stuff

Joseph Redmon

2015-05-08 0cbfa4646128206300b9a30586615c3698abfb76

stuff

 Makefile

@@ -25,7 +25,7 @@
LDFLAGS+= -L/usr/local/cuda/lib64 -lcuda -lcudart -lcublas -lcurand
endif

OBJ=gemm.o utils.o cuda.o deconvolutional_layer.o convolutional_layer.o list.o image.o activations.o im2col.o col2im.o blas.o crop_layer.o dropout_layer.o maxpool_layer.o softmax_layer.o data.o matrix.o network.o connected_layer.o cost_layer.o normalization_layer.o parser.o option_list.o darknet.o detection_layer.o imagenet.o captcha.o detection.o
OBJ=gemm.o utils.o cuda.o deconvolutional_layer.o convolutional_layer.o list.o image.o activations.o im2col.o col2im.o blas.o crop_layer.o dropout_layer.o maxpool_layer.o softmax_layer.o data.o matrix.o network.o connected_layer.o cost_layer.o normalization_layer.o parser.o option_list.o darknet.o detection_layer.o imagenet.o captcha.o detection.o route_layer.o
ifeq ($(GPU), 1) 
OBJ+=convolutional_kernels.o deconvolutional_kernels.o activation_kernels.o im2col_kernels.o col2im_kernels.o blas_kernels.o crop_layer_kernels.o dropout_layer_kernels.o maxpool_layer_kernels.o softmax_layer_kernels.o network_kernels.o
endif

 src/data.c

@@ -167,8 +167,10 @@
        h = constrain(0, 1, h);
        if (w == 0 || h == 0) continue;
        if(1){
            w = sqrt(w);
            h = sqrt(h);
            //w = sqrt(w);
            //h = sqrt(h);
            w = pow(w, 1./2.);
            h = pow(h, 1./2.);
        }

        int index = (i+j*num_boxes)*(4+classes+background);

 src/detection.c

@@ -308,8 +308,8 @@
                float y = (pred.vals[j][ci + 1] + row)/num_boxes;
                float w = pred.vals[j][ci + 2]; //* distance_from_edge(row, num_boxes);
                float h = pred.vals[j][ci + 3]; //* distance_from_edge(col, num_boxes);
                w = w*w;
                h = h*h;
                w = pow(w, 2);
                h = pow(h, 2);
                float prob = scale*pred.vals[j][k+class+background+nuisance];
                if(prob < threshold) continue;
                printf("%d %d %f %f %f %f %f\n", offset +  j, class, prob, x, y, w, h);

 src/network.c

@@ -4,7 +4,6 @@
#include "image.h"
#include "data.h"
#include "utils.h"
#include "params.h"

#include "crop_layer.h"
#include "connected_layer.h"
@@ -16,6 +15,7 @@
#include "normalization_layer.h"
#include "softmax_layer.h"
#include "dropout_layer.h"
#include "route_layer.h"

char *get_layer_string(LAYER_TYPE a)
{
@@ -40,6 +40,8 @@
            return "crop";
        case COST:
            return "cost";
        case ROUTE:
            return "route";
        default:
            break;
    }
@@ -99,6 +101,9 @@
        else if(net.types[i] == DROPOUT){
            forward_dropout_layer(*(dropout_layer *)net.layers[i], state);
        }
        else if(net.types[i] == ROUTE){
            forward_route_layer(*(route_layer *)net.layers[i], net);
        }
        state.input = get_network_output_layer(net, i);
    }
}
@@ -143,6 +148,8 @@
        return ((crop_layer *)net.layers[i]) -> output;
    } else if(net.types[i] == NORMALIZATION){
        return ((normalization_layer *)net.layers[i]) -> output;
    } else if(net.types[i] == ROUTE){
        return ((route_layer *)net.layers[i]) -> output;
    }
    return 0;
}
@@ -177,6 +184,8 @@
    } else if(net.types[i] == CONNECTED){
        connected_layer layer = *(connected_layer *)net.layers[i];
        return layer.delta;
    } else if(net.types[i] == ROUTE){
        return ((route_layer *)net.layers[i]) -> delta;
    }
    return 0;
}
@@ -247,10 +256,12 @@
        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){
        } else if(net.types[i] == COST){
            cost_layer layer = *(cost_layer *)net.layers[i];
            backward_cost_layer(layer, state);
        } else if(net.types[i] == ROUTE){
            route_layer layer = *(route_layer *)net.layers[i];
            backward_route_layer(layer, net);
        }
    }
}
@@ -369,6 +380,10 @@
            crop_layer *layer = (crop_layer *)net->layers[i];
            layer->batch = b;
        }
        else if(net->types[i] == ROUTE){
            route_layer *layer = (route_layer *)net->layers[i];
            layer->batch = b;
        }
    }
}

@@ -445,12 +460,17 @@
        softmax_layer layer = *(softmax_layer *)net.layers[i];
        return layer.inputs;
    }
    else if(net.types[i] == ROUTE){
        route_layer layer = *(route_layer *)net.layers[i];
        return layer.outputs;
    }
    fprintf(stderr, "Can't find output size\n");
    return 0;
}

int resize_network(network net, int h, int w, int c)
{
    fprintf(stderr, "Might be broken, careful!!");
    int i;
    for (i = 0; i < net.n; ++i){
        if(net.types[i] == CONVOLUTIONAL){
@@ -540,6 +560,10 @@
        crop_layer layer = *(crop_layer *)net.layers[i];
        return get_crop_image(layer);
    }
    else if(net.types[i] == ROUTE){
        route_layer layer = *(route_layer *)net.layers[i];
        return get_network_image_layer(net, layer.input_layers[0]);
    }
    return make_empty_image(0,0,0);
}


 src/network.h

@@ -4,7 +4,6 @@

#include "image.h"
#include "detection_layer.h"
#include "params.h"
#include "data.h"

typedef enum {
@@ -17,6 +16,7 @@
    NORMALIZATION,
    DROPOUT,
    CROP,
    ROUTE,
    COST
} LAYER_TYPE;


 src/network_kernels.cu

@@ -18,11 +18,12 @@
#include "normalization_layer.h"
#include "softmax_layer.h"
#include "dropout_layer.h"
#include "route_layer.h"
}

float * get_network_output_gpu_layer(network net, int i);
float * get_network_delta_gpu_layer(network net, int i);
float *get_network_output_gpu(network net);
float * get_network_output_gpu(network net);

void forward_network_gpu(network net, network_state state)
{
@@ -55,6 +56,9 @@
        else if(net.types[i] == CROP){
            forward_crop_layer_gpu(*(crop_layer *)net.layers[i], state);
        }
        else if(net.types[i] == ROUTE){
            forward_route_layer_gpu(*(route_layer *)net.layers[i], net);
        }
        state.input = get_network_output_gpu_layer(net, i);
    }
}
@@ -96,6 +100,9 @@
        else if(net.types[i] == SOFTMAX){
            backward_softmax_layer_gpu(*(softmax_layer *)net.layers[i], state);
        }
        else if(net.types[i] == ROUTE){
            backward_route_layer_gpu(*(route_layer *)net.layers[i], net);
        }
    }
}

@@ -142,6 +149,9 @@
    else if(net.types[i] == SOFTMAX){
        return ((softmax_layer *)net.layers[i]) -> output_gpu;
    }
    else if(net.types[i] == ROUTE){
        return ((route_layer *)net.layers[i]) -> output_gpu;
    }
    else if(net.types[i] == DROPOUT){
        return get_network_output_gpu_layer(net, i-1);
    }
@@ -170,6 +180,10 @@
        maxpool_layer layer = *(maxpool_layer *)net.layers[i];
        return layer.delta_gpu;
    }
    else if(net.types[i] == ROUTE){
        route_layer layer = *(route_layer *)net.layers[i];
        return layer.delta_gpu;
    }
    else if(net.types[i] == SOFTMAX){
        softmax_layer layer = *(softmax_layer *)net.layers[i];
        return layer.delta_gpu;

 src/parser.c

@@ -14,6 +14,7 @@
#include "softmax_layer.h"
#include "dropout_layer.h"
#include "detection_layer.h"
#include "route_layer.h"
#include "list.h"
#include "option_list.h"
#include "utils.h"
@@ -34,6 +35,7 @@
int is_cost(section *s);
int is_detection(section *s);
int is_normalization(section *s);
int is_route(section *s);
list *read_cfg(char *filename);

void free_section(section *s)
@@ -246,6 +248,32 @@
    return layer;
}

route_layer *parse_route(list *options, size_params params, network net)
{
    char *l = option_find(options, "layers");   
    int len = strlen(l);
    if(!l) error("Route Layer must specify input layers");
    int n = 1;
    int i;
    for(i = 0; i < len; ++i){
        if (l[i] == ',') ++n;
    }

    int *layers = calloc(n, sizeof(int));
    int *sizes = calloc(n, sizeof(int));
    for(i = 0; i < n; ++i){
        int index = atoi(l);
        l = strchr(l, ',')+1;
        layers[i] = index;
        sizes[i] = get_network_output_size_layer(net, index);
    }
    int batch = params.batch;

    route_layer *layer = make_route_layer(batch, n, layers, sizes);
    option_unused(options);
    return layer;
}

void parse_net_options(list *options, network *net)
{
    net->batch = option_find_int(options, "batch",1);
@@ -326,6 +354,10 @@
            normalization_layer *layer = parse_normalization(options, params);
            net.types[count] = NORMALIZATION;
            net.layers[count] = layer;
        }else if(is_route(s)){
            route_layer *layer = parse_route(options, params, net);
            net.types[count] = ROUTE;
            net.layers[count] = layer;
        }else if(is_dropout(s)){
            dropout_layer *layer = parse_dropout(options, params);
            net.types[count] = DROPOUT;
@@ -402,6 +434,10 @@
    return (strcmp(s->type, "[lrnorm]")==0
            || strcmp(s->type, "[localresponsenormalization]")==0);
}
int is_route(section *s)
{
    return (strcmp(s->type, "[route]")==0);
}

int read_option(char *s, list *options)
{