From 5ef74c2031a040f30a670dc7d60790fc6a9ec720 Mon Sep 17 00:00:00 2001
From: Joseph Redmon <pjreddie@gmail.com>
Date: Fri, 02 May 2014 22:20:34 +0000
Subject: [PATCH] Slowly refactoring and pushing to GPU
---
src/convolutional_layer.c | 175 +++++++++++++++-------------------------------------------
1 files changed, 46 insertions(+), 129 deletions(-)
diff --git a/src/convolutional_layer.c b/src/convolutional_layer.c
index f7c9c10..31a4af6 100644
--- a/src/convolutional_layer.c
+++ b/src/convolutional_layer.c
@@ -96,33 +96,14 @@
convolutional_out_width(layer)*
layer.batch;
- memset(layer.output, 0, m*n*sizeof(float));
-
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));
+ im2col_gpu(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;
- 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);
- }
+ 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)
@@ -143,13 +124,13 @@
void learn_convolutional_layer(convolutional_layer layer)
{
- gradient_delta_convolutional_layer(layer);
- learn_bias_convolutional_layer(layer);
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);
float *a = layer.delta;
float *b = layer.col_image;
@@ -171,9 +152,7 @@
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);
+ 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){
@@ -183,72 +162,14 @@
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;
- 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, step, layer.bias_updates, 1, layer.biases, 1);
+ scal_cpu(layer.n, 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);
}
-/*
-
-void backward_convolutional_layer2(convolutional_layer layer, float *input, float *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){
- rotate_image(layer.kernels[i]);
- }
-
- zero_image(in_delta);
- upsample_image(out_delta, layer.stride, layer.upsampled);
- for(j = 0; j < in_delta.c; ++j){
- for(i = 0; i < layer.n; ++i){
- two_d_convolve(layer.upsampled, i, layer.kernels[i], j, 1, in_delta, j, layer.edge);
- }
- }
-
- for(i = 0; i < layer.n; ++i){
- rotate_image(layer.kernels[i]);
- }
-}
-
-
-void learn_convolutional_layer(convolutional_layer layer, float *input)
-{
- int i;
- 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){
- kernel_update(in_image, layer.kernel_updates[i], layer.stride, i, out_delta, layer.edge);
- layer.bias_updates[i] += avg_image_layer(out_delta, i);
- }
-}
-
-void update_convolutional_layer(convolutional_layer layer, float step, float momentum, float decay)
-{
- int i,j;
- for(i = 0; i < layer.n; ++i){
- layer.bias_momentum[i] = step*(layer.bias_updates[i])
- + momentum*layer.bias_momentum[i];
- layer.biases[i] += layer.bias_momentum[i];
- layer.bias_updates[i] = 0;
- int pixels = layer.kernels[i].h*layer.kernels[i].w*layer.kernels[i].c;
- for(j = 0; j < pixels; ++j){
- layer.kernel_momentum[i].data[j] = step*(layer.kernel_updates[i].data[j] - decay*layer.kernels[i].data[j])
- + momentum*layer.kernel_momentum[i].data[j];
- layer.kernels[i].data[j] += layer.kernel_momentum[i].data[j];
- }
- zero_image(layer.kernel_updates[i]);
- }
-}
-*/
void test_convolutional_layer()
{
@@ -285,52 +206,48 @@
return float_to_image(h,w,c,layer.filters+i*h*w*c);
}
-void visualize_convolutional_layer(convolutional_layer layer, char *window)
+image *weighted_sum_filters(convolutional_layer layer, image *prev_filters)
{
- int color = 1;
- int border = 1;
- int h,w,c;
- int size = layer.size;
- h = size;
- w = (size + border) * layer.n - border;
- c = layer.c;
- if(c != 3 || !color){
- h = (h+border)*c - border;
- c = 1;
+ image *filters = calloc(layer.n, sizeof(image));
+ int i,j,k,c;
+ if(!prev_filters){
+ for(i = 0; i < layer.n; ++i){
+ filters[i] = copy_image(get_convolutional_filter(layer, i));
+ }
}
-
- image filters = make_image(h,w,c);
- int i,j;
- for(i = 0; i < layer.n; ++i){
- int w_offset = i*(size+border);
- image k = get_convolutional_filter(layer, i);
- //printf("%f ** ", layer.biases[i]);
- //print_image(k);
- image copy = copy_image(k);
- normalize_image(copy);
- for(j = 0; j < k.c; ++j){
- //set_pixel(copy,0,0,j,layer.biases[i]);
- }
- if(c == 3 && color){
- embed_image(copy, filters, 0, w_offset);
- }
- else{
- for(j = 0; j < k.c; ++j){
- int h_offset = j*(size+border);
- image layer = get_image_layer(k, j);
- embed_image(layer, filters, h_offset, w_offset);
- free_image(layer);
+ else{
+ image base = prev_filters[0];
+ for(i = 0; i < layer.n; ++i){
+ image filter = get_convolutional_filter(layer, i);
+ filters[i] = make_image(base.h, base.w, base.c);
+ for(j = 0; j < layer.size; ++j){
+ for(k = 0; k < layer.size; ++k){
+ for(c = 0; c < layer.c; ++c){
+ float weight = get_pixel(filter, j, k, c);
+ image prev_filter = copy_image(prev_filters[c]);
+ scale_image(prev_filter, weight);
+ add_into_image(prev_filter, filters[i], 0,0);
+ free_image(prev_filter);
+ }
+ }
}
}
- free_image(copy);
}
- image delta = get_convolutional_delta(layer);
+ return filters;
+}
+
+image *visualize_convolutional_layer(convolutional_layer layer, char *window, image *prev_filters)
+{
+ image *single_filters = weighted_sum_filters(layer, 0);
+ show_images(single_filters, layer.n, window);
+
+ image delta = get_convolutional_image(layer);
image dc = collapse_image_layers(delta, 1);
char buff[256];
- sprintf(buff, "%s: Delta", window);
+ sprintf(buff, "%s: Output", window);
show_image(dc, buff);
+ save_image(dc, buff);
free_image(dc);
- show_image(filters, window);
- free_image(filters);
+ return single_filters;
}
--
Gitblit v1.10.0