From c53e03348c65462bcba33f6352087dd6b9268e8f Mon Sep 17 00:00:00 2001
From: Joseph Redmon <pjreddie@gmail.com>
Date: Wed, 16 Sep 2015 21:12:10 +0000
Subject: [PATCH] yolo working w/ regions
---
src/region_layer.c | 197 ++++++++++++++++++++++++++++---------------------
1 files changed, 113 insertions(+), 84 deletions(-)
diff --git a/src/region_layer.c b/src/region_layer.c
index 7c34b5c..d65c1a8 100644
--- a/src/region_layer.c
+++ b/src/region_layer.c
@@ -6,34 +6,32 @@
#include "cuda.h"
#include "utils.h"
#include <stdio.h>
+#include <assert.h>
#include <string.h>
#include <stdlib.h>
-int get_region_layer_locations(region_layer l)
-{
- return l.inputs / (l.classes+l.coords);
-}
-
-region_layer make_region_layer(int batch, int inputs, int n, int classes, int coords, int rescore)
+region_layer make_region_layer(int batch, int inputs, int n, int side, int classes, int coords, int rescore)
{
region_layer l = {0};
l.type = REGION;
-
+
l.n = n;
l.batch = batch;
l.inputs = inputs;
l.classes = classes;
l.coords = coords;
l.rescore = rescore;
+ l.side = side;
+ assert(side*side*((1 + l.coords)*l.n + l.classes) == inputs);
l.cost = calloc(1, sizeof(float));
- int outputs = inputs;
- l.outputs = outputs;
- l.output = calloc(batch*outputs, sizeof(float));
- l.delta = calloc(batch*outputs, sizeof(float));
- #ifdef GPU
- l.output_gpu = cuda_make_array(0, batch*outputs);
- l.delta_gpu = cuda_make_array(0, batch*outputs);
- #endif
+ l.outputs = l.inputs;
+ l.truths = l.side*l.side*(1+l.coords+l.classes);
+ l.output = calloc(batch*l.outputs, sizeof(float));
+ l.delta = calloc(batch*l.outputs, sizeof(float));
+#ifdef GPU
+ l.output_gpu = cuda_make_array(l.output, batch*l.outputs);
+ l.delta_gpu = cuda_make_array(l.delta, batch*l.outputs);
+#endif
fprintf(stderr, "Region Layer\n");
srand(0);
@@ -43,91 +41,121 @@
void forward_region_layer(const region_layer l, network_state state)
{
- int locations = get_region_layer_locations(l);
+ int locations = l.side*l.side;
int i,j;
+ memcpy(l.output, state.input, l.outputs*l.batch*sizeof(float));
for(i = 0; i < l.batch*locations; ++i){
- int index = i*(l.classes + l.coords);
- int mask = (!state.truth || !state.truth[index]);
-
- for(j = 0; j < l.classes; ++j){
- l.output[index+j] = state.input[index+j];
- }
-
- softmax_array(l.output + index, l.classes, l.output + index);
- index += l.classes;
-
- for(j = 0; j < l.coords; ++j){
- l.output[index+j] = mask*state.input[index+j];
+ int index = i*((1+l.coords)*l.n + l.classes);
+ if(l.softmax){
+ activate_array(l.output + index, l.n*(1+l.coords), LOGISTIC);
+ int offset = l.n*(1+l.coords);
+ softmax_array(l.output + index + offset, l.classes,
+ l.output + index + offset);
}
}
if(state.train){
float avg_iou = 0;
+ float avg_cat = 0;
+ float avg_obj = 0;
+ float avg_anyobj = 0;
int count = 0;
*(l.cost) = 0;
- int size = l.outputs * l.batch;
+ int size = l.inputs * l.batch;
memset(l.delta, 0, size * sizeof(float));
for (i = 0; i < l.batch*locations; ++i) {
- int offset = i*(l.classes+l.coords);
- int bg = state.truth[offset];
- for (j = offset; j < offset+l.classes; ++j) {
- //*(l.cost) += pow(state.truth[j] - l.output[j], 2);
- //l.delta[j] = state.truth[j] - l.output[j];
- }
-
- box anchor = {0,0,.5,.5};
- box truth_code = {state.truth[j+0], state.truth[j+1], state.truth[j+2], state.truth[j+3]};
- box out_code = {l.output[j+0], l.output[j+1], l.output[j+2], l.output[j+3]};
- box out = decode_box(out_code, anchor);
- box truth = decode_box(truth_code, anchor);
-
- if(bg) continue;
- //printf("Box: %f %f %f %f\n", truth.x, truth.y, truth.w, truth.h);
- //printf("Code: %f %f %f %f\n", truth_code.x, truth_code.y, truth_code.w, truth_code.h);
- //printf("Pred : %f %f %f %f\n", out.x, out.y, out.w, out.h);
- // printf("Pred Code: %f %f %f %f\n", out_code.x, out_code.y, out_code.w, out_code.h);
- float iou = box_iou(out, truth);
- avg_iou += iou;
- ++count;
-
- /*
- *(l.cost) += pow((1-iou), 2);
- l.delta[j+0] = (state.truth[j+0] - l.output[j+0]);
- l.delta[j+1] = (state.truth[j+1] - l.output[j+1]);
- l.delta[j+2] = (state.truth[j+2] - l.output[j+2]);
- l.delta[j+3] = (state.truth[j+3] - l.output[j+3]);
- */
-
- for (j = offset+l.classes; j < offset+l.classes+l.coords; ++j) {
- //*(l.cost) += pow(state.truth[j] - l.output[j], 2);
- //l.delta[j] = state.truth[j] - l.output[j];
- float diff = state.truth[j] - l.output[j];
- if (fabs(diff) < 1){
- l.delta[j] = diff;
- *(l.cost) += .5*pow(state.truth[j] - l.output[j], 2);
- } else {
- l.delta[j] = (diff > 0) ? 1 : -1;
- *(l.cost) += fabs(diff) - .5;
+ int index = i*((1+l.coords)*l.n + l.classes);
+ for(j = 0; j < l.n; ++j){
+ int prob_index = index + j*(1 + l.coords);
+ l.delta[prob_index] = (1./l.n)*(0-l.output[prob_index]);
+ if(l.softmax){
+ l.delta[prob_index] = 1./(l.n*l.side)*(0-l.output[prob_index]);
}
- //l.delta[j] = state.truth[j] - l.output[j];
+ *(l.cost) += (1./l.n)*pow(l.output[prob_index], 2);
+ //printf("%f\n", l.output[prob_index]);
+ avg_anyobj += l.output[prob_index];
}
- /*
- if(l.rescore){
- for (j = offset; j < offset+l.classes; ++j) {
- if(state.truth[j]) state.truth[j] = iou;
- l.delta[j] = state.truth[j] - l.output[j];
- }
- }
- */
+ int truth_index = i*(1 + l.coords + l.classes);
+ int best_index = -1;
+ float best_iou = 0;
+ float best_rmse = 4;
+
+ int bg = !state.truth[truth_index];
+ if(bg) {
+ continue;
+ }
+
+ int class_index = index + l.n*(1+l.coords);
+ for(j = 0; j < l.classes; ++j) {
+ l.delta[class_index+j] = state.truth[truth_index+1+j] - l.output[class_index+j];
+ *(l.cost) += pow(state.truth[truth_index+1+j] - l.output[class_index+j], 2);
+ if(state.truth[truth_index + 1 + j]) avg_cat += l.output[class_index+j];
+ }
+ truth_index += l.classes + 1;
+ box truth = {state.truth[truth_index+0], state.truth[truth_index+1], state.truth[truth_index+2], state.truth[truth_index+3]};
+ truth.x /= l.side;
+ truth.y /= l.side;
+
+ for(j = 0; j < l.n; ++j){
+ int out_index = index + j*(1+l.coords);
+ box out = {l.output[out_index+1], l.output[out_index+2], l.output[out_index+3], l.output[out_index+4]};
+
+ out.x /= l.side;
+ out.y /= l.side;
+ if (l.sqrt){
+ out.w = out.w*out.w;
+ out.h = out.h*out.h;
+ }
+
+ float iou = box_iou(out, truth);
+ float rmse = box_rmse(out, truth);
+ if(best_iou > 0 || iou > 0){
+ if(iou > best_iou){
+ best_iou = iou;
+ best_index = j;
+ }
+ }else{
+ if(rmse < best_rmse){
+ best_rmse = rmse;
+ best_index = j;
+ }
+ }
+ }
+ //printf("%d", best_index);
+ int in_index = index + best_index*(1+l.coords);
+ *(l.cost) -= pow(l.output[in_index], 2);
+ *(l.cost) += pow(1-l.output[in_index], 2);
+ avg_obj += l.output[in_index];
+ l.delta[in_index+0] = (1.-l.output[in_index]);
+ if(l.softmax){
+ l.delta[in_index+0] = 5*(1.-l.output[in_index]);
+ }
+ //printf("%f\n", l.output[in_index]);
+
+ l.delta[in_index+1] = 5*(state.truth[truth_index+0] - l.output[in_index+1]);
+ l.delta[in_index+2] = 5*(state.truth[truth_index+1] - l.output[in_index+2]);
+ if(l.sqrt){
+ l.delta[in_index+3] = 5*(sqrt(state.truth[truth_index+2]) - l.output[in_index+3]);
+ l.delta[in_index+4] = 5*(sqrt(state.truth[truth_index+3]) - l.output[in_index+4]);
+ }else{
+ l.delta[in_index+3] = 5*(state.truth[truth_index+2] - l.output[in_index+3]);
+ l.delta[in_index+4] = 5*(state.truth[truth_index+3] - l.output[in_index+4]);
+ }
+
+ *(l.cost) += pow(1-best_iou, 2);
+ avg_iou += best_iou;
+ ++count;
+ if(l.softmax){
+ gradient_array(l.output + index, l.n*(1+l.coords), LOGISTIC, l.delta + index);
+ }
}
- printf("Avg IOU: %f\n", avg_iou/count);
+ printf("Avg IOU: %f, Avg Cat Pred: %f, Avg Obj: %f, Avg Any: %f, count: %d\n", avg_iou/count, avg_cat/count, avg_obj/count, avg_anyobj/(l.batch*locations*l.n), count);
}
}
void backward_region_layer(const region_layer l, network_state state)
{
- axpy_cpu(l.batch*l.inputs, 1, l.delta_gpu, 1, state.delta, 1);
- //copy_cpu(l.batch*l.inputs, l.delta_gpu, 1, state.delta, 1);
+ axpy_cpu(l.batch*l.inputs, 1, l.delta, 1, state.delta, 1);
}
#ifdef GPU
@@ -137,8 +165,9 @@
float *in_cpu = calloc(l.batch*l.inputs, sizeof(float));
float *truth_cpu = 0;
if(state.truth){
- truth_cpu = calloc(l.batch*l.outputs, sizeof(float));
- cuda_pull_array(state.truth, truth_cpu, l.batch*l.outputs);
+ int num_truth = l.batch*l.side*l.side*(1+l.coords+l.classes);
+ truth_cpu = calloc(num_truth, sizeof(float));
+ cuda_pull_array(state.truth, truth_cpu, num_truth);
}
cuda_pull_array(state.input, in_cpu, l.batch*l.inputs);
network_state cpu_state;
@@ -147,7 +176,7 @@
cpu_state.input = in_cpu;
forward_region_layer(l, cpu_state);
cuda_push_array(l.output_gpu, l.output, l.batch*l.outputs);
- cuda_push_array(l.delta_gpu, l.delta, l.batch*l.outputs);
+ cuda_push_array(l.delta_gpu, l.delta, l.batch*l.inputs);
free(cpu_state.input);
if(cpu_state.truth) free(cpu_state.truth);
}
--
Gitblit v1.10.0