Better imagenet distributed training

Joseph Redmon

2014-12-11 79fffcce3ce495bd415dc1284224c915d7194d4c

 src/connected_layer.c

@@ -25,8 +25,8 @@

    layer->weight_updates = calloc(inputs*outputs, sizeof(float));
    layer->weights = calloc(inputs*outputs, sizeof(float));
    float scale = 1./inputs;
    scale = .01;
    float scale = 1./sqrt(inputs);
    //scale = .01;
    for(i = 0; i < inputs*outputs; ++i){
        layer->weights[i] = scale*rand_normal();
    }
@@ -34,7 +34,7 @@
    layer->bias_updates = calloc(outputs, sizeof(float));
    layer->biases = calloc(outputs, sizeof(float));
    for(i = 0; i < outputs; ++i){
        layer->biases[i] = .01;
        layer->biases[i] = scale;
    }

#ifdef GPU