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/nightmare.c | 147 ++++++++++++++++++++++++++++++++++++++++++++----
1 files changed, 133 insertions(+), 14 deletions(-)
diff --git a/src/nightmare.c b/src/nightmare.c
index 882c0eb..ec7166c 100644
--- a/src/nightmare.c
+++ b/src/nightmare.c
@@ -4,12 +4,18 @@
#include "blas.h"
#include "utils.h"
+#ifdef OPENCV
+#include "opencv2/highgui/highgui_c.h"
+#endif
+
+// ./darknet nightmare cfg/extractor.recon.cfg ~/trained/yolo-coco.conv frame6.png -reconstruct -iters 500 -i 3 -lambda .1 -rate .01 -smooth 2
+
float abs_mean(float *x, int n)
{
int i;
float sum = 0;
for (i = 0; i < n; ++i){
- sum += abs(x[i]);
+ sum += fabs(x[i]);
}
return sum/n;
}
@@ -25,10 +31,10 @@
}
}
-void optimize_picture(network *net, image orig, int max_layer, float scale, float rate, float thresh)
+void optimize_picture(network *net, image orig, int max_layer, float scale, float rate, float thresh, int norm)
{
- scale_image(orig, 2);
- translate_image(orig, -1);
+ //scale_image(orig, 2);
+ //translate_image(orig, -1);
net->n = max_layer + 1;
int dx = rand()%16 - 8;
@@ -49,7 +55,7 @@
#ifdef GPU
state.input = cuda_make_array(im.data, im.w*im.h*im.c);
- state.delta = cuda_make_array(0, im.w*im.h*im.c);
+ state.delta = cuda_make_array(im.data, im.w*im.h*im.c);
forward_network_gpu(*net, state);
copy_ongpu(last.outputs, last.output_gpu, 1, last.delta_gpu, 1);
@@ -85,7 +91,7 @@
//rate = rate / abs_mean(out.data, out.w*out.h*out.c);
- normalize_array(out.data, out.w*out.h*out.c);
+ if(norm) normalize_array(out.data, out.w*out.h*out.c);
axpy_cpu(orig.w*orig.h*orig.c, rate, out.data, 1, orig.data, 1);
/*
@@ -94,8 +100,8 @@
translate_image(orig, mean);
*/
- translate_image(orig, 1);
- scale_image(orig, .5);
+ //translate_image(orig, 1);
+ //scale_image(orig, .5);
//normalize_image(orig);
constrain_image(orig);
@@ -108,6 +114,70 @@
}
+void smooth(image recon, image update, float lambda, int num)
+{
+ int i, j, k;
+ int ii, jj;
+ for(k = 0; k < recon.c; ++k){
+ for(j = 0; j < recon.h; ++j){
+ for(i = 0; i < recon.w; ++i){
+ int out_index = i + recon.w*(j + recon.h*k);
+ for(jj = j-num; jj <= j + num && jj < recon.h; ++jj){
+ if (jj < 0) continue;
+ for(ii = i-num; ii <= i + num && ii < recon.w; ++ii){
+ if (ii < 0) continue;
+ int in_index = ii + recon.w*(jj + recon.h*k);
+ update.data[out_index] += lambda * (recon.data[in_index] - recon.data[out_index]);
+ }
+ }
+ }
+ }
+ }
+}
+
+void reconstruct_picture(network net, float *features, image recon, image update, float rate, float momentum, float lambda, int smooth_size, int iters)
+{
+ int iter = 0;
+ for (iter = 0; iter < iters; ++iter) {
+ image delta = make_image(recon.w, recon.h, recon.c);
+
+ network_state state = {0};
+#ifdef GPU
+ state.input = cuda_make_array(recon.data, recon.w*recon.h*recon.c);
+ state.delta = cuda_make_array(delta.data, delta.w*delta.h*delta.c);
+ state.truth = cuda_make_array(features, get_network_output_size(net));
+
+ forward_network_gpu(net, state);
+ backward_network_gpu(net, state);
+
+ cuda_pull_array(state.delta, delta.data, delta.w*delta.h*delta.c);
+
+ cuda_free(state.input);
+ cuda_free(state.delta);
+ cuda_free(state.truth);
+#else
+ state.input = recon.data;
+ state.delta = delta.data;
+ state.truth = features;
+
+ forward_network(net, state);
+ backward_network(net, state);
+#endif
+
+ axpy_cpu(recon.w*recon.h*recon.c, 1, delta.data, 1, update.data, 1);
+ smooth(recon, update, lambda, smooth_size);
+
+ axpy_cpu(recon.w*recon.h*recon.c, rate, update.data, 1, recon.data, 1);
+ scal_cpu(recon.w*recon.h*recon.c, momentum, update.data, 1);
+
+ //float mag = mag_array(recon.data, recon.w*recon.h*recon.c);
+ //scal_cpu(recon.w*recon.h*recon.c, 600/mag, recon.data, 1);
+
+ constrain_image(recon);
+ free_image(delta);
+ }
+}
+
void run_nightmare(int argc, char **argv)
{
@@ -123,6 +193,7 @@
int max_layer = atoi(argv[5]);
int range = find_int_arg(argc, argv, "-range", 1);
+ int norm = find_int_arg(argc, argv, "-norm", 1);
int rounds = find_int_arg(argc, argv, "-rounds", 1);
int iters = find_int_arg(argc, argv, "-iters", 10);
int octaves = find_int_arg(argc, argv, "-octaves", 4);
@@ -130,6 +201,11 @@
float rate = find_float_arg(argc, argv, "-rate", .04);
float thresh = find_float_arg(argc, argv, "-thresh", 1.);
float rotate = find_float_arg(argc, argv, "-rotate", 0);
+ float momentum = find_float_arg(argc, argv, "-momentum", .9);
+ float lambda = find_float_arg(argc, argv, "-lambda", .01);
+ char *prefix = find_char_arg(argc, argv, "-prefix", 0);
+ int reconstruct = find_arg(argc, argv, "-reconstruct");
+ int smooth_size = find_int_arg(argc, argv, "-smooth", 1);
network net = parse_network_cfg(cfg);
load_weights(&net, weights);
@@ -149,17 +225,56 @@
im = resized;
}
+ float *features = 0;
+ image update;
+ if (reconstruct){
+ resize_network(&net, im.w, im.h);
+
+ int zz = 0;
+ network_predict(net, im.data);
+ image out_im = get_network_image(net);
+ image crop = crop_image(out_im, zz, zz, out_im.w-2*zz, out_im.h-2*zz);
+ //flip_image(crop);
+ image f_im = resize_image(crop, out_im.w, out_im.h);
+ free_image(crop);
+ printf("%d features\n", out_im.w*out_im.h*out_im.c);
+
+
+ im = resize_image(im, im.w, im.h);
+ f_im = resize_image(f_im, f_im.w, f_im.h);
+ features = f_im.data;
+
+ int i;
+ for(i = 0; i < 14*14*512; ++i){
+ features[i] += rand_uniform(-.19, .19);
+ }
+
+ free_image(im);
+ im = make_random_image(im.w, im.h, im.c);
+ update = make_image(im.w, im.h, im.c);
+
+ }
+
int e;
int n;
for(e = 0; e < rounds; ++e){
- fprintf(stderr, "Iteration: ");
- fflush(stderr);
+ fprintf(stderr, "Iteration: ");
+ fflush(stderr);
for(n = 0; n < iters; ++n){
fprintf(stderr, "%d, ", n);
fflush(stderr);
- int layer = max_layer + rand()%range - range/2;
- int octave = rand()%octaves;
- optimize_picture(&net, im, layer, 1/pow(1.33333333, octave), rate, thresh);
+ if(reconstruct){
+ reconstruct_picture(net, features, im, update, rate, momentum, lambda, smooth_size, 1);
+ //if ((n+1)%30 == 0) rate *= .5;
+ show_image(im, "reconstruction");
+#ifdef OPENCV
+ cvWaitKey(10);
+#endif
+ }else{
+ int layer = max_layer + rand()%range - range/2;
+ int octave = rand()%octaves;
+ optimize_picture(&net, im, layer, 1/pow(1.33333333, octave), rate, thresh, norm);
+ }
}
fprintf(stderr, "done\n");
if(0){
@@ -168,7 +283,11 @@
im = g;
}
char buff[256];
- sprintf(buff, "%s_%s_%d_%06d",imbase, cfgbase, max_layer, e);
+ if (prefix){
+ sprintf(buff, "%s/%s_%s_%d_%06d",prefix, imbase, cfgbase, max_layer, e);
+ }else{
+ sprintf(buff, "%s_%s_%d_%06d",imbase, cfgbase, max_layer, e);
+ }
printf("%d %s\n", e, buff);
save_image(im, buff);
//show_image(im, buff);
--
Gitblit v1.10.0