some changes, fix opencv

Joseph Redmon

2016-06-09 8a6ba2fff3ee1c14bca0aa0e0a909aba7057cc94

 src/network.c

@@ -4,29 +4,96 @@
#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 "activation_layer.h"
#include "deconvolutional_layer.h"
#include "detection_layer.h"
#include "normalization_layer.h"
#include "batchnorm_layer.h"
#include "maxpool_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"

int get_current_batch(network net)
{
    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
        if(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;
    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) 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);
        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 AVGPOOL:
@@ -43,8 +110,12 @@
            return "cost";
        case ROUTE:
            return "route";
        case SHORTCUT:
            return "shortcut";
        case NORMALIZATION:
            return "normalization";
        case BATCHNORM:
            return "batchnorm";
        default:
            break;
    }
@@ -56,6 +127,7 @@
    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 *));
@@ -65,8 +137,10 @@

void forward_network(network net, network_state state)
{
    state.workspace = net.workspace;
    int i;
    for(i = 0; i < net.n; ++i){
        state.index = i;
        layer l = net.layers[i];
        if(l.delta){
            scal_cpu(l.outputs * l.batch, 0, l.delta, 1);
@@ -75,12 +149,24 @@
            forward_convolutional_layer(l, state);
        } else if(l.type == DECONVOLUTIONAL){
            forward_deconvolutional_layer(l, state);
        } else if(l.type == ACTIVE){
            forward_activation_layer(l, state);
        } else if(l.type == LOCAL){
            forward_local_layer(l, state);
        } else if(l.type == NORMALIZATION){
            forward_normalization_layer(l, state);
        } else if(l.type == BATCHNORM){
            forward_batchnorm_layer(l, state);
        } else if(l.type == DETECTION){
            forward_detection_layer(l, state);
        } else if(l.type == CONNECTED){
            forward_connected_layer(l, state);
        } else if(l.type == RNN){
            forward_rnn_layer(l, state);
        } else if(l.type == GRU){
            forward_gru_layer(l, state);
        } else if(l.type == CRNN){
            forward_crnn_layer(l, state);
        } else if(l.type == CROP){
            forward_crop_layer(l, state);
        } else if(l.type == COST){
@@ -95,6 +181,8 @@
            forward_dropout_layer(l, state);
        } else if(l.type == ROUTE){
            forward_route_layer(l, net);
        } else if(l.type == SHORTCUT){
            forward_shortcut_layer(l, state);
        }
        state.input = l.output;
    }
@@ -104,20 +192,32 @@
{
    int i;
    int update_batch = net.batch*net.subdivisions;
    float rate = get_current_rate(net);
    for(i = 0; i < net.n; ++i){
        layer l = net.layers[i];
        if(l.type == CONVOLUTIONAL){
            update_convolutional_layer(l, update_batch, net.learning_rate, net.momentum, net.decay);
            update_convolutional_layer(l, update_batch, rate, net.momentum, net.decay);
        } else if(l.type == DECONVOLUTIONAL){
            update_deconvolutional_layer(l, net.learning_rate, net.momentum, net.decay);
            update_deconvolutional_layer(l, rate, net.momentum, net.decay);
        } else if(l.type == CONNECTED){
            update_connected_layer(l, update_batch, net.learning_rate, net.momentum, net.decay);
            update_connected_layer(l, update_batch, rate, net.momentum, net.decay);
        } else if(l.type == RNN){
            update_rnn_layer(l, update_batch, rate, net.momentum, net.decay);
        } else if(l.type == GRU){
            update_gru_layer(l, update_batch, rate, net.momentum, net.decay);
        } else if(l.type == CRNN){
            update_crnn_layer(l, update_batch, rate, net.momentum, net.decay);
        } else if(l.type == LOCAL){
            update_local_layer(l, update_batch, rate, net.momentum, net.decay);
        }
    }
}

float *get_network_output(network net)
{
    #ifdef GPU
        return get_network_output_gpu(net);
    #endif 
    int i;
    for(i = net.n-1; i > 0; --i) if(net.layers[i].type != COST) break;
    return net.layers[i].output;
@@ -129,12 +229,12 @@
    float sum = 0;
    int count = 0;
    for(i = 0; i < net.n; ++i){
        if(net.layers[net.n-1].type == COST){
            sum += net.layers[net.n-1].output[0];
        if(net.layers[i].type == COST){
            sum += net.layers[i].cost[0];
            ++count;
        }
        if(net.layers[net.n-1].type == DETECTION){
            sum += net.layers[net.n-1].cost[0];
        if(net.layers[i].type == DETECTION){
            sum += net.layers[i].cost[0];
            ++count;
        }
    }
@@ -153,7 +253,9 @@
    int i;
    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){
            state.input = original_input;
            state.delta = original_delta;
@@ -167,8 +269,12 @@
            backward_convolutional_layer(l, state);
        } else if(l.type == DECONVOLUTIONAL){
            backward_deconvolutional_layer(l, state);
        } else if(l.type == ACTIVE){
            backward_activation_layer(l, state);
        } else if(l.type == NORMALIZATION){
            backward_normalization_layer(l, state);
        } else if(l.type == BATCHNORM){
            backward_batchnorm_layer(l, state);
        } else if(l.type == MAXPOOL){
            if(i != 0) backward_maxpool_layer(l, state);
        } else if(l.type == AVGPOOL){
@@ -181,20 +287,33 @@
            if(i != 0) backward_softmax_layer(l, state);
        } else if(l.type == CONNECTED){
            backward_connected_layer(l, state);
        } else if(l.type == RNN){
            backward_rnn_layer(l, state);
        } else if(l.type == GRU){
            backward_gru_layer(l, state);
        } else if(l.type == CRNN){
            backward_crnn_layer(l, state);
        } else if(l.type == LOCAL){
            backward_local_layer(l, state);
        } else if(l.type == COST){
            backward_cost_layer(l, state);
        } else if(l.type == ROUTE){
            backward_route_layer(l, net);
        } else if(l.type == SHORTCUT){
            backward_shortcut_layer(l, state);
        }
    }
}

float train_network_datum(network net, float *x, float *y)
{
    *net.seen += net.batch;
#ifdef GPU
    if(gpu_index >= 0) return train_network_datum_gpu(net, x, y);
#endif
    network_state state;
    state.index = 0;
    state.net = net;
    state.input = x;
    state.delta = 0;
    state.truth = y;
@@ -202,7 +321,7 @@
    forward_network(net, state);
    backward_network(net, state);
    float error = get_network_cost(net);
    if((net.seen/net.batch)%net.subdivisions == 0) update_network(net);
    if(((*net.seen)/net.batch)%net.subdivisions == 0) update_network(net);
    return error;
}

@@ -215,7 +334,6 @@
    int i;
    float sum = 0;
    for(i = 0; i < n; ++i){
        net.seen += batch;
        get_random_batch(d, batch, X, y);
        float err = train_network_datum(net, X, y);
        sum += err;
@@ -236,7 +354,6 @@
    float sum = 0;
    for(i = 0; i < n; ++i){
        get_next_batch(d, batch, i*batch, X, y);
        net.seen += batch;
        float err = train_network_datum(net, X, y);
        sum += err;
    }
@@ -249,6 +366,8 @@
{
    int i,j;
    network_state state;
    state.index = 0;
    state.net = net;
    state.train = 1;
    state.delta = 0;
    float sum = 0;
@@ -282,26 +401,41 @@
    //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...", w, h);
    //fflush(stderr);
    for (i = 0; i < net->n; ++i){
        layer l = net->layers[i];
        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 == AVGPOOL){
            resize_avgpool_layer(&l, w, h);
            break;
        }else if(l.type == NORMALIZATION){
            resize_normalization_layer(&l, w, h);
        }else if(l.type == COST){
            resize_cost_layer(&l, inputs);
        }else{
            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
        cuda_free(net->workspace);
        net->workspace = cuda_make_array(0, (workspace_size-1)/sizeof(float)+1);
#else
        free(net->workspace);
        net->workspace = calloc(1, (workspace_size-1)/sizeof(float)+1);
#endif
    //fprintf(stderr, " Done!\n");
    return 0;
}
@@ -381,6 +515,8 @@
#endif

    network_state state;
    state.net = net;
    state.index = 0;
    state.input = input;
    state.truth = 0;
    state.train = 0;
@@ -488,12 +624,12 @@
    return acc;
}

float *network_accuracies(network net, data d)
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,5);
    acc[0] = matrix_topk_accuracy(d.y, guess, 1);
    acc[1] = matrix_topk_accuracy(d.y, guess, n);
    free_matrix(guess);
    return acc;
}
@@ -507,4 +643,17 @@
    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(*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);
    #endif
}