Changes to make routing work better

Joseph Redmon

2015-07-21 d00f0a1ccd2a9b1c332bbf7754f291dd61dee14f

 src/detection_layer.c

@@ -141,20 +141,20 @@
        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++];
        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++];
            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++];
            state.delta[in_i++] += l.delta[out_i++];
        }
        if(l.joint) state.delta[in_i-l.coords-l.classes-l.joint] = latent_delta;
        if(l.joint) state.delta[in_i-l.coords-l.classes-l.joint] += latent_delta;
    }
}

@@ -199,6 +199,7 @@
    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);