MNIST connected network showing off matrices

Joseph Redmon

2014-01-24 ace5aeb0f59fdceb99e607af9780added20da37c

 src/convolutional_layer.c

@@ -39,7 +39,7 @@
    layer->w = w;
    layer->c = c;
    layer->n = n;
    layer->edge = 1;
    layer->edge = 0;
    layer->stride = stride;
    layer->kernels = calloc(n, sizeof(image));
    layer->kernel_updates = calloc(n, sizeof(image));
@@ -143,26 +143,22 @@
    for(i = 0; i < layer.n; ++i){
        kernel_update(in_image, layer.kernel_updates[i], layer.stride, i, out_delta, layer.edge);
        layer.bias_updates[i] += avg_image_layer(out_delta, i);
        //printf("%30.20lf\n", layer.bias_updates[i]);
    }
}

void update_convolutional_layer(convolutional_layer layer, double step, double momentum, double decay)
{
    //step = .01;
    int i,j;
    for(i = 0; i < layer.n; ++i){
        layer.bias_momentum[i] = step*(layer.bias_updates[i]) 
                                + momentum*layer.bias_momentum[i];
        layer.biases[i] += layer.bias_momentum[i];
        //layer.biases[i] = constrain(layer.biases[i],1.);
        layer.bias_updates[i] = 0;
        int pixels = layer.kernels[i].h*layer.kernels[i].w*layer.kernels[i].c;
        for(j = 0; j < pixels; ++j){
            layer.kernel_momentum[i].data[j] = step*(layer.kernel_updates[i].data[j] - decay*layer.kernels[i].data[j]) 
                                                + momentum*layer.kernel_momentum[i].data[j];
            layer.kernels[i].data[j] += layer.kernel_momentum[i].data[j];
            //layer.kernels[i].data[j] = constrain(layer.kernels[i].data[j], 1.);
        }
        zero_image(layer.kernel_updates[i]);
    }
@@ -188,14 +184,6 @@
        int w_offset = i*(size+border);
        image k = layer.kernels[i];
        image copy = copy_image(k);
        /*
        printf("Kernel %d - Bias: %f, Channels:",i,layer.biases[i]);
        for(j = 0; j < k.c; ++j){
            double a = avg_image_layer(k, j);
            printf("%f, ", a);
        }
        printf("\n");
        */
        normalize_image(copy);
        for(j = 0; j < k.c; ++j){
            set_pixel(copy,0,0,j,layer.biases[i]);
@@ -227,7 +215,6 @@
{
    int i;
    char buff[256];
    //image vis = make_image(layer.n*layer.size, layer.size*layer.kernels[0].c, 3);
    for(i = 0; i < layer.n; ++i){
        image k = layer.kernels[i];
        sprintf(buff, "Kernel %d", i);