From d9f1b0b16edeb59281355a855e18a8be343fc33c Mon Sep 17 00:00:00 2001
From: Joseph Redmon <pjreddie@gmail.com>
Date: Fri, 08 Aug 2014 19:04:15 +0000
Subject: [PATCH] probably how maxpool layers should be
---
src/connected_layer.c | 124 +++++++++++++++++++++++------------------
1 files changed, 69 insertions(+), 55 deletions(-)
diff --git a/src/connected_layer.c b/src/connected_layer.c
index 99f146b..368fb63 100644
--- a/src/connected_layer.c
+++ b/src/connected_layer.c
@@ -1,32 +1,42 @@
#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 activation)
+connected_layer *make_connected_layer(int batch, int inputs, int outputs, ACTIVATION activation, float learning_rate, float momentum, float decay)
{
- printf("Connected Layer: %d inputs, %d outputs\n", inputs, outputs);
+ 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->delta = 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->weight_momentum = calloc(inputs*outputs, sizeof(double));
- layer->weights = calloc(inputs*outputs, sizeof(double));
- double scale = 2./inputs;
+ 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] = rand_normal()*scale;
+ layer->weights[i] = scale*(rand_uniform()-.5);
- layer->bias_updates = calloc(outputs, sizeof(double));
- layer->bias_momentum = 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] = rand_normal()*scale + scale;
layer->biases[i] = 1;
@@ -35,56 +45,60 @@
return layer;
}
-void forward_connected_layer(connected_layer layer, double *input)
+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] = activate(layer.output[i], layer.activation);
- }
-}
-
-void learn_connected_layer(connected_layer layer, double *input)
-{
- int i, j;
- for(i = 0; i < layer.outputs; ++i){
- layer.delta[i] *= gradient(layer.output[i], layer.activation);
- layer.bias_updates[i] += layer.delta[i];
- for(j = 0; j < layer.inputs; ++j){
- layer.weight_updates[i*layer.inputs + j] += layer.delta[i]*input[j];
- }
- }
-}
-
-void update_connected_layer(connected_layer layer, double step, double momentum, double decay)
-{
- int i,j;
- for(i = 0; i < layer.outputs; ++i){
- layer.bias_momentum[i] = step*(layer.bias_updates[i]) + momentum*layer.bias_momentum[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(j = 0; j < layer.inputs; ++j){
- int index = i*layer.inputs+j;
- layer.weight_momentum[index] = step*(layer.weight_updates[index] - decay*layer.weights[index]) + momentum*layer.weight_momentum[index];
- layer.weights[index] += layer.weight_momentum[index];
- //layer.weights[index] = constrain(layer.weights[index], 100.);
- }
}
- memset(layer.bias_updates, 0, layer.outputs*sizeof(double));
- memset(layer.weight_updates, 0, layer.outputs*layer.inputs*sizeof(double));
+ 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 backward_connected_layer(connected_layer layer, double *input, double *delta)
+void forward_connected_layer(connected_layer layer, float *input)
{
- int i, j;
-
- for(j = 0; j < layer.inputs; ++j){
- delta[j] = 0;
- for(i = 0; i < layer.outputs; ++i){
- delta[j] += layer.delta[i]*layer.weights[i*layer.inputs + j];
- }
+ 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 backward_connected_layer(connected_layer layer, float *input, float *delta)
+{
+ 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];
+ }
+ 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);
+
+ m = layer.batch;
+ k = layer.outputs;
+ n = layer.inputs;
+
+ a = layer.delta;
+ b = layer.weights;
+ c = delta;
+
+ if(c) gemm(0,1,m,n,k,1,a,k,b,k,0,c,n);
}
--
Gitblit v1.10.0