If there is excessive GPU-RAM consumption by CUDNN then then do not use Wor...

AlexeyAB

2018-02-21 033e934ce82826c73d851098baf7ce4b1a27c89a

 src/network.c

@@ -1,334 +1,199 @@
#include <stdio.h>
#include <time.h>
#include <assert.h>
#include "network.h"
#include "image.h"
#include "data.h"
#include "utils.h"
#include "blas.h"

#include "crop_layer.h"
#include "connected_layer.h"
#include "gru_layer.h"
#include "rnn_layer.h"
#include "crnn_layer.h"
#include "local_layer.h"
#include "convolutional_layer.h"
#include "maxpool_layer.h"
#include "cost_layer.h"
#include "activation_layer.h"
#include "detection_layer.h"
#include "region_layer.h"
#include "normalization_layer.h"
#include "freeweight_layer.h"
#include "batchnorm_layer.h"
#include "maxpool_layer.h"
#include "reorg_layer.h"
#include "avgpool_layer.h"
#include "cost_layer.h"
#include "softmax_layer.h"
#include "dropout_layer.h"
#include "route_layer.h"
#include "shortcut_layer.h"

network make_network(int n, int batch)
int get_current_batch(network net)
{
    network net;
    net.n = n;
    net.batch = batch;
    net.layers = calloc(net.n, sizeof(void *));
    net.types = calloc(net.n, sizeof(LAYER_TYPE));
    net.outputs = 0;
    net.output = 0;
    int batch_num = (*net.seen)/(net.batch*net.subdivisions);
    return batch_num;
}

void reset_momentum(network net)
{
    if (net.momentum == 0) return;
    net.learning_rate = 0;
    net.momentum = 0;
    net.decay = 0;
    #ifdef GPU
    net.input_cl = calloc(1, sizeof(cl_mem));
    net.truth_cl = calloc(1, sizeof(cl_mem));
        //if(net.gpu_index >= 0) update_network_gpu(net);
    #endif
}

float get_current_rate(network net)
{
    int batch_num = get_current_batch(net);
    int i;
    float rate;
   if (batch_num < net.burn_in) return net.learning_rate * pow((float)batch_num / net.burn_in, net.power);
    switch (net.policy) {
        case CONSTANT:
            return net.learning_rate;
        case STEP:
            return net.learning_rate * pow(net.scale, batch_num/net.step);
        case STEPS:
            rate = net.learning_rate;
            for(i = 0; i < net.num_steps; ++i){
                if(net.steps[i] > batch_num) return rate;
                rate *= net.scales[i];
                //if(net.steps[i] > batch_num - 1 && net.scales[i] > 1) reset_momentum(net);
            }
            return rate;
        case EXP:
            return net.learning_rate * pow(net.gamma, batch_num);
        case POLY:
         return net.learning_rate * pow(1 - (float)batch_num / net.max_batches, net.power);
            //if (batch_num < net.burn_in) return net.learning_rate * pow((float)batch_num / net.burn_in, net.power);
            //return net.learning_rate * pow(1 - (float)batch_num / net.max_batches, net.power);
        case RANDOM:
            return net.learning_rate * pow(rand_uniform(0,1), net.power);
        case SIG:
            return net.learning_rate * (1./(1.+exp(net.gamma*(batch_num - net.step))));
        default:
            fprintf(stderr, "Policy is weird!\n");
            return net.learning_rate;
    }
}

char *get_layer_string(LAYER_TYPE a)
{
    switch(a){
        case CONVOLUTIONAL:
            return "convolutional";
        case ACTIVE:
            return "activation";
        case LOCAL:
            return "local";
        case DECONVOLUTIONAL:
            return "deconvolutional";
        case CONNECTED:
            return "connected";
        case RNN:
            return "rnn";
        case GRU:
            return "gru";
        case CRNN:
            return "crnn";
        case MAXPOOL:
            return "maxpool";
        case REORG:
            return "reorg";
        case AVGPOOL:
            return "avgpool";
        case SOFTMAX:
            return "softmax";
        case DETECTION:
            return "detection";
        case REGION:
            return "region";
        case DROPOUT:
            return "dropout";
        case CROP:
            return "crop";
        case COST:
            return "cost";
        case ROUTE:
            return "route";
        case SHORTCUT:
            return "shortcut";
        case NORMALIZATION:
            return "normalization";
        case BATCHNORM:
            return "batchnorm";
        default:
            break;
    }
    return "none";
}

network make_network(int n)
{
    network net = {0};
    net.n = n;
    net.layers = calloc(net.n, sizeof(layer));
    net.seen = calloc(1, sizeof(int));
    #ifdef GPU
    net.input_gpu = calloc(1, sizeof(float *));
    net.truth_gpu = calloc(1, sizeof(float *));
    #endif
    return net;
}

#ifdef GPU

void forward_network_gpu(network net, cl_mem input, cl_mem truth, int train)
void forward_network(network net, network_state state)
{
    //printf("start\n");
    state.workspace = net.workspace;
    int i;
    for(i = 0; i < net.n; ++i){
        if(net.types[i] == CONVOLUTIONAL){
            convolutional_layer layer = *(convolutional_layer *)net.layers[i];
            forward_convolutional_layer_gpu(layer, input);
            input = layer.output_cl;
        state.index = i;
        layer l = net.layers[i];
        if(l.delta){
            scal_cpu(l.outputs * l.batch, 0, l.delta, 1);
        }
        else if(net.types[i] == COST){
            cost_layer layer = *(cost_layer *)net.layers[i];
            forward_cost_layer_gpu(layer, input, truth);
        }
        else if(net.types[i] == CONNECTED){
            connected_layer layer = *(connected_layer *)net.layers[i];
            forward_connected_layer_gpu(layer, input);
            input = layer.output_cl;
        }
        else if(net.types[i] == MAXPOOL){
            maxpool_layer layer = *(maxpool_layer *)net.layers[i];
            forward_maxpool_layer_gpu(layer, input);
            input = layer.output_cl;
        }
        else if(net.types[i] == SOFTMAX){
            softmax_layer layer = *(softmax_layer *)net.layers[i];
            forward_softmax_layer_gpu(layer, input);
            input = layer.output_cl;
        }
        /*
           else if(net.types[i] == CROP){
           crop_layer layer = *(crop_layer *)net.layers[i];
           forward_crop_layer(layer, input);
           input = layer.output;
           }
           else if(net.types[i] == NORMALIZATION){
           normalization_layer layer = *(normalization_layer *)net.layers[i];
           forward_normalization_layer(layer, input);
           input = layer.output;
           }
         */
    }
}

void backward_network_gpu(network net, cl_mem input)
{
    int i;
    cl_mem prev_input;
    cl_mem prev_delta;
    for(i = net.n-1; i >= 0; --i){
        if(i == 0){
            prev_input = input;
            prev_delta = 0;
        }else{
            prev_input = get_network_output_cl_layer(net, i-1);
            prev_delta = get_network_delta_cl_layer(net, i-1);
        }
        if(net.types[i] == CONVOLUTIONAL){
            convolutional_layer layer = *(convolutional_layer *)net.layers[i];
            backward_convolutional_layer_gpu(layer, prev_delta);
        }
        else if(net.types[i] == COST){
            cost_layer layer = *(cost_layer *)net.layers[i];
            backward_cost_layer_gpu(layer, prev_input, prev_delta);
        }
        else if(net.types[i] == CONNECTED){
            connected_layer layer = *(connected_layer *)net.layers[i];
            backward_connected_layer_gpu(layer, prev_input, prev_delta);
        }
        else if(net.types[i] == MAXPOOL){
            maxpool_layer layer = *(maxpool_layer *)net.layers[i];
            backward_maxpool_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);
        }
    }
}

void update_network_gpu(network net)
{
    int i;
    for(i = 0; i < net.n; ++i){
        if(net.types[i] == CONVOLUTIONAL){
            convolutional_layer layer = *(convolutional_layer *)net.layers[i];
            update_convolutional_layer_gpu(layer);
        }
        else if(net.types[i] == CONNECTED){
            connected_layer layer = *(connected_layer *)net.layers[i];
            update_connected_layer_gpu(layer);
        }
    }
}

cl_mem get_network_output_cl_layer(network net, int i)
{
    if(net.types[i] == CONVOLUTIONAL){
        convolutional_layer layer = *(convolutional_layer *)net.layers[i];
        return layer.output_cl;
    }
    else if(net.types[i] == CONNECTED){
        connected_layer layer = *(connected_layer *)net.layers[i];
        return layer.output_cl;
    }
    else if(net.types[i] == MAXPOOL){
        maxpool_layer layer = *(maxpool_layer *)net.layers[i];
        return layer.output_cl;
    }
    else if(net.types[i] == SOFTMAX){
        softmax_layer layer = *(softmax_layer *)net.layers[i];
        return layer.output_cl;
    }
    return 0;
}

cl_mem get_network_delta_cl_layer(network net, int i)
{
    if(net.types[i] == CONVOLUTIONAL){
        convolutional_layer layer = *(convolutional_layer *)net.layers[i];
        return layer.delta_cl;
    }
    else if(net.types[i] == CONNECTED){
        connected_layer layer = *(connected_layer *)net.layers[i];
        return layer.delta_cl;
    }
    else if(net.types[i] == MAXPOOL){
        maxpool_layer layer = *(maxpool_layer *)net.layers[i];
        return layer.delta_cl;
    }
    else if(net.types[i] == SOFTMAX){
        softmax_layer layer = *(softmax_layer *)net.layers[i];
        return layer.delta_cl;
    }
    return 0;
}

#endif

void forward_network(network net, float *input, float *truth, int train)
{
    int i;
    for(i = 0; i < net.n; ++i){
        if(net.types[i] == CONVOLUTIONAL){
            convolutional_layer layer = *(convolutional_layer *)net.layers[i];
            forward_convolutional_layer(layer, input);
            input = layer.output;
        }
        else if(net.types[i] == CONNECTED){
            connected_layer layer = *(connected_layer *)net.layers[i];
            forward_connected_layer(layer, input);
            input = layer.output;
        }
        else if(net.types[i] == CROP){
            crop_layer layer = *(crop_layer *)net.layers[i];
            forward_crop_layer(layer, input);
            input = layer.output;
        }
        else if(net.types[i] == COST){
            cost_layer layer = *(cost_layer *)net.layers[i];
            forward_cost_layer(layer, input, truth);
        }
        else if(net.types[i] == SOFTMAX){
            softmax_layer layer = *(softmax_layer *)net.layers[i];
            forward_softmax_layer(layer, input);
            input = layer.output;
        }
        else if(net.types[i] == MAXPOOL){
            maxpool_layer layer = *(maxpool_layer *)net.layers[i];
            forward_maxpool_layer(layer, input);
            input = layer.output;
        }
        else if(net.types[i] == NORMALIZATION){
            normalization_layer layer = *(normalization_layer *)net.layers[i];
            forward_normalization_layer(layer, input);
            input = layer.output;
        }
        else if(net.types[i] == DROPOUT){
            if(!train) continue;
            dropout_layer layer = *(dropout_layer *)net.layers[i];
            forward_dropout_layer(layer, input);
        }
        else if(net.types[i] == FREEWEIGHT){
            if(!train) continue;
            freeweight_layer layer = *(freeweight_layer *)net.layers[i];
            forward_freeweight_layer(layer, input);
        }
        l.forward(l, state);
        state.input = l.output;
    }
}

void update_network(network net)
{
    int i;
    int update_batch = net.batch*net.subdivisions;
    float rate = get_current_rate(net);
    for(i = 0; i < net.n; ++i){
        if(net.types[i] == CONVOLUTIONAL){
            convolutional_layer layer = *(convolutional_layer *)net.layers[i];
            update_convolutional_layer(layer);
        }
        else if(net.types[i] == MAXPOOL){
            //maxpool_layer layer = *(maxpool_layer *)net.layers[i];
        }
        else if(net.types[i] == SOFTMAX){
            //maxpool_layer layer = *(maxpool_layer *)net.layers[i];
        }
        else if(net.types[i] == NORMALIZATION){
            //maxpool_layer layer = *(maxpool_layer *)net.layers[i];
        }
        else if(net.types[i] == CONNECTED){
            connected_layer layer = *(connected_layer *)net.layers[i];
            update_connected_layer(layer);
        layer l = net.layers[i];
        if(l.update){
            l.update(l, update_batch, rate, net.momentum, net.decay);
        }
    }
}

float *get_network_output_layer(network net, int i)
{
    if(net.types[i] == CONVOLUTIONAL){
        convolutional_layer layer = *(convolutional_layer *)net.layers[i];
        return layer.output;
    } else if(net.types[i] == MAXPOOL){
        maxpool_layer layer = *(maxpool_layer *)net.layers[i];
        return layer.output;
    } else if(net.types[i] == SOFTMAX){
        softmax_layer layer = *(softmax_layer *)net.layers[i];
        return layer.output;
    } else if(net.types[i] == DROPOUT){
        return get_network_output_layer(net, i-1);
    } else if(net.types[i] == FREEWEIGHT){
        return get_network_output_layer(net, i-1);
    } else if(net.types[i] == CONNECTED){
        connected_layer layer = *(connected_layer *)net.layers[i];
        return layer.output;
    } else if(net.types[i] == NORMALIZATION){
        normalization_layer layer = *(normalization_layer *)net.layers[i];
        return layer.output;
    }
    return 0;
}
float *get_network_output(network net)
{
#ifdef GPU
    if (gpu_index >= 0) return get_network_output_gpu(net);
#endif 
    int i;
    for(i = net.n-1; i > 0; --i) if(net.types[i] != COST) break;
    return get_network_output_layer(net, i);
}

float *get_network_delta_layer(network net, int i)
{
    if(net.types[i] == CONVOLUTIONAL){
        convolutional_layer layer = *(convolutional_layer *)net.layers[i];
        return layer.delta;
    } else if(net.types[i] == MAXPOOL){
        maxpool_layer layer = *(maxpool_layer *)net.layers[i];
        return layer.delta;
    } else if(net.types[i] == SOFTMAX){
        softmax_layer layer = *(softmax_layer *)net.layers[i];
        return layer.delta;
    } else if(net.types[i] == DROPOUT){
        return get_network_delta_layer(net, i-1);
    } else if(net.types[i] == FREEWEIGHT){
        return get_network_delta_layer(net, i-1);
    } else if(net.types[i] == CONNECTED){
        connected_layer layer = *(connected_layer *)net.layers[i];
        return layer.delta;
    }
    return 0;
    for(i = net.n-1; i > 0; --i) if(net.layers[i].type != COST) break;
    return net.layers[i].output;
}

float get_network_cost(network net)
{
    if(net.types[net.n-1] == COST){
        return ((cost_layer *)net.layers[net.n-1])->output[0];
    }
    return 0;
}

float *get_network_delta(network net)
{
    return get_network_delta_layer(net, net.n-1);
}

float calculate_error_network(network net, float *truth)
{
    float sum = 0;
    float *delta = get_network_delta(net);
    float *out = get_network_output(net);
    int i;
    for(i = 0; i < get_network_output_size(net)*net.batch; ++i){
        //if(i %get_network_output_size(net) == 0) printf("\n");
        //printf("%5.2f %5.2f, ", out[i], truth[i]);
        //if(i == get_network_output_size(net)) printf("\n");
        delta[i] = truth[i] - out[i];
        //printf("%.10f, ", out[i]);
        sum += delta[i]*delta[i];
    float sum = 0;
    int count = 0;
    for(i = 0; i < net.n; ++i){
        if(net.layers[i].cost){
            sum += net.layers[i].cost[0];
            ++count;
        }
    }
    //printf("\n");
    return sum;
    return sum/count;
}

int get_predicted_class_network(network net)
@@ -338,94 +203,45 @@
    return max_index(out, k);
}

void backward_network(network net, float *input)
void backward_network(network net, network_state state)
{
    int i;
    float *prev_input;
    float *prev_delta;
    float *original_input = state.input;
    float *original_delta = state.delta;
    state.workspace = net.workspace;
    for(i = net.n-1; i >= 0; --i){
        state.index = i;
        if(i == 0){
            prev_input = input;
            prev_delta = 0;
            state.input = original_input;
            state.delta = original_delta;
        }else{
            prev_input = get_network_output_layer(net, i-1);
            prev_delta = get_network_delta_layer(net, i-1);
            layer prev = net.layers[i-1];
            state.input = prev.output;
            state.delta = prev.delta;
        }
        if(net.types[i] == CONVOLUTIONAL){
            convolutional_layer layer = *(convolutional_layer *)net.layers[i];
            backward_convolutional_layer(layer, prev_delta);
        }
        else if(net.types[i] == MAXPOOL){
            maxpool_layer layer = *(maxpool_layer *)net.layers[i];
            if(i != 0) backward_maxpool_layer(layer, prev_delta);
        }
        else if(net.types[i] == NORMALIZATION){
            normalization_layer layer = *(normalization_layer *)net.layers[i];
            if(i != 0) backward_normalization_layer(layer, prev_input, prev_delta);
        }
        else if(net.types[i] == SOFTMAX){
            softmax_layer layer = *(softmax_layer *)net.layers[i];
            if(i != 0) backward_softmax_layer(layer, prev_delta);
        }
        else if(net.types[i] == CONNECTED){
            connected_layer layer = *(connected_layer *)net.layers[i];
            backward_connected_layer(layer, prev_input, prev_delta);
        }
        else if(net.types[i] == COST){
            cost_layer layer = *(cost_layer *)net.layers[i];
            backward_cost_layer(layer, prev_input, prev_delta);
        }
        layer l = net.layers[i];
        if (l.stopbackward) break;
        l.backward(l, state);
    }
}


#ifdef GPU
float train_network_datum_gpu(network net, float *x, float *y)
{
    int x_size = get_network_input_size(net)*net.batch;
    int y_size = get_network_output_size(net)*net.batch;
    if(!*net.input_cl){
        *net.input_cl = cl_make_array(x, x_size);
        *net.truth_cl = cl_make_array(y, y_size);
    }else{
        cl_write_array(*net.input_cl, x, x_size);
        cl_write_array(*net.truth_cl, y, y_size);
    }
    forward_network_gpu(net, *net.input_cl, *net.truth_cl, 1);
    backward_network_gpu(net, *net.input_cl);
    float error = get_network_cost(net);
    update_network_gpu(net);
    return error;
}

float train_network_sgd_gpu(network net, data d, int n)
{
    int batch = net.batch;
    float *X = calloc(batch*d.X.cols, sizeof(float));
    float *y = calloc(batch*d.y.cols, sizeof(float));

    int i;
    float sum = 0;
    for(i = 0; i < n; ++i){
        get_batch(d, batch, X, y);
        float err = train_network_datum_gpu(net, X, y);
        sum += err;
    }
    free(X);
    free(y);
    return (float)sum/(n*batch);
}
#endif


float train_network_datum(network net, float *x, float *y)
{
    forward_network(net, x, y, 1);
    //int class = get_predicted_class_network(net);
    backward_network(net, x);
#ifdef GPU
    if(gpu_index >= 0) return train_network_datum_gpu(net, x, y);
#endif
    network_state state;
    *net.seen += net.batch;
    state.index = 0;
    state.net = net;
    state.input = x;
    state.delta = 0;
    state.truth = y;
    state.train = 1;
    forward_network(net, state);
    backward_network(net, state);
    float error = get_network_cost(net);
    update_network(net);
    //return (y[class]?1:0);
    if(((*net.seen)/net.batch)%net.subdivisions == 0) update_network(net);
    return error;
}

@@ -438,7 +254,7 @@
    int i;
    float sum = 0;
    for(i = 0; i < n; ++i){
        get_batch(d, batch, X, y);
        get_random_batch(d, batch, X, y);
        float err = train_network_datum(net, X, y);
        sum += err;
    }
@@ -446,18 +262,45 @@
    free(y);
    return (float)sum/(n*batch);
}

float train_network(network net, data d)
{
    assert(d.X.rows % net.batch == 0);
    int batch = net.batch;
    int n = d.X.rows / batch;
    float *X = calloc(batch*d.X.cols, sizeof(float));
    float *y = calloc(batch*d.y.cols, sizeof(float));

    int i;
    float sum = 0;
    for(i = 0; i < n; ++i){
        get_next_batch(d, batch, i*batch, X, y);
        float err = train_network_datum(net, X, y);
        sum += err;
    }
    free(X);
    free(y);
    return (float)sum/(n*batch);
}


float train_network_batch(network net, data d, int n)
{
    int i,j;
    network_state state;
    state.index = 0;
    state.net = net;
    state.train = 1;
    state.delta = 0;
    float sum = 0;
    int batch = 2;
    for(i = 0; i < n; ++i){
        for(j = 0; j < batch; ++j){
            int index = rand()%d.X.rows;
            float *x = d.X.vals[index];
            float *y = d.y.vals[index];
            forward_network(net, x, y, 1);
            backward_network(net, x);
            state.input = d.X.vals[index];
            state.truth = d.y.vals[index];
            forward_network(net, state);
            backward_network(net, state);
            sum += get_network_cost(net);
        }
        update_network(net);
@@ -465,145 +308,134 @@
    return (float)sum/(n*batch);
}


void train_network(network net, data d)
void set_batch_network(network *net, int b)
{
    net->batch = b;
    int i;
    int correct = 0;
    for(i = 0; i < d.X.rows; ++i){
        correct += train_network_datum(net, d.X.vals[i], d.y.vals[i]);
        if(i%100 == 0){
            visualize_network(net);
            cvWaitKey(10);
    for(i = 0; i < net->n; ++i){
        net->layers[i].batch = b;
#ifdef CUDNN
        if(net->layers[i].type == CONVOLUTIONAL){
         layer *l = net->layers + i;
            cudnn_convolutional_setup(l, cudnn_fastest);
         // check for excessive memory consumption 
         size_t free_byte;
         size_t total_byte;
         check_error(cudaMemGetInfo(&free_byte, &total_byte));
         if (l->workspace_size > free_byte || l->workspace_size >= total_byte / 2) {
            printf(" used slow CUDNN algo without Workspace! \n");
            cudnn_convolutional_setup(l, cudnn_smallest);
            l->workspace_size = get_workspace_size(*l);
         }
        }
#endif
    }
    visualize_network(net);
    cvWaitKey(100);
    fprintf(stderr, "Accuracy: %f\n", (float)correct/d.X.rows);
}

int get_network_input_size_layer(network net, int i)
int resize_network(network *net, int w, int h)
{
    if(net.types[i] == CONVOLUTIONAL){
        convolutional_layer layer = *(convolutional_layer *)net.layers[i];
        return layer.h*layer.w*layer.c;
#ifdef GPU
    cuda_set_device(net->gpu_index);
    if(gpu_index >= 0){
        cuda_free(net->workspace);
      if (net->input_gpu) {
         cuda_free(*net->input_gpu);
         *net->input_gpu = 0;
         cuda_free(*net->truth_gpu);
         *net->truth_gpu = 0;
      }
    }
    else if(net.types[i] == MAXPOOL){
        maxpool_layer layer = *(maxpool_layer *)net.layers[i];
        return layer.h*layer.w*layer.c;
    }
    else if(net.types[i] == CONNECTED){
        connected_layer layer = *(connected_layer *)net.layers[i];
        return layer.inputs;
    } else if(net.types[i] == DROPOUT){
        dropout_layer layer = *(dropout_layer *) net.layers[i];
        return layer.inputs;
    }
    else if(net.types[i] == FREEWEIGHT){
        freeweight_layer layer = *(freeweight_layer *) net.layers[i];
        return layer.inputs;
    }
    else if(net.types[i] == SOFTMAX){
        softmax_layer layer = *(softmax_layer *)net.layers[i];
        return layer.inputs;
    }
    return 0;
}

int get_network_output_size_layer(network net, int i)
{
    if(net.types[i] == CONVOLUTIONAL){
        convolutional_layer layer = *(convolutional_layer *)net.layers[i];
        image output = get_convolutional_image(layer);
        return output.h*output.w*output.c;
    }
    else if(net.types[i] == MAXPOOL){
        maxpool_layer layer = *(maxpool_layer *)net.layers[i];
        image output = get_maxpool_image(layer);
        return output.h*output.w*output.c;
    }
    else if(net.types[i] == CONNECTED){
        connected_layer layer = *(connected_layer *)net.layers[i];
        return layer.outputs;
    }
    else if(net.types[i] == DROPOUT){
        dropout_layer layer = *(dropout_layer *) net.layers[i];
        return layer.inputs;
    }
    else if(net.types[i] == FREEWEIGHT){
        freeweight_layer layer = *(freeweight_layer *) net.layers[i];
        return layer.inputs;
    }
    else if(net.types[i] == SOFTMAX){
        softmax_layer layer = *(softmax_layer *)net.layers[i];
        return layer.inputs;
    }
    return 0;
}

int resize_network(network net, int h, int w, int c)
{
#endif
    int i;
    for (i = 0; i < net.n; ++i){
        if(net.types[i] == CONVOLUTIONAL){
            convolutional_layer *layer = (convolutional_layer *)net.layers[i];
            resize_convolutional_layer(layer, h, w, c);
            image output = get_convolutional_image(*layer);
            h = output.h;
            w = output.w;
            c = output.c;
        }else if(net.types[i] == MAXPOOL){
            maxpool_layer *layer = (maxpool_layer *)net.layers[i];
            resize_maxpool_layer(layer, h, w, c);
            image output = get_maxpool_image(*layer);
            h = output.h;
            w = output.w;
            c = output.c;
        }else if(net.types[i] == NORMALIZATION){
            normalization_layer *layer = (normalization_layer *)net.layers[i];
            resize_normalization_layer(layer, h, w, c);
            image output = get_normalization_image(*layer);
            h = output.h;
            w = output.w;
            c = output.c;
    //if(w == net->w && h == net->h) return 0;
    net->w = w;
    net->h = h;
    int inputs = 0;
    size_t workspace_size = 0;
    //fprintf(stderr, "Resizing to %d x %d...\n", w, h);
    //fflush(stderr);
    for (i = 0; i < net->n; ++i){
        layer l = net->layers[i];
      printf(" %d: layer = %d,", i, l.type);
        if(l.type == CONVOLUTIONAL){
            resize_convolutional_layer(&l, w, h);
        }else if(l.type == CROP){
            resize_crop_layer(&l, w, h);
        }else if(l.type == MAXPOOL){
            resize_maxpool_layer(&l, w, h);
        }else if(l.type == REGION){
            resize_region_layer(&l, w, h);
        }else if(l.type == ROUTE){
            resize_route_layer(&l, net);
        }else if(l.type == REORG){
            resize_reorg_layer(&l, w, h);
        }else if(l.type == AVGPOOL){
            resize_avgpool_layer(&l, w, h);
        }else if(l.type == NORMALIZATION){
            resize_normalization_layer(&l, w, h);
        }else if(l.type == COST){
            resize_cost_layer(&l, inputs);
        }else{
         fprintf(stderr, "Resizing type %d \n", (int)l.type);
            error("Cannot resize this type of layer");
        }
        if(l.workspace_size > workspace_size) workspace_size = l.workspace_size;
        inputs = l.outputs;
        net->layers[i] = l;
        w = l.out_w;
        h = l.out_h;
        if(l.type == AVGPOOL) break;
    }
#ifdef GPU
    if(gpu_index >= 0){
      printf(" try to allocate workspace = %zu * sizeof(float), ", (workspace_size - 1) / sizeof(float) + 1);
        net->workspace = cuda_make_array(0, (workspace_size-1)/sizeof(float)+1);
      printf(" CUDA allocate done! \n");
    }else {
        free(net->workspace);
        net->workspace = calloc(1, workspace_size);
    }
#else
    free(net->workspace);
    net->workspace = calloc(1, workspace_size);
#endif
    //fprintf(stderr, " Done!\n");
    return 0;
}

int get_network_output_size(network net)
{
    int i;
    for(i = net.n-1; i > 0; --i) if(net.types[i] != COST) break;
    return get_network_output_size_layer(net, i);
    for(i = net.n-1; i > 0; --i) if(net.layers[i].type != COST) break;
    return net.layers[i].outputs;
}

int get_network_input_size(network net)
{
    return get_network_input_size_layer(net, 0);
    return net.layers[0].inputs;
}

detection_layer get_network_detection_layer(network net)
{
    int i;
    for(i = 0; i < net.n; ++i){
        if(net.layers[i].type == DETECTION){
            return net.layers[i];
        }
    }
    fprintf(stderr, "Detection layer not found!!\n");
    detection_layer l = {0};
    return l;
}

image get_network_image_layer(network net, int i)
{
    if(net.types[i] == CONVOLUTIONAL){
        convolutional_layer layer = *(convolutional_layer *)net.layers[i];
        return get_convolutional_image(layer);
    layer l = net.layers[i];
    if (l.out_w && l.out_h && l.out_c){
        return float_to_image(l.out_w, l.out_h, l.out_c, l.output);
    }
    else if(net.types[i] == MAXPOOL){
        maxpool_layer layer = *(maxpool_layer *)net.layers[i];
        return get_maxpool_image(layer);
    }
    else if(net.types[i] == NORMALIZATION){
        normalization_layer layer = *(normalization_layer *)net.layers[i];
        return get_normalization_image(layer);
    }
    else if(net.types[i] == CROP){
        crop_layer layer = *(crop_layer *)net.layers[i];
        return get_crop_image(layer);
    }
    return make_empty_image(0,0,0);
    image def = {0};
    return def;
}

image get_network_image(network net)
@@ -613,7 +445,8 @@
        image m = get_network_image_layer(net, i);
        if(m.h != 0) return m;
    }
    return make_empty_image(0,0,0);
    image def = {0};
    return def;
}

void visualize_network(network net)
@@ -621,23 +454,37 @@
    image *prev = 0;
    int i;
    char buff[256];
    show_image(get_network_image_layer(net, 0), "Crop");
    for(i = 0; i < net.n; ++i){
        sprintf(buff, "Layer %d", i);
        if(net.types[i] == CONVOLUTIONAL){
            convolutional_layer layer = *(convolutional_layer *)net.layers[i];
            prev = visualize_convolutional_layer(layer, buff, prev);
        }
        if(net.types[i] == NORMALIZATION){
            normalization_layer layer = *(normalization_layer *)net.layers[i];
            visualize_normalization_layer(layer, buff);
        layer l = net.layers[i];
        if(l.type == CONVOLUTIONAL){
            prev = visualize_convolutional_layer(l, buff, prev);
        }
    } 
}

void top_predictions(network net, int k, int *index)
{
    int size = get_network_output_size(net);
    float *out = get_network_output(net);
    top_k(out, size, k, index);
}


float *network_predict(network net, float *input)
{
    forward_network(net, input, 0, 0);
#ifdef GPU
    if(gpu_index >= 0)  return network_predict_gpu(net, input);
#endif

    network_state state;
    state.net = net;
    state.index = 0;
    state.input = input;
    state.truth = 0;
    state.train = 0;
    state.delta = 0;
    forward_network(net, state);
    float *out = get_network_output(net);
    return out;
}
@@ -672,7 +519,7 @@
    int i,j,b;
    int k = get_network_output_size(net);
    matrix pred = make_matrix(test.X.rows, k);
    float *X = calloc(net.batch*test.X.rows, sizeof(float));
    float *X = calloc(net.batch*test.X.cols, sizeof(float));
    for(i = 0; i < test.X.rows; i += net.batch){
        for(b = 0; b < net.batch; ++b){
            if(i+b == test.X.rows) break;
@@ -694,36 +541,9 @@
{
    int i,j;
    for(i = 0; i < net.n; ++i){
        float *output = 0;
        int n = 0;
        if(net.types[i] == CONVOLUTIONAL){
            convolutional_layer layer = *(convolutional_layer *)net.layers[i];
            output = layer.output;
            image m = get_convolutional_image(layer);
            n = m.h*m.w*m.c;
        }
        else if(net.types[i] == MAXPOOL){
            maxpool_layer layer = *(maxpool_layer *)net.layers[i];
            output = layer.output;
            image m = get_maxpool_image(layer);
            n = m.h*m.w*m.c;
        }
        else if(net.types[i] == CROP){
            crop_layer layer = *(crop_layer *)net.layers[i];
            output = layer.output;
            image m = get_crop_image(layer);
            n = m.h*m.w*m.c;
        }
        else if(net.types[i] == CONNECTED){
            connected_layer layer = *(connected_layer *)net.layers[i];
            output = layer.output;
            n = layer.outputs;
        }
        else if(net.types[i] == SOFTMAX){
            softmax_layer layer = *(softmax_layer *)net.layers[i];
            output = layer.output;
            n = layer.inputs;
        }
        layer l = net.layers[i];
        float *output = l.output;
        int n = l.outputs;
        float mean = mean_array(output, n);
        float vari = variance_array(output, n);
        fprintf(stderr, "Layer %d - Mean: %f, Variance: %f\n",i,mean, vari);
@@ -734,10 +554,45 @@
    }
}

void compare_networks(network n1, network n2, data test)
{
    matrix g1 = network_predict_data(n1, test);
    matrix g2 = network_predict_data(n2, test);
    int i;
    int a,b,c,d;
    a = b = c = d = 0;
    for(i = 0; i < g1.rows; ++i){
        int truth = max_index(test.y.vals[i], test.y.cols);
        int p1 = max_index(g1.vals[i], g1.cols);
        int p2 = max_index(g2.vals[i], g2.cols);
        if(p1 == truth){
            if(p2 == truth) ++d;
            else ++c;
        }else{
            if(p2 == truth) ++b;
            else ++a;
        }
    }
    printf("%5d %5d\n%5d %5d\n", a, b, c, d);
    float num = pow((abs(b - c) - 1.), 2.);
    float den = b + c;
    printf("%f\n", num/den); 
}

float network_accuracy(network net, data d)
{
    matrix guess = network_predict_data(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(network net, data d, int n)
{
    static float acc[2];
    matrix guess = network_predict_data(net, d);
    acc[0] = matrix_topk_accuracy(d.y, guess, 1);
    acc[1] = matrix_topk_accuracy(d.y, guess, n);
    free_matrix(guess);
    return acc;
}
@@ -745,9 +600,26 @@
float network_accuracy_multi(network net, data d, int n)
{
    matrix guess = network_predict_data_multi(net, d, n);
    float acc = matrix_accuracy(d.y, guess);
    float acc = matrix_topk_accuracy(d.y, guess,1);
    free_matrix(guess);
    return acc;
}


void free_network(network net)
{
   int i;
   for (i = 0; i < net.n; ++i) {
      free_layer(net.layers[i]);
   }
   free(net.layers);
#ifdef GPU
   if (gpu_index >= 0) cuda_free(net.workspace);
   else free(net.workspace);
   if (*net.input_gpu) cuda_free(*net.input_gpu);
   if (*net.truth_gpu) cuda_free(*net.truth_gpu);
   if (net.input_gpu) free(net.input_gpu);
   if (net.truth_gpu) free(net.truth_gpu);
#else
   free(net.workspace);
#endif
}