probably how maxpool layers should be

Joseph Redmon

2014-08-08 d9f1b0b16edeb59281355a855e18a8be343fc33c

 src/connected_layer.c

@@ -1,96 +1,104 @@
#include "connected_layer.h"
#include "utils.h"
#include "mini_blas.h"

#include <math.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>

connected_layer *make_connected_layer(int inputs, int outputs, ACTIVATION activator)
connected_layer *make_connected_layer(int batch, int inputs, int outputs, ACTIVATION activation, float learning_rate, float momentum, float decay)
{
    fprintf(stderr, "Connected Layer: %d inputs, %d outputs\n", inputs, outputs);
    int i;
    connected_layer *layer = calloc(1, sizeof(connected_layer));

    layer->learning_rate = learning_rate;
    layer->momentum = momentum;
    layer->decay = decay;

    layer->inputs = inputs;
    layer->outputs = outputs;
    layer->batch=batch;

    layer->output = calloc(outputs, sizeof(double*));
    layer->output = calloc(batch*outputs, sizeof(float*));
    layer->delta = calloc(batch*outputs, sizeof(float*));

    layer->weight_updates = calloc(inputs*outputs, sizeof(double));
    layer->weights = calloc(inputs*outputs, sizeof(double));
    layer->weight_updates = calloc(inputs*outputs, sizeof(float));
    layer->weight_adapt = calloc(inputs*outputs, sizeof(float));
    layer->weight_momentum = calloc(inputs*outputs, sizeof(float));
    layer->weights = calloc(inputs*outputs, sizeof(float));
    float scale = 1./inputs;
    //scale = .01;
    for(i = 0; i < inputs*outputs; ++i)
        layer->weights[i] = .5 - (double)rand()/RAND_MAX;
        layer->weights[i] = scale*(rand_uniform()-.5);

    layer->bias_updates = calloc(outputs, sizeof(double));
    layer->biases = calloc(outputs, sizeof(double));
    layer->bias_updates = calloc(outputs, sizeof(float));
    layer->bias_adapt = calloc(outputs, sizeof(float));
    layer->bias_momentum = calloc(outputs, sizeof(float));
    layer->biases = calloc(outputs, sizeof(float));
    for(i = 0; i < outputs; ++i)
        layer->biases[i] = (double)rand()/RAND_MAX;
        //layer->biases[i] = rand_normal()*scale + scale;
        layer->biases[i] = 1;

    if(activator == SIGMOID){
        layer->activation = sigmoid_activation;
        layer->gradient = sigmoid_gradient;
    }else if(activator == RELU){
        layer->activation = relu_activation;
        layer->gradient = relu_gradient;
    }else if(activator == IDENTITY){
        layer->activation = identity_activation;
        layer->gradient = identity_gradient;
    }

    layer->activation = activation;
    return layer;
}

void run_connected_layer(double *input, connected_layer layer)
void update_connected_layer(connected_layer layer)
{
    int i, j;
    int i;
    for(i = 0; i < layer.outputs; ++i){
        layer.output[i] = layer.biases[i];
        for(j = 0; j < layer.inputs; ++j){
            layer.output[i] += input[j]*layer.weights[i*layer.inputs + j];
        }
        layer.output[i] = layer.activation(layer.output[i]);
        layer.bias_momentum[i] = layer.learning_rate*(layer.bias_updates[i]) + layer.momentum*layer.bias_momentum[i];
        layer.biases[i] += layer.bias_momentum[i];
    }
    for(i = 0; i < layer.outputs*layer.inputs; ++i){
        layer.weight_momentum[i] = layer.learning_rate*(layer.weight_updates[i] - layer.decay*layer.weights[i]) + layer.momentum*layer.weight_momentum[i];
        layer.weights[i] += layer.weight_momentum[i];
    }
    memset(layer.bias_updates, 0, layer.outputs*sizeof(float));
    memset(layer.weight_updates, 0, layer.outputs*layer.inputs*sizeof(float));
}

void learn_connected_layer(double *input, connected_layer layer)
void forward_connected_layer(connected_layer layer, float *input)
{
    calculate_update_connected_layer(input, layer);
    backpropagate_connected_layer(input, layer);
    int i;
    for(i = 0; i < layer.batch; ++i){
        memcpy(layer.output+i*layer.outputs, layer.biases, layer.outputs*sizeof(float));
    }
    int m = layer.batch;
    int k = layer.inputs;
    int n = layer.outputs;
    float *a = input;
    float *b = layer.weights;
    float *c = layer.output;
    gemm(0,0,m,n,k,1,a,k,b,n,1,c,n);
    activate_array(layer.output, layer.outputs*layer.batch, layer.activation);
}

void update_connected_layer(connected_layer layer, double step)
void backward_connected_layer(connected_layer layer, float *input, float *delta)
{
    int i,j;
    for(i = 0; i < layer.outputs; ++i){
        layer.biases[i] += step*layer.bias_updates[i];
        for(j = 0; j < layer.inputs; ++j){
            int index = i*layer.inputs+j;
            layer.weights[index] += step*layer.weight_updates[index];
        }
    int i;
    for(i = 0; i < layer.outputs*layer.batch; ++i){
        layer.delta[i] *= gradient(layer.output[i], layer.activation);
        layer.bias_updates[i%layer.outputs] += layer.delta[i];
    }
    memset(layer.bias_updates, 0, layer.outputs*sizeof(double));
    memset(layer.weight_updates, 0, layer.outputs*layer.inputs*sizeof(double));
}
    int m = layer.inputs;
    int k = layer.batch;
    int n = layer.outputs;
    float *a = input;
    float *b = layer.delta;
    float *c = layer.weight_updates;
    gemm(1,0,m,n,k,1,a,m,b,n,1,c,n);

void calculate_update_connected_layer(double *input, connected_layer layer)
{
    int i, j;
    for(i = 0; i < layer.outputs; ++i){
        layer.bias_updates[i] += layer.output[i];
        for(j = 0; j < layer.inputs; ++j){
            layer.weight_updates[i*layer.inputs + j] += layer.output[i]*input[j];
        }
    }
}
    m = layer.batch;
    k = layer.outputs;
    n = layer.inputs;

void backpropagate_connected_layer(double *input, connected_layer layer)
{
    int i, j;
    a = layer.delta;
    b = layer.weights;
    c = delta;

    for(j = 0; j < layer.inputs; ++j){
        double grad = layer.gradient(input[j]);
        input[j] = 0;
        for(i = 0; i < layer.outputs; ++i){
            input[j] += layer.output[i]*layer.weights[i*layer.inputs + j];
        }
        input[j] *= grad;
    }
    if(c) gemm(0,1,m,n,k,1,a,k,b,k,0,c,n);
}