From 028696bf15efeca3acb3db8c42a96f7b9e0f55ff Mon Sep 17 00:00:00 2001
From: iovodov <b@ovdv.ru>
Date: Thu, 03 May 2018 13:33:46 +0000
Subject: [PATCH] Output improvements for detector results: When printing detector results, output was done in random order, obfuscating results for interpreting. Now: 1. Text output includes coordinates of rects in (left,right,top,bottom in pixels) along with label and score 2. Text output is sorted by rect lefts to simplify finding appropriate rects on image 3. If several class probs are > thresh for some detection, the most probable is written first and coordinates for others are not repeated 4. Rects are imprinted in image in order by their best class prob, so most probable rects are always on top and not overlayed by less probable ones 5. Most probable label for rect is always written first Also: 6. Message about low GPU memory include required amount
---
src/yolo_layer.c | 83 +++++++++++++++++++++++++++++++++--------
1 files changed, 67 insertions(+), 16 deletions(-)
diff --git a/src/yolo_layer.c b/src/yolo_layer.c
index 46846ef..b2443ba 100644
--- a/src/yolo_layer.c
+++ b/src/yolo_layer.c
@@ -10,7 +10,7 @@
#include <string.h>
#include <stdlib.h>
-layer make_yolo_layer(int batch, int w, int h, int n, int total, int *mask, int classes)
+layer make_yolo_layer(int batch, int w, int h, int n, int total, int *mask, int classes, int max_boxes)
{
int i;
layer l = {0};
@@ -38,7 +38,8 @@
l.bias_updates = calloc(n*2, sizeof(float));
l.outputs = h*w*n*(classes + 4 + 1);
l.inputs = l.outputs;
- l.truths = 90*(4 + 1);
+ l.max_boxes = max_boxes;
+ l.truths = l.max_boxes*(4 + 1); // 90*(4 + 1);
l.delta = calloc(batch*l.outputs, sizeof(float));
l.output = calloc(batch*l.outputs, sizeof(float));
for(i = 0; i < total*2; ++i){
@@ -108,18 +109,41 @@
}
-void delta_yolo_class(float *output, float *delta, int index, int class, int classes, int stride, float *avg_cat)
+void delta_yolo_class(float *output, float *delta, int index, int class_id, int classes, int stride, float *avg_cat, int focal_loss)
{
int n;
if (delta[index]){
- delta[index + stride*class] = 1 - output[index + stride*class];
- if(avg_cat) *avg_cat += output[index + stride*class];
+ delta[index + stride*class_id] = 1 - output[index + stride*class_id];
+ if(avg_cat) *avg_cat += output[index + stride*class_id];
return;
}
- for(n = 0; n < classes; ++n){
- delta[index + stride*n] = ((n == class)?1 : 0) - output[index + stride*n];
- if(n == class && avg_cat) *avg_cat += output[index + stride*n];
- }
+ // Focal loss
+ if (focal_loss) {
+ // Focal Loss
+ float alpha = 0.5; // 0.25 or 0.5
+ //float gamma = 2; // hardcoded in many places of the grad-formula
+
+ int ti = index + stride*class_id;
+ float pt = output[ti] + 0.000000000000001F;
+ // http://fooplot.com/#W3sidHlwZSI6MCwiZXEiOiItKDEteCkqKDIqeCpsb2coeCkreC0xKSIsImNvbG9yIjoiIzAwMDAwMCJ9LHsidHlwZSI6MTAwMH1d
+ float grad = -(1 - pt) * (2 * pt*logf(pt) + pt - 1); // http://blog.csdn.net/linmingan/article/details/77885832
+ //float grad = (1 - pt) * (2 * pt*logf(pt) + pt - 1); // https://github.com/unsky/focal-loss
+
+ for (n = 0; n < classes; ++n) {
+ delta[index + stride*n] = (((n == class_id) ? 1 : 0) - output[index + stride*n]);
+
+ delta[index + stride*n] *= alpha*grad;
+
+ if (n == class_id) *avg_cat += output[index + stride*n];
+ }
+ }
+ else {
+ // default
+ for (n = 0; n < classes; ++n) {
+ delta[index + stride*n] = ((n == class_id) ? 1 : 0) - output[index + stride*n];
+ if (n == class_id && avg_cat) *avg_cat += output[index + stride*n];
+ }
+ }
}
static int entry_index(layer l, int batch, int location, int entry)
@@ -129,6 +153,16 @@
return batch*l.outputs + n*l.w*l.h*(4+l.classes+1) + entry*l.w*l.h + loc;
}
+static box float_to_box_stride(float *f, int stride)
+{
+ box b = { 0 };
+ b.x = f[0];
+ b.y = f[1 * stride];
+ b.w = f[2 * stride];
+ b.h = f[3 * stride];
+ return b;
+}
+
void forward_yolo_layer(const layer l, network_state state)
{
int i,j,b,t,n;
@@ -165,7 +199,7 @@
float best_iou = 0;
int best_t = 0;
for(t = 0; t < l.max_boxes; ++t){
- box truth = float_to_box(state.truth + t*(4 + 1) + b*l.truths, 1);
+ box truth = float_to_box_stride(state.truth + t*(4 + 1) + b*l.truths, 1);
if(!truth.x) break;
float iou = box_iou(pred, truth);
if (iou > best_iou) {
@@ -185,15 +219,15 @@
int class = state.truth[best_t*(4 + 1) + b*l.truths + 4];
if (l.map) class = l.map[class];
int class_index = entry_index(l, b, n*l.w*l.h + j*l.w + i, 4 + 1);
- delta_yolo_class(l.output, l.delta, class_index, class, l.classes, l.w*l.h, 0);
- box truth = float_to_box(state.truth + best_t*(4 + 1) + b*l.truths, 1);
+ delta_yolo_class(l.output, l.delta, class_index, class, l.classes, l.w*l.h, 0, l.focal_loss);
+ box truth = float_to_box_stride(state.truth + best_t*(4 + 1) + b*l.truths, 1);
delta_yolo_box(truth, l.output, l.biases, l.mask[n], box_index, i, j, l.w, l.h, state.net.w, state.net.h, l.delta, (2-truth.w*truth.h), l.w*l.h);
}
}
}
}
for(t = 0; t < l.max_boxes; ++t){
- box truth = float_to_box(state.truth + t*(4 + 1) + b*l.truths, 1);
+ box truth = float_to_box_stride(state.truth + t*(4 + 1) + b*l.truths, 1);
if(!truth.x) break;
float best_iou = 0;
@@ -225,7 +259,7 @@
int class = state.truth[t*(4 + 1) + b*l.truths + 4];
if (l.map) class = l.map[class];
int class_index = entry_index(l, b, mask_n*l.w*l.h + j*l.w + i, 4 + 1);
- delta_yolo_class(l.output, l.delta, class_index, class, l.classes, l.w*l.h, &avg_cat);
+ delta_yolo_class(l.output, l.delta, class_index, class, l.classes, l.w*l.h, &avg_cat, l.focal_loss);
++count;
++class_count;
@@ -368,9 +402,26 @@
return;
}
- cuda_pull_array(l.output_gpu, state.input, l.batch*l.inputs);
- forward_yolo_layer(l, state);
+ //cuda_pull_array(l.output_gpu, state.input, l.batch*l.inputs);
+ float *in_cpu = calloc(l.batch*l.inputs, sizeof(float));
+ cuda_pull_array(l.output_gpu, in_cpu, l.batch*l.inputs);
+ float *truth_cpu = 0;
+ if (state.truth) {
+ int num_truth = l.batch*l.truths;
+ truth_cpu = calloc(num_truth, sizeof(float));
+ cuda_pull_array(state.truth, truth_cpu, num_truth);
+ }
+ network_state cpu_state = state;
+ cpu_state.net = state.net;
+ cpu_state.index = state.index;
+ cpu_state.train = state.train;
+ cpu_state.truth = truth_cpu;
+ cpu_state.input = in_cpu;
+ forward_yolo_layer(l, cpu_state);
+ //forward_yolo_layer(l, state);
cuda_push_array(l.delta_gpu, l.delta, l.batch*l.outputs);
+ free(in_cpu);
+ if (cpu_state.truth) free(cpu_state.truth);
}
void backward_yolo_layer_gpu(const layer l, network_state state)
--
Gitblit v1.10.0