From f98bf6bbdb5ed81f2ea2071ad8e705130f7ba596 Mon Sep 17 00:00:00 2001
From: Joseph Redmon <pjreddie@gmail.com>
Date: Sat, 28 Mar 2015 23:11:37 +0000
Subject: [PATCH] We do our OWN resizing!
---
src/convolutional_layer.c | 227 +++++++++++++++++++++++++++++---------------------------
1 files changed, 119 insertions(+), 108 deletions(-)
diff --git a/src/convolutional_layer.c b/src/convolutional_layer.c
index 45bb54a..e20a41c 100644
--- a/src/convolutional_layer.c
+++ b/src/convolutional_layer.c
@@ -1,16 +1,26 @@
#include "convolutional_layer.h"
#include "utils.h"
-#include "mini_blas.h"
+#include "im2col.h"
+#include "col2im.h"
+#include "blas.h"
+#include "gemm.h"
#include <stdio.h>
+#include <time.h>
int convolutional_out_height(convolutional_layer layer)
{
- return (layer.h-layer.size)/layer.stride + 1;
+ int h = layer.h;
+ if (!layer.pad) h -= layer.size;
+ else h -= 1;
+ return h/layer.stride + 1;
}
int convolutional_out_width(convolutional_layer layer)
{
- return (layer.w-layer.size)/layer.stride + 1;
+ int w = layer.w;
+ if (!layer.pad) w -= layer.size;
+ else w -= 1;
+ return w/layer.stride + 1;
}
image get_convolutional_image(convolutional_layer layer)
@@ -31,11 +41,11 @@
return float_to_image(h,w,c,layer.delta);
}
-convolutional_layer *make_convolutional_layer(int batch, 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, int pad, ACTIVATION activation)
{
int i;
- 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;
@@ -43,163 +53,164 @@
layer->batch = batch;
layer->stride = stride;
layer->size = size;
+ layer->pad = pad;
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(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());
+ float scale = 1./sqrt(size*size*c);
+ for(i = 0; i < c*n*size*size; ++i) layer->filters[i] = scale*rand_normal();
for(i = 0; i < n; ++i){
- //layer->biases[i] = rand_normal()*scale + scale;
- layer->biases[i] = 0;
+ layer->biases[i] = scale;
}
int out_h = convolutional_out_height(*layer);
int out_w = convolutional_out_width(*layer);
- layer->col_image = calloc(layer->batch*out_h*out_w*size*size*c, sizeof(float));
+ layer->col_image = calloc(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));
+
+ #ifdef GPU
+ layer->filters_gpu = cuda_make_array(layer->filters, c*n*size*size);
+ layer->filter_updates_gpu = cuda_make_array(layer->filter_updates, c*n*size*size);
+
+ layer->biases_gpu = cuda_make_array(layer->biases, n);
+ layer->bias_updates_gpu = cuda_make_array(layer->bias_updates, n);
+
+ layer->col_image_gpu = cuda_make_array(layer->col_image, out_h*out_w*size*size*c);
+ layer->delta_gpu = cuda_make_array(layer->delta, layer->batch*out_h*out_w*n);
+ layer->output_gpu = cuda_make_array(layer->output, layer->batch*out_h*out_w*n);
+ #endif
layer->activation = activation;
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);
return layer;
}
-void resize_convolutional_layer(convolutional_layer *layer, int h, int w, int c)
+void resize_convolutional_layer(convolutional_layer *layer, int h, int w)
{
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));
+ 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));
+
+ #ifdef GPU
+ cuda_free(layer->col_image_gpu);
+ cuda_free(layer->delta_gpu);
+ cuda_free(layer->output_gpu);
+
+ layer->col_image_gpu = cuda_make_array(layer->col_image, out_h*out_w*layer->size*layer->size*layer->c);
+ layer->delta_gpu = cuda_make_array(layer->delta, layer->batch*out_h*out_w*layer->n);
+ layer->output_gpu = cuda_make_array(layer->output, layer->batch*out_h*out_w*layer->n);
+ #endif
}
-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 = convolutional_out_height(layer)*
- convolutional_out_width(layer)*
- layer.batch;
-
- 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,0,c,n);
- activate_array(layer.output, m*n, layer.activation);
-}
-
-void learn_bias_convolutional_layer(convolutional_layer layer)
+void bias_output(float *output, float *biases, int batch, int n, int 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(b = 0; b < batch; ++b){
+ for(i = 0; i < n; ++i){
for(j = 0; j < size; ++j){
- sum += layer.delta[j+size*(i+b*layer.n)];
+ output[(b*n + i)*size + j] = biases[i];
}
- layer.bias_updates[i] += sum/size;
}
}
}
-void learn_convolutional_layer(convolutional_layer layer)
+void backward_bias(float *bias_updates, float *delta, int batch, int n, int size)
{
+ int i,b;
+ for(b = 0; b < batch; ++b){
+ for(i = 0; i < n; ++i){
+ bias_updates[i] += sum_array(delta+size*(i+b*n), size);
+ }
+ }
+}
+
+
+void forward_convolutional_layer(const convolutional_layer layer, network_state state)
+{
+ int out_h = convolutional_out_height(layer);
+ int out_w = convolutional_out_width(layer);
+ int i;
+
+ bias_output(layer.output, layer.biases, layer.batch, layer.n, out_h*out_w);
+
+ int m = layer.n;
+ int k = layer.size*layer.size*layer.c;
+ int n = out_h*out_w;
+
+ float *a = layer.filters;
+ float *b = layer.col_image;
+ float *c = layer.output;
+
+ for(i = 0; i < layer.batch; ++i){
+ im2col_cpu(state.input, layer.c, layer.h, layer.w,
+ layer.size, layer.stride, layer.pad, b);
+ gemm(0,0,m,n,k,1,a,k,b,n,1,c,n);
+ c += n*m;
+ state.input += layer.c*layer.h*layer.w;
+ }
+ activate_array(layer.output, m*n*layer.batch, layer.activation);
+}
+
+void backward_convolutional_layer(convolutional_layer layer, network_state state)
+{
+ int i;
int m = layer.n;
int n = layer.size*layer.size*layer.c;
int k = convolutional_out_height(layer)*
- convolutional_out_width(layer)*
- layer.batch;
- gradient_array(layer.output, m*k, layer.activation, layer.delta);
- learn_bias_convolutional_layer(layer);
+ convolutional_out_width(layer);
- float *a = layer.delta;
- float *b = layer.col_image;
- float *c = layer.filter_updates;
+ gradient_array(layer.output, m*k*layer.batch, layer.activation, layer.delta);
+ backward_bias(layer.bias_updates, layer.delta, layer.batch, layer.n, k);
- gemm(0,1,m,n,k,1,a,k,b,k,1,c,n);
-}
+ if(state.delta) memset(state.delta, 0, layer.batch*layer.h*layer.w*layer.c*sizeof(float));
-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;
-
- gemm(1,0,m,n,k,1,a,m,b,n,0,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);
+ float *a = layer.delta + i*m*k;
+ float *b = layer.col_image;
+ float *c = layer.filter_updates;
+
+ float *im = state.input+i*layer.c*layer.h*layer.w;
+
+ im2col_cpu(im, layer.c, layer.h, layer.w,
+ layer.size, layer.stride, layer.pad, b);
+ gemm(0,1,m,n,k,1,a,k,b,k,1,c,n);
+
+ if(state.delta){
+ a = layer.filters;
+ b = layer.delta + i*m*k;
+ c = layer.col_image;
+
+ gemm(1,0,n,k,m,1,a,n,b,k,0,c,k);
+
+ col2im_cpu(layer.col_image, layer.c, layer.h, layer.w, layer.size, layer.stride, layer.pad, state.delta+i*layer.c*layer.h*layer.w);
+ }
}
}
-void update_convolutional_layer(convolutional_layer layer, float step, float momentum, float decay)
+void update_convolutional_layer(convolutional_layer layer, int batch, float learning_rate, float momentum, float decay)
{
- int i;
int size = layer.size*layer.size*layer.c*layer.n;
- for(i = 0; i < layer.n; ++i){
- layer.biases[i] += step*layer.bias_updates[i];
- layer.bias_updates[i] *= momentum;
- }
- for(i = 0; i < size; ++i){
- layer.filters[i] += step*(layer.filter_updates[i] - decay*layer.filters[i]);
- layer.filter_updates[i] *= momentum;
- }
+ axpy_cpu(layer.n, learning_rate/batch, layer.bias_updates, 1, layer.biases, 1);
+ scal_cpu(layer.n, momentum, layer.bias_updates, 1);
+
+ axpy_cpu(size, -decay*batch, layer.filters, 1, layer.filter_updates, 1);
+ axpy_cpu(size, learning_rate/batch, layer.filter_updates, 1, layer.filters, 1);
+ scal_cpu(size, momentum, layer.filter_updates, 1);
}
-void test_convolutional_layer()
-{
- 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};
- float filter[] = {.5, 0, .3,
- 0 , 1, 0,
- .2 , 0, 1};
- 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 = 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)
{
@@ -248,8 +259,8 @@
image dc = collapse_image_layers(delta, 1);
char buff[256];
sprintf(buff, "%s: Output", window);
- show_image(dc, buff);
- save_image(dc, buff);
+ //show_image(dc, buff);
+ //save_image(dc, buff);
free_image(dc);
return single_filters;
}
--
Gitblit v1.10.0