From 2ea63c0e99a5358eaf38785ea83b9c5923fcc9cd Mon Sep 17 00:00:00 2001
From: Joseph Redmon <pjreddie@gmail.com>
Date: Thu, 13 Mar 2014 04:57:34 +0000
Subject: [PATCH] Better VOC handling and resizing
---
src/convolutional_layer.c | 184 +++++++++++++++++++++++++++++++--------------
1 files changed, 127 insertions(+), 57 deletions(-)
diff --git a/src/convolutional_layer.c b/src/convolutional_layer.c
index 53eb7bf..f7c9c10 100644
--- a/src/convolutional_layer.c
+++ b/src/convolutional_layer.c
@@ -3,59 +3,67 @@
#include "mini_blas.h"
#include <stdio.h>
+int convolutional_out_height(convolutional_layer layer)
+{
+ return (layer.h-layer.size)/layer.stride + 1;
+}
+
+int convolutional_out_width(convolutional_layer layer)
+{
+ return (layer.w-layer.size)/layer.stride + 1;
+}
+
image get_convolutional_image(convolutional_layer layer)
{
int h,w,c;
- h = layer.out_h;
- w = layer.out_w;
+ h = convolutional_out_height(layer);
+ w = convolutional_out_width(layer);
c = layer.n;
- return double_to_image(h,w,c,layer.output);
+ return float_to_image(h,w,c,layer.output);
}
image get_convolutional_delta(convolutional_layer layer)
{
int h,w,c;
- h = layer.out_h;
- w = layer.out_w;
+ h = convolutional_out_height(layer);
+ w = convolutional_out_width(layer);
c = layer.n;
- return double_to_image(h,w,c,layer.delta);
+ return float_to_image(h,w,c,layer.delta);
}
-convolutional_layer *make_convolutional_layer(int h, int w, int c, int n, int size, int stride, ACTIVATION activation)
+convolutional_layer *make_convolutional_layer(int batch, int h, int w, int c, int n, int size, int stride, ACTIVATION activation)
{
int i;
- int out_h,out_w;
size = 2*(size/2)+1; //HA! And you thought you'd use an even sized filter...
convolutional_layer *layer = calloc(1, sizeof(convolutional_layer));
layer->h = h;
layer->w = w;
layer->c = c;
layer->n = n;
+ layer->batch = batch;
layer->stride = stride;
layer->size = size;
- layer->filters = calloc(c*n*size*size, sizeof(double));
- layer->filter_updates = calloc(c*n*size*size, sizeof(double));
- layer->filter_momentum = calloc(c*n*size*size, sizeof(double));
+ layer->filters = calloc(c*n*size*size, sizeof(float));
+ layer->filter_updates = calloc(c*n*size*size, sizeof(float));
+ layer->filter_momentum = calloc(c*n*size*size, sizeof(float));
- layer->biases = calloc(n, sizeof(double));
- layer->bias_updates = calloc(n, sizeof(double));
- layer->bias_momentum = calloc(n, sizeof(double));
- double scale = 2./(size*size);
- for(i = 0; i < c*n*size*size; ++i) layer->filters[i] = rand_normal()*scale;
+ layer->biases = calloc(n, sizeof(float));
+ layer->bias_updates = calloc(n, sizeof(float));
+ layer->bias_momentum = calloc(n, sizeof(float));
+ float scale = 1./(size*size*c);
+ for(i = 0; i < c*n*size*size; ++i) layer->filters[i] = scale*(rand_uniform());
for(i = 0; i < n; ++i){
//layer->biases[i] = rand_normal()*scale + scale;
layer->biases[i] = 0;
}
- out_h = (h-size)/stride + 1;
- out_w = (w-size)/stride + 1;
+ int out_h = convolutional_out_height(*layer);
+ int out_w = convolutional_out_width(*layer);
- layer->col_image = calloc(out_h*out_w*size*size*c, sizeof(double));
- layer->output = calloc(out_h * out_w * n, sizeof(double));
- layer->delta = calloc(out_h * out_w * n, sizeof(double));
+ layer->col_image = calloc(layer->batch*out_h*out_w*size*size*c, sizeof(float));
+ layer->output = calloc(layer->batch*out_h * out_w * n, sizeof(float));
+ layer->delta = calloc(layer->batch*out_h * out_w * n, sizeof(float));
layer->activation = activation;
- layer->out_h = out_h;
- layer->out_w = out_w;
fprintf(stderr, "Convolutional Layer: %d x %d x %d image, %d filters -> %d x %d x %d image\n", h,w,c,n, out_h, out_w, n);
srand(0);
@@ -63,42 +71,73 @@
return layer;
}
-void forward_convolutional_layer(const convolutional_layer layer, double *in)
+void resize_convolutional_layer(convolutional_layer *layer, int h, int w, int c)
{
+ layer->h = h;
+ layer->w = w;
+ layer->c = c;
+ int out_h = convolutional_out_height(*layer);
+ int out_w = convolutional_out_width(*layer);
+
+ layer->col_image = realloc(layer->col_image,
+ layer->batch*out_h*out_w*layer->size*layer->size*layer->c*sizeof(float));
+ layer->output = realloc(layer->output,
+ layer->batch*out_h * out_w * layer->n*sizeof(float));
+ layer->delta = realloc(layer->delta,
+ layer->batch*out_h * out_w * layer->n*sizeof(float));
+}
+
+void forward_convolutional_layer(const convolutional_layer layer, float *in)
+{
+ int i;
int m = layer.n;
int k = layer.size*layer.size*layer.c;
- int n = ((layer.h-layer.size)/layer.stride + 1)*
- ((layer.w-layer.size)/layer.stride + 1);
+ int n = convolutional_out_height(layer)*
+ convolutional_out_width(layer)*
+ layer.batch;
- memset(layer.output, 0, m*n*sizeof(double));
+ memset(layer.output, 0, m*n*sizeof(float));
- double *a = layer.filters;
- double *b = layer.col_image;
- double *c = layer.output;
-
- im2col_cpu(in, layer.c, layer.h, layer.w, layer.size, layer.stride, b);
+ float *a = layer.filters;
+ float *b = layer.col_image;
+ float *c = layer.output;
+ for(i = 0; i < layer.batch; ++i){
+ im2col_cpu(in+i*(n/layer.batch), layer.c, layer.h, layer.w, layer.size, layer.stride, b+i*(n/layer.batch));
+ }
gemm(0,0,m,n,k,1,a,k,b,n,1,c,n);
+ for(i = 0; i < m*n; ++i){
+ layer.output[i] = activate(layer.output[i], layer.activation);
+ }
+ //for(i = 0; i < m*n; ++i) if(i%(m*n/10+1)==0) printf("%f, ", layer.output[i]); printf("\n");
+
}
void gradient_delta_convolutional_layer(convolutional_layer layer)
{
int i;
- for(i = 0; i < layer.out_h*layer.out_w*layer.n; ++i){
+ int size = convolutional_out_height(layer)*
+ convolutional_out_width(layer)*
+ layer.n*
+ layer.batch;
+ for(i = 0; i < size; ++i){
layer.delta[i] *= gradient(layer.output[i], layer.activation);
}
}
void learn_bias_convolutional_layer(convolutional_layer layer)
{
- int i,j;
- int size = layer.out_h*layer.out_w;
- for(i = 0; i < layer.n; ++i){
- double sum = 0;
- for(j = 0; j < size; ++j){
- sum += layer.delta[j+i*size];
+ int i,j,b;
+ int size = convolutional_out_height(layer)
+ *convolutional_out_width(layer);
+ for(b = 0; b < layer.batch; ++b){
+ for(i = 0; i < layer.n; ++i){
+ float sum = 0;
+ for(j = 0; j < size; ++j){
+ sum += layer.delta[j+size*(i+b*layer.n)];
+ }
+ layer.bias_updates[i] += sum/size;
}
- layer.bias_updates[i] += sum/size;
}
}
@@ -108,17 +147,41 @@
learn_bias_convolutional_layer(layer);
int m = layer.n;
int n = layer.size*layer.size*layer.c;
- int k = ((layer.h-layer.size)/layer.stride + 1)*
- ((layer.w-layer.size)/layer.stride + 1);
+ int k = convolutional_out_height(layer)*
+ convolutional_out_width(layer)*
+ layer.batch;
- double *a = layer.delta;
- double *b = layer.col_image;
- double *c = layer.filter_updates;
+ float *a = layer.delta;
+ float *b = layer.col_image;
+ float *c = layer.filter_updates;
gemm(0,1,m,n,k,1,a,k,b,k,1,c,n);
}
-void update_convolutional_layer(convolutional_layer layer, double step, double momentum, double decay)
+void backward_convolutional_layer(convolutional_layer layer, float *delta)
+{
+ int i;
+ int m = layer.size*layer.size*layer.c;
+ int k = layer.n;
+ int n = convolutional_out_height(layer)*
+ convolutional_out_width(layer)*
+ layer.batch;
+
+ float *a = layer.filters;
+ float *b = layer.delta;
+ float *c = layer.col_image;
+
+
+ memset(c, 0, m*n*sizeof(float));
+ gemm(1,0,m,n,k,1,a,m,b,n,1,c,n);
+
+ memset(delta, 0, layer.batch*layer.h*layer.w*layer.c*sizeof(float));
+ for(i = 0; i < layer.batch; ++i){
+ col2im_cpu(c+i*n/layer.batch, layer.c, layer.h, layer.w, layer.size, layer.stride, delta+i*n/layer.batch);
+ }
+}
+
+void update_convolutional_layer(convolutional_layer layer, float step, float momentum, float decay)
{
int i;
int size = layer.size*layer.size*layer.c*layer.n;
@@ -133,9 +196,9 @@
}
/*
-void backward_convolutional_layer2(convolutional_layer layer, double *input, double *delta)
+void backward_convolutional_layer2(convolutional_layer layer, float *input, float *delta)
{
- image in_delta = double_to_image(layer.h, layer.w, layer.c, delta);
+ image in_delta = float_to_image(layer.h, layer.w, layer.c, delta);
image out_delta = get_convolutional_delta(layer);
int i,j;
for(i = 0; i < layer.n; ++i){
@@ -156,10 +219,10 @@
}
-void learn_convolutional_layer(convolutional_layer layer, double *input)
+void learn_convolutional_layer(convolutional_layer layer, float *input)
{
int i;
- image in_image = double_to_image(layer.h, layer.w, layer.c, input);
+ image in_image = float_to_image(layer.h, layer.w, layer.c, input);
image out_delta = get_convolutional_delta(layer);
gradient_delta_convolutional_layer(layer);
for(i = 0; i < layer.n; ++i){
@@ -168,7 +231,7 @@
}
}
-void update_convolutional_layer(convolutional_layer layer, double step, double momentum, double decay)
+void update_convolutional_layer(convolutional_layer layer, float step, float momentum, float decay)
{
int i,j;
for(i = 0; i < layer.n; ++i){
@@ -189,22 +252,29 @@
void test_convolutional_layer()
{
- convolutional_layer l = *make_convolutional_layer(4,4,1,1,3,1,LINEAR);
- double input[] = {1,2,3,4,
+ convolutional_layer l = *make_convolutional_layer(1,4,4,1,1,3,1,LINEAR);
+ float input[] = {1,2,3,4,
5,6,7,8,
9,10,11,12,
13,14,15,16};
- double filter[] = {.5, 0, .3,
+ float filter[] = {.5, 0, .3,
0 , 1, 0,
.2 , 0, 1};
- double delta[] = {1, 2,
+ float delta[] = {1, 2,
3, 4};
+ float in_delta[] = {.5,1,.3,.6,
+ 5,6,7,8,
+ 9,10,11,12,
+ 13,14,15,16};
l.filters = filter;
forward_convolutional_layer(l, input);
l.delta = delta;
learn_convolutional_layer(l);
- image filter_updates = double_to_image(3,3,1,l.filter_updates);
+ image filter_updates = float_to_image(3,3,1,l.filter_updates);
print_image(filter_updates);
+ printf("Delta:\n");
+ backward_convolutional_layer(l, in_delta);
+ pm(4,4,in_delta);
}
image get_convolutional_filter(convolutional_layer layer, int i)
@@ -212,7 +282,7 @@
int h = layer.size;
int w = layer.size;
int c = layer.c;
- return double_to_image(h,w,c,layer.filters+i*h*w*c);
+ return float_to_image(h,w,c,layer.filters+i*h*w*c);
}
void visualize_convolutional_layer(convolutional_layer layer, char *window)
--
Gitblit v1.10.0