stable

Joseph Redmon

2015-09-09 393dc8eb6f3a9dd92ec665200444186c1addc5d2

 src/detection_layer.c

@@ -39,8 +39,8 @@
    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);
    l.output_gpu = cuda_make_array(l.output, batch*outputs);
    l.delta_gpu  = cuda_make_array(l.delta,  batch*outputs);
    #endif

    fprintf(stderr, "Detection Layer\n");
@@ -85,11 +85,12 @@
        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 classes = (l.objectness || l.background)+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];
                if(l.background && j == offset) l.delta[j] *= .1;
            }

            box truth;
@@ -97,6 +98,7 @@
            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;
@@ -107,13 +109,6 @@
            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]);
@@ -121,9 +116,15 @@
            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];
                if(l.objectness){
                    state.truth[offset] = iou;
                    l.delta[offset] = state.truth[offset] - l.output[offset];
                }
                else{
                    for (j = offset; j < offset+classes; ++j) {
                        if(state.truth[j]) state.truth[j] = iou;
                        l.delta[j] =  state.truth[j] - l.output[j];
                    }
                }
            }
        }
@@ -151,7 +152,7 @@
        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){
        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;
@@ -204,6 +205,7 @@
    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);
}