolutional_layer(*(convolutional_layer *)net.layers[i], state);
        layer l = net.layers[i];
        if(l.delta){
            scal_cpu(l.outputs * l.batch, 0, l.delta, 1);
        }
        else if(net.types[i] == DECONVOLUTIONAL){
            forward_deconvolutional_layer(*(deconvolutional_layer *)net.layers[i], state);
        if(l.type == CONVOLUTIONAL){
            forward_convolutional_layer(l, state);
        } else if(l.type == DECONVOLUTIONAL){
            forward_deconvolutional_layer(l, state);
        } else if(l.type == NORMALIZATION){
            forward_normalization_layer(l, state);
        } else if(l.type == DETECTION){
            forward_detection_layer(l, state);
        } else if(l.type == CONNECTED){
            forward_connected_layer(l, state);
        } else if(l.type == CROP){
            forward_crop_layer(l, state);
        } else if(l.type == COST){
            forward_cost_layer(l, state);
        } else if(l.type == SOFTMAX){
            forward_softmax_layer(l, state);
        } else if(l.type == MAXPOOL){
            forward_maxpool_layer(l, state);
        } else if(l.type == AVGPOOL){
            forward_avgpool_layer(l, state);
        } else if(l.type == DROPOUT){
            forward_dropout_layer(l, state);
        } else if(l.type == ROUTE){
            forward_route_layer(l, net);
        }
        else if(net.types[i] == DETECTION){
            forward_detection_layer(*(detection_layer *)net.layers[i], state);
        }
        else if(net.types[i] == CONNECTED){
            forward_connected_layer(*(connected_layer *)net.layers[i], state);
        }
        else if(net.types[i] == CROP){
            forward_crop_layer(*(crop_layer *)net.layers[i], state);
        }
        else if(net.types[i] == COST){
            forward_cost_layer(*(cost_layer *)net.layers[i], state);
        }
        else if(net.types[i] == SOFTMAX){
            forward_softmax_layer(*(softmax_layer *)net.layers[i], state);
        }
        else if(net.types[i] == MAXPOOL){
            forward_maxpool_layer(*(maxpool_layer *)net.layers[i], state);
        }
        else if(net.types[i] == NORMALIZATION){
            forward_normalization_layer(*(normalization_layer *)net.layers[i], state);
        }
        else if(net.types[i] == DROPOUT){
            forward_dropout_layer(*(dropout_layer *)net.layers[i], state);
        }
        state.input = get_network_output_layer(net, i);
        state.input = l.output;
    }
}

@@ -108,108 +106,40 @@
    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_batch, net.learning_rate, net.momentum, net.decay);
        }
        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_batch, net.learning_rate, net.momentum, net.decay);
        layer l = net.layers[i];
        if(l.type == CONVOLUTIONAL){
            update_convolutional_layer(l, update_batch, net.learning_rate, net.momentum, net.decay);
        } else if(l.type == DECONVOLUTIONAL){
            update_deconvolutional_layer(l, net.learning_rate, net.momentum, net.decay);
        } else if(l.type == CONNECTED){
            update_connected_layer(l, update_batch, net.learning_rate, net.momentum, net.decay);
        }
    }
}

float *get_network_output_layer(network net, int i)
{
    if(net.types[i] == CONVOLUTIONAL){
        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){
        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){
        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] == CONNECTED){
        return ((connected_layer *)net.layers[i]) -> output;
    } else if(net.types[i] == CROP){
        return ((crop_layer *)net.layers[i]) -> output;
    } else if(net.types[i] == NORMALIZATION){
        return ((normalization_layer *)net.layers[i]) -> output;
    }
    return 0;
}

float *get_network_output(network net)
{
    int i;
    for(i = net.n-1; i > 0; --i) if(net.types[i] != COST) break;
    return get_network_output_layer(net, i);
}

float *get_network_delta_layer(network net, int i)
{
    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){
        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];
        return layer.delta;
    }
    return 0;
    for(i = net.n-1; i > 0; --i) if(net.layers[i].type != COST) break;
    return net.layers[i].output;
}

float get_network_cost(network net)
{
    if(net.types[net.n-1] == COST){
        return ((cost_layer *)net.layers[net.n-1])->output[0];
    }
    return 0;
}

float *get_network_delta(network net)
{
    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];
    float sum = 0;
    int count = 0;
    for(i = 0; i < net.n; ++i){
        if(net.layers[net.n-1].type == COST){
            sum += net.layers[net.n-1].output[0];
            ++count;
        }
        if(net.layers[net.n-1].type == DETECTION){
            sum += net.layers[net.n-1].cost[0];
            ++count;
        }
    }
    //printf("\n");
    return sum;
    return sum/count;
}

int get_predicted_class_network(network net)
@@ -223,60 +153,51 @@
{
    int i;
    float *original_input = state.input;
    float *original_delta = state.delta;
    for(i = net.n-1; i >= 0; --i){
        if(i == 0){
            state.input = original_input;
            state.delta = 0;
            state.delta = original_delta;
        }else{
            state.input = get_network_output_layer(net, i-1);
            state.delta = get_network_delta_layer(net, i-1);
            layer prev = net.layers[i-1];
            state.input = prev.output;
            state.delta = prev.delta;
        }

        if(net.types[i] == CONVOLUTIONAL){
            convolutional_layer layer = *(convolutional_layer *)net.layers[i];
            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, state);
        }
        else if(net.types[i] == DROPOUT){
            dropout_layer layer = *(dropout_layer *)net.layers[i];
            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, state);
        }
        else if(net.types[i] == SOFTMAX){
            softmax_layer layer = *(softmax_layer *)net.layers[i];
            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, state);
        }
        else if(net.types[i] == COST){
            cost_layer layer = *(cost_layer *)net.layers[i];
            backward_cost_layer(layer, state);
        layer l = net.layers[i];
        if(l.type == CONVOLUTIONAL){
            backward_convolutional_layer(l, state);
        } else if(l.type == DECONVOLUTIONAL){
            backward_deconvolutional_layer(l, state);
        } else if(l.type == NORMALIZATION){
            backward_normalization_layer(l, state);
        } else if(l.type == MAXPOOL){
            if(i != 0) backward_maxpool_layer(l, state);
        } else if(l.type == AVGPOOL){
            backward_avgpool_layer(l, state);
        } else if(l.type == DROPOUT){
            backward_dropout_layer(l, state);
        } else if(l.type == DETECTION){
            backward_detection_layer(l, state);
        } else if(l.type == SOFTMAX){
            if(i != 0) backward_softmax_layer(l, state);
        } else if(l.type == CONNECTED){
            backward_connected_layer(l, state);
        } else if(l.type == COST){
            backward_cost_layer(l, state);
        } else if(l.type == ROUTE){
            backward_route_layer(l, net);
        }
    }
}

float train_network_datum(network net, float *x, float *y)
{
    #ifdef GPU
#ifdef GPU
    if(gpu_index >= 0) return train_network_datum_gpu(net, x, y);
    #endif
#endif
    network_state state;
    state.input = x;
    state.delta = 0;
    state.truth = y;
    state.train = 1;
    forward_network(net, state);
@@ -330,6 +251,7 @@
    int i,j;
    network_state state;
    state.train = 1;
    state.delta = 0;
    float sum = 0;
    int batch = 2;
    for(i = 0; i < n; ++i){
@@ -351,199 +273,73 @@
    net->batch = b;
    int i;
    for(i = 0; i < net->n; ++i){
        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];
            layer->batch = b;
        }
        else if(net->types[i] == CONNECTED){
            connected_layer *layer = (connected_layer *)net->layers[i];
            layer->batch = b;
        } else if(net->types[i] == DROPOUT){
            dropout_layer *layer = (dropout_layer *) net->layers[i];
            layer->batch = b;
        } else if(net->types[i] == DETECTION){
            detection_layer *layer = (detection_layer *) net->layers[i];
            layer->batch = b;
        }
        else if(net->types[i] == SOFTMAX){
            softmax_layer *layer = (softmax_layer *)net->layers[i];
            layer->batch = b;
        }
        else if(net->types[i] == COST){
            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;
        }
        net->layers[i].batch = b;
    }
}


int get_network_input_size_layer(network net, int i)
{
    if(net.types[i] == CONVOLUTIONAL){
        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;
    }
    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] == 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] == SOFTMAX){
        softmax_layer layer = *(softmax_layer *)net.layers[i];
        return layer.inputs;
    }
    fprintf(stderr, "Can't find input size\n");
    return 0;
}

int get_network_output_size_layer(network net, int i)
{
    if(net.types[i] == CONVOLUTIONAL){
        convolutional_layer layer = *(convolutional_layer *)net.layers[i];
        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){
        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;
    }
    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];
        return layer.inputs;
    }
    fprintf(stderr, "Can't find output size\n");
    return 0;
}

int resize_network(network net, int h, int w, int c)
int resize_network(network *net, int w, int h)
{
    int i;
    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);
            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);
            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);
            image output = get_normalization_image(*layer);
            h = output.h;
            w = output.w;
            c = output.c;
    //if(w == net->w && h == net->h) return 0;
    net->w = w;
    net->h = h;
    //fprintf(stderr, "Resizing to %d x %d...", w, h);
    //fflush(stderr);
    for (i = 0; i < net->n; ++i){
        layer l = net->layers[i];
        if(l.type == CONVOLUTIONAL){
            resize_convolutional_layer(&l, w, h);
        }else if(l.type == MAXPOOL){
            resize_maxpool_layer(&l, w, h);
        }else if(l.type == AVGPOOL){
            resize_avgpool_layer(&l, w, h);
            break;
        }else if(l.type == NORMALIZATION){
            resize_normalization_layer(&l, w, h);
        }else{
            error("Cannot resize this type of layer");
        }
        net->layers[i] = l;
        w = l.out_w;
        h = l.out_h;
    }
    //fprintf(stderr, " Done!\n");
    return 0;
}

int get_network_output_size(network net)
{
    int i;
    for(i = net.n-1; i > 0; --i) if(net.types[i] != COST) break;
    return get_network_output_size_layer(net, i);
    for(i = net.n-1; i > 0; --i) if(net.layers[i].type != COST) break;
    return net.layers[i].outputs;
}

int get_network_input_size(network net)
{
    return get_network_input_size_layer(net, 0);
    return net.layers[0].inputs;
}

detection_layer get_network_detection_layer(network net)
{
    int i;
    for(i = 0; i < net.n; ++i){
        if(net.layers[i].type == DETECTION){
            return net.layers[i];
        }
    }
    fprintf(stderr, "Detection layer not found!!\n");
    detection_layer l = {0};
    return l;
}

image get_network_image_layer(network net, int i)
{
    if(net.types[i] == CONVOLUTIONAL){
        convolutional_layer layer = *(convolutional_layer *)net.layers[i];
        return get_convolutional_image(layer);
    layer l = net.layers[i];
    if (l.out_w && l.out_h && l.out_c){
        return float_to_image(l.out_w, l.out_h, l.out_c, l.output);
    }
    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);
    }
    else if(net.types[i] == NORMALIZATION){
        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);
    }
    return make_empty_image(0,0,0);
    image def = {0};
    return def;
}

image get_network_image(network net)
@@ -553,7 +349,8 @@
        image m = get_network_image_layer(net, i);
        if(m.h != 0) return m;
    }
    return make_empty_image(0,0,0);
    image def = {0};
    return def;
}

void visualize_network(network net)
@@ -561,16 +358,11 @@
    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){
            convolutional_layer layer = *(convolutional_layer *)net.layers[i];
            prev = visualize_convolutional_layer(layer, buff, prev);
        }
        if(net.types[i] == NORMALIZATION){
            normalization_layer layer = *(normalization_layer *)net.layers[i];
            visualize_normalization_layer(layer, buff);
        layer l = net.layers[i];
        if(l.type == CONVOLUTIONAL){
            prev = visualize_convolutional_layer(l, buff, prev);
        }
    } 
}
@@ -651,36 +443,9 @@
{
    int i,j;
    for(i = 0; i < net.n; ++i){
        float *output = 0;
        int n = 0;
        if(net.types[i] == CONVOLUTIONAL){
            convolutional_layer layer = *(convolutional_layer *)net.layers[i];
            output = layer.output;
            image m = get_convolutional_image(layer);
            n = m.h*m.w*m.c;
        }
        else if(net.types[i] == MAXPOOL){
            maxpool_layer layer = *(maxpool_layer *)net.layers[i];
            output = layer.output;
            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;
            n = layer.outputs;
        }
        else if(net.types[i] == SOFTMAX){
            softmax_layer layer = *(softmax_layer *)net.layers[i];
            output = layer.output;
            n = layer.inputs;
        }
        layer l = net.layers[i];
        float *output = l.output;
        int n = l.outputs;
        float mean = mean_array(output, n);
        float vari = variance_array(output, n);
        fprintf(stderr, "Layer %d - Mean: %f, Variance: %f\n",i,mean, vari);