fixed dropout ><

Joseph Redmon

2014-12-13 90d354a2a5a3ba76071337d8794cfc00f7bc5fab

 src/network_gpu.c

@@ -22,7 +22,9 @@
{
    //printf("start\n");
    int i;
   // printf("Truth: %f\n", cl_checksum(truth, 1000*net.batch));
    for(i = 0; i < net.n; ++i){
        //printf("Truth %i: %f\n", i, cl_checksum(truth, 1000*net.batch));
        //clock_t time = clock();
        if(net.types[i] == CONVOLUTIONAL){
            convolutional_layer layer = *(convolutional_layer *)net.layers[i];
@@ -48,6 +50,11 @@
            forward_softmax_layer_gpu(layer, input);
            input = layer.output_cl;
        }
        else if(net.types[i] == DROPOUT){
            if(!train) continue;
            dropout_layer layer = *(dropout_layer *)net.layers[i];
            forward_dropout_layer_gpu(layer, input);
        }
        //printf("%d %f\n", i, sec(clock()-time));
        /*
           else if(net.types[i] == CROP){
@@ -80,7 +87,7 @@
        }
        if(net.types[i] == CONVOLUTIONAL){
            convolutional_layer layer = *(convolutional_layer *)net.layers[i];
            backward_convolutional_layer_gpu(layer, prev_delta);
            backward_convolutional_layer_gpu(layer, prev_input, prev_delta);
        }
        else if(net.types[i] == COST){
            cost_layer layer = *(cost_layer *)net.layers[i];
@@ -94,6 +101,10 @@
            maxpool_layer layer = *(maxpool_layer *)net.layers[i];
            backward_maxpool_layer_gpu(layer, prev_delta);
        }
        else if(net.types[i] == DROPOUT){
            dropout_layer layer = *(dropout_layer *)net.layers[i];
            backward_dropout_layer_gpu(layer, prev_delta);
        }
        else if(net.types[i] == SOFTMAX){
            softmax_layer layer = *(softmax_layer *)net.layers[i];
            backward_softmax_layer_gpu(layer, prev_delta);
@@ -134,6 +145,8 @@
    else if(net.types[i] == SOFTMAX){
        softmax_layer layer = *(softmax_layer *)net.layers[i];
        return layer.output_cl;
    } else if(net.types[i] == DROPOUT){
        return get_network_output_cl_layer(net, i-1);
    }
    return 0;
}
@@ -155,6 +168,8 @@
    else if(net.types[i] == SOFTMAX){
        softmax_layer layer = *(softmax_layer *)net.layers[i];
        return layer.delta_cl;
    } else if(net.types[i] == DROPOUT){
        return get_network_delta_cl_layer(net, i-1);
    }
    return 0;
}
@@ -171,18 +186,10 @@
        cl_write_array(*net.input_cl, x, x_size);
        cl_write_array(*net.truth_cl, y, y_size);
    }
    //printf("trans %f\n", sec(clock()-time));
    //time = clock();
    forward_network_gpu(net, *net.input_cl, *net.truth_cl, 1);
    //printf("forw %f\n", sec(clock()-time));
    //time = clock();
    backward_network_gpu(net, *net.input_cl);
    //printf("back %f\n", sec(clock()-time));
    //time = clock();
    update_network_gpu(net);
    float error = get_network_cost(net);
    //printf("updt %f\n", sec(clock()-time));
    //time = clock();
    return error;
}

@@ -287,11 +294,29 @@
float network_accuracy_gpu(network net, data d)
{
    matrix guess = network_predict_data_gpu(net, d);
    float acc = matrix_accuracy(d.y, guess);
    float acc = matrix_topk_accuracy(d.y, guess,1);
    free_matrix(guess);
    return acc;
}

float *network_accuracies_gpu(network net, data d)
{
    static float acc[2];
    matrix guess = network_predict_data_gpu(net, d);
    acc[0] = matrix_topk_accuracy(d.y, guess,1);
    acc[1] = matrix_topk_accuracy(d.y, guess,5);
    free_matrix(guess);
    return acc;
}


#else
void forward_network_gpu(network net, cl_mem input, cl_mem truth, int train){}
void backward_network_gpu(network net, cl_mem input){}
void update_network_gpu(network net){}
float train_network_sgd_gpu(network net, data d, int n){return 0;}
float train_network_data_gpu(network net, data d, int n){return 0;}
float *network_predict_gpu(network net, float *input){return 0;}
float network_accuracy_gpu(network net, data d){return 0;}

#endif