From c6ecf1e0420737eafeb99b27b1d716b46a6cbb7a Mon Sep 17 00:00:00 2001
From: Jud White <github@judsonwhite.com>
Date: Sun, 25 Mar 2018 20:41:48 +0000
Subject: [PATCH] README.md: add notes to How to compile on Windows
---
src/region_layer.c | 42 +++++++++++++++++++++++++++++++-----------
1 files changed, 31 insertions(+), 11 deletions(-)
diff --git a/src/region_layer.c b/src/region_layer.c
index 19da30c..f179906 100644
--- a/src/region_layer.c
+++ b/src/region_layer.c
@@ -110,7 +110,7 @@
return iou;
}
-void delta_region_class(float *output, float *delta, int index, int class_id, int classes, tree *hier, float scale, float *avg_cat)
+void delta_region_class(float *output, float *delta, int index, int class_id, int classes, tree *hier, float scale, float *avg_cat, int focal_loss)
{
int i, n;
if(hier){
@@ -127,11 +127,31 @@
class_id = hier->parent[class_id];
}
*avg_cat += pred;
- } else {
- for(n = 0; n < classes; ++n){
- delta[index + n] = scale * (((n == class_id)?1 : 0) - output[index + n]);
- if(n == class_id) *avg_cat += output[index + n];
- }
+ } else {
+ // Focal loss
+ if (focal_loss) {
+ // Focal Loss for Dense Object Detection: http://blog.csdn.net/linmingan/article/details/77885832
+ float alpha = 0.5; // 0.25 or 0.5
+ //float gamma = 2; // hardcoded in many places of the grad-formula
+
+ int ti = index + class_id;
+ float grad = -2 * (1 - output[ti])*logf(fmaxf(output[ti], 0.0000001))*output[ti] + (1 - output[ti])*(1 - output[ti]);
+
+ for (n = 0; n < classes; ++n) {
+ delta[index + n] = scale * (((n == class_id) ? 1 : 0) - output[index + n]);
+
+ delta[index + n] *= alpha*grad;
+
+ if (n == class_id) *avg_cat += output[index + n];
+ }
+ }
+ else {
+ // default
+ for (n = 0; n < classes; ++n) {
+ delta[index + n] = scale * (((n == class_id) ? 1 : 0) - output[index + n]);
+ if (n == class_id) *avg_cat += output[index + n];
+ }
+ }
}
}
@@ -209,7 +229,7 @@
}
}
int index = size*maxi + b*l.outputs + 5;
- delta_region_class(l.output, l.delta, index, class_id, l.classes, l.softmax_tree, l.class_scale, &avg_cat);
+ delta_region_class(l.output, l.delta, index, class_id, l.classes, l.softmax_tree, l.class_scale, &avg_cat, l.focal_loss);
++class_count;
onlyclass_id = 1;
break;
@@ -240,7 +260,7 @@
if (best_iou > l.thresh) {
l.delta[index + 4] = 0;
if(l.classfix > 0){
- delta_region_class(l.output, l.delta, index + 5, best_class_id, l.classes, l.softmax_tree, l.class_scale*(l.classfix == 2 ? l.output[index + 4] : 1), &avg_cat);
+ delta_region_class(l.output, l.delta, index + 5, best_class_id, l.classes, l.softmax_tree, l.class_scale*(l.classfix == 2 ? l.output[index + 4] : 1), &avg_cat, l.focal_loss);
++class_count;
}
}
@@ -312,7 +332,7 @@
int class_id = state.truth[t*5 + b*l.truths + 4];
if (l.map) class_id = l.map[class_id];
- delta_region_class(l.output, l.delta, best_index + 5, class_id, l.classes, l.softmax_tree, l.class_scale, &avg_cat);
+ delta_region_class(l.output, l.delta, best_index + 5, class_id, l.classes, l.softmax_tree, l.class_scale, &avg_cat, l.focal_loss);
++count;
++class_count;
}
@@ -414,7 +434,7 @@
cuda_pull_array(state.truth, truth_cpu, num_truth);
}
cuda_pull_array(l.output_gpu, in_cpu, l.batch*l.inputs);
- cudaStreamSynchronize(get_cuda_stream());
+ //cudaStreamSynchronize(get_cuda_stream());
network_state cpu_state = state;
cpu_state.train = state.train;
cpu_state.truth = truth_cpu;
@@ -424,7 +444,7 @@
free(cpu_state.input);
if(!state.train) return;
cuda_push_array(l.delta_gpu, l.delta, l.batch*l.outputs);
- cudaStreamSynchronize(get_cuda_stream());
+ //cudaStreamSynchronize(get_cuda_stream());
if(cpu_state.truth) free(cpu_state.truth);
}
--
Gitblit v1.10.0