From 989ab8c38a02fa7ea9c25108151736c62e81c972 Mon Sep 17 00:00:00 2001
From: Joseph Redmon <pjreddie@gmail.com>
Date: Fri, 24 Apr 2015 17:27:50 +0000
Subject: [PATCH] IOU loss function
---
src/convolutional_layer.c | 92 ++++++++++++++++++---------------------------
1 files changed, 37 insertions(+), 55 deletions(-)
diff --git a/src/convolutional_layer.c b/src/convolutional_layer.c
index 2e25844..cd357d3 100644
--- a/src/convolutional_layer.c
+++ b/src/convolutional_layer.c
@@ -29,7 +29,7 @@
h = convolutional_out_height(layer);
w = convolutional_out_width(layer);
c = layer.n;
- return float_to_image(h,w,c,layer.output);
+ return float_to_image(w,h,c,layer.output);
}
image get_convolutional_delta(convolutional_layer layer)
@@ -38,19 +38,14 @@
h = convolutional_out_height(layer);
w = convolutional_out_width(layer);
c = layer.n;
- return float_to_image(h,w,c,layer.delta);
+ return float_to_image(w,h,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, float learning_rate, float momentum, float decay)
+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->learning_rate = learning_rate;
- layer->momentum = momentum;
- layer->decay = decay;
-
layer->h = h;
layer->w = w;
layer->c = c;
@@ -66,7 +61,7 @@
layer->biases = calloc(n, sizeof(float));
layer->bias_updates = calloc(n, sizeof(float));
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 < c*n*size*size; ++i) layer->filters[i] = 2*scale*rand_uniform() - scale;
for(i = 0; i < n; ++i){
layer->biases[i] = scale;
}
@@ -95,11 +90,10 @@
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);
@@ -109,6 +103,16 @@
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 bias_output(float *output, float *biases, int batch, int n, int size)
@@ -125,17 +129,16 @@
void backward_bias(float *bias_updates, float *delta, int batch, int n, int size)
{
- float alpha = 1./batch;
int i,b;
for(b = 0; b < batch; ++b){
for(i = 0; i < n; ++i){
- bias_updates[i] += alpha * sum_array(delta+size*(i+b*n), size);
+ bias_updates[i] += sum_array(delta+size*(i+b*n), size);
}
}
}
-void forward_convolutional_layer(const convolutional_layer layer, float *in)
+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);
@@ -152,18 +155,17 @@
float *c = layer.output;
for(i = 0; i < layer.batch; ++i){
- im2col_cpu(in, layer.c, layer.h, layer.w,
+ 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;
- in += layer.c*layer.h*layer.w;
+ 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, float *in, float *delta)
+void backward_convolutional_layer(convolutional_layer layer, network_state state)
{
- float alpha = 1./layer.batch;
int i;
int m = layer.n;
int n = layer.size*layer.size*layer.c;
@@ -173,40 +175,40 @@
gradient_array(layer.output, m*k*layer.batch, layer.activation, layer.delta);
backward_bias(layer.bias_updates, layer.delta, layer.batch, layer.n, k);
- if(delta) memset(delta, 0, layer.batch*layer.h*layer.w*layer.c*sizeof(float));
+ if(state.delta) memset(state.delta, 0, layer.batch*layer.h*layer.w*layer.c*sizeof(float));
for(i = 0; i < layer.batch; ++i){
float *a = layer.delta + i*m*k;
float *b = layer.col_image;
float *c = layer.filter_updates;
- float *im = in+i*layer.c*layer.h*layer.w;
+ 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,alpha,a,k,b,k,1,c,n);
+ gemm(0,1,m,n,k,1,a,k,b,k,1,c,n);
- if(delta){
+ 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, delta+i*layer.c*layer.h*layer.w);
+ 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)
+void update_convolutional_layer(convolutional_layer layer, int batch, float learning_rate, float momentum, float decay)
{
int size = layer.size*layer.size*layer.c*layer.n;
- axpy_cpu(layer.n, layer.learning_rate, layer.bias_updates, 1, layer.biases, 1);
- scal_cpu(layer.n, layer.momentum, layer.bias_updates, 1);
+ 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, -layer.decay, layer.filters, 1, layer.filter_updates, 1);
- axpy_cpu(size, layer.learning_rate, layer.filter_updates, 1, layer.filters, 1);
- scal_cpu(size, layer.momentum, layer.filter_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);
}
@@ -215,42 +217,22 @@
int h = layer.size;
int w = layer.size;
int c = layer.c;
- return float_to_image(h,w,c,layer.filters+i*h*w*c);
+ return float_to_image(w,h,c,layer.filters+i*h*w*c);
}
-image *weighted_sum_filters(convolutional_layer layer, image *prev_filters)
+image *get_filters(convolutional_layer layer)
{
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));
- }
- }
- 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);
- }
- }
- }
- }
+ int i;
+ for(i = 0; i < layer.n; ++i){
+ filters[i] = copy_image(get_convolutional_filter(layer, i));
}
return filters;
}
image *visualize_convolutional_layer(convolutional_layer layer, char *window, image *prev_filters)
{
- image *single_filters = weighted_sum_filters(layer, 0);
+ image *single_filters = get_filters(layer);
show_images(single_filters, layer.n, window);
image delta = get_convolutional_image(layer);
--
Gitblit v1.10.0