idk

Joseph Redmon

2016-06-25 a6b2511a566f77a0838dc1dd0d5f3e3c49a8faa0

 src/detection_layer.c

@@ -6,42 +6,32 @@
#include "cuda.h"
#include "utils.h"
#include <stdio.h>
#include <assert.h>
#include <string.h>
#include <stdlib.h>

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

int get_detection_layer_output_size(detection_layer l)
{
    return get_detection_layer_locations(l)*((l.background || l.objectness) + l.classes + l.coords);
}

detection_layer make_detection_layer(int batch, int inputs, int classes, int coords, int joint, int rescore, int background, int objectness)
detection_layer make_detection_layer(int batch, int inputs, int n, int side, int classes, int coords, int rescore)
{
    detection_layer l = {0};
    l.type = DETECTION;
    

    l.n = n;
    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.side = side;
    assert(side*side*((1 + l.coords)*l.n + l.classes) == inputs);
    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
    l.outputs = l.inputs;
    l.truths = l.side*l.side*(1+l.coords+l.classes);
    l.output = calloc(batch*l.outputs, sizeof(float));
    l.delta = calloc(batch*l.outputs, sizeof(float));
#ifdef GPU
    l.output_gpu = cuda_make_array(l.output, batch*l.outputs);
    l.delta_gpu = cuda_make_array(l.delta, batch*l.outputs);
#endif

    fprintf(stderr, "Detection Layer\n");
    srand(0);
@@ -51,162 +41,209 @@

void forward_detection_layer(const detection_layer l, network_state state)
{
    int in_i = 0;
    int out_i = 0;
    int locations = get_detection_layer_locations(l);
    int locations = l.side*l.side;
    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++];
    memcpy(l.output, state.input, l.outputs*l.batch*sizeof(float));
    int b;
    if (l.softmax){
        for(b = 0; b < l.batch; ++b){
            int index = b*l.inputs;
            for (i = 0; i < locations; ++i) {
                int offset = i*l.classes;
                softmax_array(l.output + index + offset, l.classes, 1,
                        l.output + index + offset);
            }
        }
    }
    float avg_iou = 0;
    int count = 0;
    if(l.does_cost && state.train){
    if(state.train){
        float avg_iou = 0;
        float avg_cat = 0;
        float avg_allcat = 0;
        float avg_obj = 0;
        float avg_anyobj = 0;
        int count = 0;
        *(l.cost) = 0;
        int size = get_detection_layer_output_size(l) * l.batch;
        int size = l.inputs * 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];
        for (b = 0; b < l.batch; ++b){
            int index = b*l.inputs;
            for (i = 0; i < locations; ++i) {
                int truth_index = (b*locations + i)*(1+l.coords+l.classes);
                int is_obj = state.truth[truth_index];
                for (j = 0; j < l.n; ++j) {
                    int p_index = index + locations*l.classes + i*l.n + j;
                    l.delta[p_index] = l.noobject_scale*(0 - l.output[p_index]);
                    *(l.cost) += l.noobject_scale*pow(l.output[p_index], 2);
                    avg_anyobj += l.output[p_index];
                }

                int best_index = -1;
                float best_iou = 0;
                float best_rmse = 20;

                if (!is_obj){
                    continue;
                }

                int class_index = index + i*l.classes;
                for(j = 0; j < l.classes; ++j) {
                    l.delta[class_index+j] = l.class_scale * (state.truth[truth_index+1+j] - l.output[class_index+j]);
                    *(l.cost) += l.class_scale * pow(state.truth[truth_index+1+j] - l.output[class_index+j], 2);
                    if(state.truth[truth_index + 1 + j]) avg_cat += l.output[class_index+j];
                    avg_allcat += l.output[class_index+j];
                }

                box truth = float_to_box(state.truth + truth_index + 1 + l.classes);
                truth.x /= l.side;
                truth.y /= l.side;

                for(j = 0; j < l.n; ++j){
                    int box_index = index + locations*(l.classes + l.n) + (i*l.n + j) * l.coords;
                    box out = float_to_box(l.output + box_index);
                    out.x /= l.side;
                    out.y /= l.side;

                    if (l.sqrt){
                        out.w = out.w*out.w;
                        out.h = out.h*out.h;
                    }

                    float iou  = box_iou(out, truth);
                    //iou = 0;
                    float rmse = box_rmse(out, truth);
                    if(best_iou > 0 || iou > 0){
                        if(iou > best_iou){
                            best_iou = iou;
                            best_index = j;
                        }
                    }else{
                        if(rmse < best_rmse){
                            best_rmse = rmse;
                            best_index = j;
                        }
                    }
                }

                if(l.forced){
                    if(truth.w*truth.h < .1){
                        best_index = 1;
                    }else{
                        best_index = 0;
                    }
                }
                if(l.random && *(state.net.seen) < 64000){
                    best_index = rand()%l.n;
                }

                int box_index = index + locations*(l.classes + l.n) + (i*l.n + best_index) * l.coords;
                int tbox_index = truth_index + 1 + l.classes;

                box out = float_to_box(l.output + box_index);
                out.x /= l.side;
                out.y /= l.side;
                if (l.sqrt) {
                    out.w = out.w*out.w;
                    out.h = out.h*out.h;
                }
                float iou  = box_iou(out, truth);

                //printf("%d,", best_index);
                int p_index = index + locations*l.classes + i*l.n + best_index;
                *(l.cost) -= l.noobject_scale * pow(l.output[p_index], 2);
                *(l.cost) += l.object_scale * pow(1-l.output[p_index], 2);
                avg_obj += l.output[p_index];
                l.delta[p_index] = l.object_scale * (1.-l.output[p_index]);

                if(l.rescore){
                    l.delta[p_index] = l.object_scale * (iou - l.output[p_index]);
                }

                l.delta[box_index+0] = l.coord_scale*(state.truth[tbox_index + 0] - l.output[box_index + 0]);
                l.delta[box_index+1] = l.coord_scale*(state.truth[tbox_index + 1] - l.output[box_index + 1]);
                l.delta[box_index+2] = l.coord_scale*(state.truth[tbox_index + 2] - l.output[box_index + 2]);
                l.delta[box_index+3] = l.coord_scale*(state.truth[tbox_index + 3] - l.output[box_index + 3]);
                if(l.sqrt){
                    l.delta[box_index+2] = l.coord_scale*(sqrt(state.truth[tbox_index + 2]) - l.output[box_index + 2]);
                    l.delta[box_index+3] = l.coord_scale*(sqrt(state.truth[tbox_index + 3]) - l.output[box_index + 3]);
                }

                *(l.cost) += pow(1-iou, 2);
                avg_iou += iou;
                ++count;
            }
        }
        printf("Avg IOU: %f\n", avg_iou/count);

        if(0){
            float *costs = calloc(l.batch*locations*l.n, sizeof(float));
            for (b = 0; b < l.batch; ++b) {
                int index = b*l.inputs;
                for (i = 0; i < locations; ++i) {
                    for (j = 0; j < l.n; ++j) {
                        int p_index = index + locations*l.classes + i*l.n + j;
                        costs[b*locations*l.n + i*l.n + j] = l.delta[p_index]*l.delta[p_index];
                    }
                }
            }
            int indexes[100];
            top_k(costs, l.batch*locations*l.n, 100, indexes);
            float cutoff = costs[indexes[99]];
            for (b = 0; b < l.batch; ++b) {
                int index = b*l.inputs;
                for (i = 0; i < locations; ++i) {
                    for (j = 0; j < l.n; ++j) {
                        int p_index = index + locations*l.classes + i*l.n + j;
                        if (l.delta[p_index]*l.delta[p_index] < cutoff) l.delta[p_index] = 0;
                    }
                }
            }
            free(costs);
        }


        *(l.cost) = pow(mag_array(l.delta, l.outputs * l.batch), 2);


        printf("Detection Avg IOU: %f, Pos Cat: %f, All Cat: %f, Pos Obj: %f, Any Obj: %f, count: %d\n", avg_iou/count, avg_cat/count, avg_allcat/(count*l.classes), avg_obj/count, avg_anyobj/(l.batch*locations*l.n), count);
    }
}

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;
    }
    axpy_cpu(l.batch*l.inputs, 1, l.delta, 1, state.delta, 1);
}

#ifdef GPU

void forward_detection_layer_gpu(const detection_layer l, network_state state)
{
    int outputs = get_detection_layer_output_size(l);
    if(!state.train){
        copy_ongpu(l.batch*l.inputs, state.input, 1, l.output_gpu, 1);
        return;
    }

    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);
        int num_truth = l.batch*l.side*l.side*(1+l.coords+l.classes);
        truth_cpu = calloc(num_truth, sizeof(float));
        cuda_pull_array(state.truth, truth_cpu, num_truth);
    }
    cuda_pull_array(state.input, in_cpu, l.batch*l.inputs);
    network_state cpu_state;
    network_state cpu_state = 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);
    cuda_push_array(l.output_gpu, l.output, l.batch*l.outputs);
    cuda_push_array(l.delta_gpu, l.delta, l.batch*l.inputs);
    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);

    if (truth_cpu) free(truth_cpu);
    free(in_cpu);
    free(delta_cpu);
    axpy_ongpu(l.batch*l.inputs, 1, l.delta_gpu, 1, state.delta, 1);
    //copy_ongpu(l.batch*l.inputs, l.delta_gpu, 1, state.delta, 1);
}
#endif