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/convolutional_layer.c | 108 ++++++++++++++++++++++++++++++++---------------------
1 files changed, 65 insertions(+), 43 deletions(-)
diff --git a/src/convolutional_layer.c b/src/convolutional_layer.c
index 7571e7a..6c7f947 100644
--- a/src/convolutional_layer.c
+++ b/src/convolutional_layer.c
@@ -37,11 +37,16 @@
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, int pad, 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, float learning_rate, float momentum, float decay)
{
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->learning_rate = learning_rate;
+ layer->momentum = momentum;
+ layer->decay = decay;
+
layer->h = h;
layer->w = w;
layer->c = c;
@@ -59,7 +64,8 @@
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());
+ //scale = .0001;
+ for(i = 0; i < c*n*size*size; ++i) layer->filters[i] = scale*(rand_uniform()-.5);
for(i = 0; i < n; ++i){
//layer->biases[i] = rand_normal()*scale + scale;
layer->biases[i] = .5;
@@ -124,46 +130,35 @@
{
int out_h = convolutional_out_height(layer);
int out_w = convolutional_out_width(layer);
+ int i;
+
+ bias_output(layer);
int m = layer.n;
int k = layer.size*layer.size*layer.c;
- int n = out_h*out_w*layer.batch;
+ int n = out_h*out_w;
float *a = layer.filters;
float *b = layer.col_image;
float *c = layer.output;
+
im2col_cpu(in, layer.batch, layer.c, layer.h, layer.w,
layer.size, layer.stride, layer.pad, b);
- bias_output(layer);
- gemm(0,0,m,n,k,1,a,k,b,n,1,c,n);
+
+ for(i = 0; i < layer.batch; ++i){
+ gemm(0,0,m,n,k,1,a,k,b,n,1,c,n);
+ c += n*m;
+ in += layer.h*layer.w*layer.c;
+ b += k*n;
+ }
/*
int i;
for(i = 0; i < m*n; ++i) printf("%f, ", layer.output[i]);
printf("\n");
*/
- activate_array(layer.output, m*n, layer.activation, 0.);
+ activate_array(layer.output, m*n*layer.batch, layer.activation);
}
-#ifdef GPU
-void forward_convolutional_layer_gpu(convolutional_layer layer, cl_mem in)
-{
- int m = layer.n;
- int k = layer.size*layer.size*layer.c;
- int n = convolutional_out_height(layer)*
- convolutional_out_width(layer)*
- layer.batch;
-
- cl_write_array(layer.filters_cl, layer.filters, m*k);
- cl_mem a = layer.filters_cl;
- cl_mem b = layer.col_image_cl;
- cl_mem c = layer.output_cl;
- im2col_ongpu(in, layer.batch, layer.c, layer.h, layer.w, layer.size, layer.stride, b);
- gemm_ongpu(0,0,m,n,k,1,a,k,b,n,0,c,n);
- activate_array_ongpu(layer.output_cl, m*n, layer.activation, 0.);
- cl_read_array(layer.output_cl, layer.output, m*n);
-}
-#endif
-
void learn_bias_convolutional_layer(convolutional_layer layer)
{
int i,b;
@@ -178,47 +173,54 @@
void backward_convolutional_layer(convolutional_layer layer, float *delta)
{
+ 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);
+ convolutional_out_width(layer);
+ gradient_array(layer.output, m*k*layer.batch, layer.activation, layer.delta);
learn_bias_convolutional_layer(layer);
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);
+ for(i = 0; i < layer.batch; ++i){
+ gemm(0,1,m,n,k,1,a,k,b,k,1,c,n);
+ a += m*k;
+ b += k*n;
+ }
if(delta){
m = layer.size*layer.size*layer.c;
k = layer.n;
n = convolutional_out_height(layer)*
- convolutional_out_width(layer)*
- layer.batch;
+ convolutional_out_width(layer);
a = layer.filters;
b = layer.delta;
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));
- col2im_cpu(c, layer.batch, layer.c, layer.h, layer.w, layer.size, layer.stride, layer.pad, delta);
+
+ for(i = 0; i < layer.batch; ++i){
+ gemm(1,0,m,n,k,1,a,m,b,n,0,c,n);
+ col2im_cpu(c, layer.c, layer.h, layer.w, layer.size, layer.stride, layer.pad, delta);
+ c += k*n;
+ delta += layer.h*layer.w*layer.c;
+ }
}
}
-void update_convolutional_layer(convolutional_layer layer, float step, float momentum, float decay)
+void update_convolutional_layer(convolutional_layer layer)
{
int size = layer.size*layer.size*layer.c*layer.n;
- axpy_cpu(layer.n, step, layer.bias_updates, 1, layer.biases, 1);
- scal_cpu(layer.n, momentum, layer.bias_updates, 1);
+ axpy_cpu(layer.n, layer.learning_rate, layer.bias_updates, 1, layer.biases, 1);
+ scal_cpu(layer.n,layer.momentum, layer.bias_updates, 1);
- scal_cpu(size, 1.-step*decay, layer.filters, 1);
- axpy_cpu(size, step, layer.filter_updates, 1, layer.filters, 1);
- scal_cpu(size, momentum, layer.filter_updates, 1);
+ scal_cpu(size, 1.-layer.learning_rate*layer.decay, layer.filters, 1);
+ axpy_cpu(size, layer.learning_rate, layer.filter_updates, 1, layer.filters, 1);
+ scal_cpu(size, layer.momentum, layer.filter_updates, 1);
}
@@ -269,9 +271,29 @@
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;
}
+#ifdef GPU
+void forward_convolutional_layer_gpu(convolutional_layer layer, cl_mem in)
+{
+ int m = layer.n;
+ int k = layer.size*layer.size*layer.c;
+ int n = convolutional_out_height(layer)*
+ convolutional_out_width(layer)*
+ layer.batch;
+
+ cl_write_array(layer.filters_cl, layer.filters, m*k);
+ cl_mem a = layer.filters_cl;
+ cl_mem b = layer.col_image_cl;
+ cl_mem c = layer.output_cl;
+ im2col_ongpu(in, layer.batch, layer.c, layer.h, layer.w, layer.size, layer.stride, b);
+ gemm_ongpu(0,0,m,n,k,1,a,k,b,n,0,c,n);
+ activate_array_ongpu(layer.output_cl, m*n, layer.activation);
+ cl_read_array(layer.output_cl, layer.output, m*n);
+}
+#endif
+
--
Gitblit v1.10.0