Changes to make routing work better

Joseph Redmon

2015-07-21 d00f0a1ccd2a9b1c332bbf7754f291dd61dee14f

 src/detection_layer.c

@@ -1,72 +1,211 @@
int detection_out_height(detection_layer layer)
#include "detection_layer.h"
#include "activations.h"
#include "softmax_layer.h"
#include "blas.h"
#include "box.h"
#include "cuda.h"
#include "utils.h"
#include <stdio.h>
#include <string.h>
#include <stdlib.h>

int get_detection_layer_locations(detection_layer l)
{
    return layer.size + layer.h*layer.stride;
    return l.inputs / (l.classes+l.coords+l.joint+(l.background || l.objectness));
}

int detection_out_width(detection_layer layer)
int get_detection_layer_output_size(detection_layer l)
{
    return layer.size + layer.w*layer.stride;
    return get_detection_layer_locations(l)*((l.background || l.objectness) + l.classes + l.coords);
}

detection_layer *make_detection_layer(int batch, int h, int w, int c, int n, int size, int stride, ACTIVATION activation)
detection_layer make_detection_layer(int batch, int inputs, int classes, int coords, int joint, int rescore, int background, int objectness)
{
    int i;
    size = 2*(size/2)+1; //HA! And you thought you'd use an even sized filter...
    detection_layer *layer = calloc(1, sizeof(detection_layer));
    layer->h = h;
    layer->w = w;
    layer->c = c;
    layer->n = n;
    layer->batch = batch;
    layer->stride = stride;
    layer->size = size;
    assert(c%n == 0);
    detection_layer l = {0};
    l.type = DETECTION;
    
    l.batch = batch;
    l.inputs = inputs;
    l.classes = classes;
    l.coords = coords;
    l.rescore = rescore;
    l.objectness = objectness;
    l.background = background;
    l.joint = joint;
    l.cost = calloc(1, sizeof(float));
    l.does_cost=1;
    int outputs = get_detection_layer_output_size(l);
    l.outputs = outputs;
    l.output = calloc(batch*outputs, sizeof(float));
    l.delta = calloc(batch*outputs, sizeof(float));
    #ifdef GPU
    l.output_gpu = cuda_make_array(0, batch*outputs);
    l.delta_gpu = cuda_make_array(0, batch*outputs);
    #endif

    layer->filters = calloc(c*size*size, sizeof(float));
    layer->filter_updates = calloc(c*size*size, sizeof(float));
    layer->filter_momentum = calloc(c*size*size, sizeof(float));

    float scale = 1./(size*size*c);
    for(i = 0; i < c*n*size*size; ++i) layer->filters[i] = scale*(rand_uniform());

    int out_h = detection_out_height(*layer);
    int out_w = detection_out_width(*layer);

    layer->output = calloc(layer->batch * out_h * out_w * n, sizeof(float));
    layer->delta  = calloc(layer->batch * out_h * out_w * n, sizeof(float));

    layer->activation = activation;

    fprintf(stderr, "Convolutional Layer: %d x %d x %d image, %d filters -> %d x %d x %d image\n", h,w,c,n, out_h, out_w, n);
    fprintf(stderr, "Detection Layer\n");
    srand(0);

    return layer;
    return l;
}

void forward_detection_layer(const detection_layer layer, float *in)
void forward_detection_layer(const detection_layer l, network_state state)
{
    int out_h = detection_out_height(layer);
    int out_w = detection_out_width(layer);
    int i,j,fh, fw,c;
    memset(layer.output, 0, layer->batch*layer->n*out_h*out_w*sizeof(float));
    for(c = 0; c < layer.c; ++c){
        for(i = 0; i < layer.h; ++i){
            for(j = 0; j < layer.w; ++j){
                float val = layer->input[j+(i + c*layer.h)*layer.w];
                for(fh = 0; fh < layer.size; ++fh){
                    for(fw = 0; fw < layer.size; ++fw){
                        int h = i*layer.stride + fh;
                        int w = j*layer.stride + fw;
                        layer.output[w+(h+c/n*out_h)*out_w] += val*layer->filters[fw+(fh+c*layer.size)*layer.size];
                    }
    int in_i = 0;
    int out_i = 0;
    int locations = get_detection_layer_locations(l);
    int i,j;
    for(i = 0; i < l.batch*locations; ++i){
        int mask = (!state.truth || state.truth[out_i + (l.background || l.objectness) + l.classes + 2]);
        float scale = 1;
        if(l.joint) scale = state.input[in_i++];
        else if(l.objectness){
            l.output[out_i++] = 1-state.input[in_i++];
            scale = mask;
        }
        else if(l.background) l.output[out_i++] = scale*state.input[in_i++];

        for(j = 0; j < l.classes; ++j){
            l.output[out_i++] = scale*state.input[in_i++];
        }
        if(l.objectness){

        }else if(l.background){
            softmax_array(l.output + out_i - l.classes-l.background, l.classes+l.background, l.output + out_i - l.classes-l.background);
            activate_array(state.input+in_i, l.coords, LOGISTIC);
        }
        for(j = 0; j < l.coords; ++j){
            l.output[out_i++] = mask*state.input[in_i++];
        }
    }
    float avg_iou = 0;
    int count = 0;
    if(l.does_cost && state.train){
        *(l.cost) = 0;
        int size = get_detection_layer_output_size(l) * l.batch;
        memset(l.delta, 0, size * sizeof(float));
        for (i = 0; i < l.batch*locations; ++i) {
            int classes = l.objectness+l.classes;
            int offset = i*(classes+l.coords);
            for (j = offset; j < offset+classes; ++j) {
                *(l.cost) += pow(state.truth[j] - l.output[j], 2);
                l.delta[j] =  state.truth[j] - l.output[j];
            }

            box truth;
            truth.x = state.truth[j+0]/7;
            truth.y = state.truth[j+1]/7;
            truth.w = pow(state.truth[j+2], 2);
            truth.h = pow(state.truth[j+3], 2);
            box out;
            out.x = l.output[j+0]/7;
            out.y = l.output[j+1]/7;
            out.w = pow(l.output[j+2], 2);
            out.h = pow(l.output[j+3], 2);

            if(!(truth.w*truth.h)) continue;
            float iou = box_iou(out, truth);
            avg_iou += iou;
            ++count;
            dbox delta = diou(out, truth);

            l.delta[j+0] = 10 * delta.dx/7;
            l.delta[j+1] = 10 * delta.dy/7;
            l.delta[j+2] = 10 * delta.dw * 2 * sqrt(out.w);
            l.delta[j+3] = 10 * delta.dh * 2 * sqrt(out.h);


            *(l.cost) += pow((1-iou), 2);
            l.delta[j+0] = 4 * (state.truth[j+0] - l.output[j+0]);
            l.delta[j+1] = 4 * (state.truth[j+1] - l.output[j+1]);
            l.delta[j+2] = 4 * (state.truth[j+2] - l.output[j+2]);
            l.delta[j+3] = 4 * (state.truth[j+3] - l.output[j+3]);
            if(l.rescore){
                for (j = offset; j < offset+classes; ++j) {
                    if(state.truth[j]) state.truth[j] = iou;
                    l.delta[j] =  state.truth[j] - l.output[j];
                }
            }
        }
        printf("Avg IOU: %f\n", avg_iou/count);
    }
}

void backward_detection_layer(const detection_layer layer, float *delta)
void backward_detection_layer(const detection_layer l, network_state state)
{
    int locations = get_detection_layer_locations(l);
    int i,j;
    int in_i = 0;
    int out_i = 0;
    for(i = 0; i < l.batch*locations; ++i){
        float scale = 1;
        float latent_delta = 0;
        if(l.joint) scale = state.input[in_i++];
        else if (l.objectness)   state.delta[in_i++] += -l.delta[out_i++];
        else if (l.background) state.delta[in_i++] += scale*l.delta[out_i++];
        for(j = 0; j < l.classes; ++j){
            latent_delta += state.input[in_i]*l.delta[out_i];
            state.delta[in_i++] += scale*l.delta[out_i++];
        }

        if (l.objectness) {

        }else if (l.background) gradient_array(l.output + out_i, l.coords, LOGISTIC, l.delta + out_i);
        for(j = 0; j < l.coords; ++j){
            state.delta[in_i++] += l.delta[out_i++];
        }
        if(l.joint) state.delta[in_i-l.coords-l.classes-l.joint] += latent_delta;
    }
}

#ifdef GPU

void forward_detection_layer_gpu(const detection_layer l, network_state state)
{
    int outputs = get_detection_layer_output_size(l);
    float *in_cpu = calloc(l.batch*l.inputs, sizeof(float));
    float *truth_cpu = 0;
    if(state.truth){
        truth_cpu = calloc(l.batch*outputs, sizeof(float));
        cuda_pull_array(state.truth, truth_cpu, l.batch*outputs);
    }
    cuda_pull_array(state.input, in_cpu, l.batch*l.inputs);
    network_state cpu_state;
    cpu_state.train = state.train;
    cpu_state.truth = truth_cpu;
    cpu_state.input = in_cpu;
    forward_detection_layer(l, cpu_state);
    cuda_push_array(l.output_gpu, l.output, l.batch*outputs);
    cuda_push_array(l.delta_gpu, l.delta, l.batch*outputs);
    free(cpu_state.input);
    if(cpu_state.truth) free(cpu_state.truth);
}

void backward_detection_layer_gpu(detection_layer l, network_state state)
{
    int outputs = get_detection_layer_output_size(l);

    float *in_cpu    = calloc(l.batch*l.inputs, sizeof(float));
    float *delta_cpu = calloc(l.batch*l.inputs, sizeof(float));
    float *truth_cpu = 0;
    if(state.truth){
        truth_cpu = calloc(l.batch*outputs, sizeof(float));
        cuda_pull_array(state.truth, truth_cpu, l.batch*outputs);
    }
    network_state cpu_state;
    cpu_state.train = state.train;
    cpu_state.input = in_cpu;
    cpu_state.truth = truth_cpu;
    cpu_state.delta = delta_cpu;

    cuda_pull_array(state.input, in_cpu,    l.batch*l.inputs);
    cuda_pull_array(state.delta, delta_cpu, l.batch*l.inputs);
    cuda_pull_array(l.delta_gpu, l.delta, l.batch*outputs);
    backward_detection_layer(l, cpu_state);
    cuda_push_array(state.delta, delta_cpu, l.batch*l.inputs);

    free(in_cpu);
    free(delta_cpu);
}
#endif