updates

Joseph Redmon

2016-07-19 9361292c429c0ba3400c31c7fa5d5e3d3cb6ab47

 src/network.c

@@ -16,6 +16,7 @@
#include "activation_layer.h"
#include "deconvolutional_layer.h"
#include "detection_layer.h"
#include "region_layer.h"
#include "normalization_layer.h"
#include "batchnorm_layer.h"
#include "maxpool_layer.h"
@@ -64,7 +65,10 @@
        case EXP:
            return net.learning_rate * pow(net.gamma, batch_num);
        case POLY:
            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:
@@ -100,6 +104,8 @@
            return "softmax";
        case DETECTION:
            return "detection";
        case REGION:
            return "region";
        case DROPOUT:
            return "dropout";
        case CROP:
@@ -135,6 +141,7 @@

void forward_network(network net, network_state state)
{
    state.workspace = net.workspace;
    int i;
    for(i = 0; i < net.n; ++i){
        state.index = i;
@@ -156,6 +163,8 @@
            forward_batchnorm_layer(l, state);
        } else if(l.type == DETECTION){
            forward_detection_layer(l, state);
        } else if(l.type == REGION){
            forward_region_layer(l, state);
        } else if(l.type == CONNECTED){
            forward_connected_layer(l, state);
        } else if(l.type == RNN){
@@ -226,11 +235,7 @@
    float sum = 0;
    int count = 0;
    for(i = 0; i < net.n; ++i){
        if(net.layers[i].type == COST){
            sum += net.layers[i].cost[0];
            ++count;
        }
        if(net.layers[i].type == DETECTION){
        if(net.layers[i].cost){
            sum += net.layers[i].cost[0];
            ++count;
        }
@@ -250,6 +255,7 @@
    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){
@@ -279,6 +285,8 @@
            backward_dropout_layer(l, state);
        } else if(l.type == DETECTION){
            backward_detection_layer(l, state);
        } else if(l.type == REGION){
            backward_region_layer(l, state);
        } else if(l.type == SOFTMAX){
            if(i != 0) backward_softmax_layer(l, state);
        } else if(l.type == CONNECTED){
@@ -388,6 +396,11 @@
    int i;
    for(i = 0; i < net->n; ++i){
        net->layers[i].batch = b;
        #ifdef CUDNN
        if(net->layers[i].type == CONVOLUTIONAL){
            cudnn_convolutional_setup(net->layers + i);
        }
        #endif
    }
}

@@ -398,6 +411,7 @@
    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){
@@ -417,12 +431,20 @@
        }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);
#endif
    //fprintf(stderr, " Done!\n");
    return 0;
}