1 #include "yolo_layer.h"
2 #include "activations.h"
3 #include "blas.h"
4 #include "box.h"
5 #include "dark_cuda.h"
6 #include "utils.h"
7
8 #include <math.h>
9 #include <stdio.h>
10 #include <assert.h>
11 #include <string.h>
12 #include <stdlib.h>
13
14 extern int check_mistakes;
15
make_yolo_layer(int batch,int w,int h,int n,int total,int * mask,int classes,int max_boxes)16 layer make_yolo_layer(int batch, int w, int h, int n, int total, int *mask, int classes, int max_boxes)
17 {
18 int i;
19 layer l = { (LAYER_TYPE)0 };
20 l.type = YOLO;
21
22 l.n = n;
23 l.total = total;
24 l.batch = batch;
25 l.h = h;
26 l.w = w;
27 l.c = n*(classes + 4 + 1);
28 l.out_w = l.w;
29 l.out_h = l.h;
30 l.out_c = l.c;
31 l.classes = classes;
32 l.cost = (float*)xcalloc(1, sizeof(float));
33 l.biases = (float*)xcalloc(total * 2, sizeof(float));
34 if(mask) l.mask = mask;
35 else{
36 l.mask = (int*)xcalloc(n, sizeof(int));
37 for(i = 0; i < n; ++i){
38 l.mask[i] = i;
39 }
40 }
41 l.bias_updates = (float*)xcalloc(n * 2, sizeof(float));
42 l.outputs = h*w*n*(classes + 4 + 1);
43 l.inputs = l.outputs;
44 l.max_boxes = max_boxes;
45 l.truths = l.max_boxes*(4 + 1); // 90*(4 + 1);
46 l.delta = (float*)xcalloc(batch * l.outputs, sizeof(float));
47 l.output = (float*)xcalloc(batch * l.outputs, sizeof(float));
48 for(i = 0; i < total*2; ++i){
49 l.biases[i] = .5;
50 }
51
52 l.forward = forward_yolo_layer;
53 l.backward = backward_yolo_layer;
54 #ifdef GPU
55 l.forward_gpu = forward_yolo_layer_gpu;
56 l.backward_gpu = backward_yolo_layer_gpu;
57 l.output_gpu = cuda_make_array(l.output, batch*l.outputs);
58 l.output_avg_gpu = cuda_make_array(l.output, batch*l.outputs);
59 l.delta_gpu = cuda_make_array(l.delta, batch*l.outputs);
60
61 free(l.output);
62 if (cudaSuccess == cudaHostAlloc(&l.output, batch*l.outputs*sizeof(float), cudaHostRegisterMapped)) l.output_pinned = 1;
63 else {
64 cudaGetLastError(); // reset CUDA-error
65 l.output = (float*)xcalloc(batch * l.outputs, sizeof(float));
66 }
67
68 free(l.delta);
69 if (cudaSuccess == cudaHostAlloc(&l.delta, batch*l.outputs*sizeof(float), cudaHostRegisterMapped)) l.delta_pinned = 1;
70 else {
71 cudaGetLastError(); // reset CUDA-error
72 l.delta = (float*)xcalloc(batch * l.outputs, sizeof(float));
73 }
74 #endif
75
76 fprintf(stderr, "yolo\n");
77 srand(time(0));
78
79 return l;
80 }
81
resize_yolo_layer(layer * l,int w,int h)82 void resize_yolo_layer(layer *l, int w, int h)
83 {
84 l->w = w;
85 l->h = h;
86
87 l->outputs = h*w*l->n*(l->classes + 4 + 1);
88 l->inputs = l->outputs;
89
90 if (!l->output_pinned) l->output = (float*)xrealloc(l->output, l->batch*l->outputs * sizeof(float));
91 if (!l->delta_pinned) l->delta = (float*)xrealloc(l->delta, l->batch*l->outputs*sizeof(float));
92
93 #ifdef GPU
94 if (l->output_pinned) {
95 CHECK_CUDA(cudaFreeHost(l->output));
96 if (cudaSuccess != cudaHostAlloc(&l->output, l->batch*l->outputs * sizeof(float), cudaHostRegisterMapped)) {
97 cudaGetLastError(); // reset CUDA-error
98 l->output = (float*)xcalloc(l->batch * l->outputs, sizeof(float));
99 l->output_pinned = 0;
100 }
101 }
102
103 if (l->delta_pinned) {
104 CHECK_CUDA(cudaFreeHost(l->delta));
105 if (cudaSuccess != cudaHostAlloc(&l->delta, l->batch*l->outputs * sizeof(float), cudaHostRegisterMapped)) {
106 cudaGetLastError(); // reset CUDA-error
107 l->delta = (float*)xcalloc(l->batch * l->outputs, sizeof(float));
108 l->delta_pinned = 0;
109 }
110 }
111
112 cuda_free(l->delta_gpu);
113 cuda_free(l->output_gpu);
114 cuda_free(l->output_avg_gpu);
115
116 l->delta_gpu = cuda_make_array(l->delta, l->batch*l->outputs);
117 l->output_gpu = cuda_make_array(l->output, l->batch*l->outputs);
118 l->output_avg_gpu = cuda_make_array(l->output, l->batch*l->outputs);
119 #endif
120 }
121
get_yolo_box(float * x,float * biases,int n,int index,int i,int j,int lw,int lh,int w,int h,int stride)122 box get_yolo_box(float *x, float *biases, int n, int index, int i, int j, int lw, int lh, int w, int h, int stride)
123 {
124 box b;
125 // ln - natural logarithm (base = e)
126 // x` = t.x * lw - i; // x = ln(x`/(1-x`)) // x - output of previous conv-layer
127 // y` = t.y * lh - i; // y = ln(y`/(1-y`)) // y - output of previous conv-layer
128 // w = ln(t.w * net.w / anchors_w); // w - output of previous conv-layer
129 // h = ln(t.h * net.h / anchors_h); // h - output of previous conv-layer
130 b.x = (i + x[index + 0*stride]) / lw;
131 b.y = (j + x[index + 1*stride]) / lh;
132 b.w = exp(x[index + 2*stride]) * biases[2*n] / w;
133 b.h = exp(x[index + 3*stride]) * biases[2*n+1] / h;
134 return b;
135 }
136
fix_nan_inf(float val)137 static inline float fix_nan_inf(float val)
138 {
139 if (isnan(val) || isinf(val)) val = 0;
140 return val;
141 }
142
clip_value(float val,const float max_val)143 static inline float clip_value(float val, const float max_val)
144 {
145 if (val > max_val) {
146 //printf("\n val = %f > max_val = %f \n", val, max_val);
147 val = max_val;
148 }
149 else if (val < -max_val) {
150 //printf("\n val = %f < -max_val = %f \n", val, -max_val);
151 val = -max_val;
152 }
153 return val;
154 }
155
delta_yolo_box(box truth,float * x,float * biases,int n,int index,int i,int j,int lw,int lh,int w,int h,float * delta,float scale,int stride,float iou_normalizer,IOU_LOSS iou_loss,int accumulate,float max_delta)156 ious delta_yolo_box(box truth, float *x, float *biases, int n, int index, int i, int j, int lw, int lh, int w, int h, float *delta, float scale, int stride, float iou_normalizer, IOU_LOSS iou_loss, int accumulate, float max_delta)
157 {
158 ious all_ious = { 0 };
159 // i - step in layer width
160 // j - step in layer height
161 // Returns a box in absolute coordinates
162 box pred = get_yolo_box(x, biases, n, index, i, j, lw, lh, w, h, stride);
163 all_ious.iou = box_iou(pred, truth);
164 all_ious.giou = box_giou(pred, truth);
165 all_ious.diou = box_diou(pred, truth);
166 all_ious.ciou = box_ciou(pred, truth);
167 // avoid nan in dx_box_iou
168 if (pred.w == 0) { pred.w = 1.0; }
169 if (pred.h == 0) { pred.h = 1.0; }
170 if (iou_loss == MSE) // old loss
171 {
172 float tx = (truth.x*lw - i);
173 float ty = (truth.y*lh - j);
174 float tw = log(truth.w*w / biases[2 * n]);
175 float th = log(truth.h*h / biases[2 * n + 1]);
176
177 //printf(" tx = %f, ty = %f, tw = %f, th = %f \n", tx, ty, tw, th);
178 //printf(" x = %f, y = %f, w = %f, h = %f \n", x[index + 0 * stride], x[index + 1 * stride], x[index + 2 * stride], x[index + 3 * stride]);
179
180 // accumulate delta
181 delta[index + 0 * stride] += scale * (tx - x[index + 0 * stride]) * iou_normalizer;
182 delta[index + 1 * stride] += scale * (ty - x[index + 1 * stride]) * iou_normalizer;
183 delta[index + 2 * stride] += scale * (tw - x[index + 2 * stride]) * iou_normalizer;
184 delta[index + 3 * stride] += scale * (th - x[index + 3 * stride]) * iou_normalizer;
185 }
186 else {
187 // https://github.com/generalized-iou/g-darknet
188 // https://arxiv.org/abs/1902.09630v2
189 // https://giou.stanford.edu/
190 all_ious.dx_iou = dx_box_iou(pred, truth, iou_loss);
191
192 // jacobian^t (transpose)
193 //float dx = (all_ious.dx_iou.dl + all_ious.dx_iou.dr);
194 //float dy = (all_ious.dx_iou.dt + all_ious.dx_iou.db);
195 //float dw = ((-0.5 * all_ious.dx_iou.dl) + (0.5 * all_ious.dx_iou.dr));
196 //float dh = ((-0.5 * all_ious.dx_iou.dt) + (0.5 * all_ious.dx_iou.db));
197
198 // jacobian^t (transpose)
199 float dx = all_ious.dx_iou.dt;
200 float dy = all_ious.dx_iou.db;
201 float dw = all_ious.dx_iou.dl;
202 float dh = all_ious.dx_iou.dr;
203
204 // predict exponential, apply gradient of e^delta_t ONLY for w,h
205 dw *= exp(x[index + 2 * stride]);
206 dh *= exp(x[index + 3 * stride]);
207
208 // normalize iou weight
209 dx *= iou_normalizer;
210 dy *= iou_normalizer;
211 dw *= iou_normalizer;
212 dh *= iou_normalizer;
213
214
215 dx = fix_nan_inf(dx);
216 dy = fix_nan_inf(dy);
217 dw = fix_nan_inf(dw);
218 dh = fix_nan_inf(dh);
219
220 if (max_delta != FLT_MAX) {
221 dx = clip_value(dx, max_delta);
222 dy = clip_value(dy, max_delta);
223 dw = clip_value(dw, max_delta);
224 dh = clip_value(dh, max_delta);
225 }
226
227
228 if (!accumulate) {
229 delta[index + 0 * stride] = 0;
230 delta[index + 1 * stride] = 0;
231 delta[index + 2 * stride] = 0;
232 delta[index + 3 * stride] = 0;
233 }
234
235 // accumulate delta
236 delta[index + 0 * stride] += dx;
237 delta[index + 1 * stride] += dy;
238 delta[index + 2 * stride] += dw;
239 delta[index + 3 * stride] += dh;
240 }
241
242 return all_ious;
243 }
244
averages_yolo_deltas(int class_index,int box_index,int stride,int classes,float * delta)245 void averages_yolo_deltas(int class_index, int box_index, int stride, int classes, float *delta)
246 {
247
248 int classes_in_one_box = 0;
249 int c;
250 for (c = 0; c < classes; ++c) {
251 if (delta[class_index + stride*c] > 0) classes_in_one_box++;
252 }
253
254 if (classes_in_one_box > 0) {
255 delta[box_index + 0 * stride] /= classes_in_one_box;
256 delta[box_index + 1 * stride] /= classes_in_one_box;
257 delta[box_index + 2 * stride] /= classes_in_one_box;
258 delta[box_index + 3 * stride] /= classes_in_one_box;
259 }
260 }
261
delta_yolo_class(float * output,float * delta,int index,int class_id,int classes,int stride,float * avg_cat,int focal_loss,float label_smooth_eps,float * classes_multipliers)262 void delta_yolo_class(float *output, float *delta, int index, int class_id, int classes, int stride, float *avg_cat, int focal_loss, float label_smooth_eps, float *classes_multipliers)
263 {
264 int n;
265 if (delta[index + stride*class_id]){
266 float y_true = 1;
267 if(label_smooth_eps) y_true = y_true * (1 - label_smooth_eps) + 0.5*label_smooth_eps;
268 float result_delta = y_true - output[index + stride*class_id];
269 if(!isnan(result_delta) && !isinf(result_delta)) delta[index + stride*class_id] = result_delta;
270 //delta[index + stride*class_id] = 1 - output[index + stride*class_id];
271
272 if (classes_multipliers) delta[index + stride*class_id] *= classes_multipliers[class_id];
273 if(avg_cat) *avg_cat += output[index + stride*class_id];
274 return;
275 }
276 // Focal loss
277 if (focal_loss) {
278 // Focal Loss
279 float alpha = 0.5; // 0.25 or 0.5
280 //float gamma = 2; // hardcoded in many places of the grad-formula
281
282 int ti = index + stride*class_id;
283 float pt = output[ti] + 0.000000000000001F;
284 // http://fooplot.com/#W3sidHlwZSI6MCwiZXEiOiItKDEteCkqKDIqeCpsb2coeCkreC0xKSIsImNvbG9yIjoiIzAwMDAwMCJ9LHsidHlwZSI6MTAwMH1d
285 float grad = -(1 - pt) * (2 * pt*logf(pt) + pt - 1); // http://blog.csdn.net/linmingan/article/details/77885832
286 //float grad = (1 - pt) * (2 * pt*logf(pt) + pt - 1); // https://github.com/unsky/focal-loss
287
288 for (n = 0; n < classes; ++n) {
289 delta[index + stride*n] = (((n == class_id) ? 1 : 0) - output[index + stride*n]);
290
291 delta[index + stride*n] *= alpha*grad;
292
293 if (n == class_id && avg_cat) *avg_cat += output[index + stride*n];
294 }
295 }
296 else {
297 // default
298 for (n = 0; n < classes; ++n) {
299 float y_true = ((n == class_id) ? 1 : 0);
300 if (label_smooth_eps) y_true = y_true * (1 - label_smooth_eps) + 0.5*label_smooth_eps;
301 float result_delta = y_true - output[index + stride*n];
302 if (!isnan(result_delta) && !isinf(result_delta)) delta[index + stride*n] = result_delta;
303
304 if (classes_multipliers && n == class_id) delta[index + stride*class_id] *= classes_multipliers[class_id];
305 if (n == class_id && avg_cat) *avg_cat += output[index + stride*n];
306 }
307 }
308 }
309
compare_yolo_class(float * output,int classes,int class_index,int stride,float objectness,int class_id,float conf_thresh)310 int compare_yolo_class(float *output, int classes, int class_index, int stride, float objectness, int class_id, float conf_thresh)
311 {
312 int j;
313 for (j = 0; j < classes; ++j) {
314 //float prob = objectness * output[class_index + stride*j];
315 float prob = output[class_index + stride*j];
316 if (prob > conf_thresh) {
317 return 1;
318 }
319 }
320 return 0;
321 }
322
entry_index(layer l,int batch,int location,int entry)323 static int entry_index(layer l, int batch, int location, int entry)
324 {
325 int n = location / (l.w*l.h);
326 int loc = location % (l.w*l.h);
327 return batch*l.outputs + n*l.w*l.h*(4+l.classes+1) + entry*l.w*l.h + loc;
328 }
329
forward_yolo_layer(const layer l,network_state state)330 void forward_yolo_layer(const layer l, network_state state)
331 {
332 int i, j, b, t, n;
333 memcpy(l.output, state.input, l.outputs*l.batch * sizeof(float));
334
335 #ifndef GPU
336 for (b = 0; b < l.batch; ++b) {
337 for (n = 0; n < l.n; ++n) {
338 int index = entry_index(l, b, n*l.w*l.h, 0);
339 activate_array(l.output + index, 2 * l.w*l.h, LOGISTIC); // x,y,
340 scal_add_cpu(2 * l.w*l.h, l.scale_x_y, -0.5*(l.scale_x_y - 1), l.output + index, 1); // scale x,y
341 index = entry_index(l, b, n*l.w*l.h, 4);
342 activate_array(l.output + index, (1 + l.classes)*l.w*l.h, LOGISTIC);
343 }
344 }
345 #endif
346
347 // delta is zeroed
348 memset(l.delta, 0, l.outputs * l.batch * sizeof(float));
349 if (!state.train) return;
350 //float avg_iou = 0;
351 float tot_iou = 0;
352 float tot_giou = 0;
353 float tot_diou = 0;
354 float tot_ciou = 0;
355 float tot_iou_loss = 0;
356 float tot_giou_loss = 0;
357 float tot_diou_loss = 0;
358 float tot_ciou_loss = 0;
359 float recall = 0;
360 float recall75 = 0;
361 float avg_cat = 0;
362 float avg_obj = 0;
363 float avg_anyobj = 0;
364 int count = 0;
365 int class_count = 0;
366 *(l.cost) = 0;
367 for (b = 0; b < l.batch; ++b) {
368 for (j = 0; j < l.h; ++j) {
369 for (i = 0; i < l.w; ++i) {
370 for (n = 0; n < l.n; ++n) {
371 const int class_index = entry_index(l, b, n*l.w*l.h + j*l.w + i, 4 + 1);
372 const int obj_index = entry_index(l, b, n*l.w*l.h + j*l.w + i, 4);
373 const int box_index = entry_index(l, b, n*l.w*l.h + j*l.w + i, 0);
374 const int stride = l.w*l.h;
375 box pred = get_yolo_box(l.output, l.biases, l.mask[n], box_index, i, j, l.w, l.h, state.net.w, state.net.h, l.w*l.h);
376 float best_match_iou = 0;
377 int best_match_t = 0;
378 float best_iou = 0;
379 int best_t = 0;
380 for (t = 0; t < l.max_boxes; ++t) {
381 box truth = float_to_box_stride(state.truth + t*(4 + 1) + b*l.truths, 1);
382 int class_id = state.truth[t*(4 + 1) + b*l.truths + 4];
383 if (class_id >= l.classes || class_id < 0) {
384 printf("\n Warning: in txt-labels class_id=%d >= classes=%d in cfg-file. In txt-labels class_id should be [from 0 to %d] \n", class_id, l.classes, l.classes - 1);
385 printf("\n truth.x = %f, truth.y = %f, truth.w = %f, truth.h = %f, class_id = %d \n", truth.x, truth.y, truth.w, truth.h, class_id);
386 if (check_mistakes) getchar();
387 continue; // if label contains class_id more than number of classes in the cfg-file and class_id check garbage value
388 }
389 if (!truth.x) break; // continue;
390
391 float objectness = l.output[obj_index];
392 if (isnan(objectness) || isinf(objectness)) l.output[obj_index] = 0;
393 int class_id_match = compare_yolo_class(l.output, l.classes, class_index, l.w*l.h, objectness, class_id, 0.25f);
394
395 float iou = box_iou(pred, truth);
396 if (iou > best_match_iou && class_id_match == 1) {
397 best_match_iou = iou;
398 best_match_t = t;
399 }
400 if (iou > best_iou) {
401 best_iou = iou;
402 best_t = t;
403 }
404 }
405
406 avg_anyobj += l.output[obj_index];
407 l.delta[obj_index] = l.cls_normalizer * (0 - l.output[obj_index]);
408 if (best_match_iou > l.ignore_thresh) {
409 const float iou_multiplier = best_match_iou*best_match_iou;// (best_match_iou - l.ignore_thresh) / (1.0 - l.ignore_thresh);
410 if (l.objectness_smooth) {
411 l.delta[obj_index] = l.cls_normalizer * (iou_multiplier - l.output[obj_index]);
412
413 int class_id = state.truth[best_match_t*(4 + 1) + b*l.truths + 4];
414 if (l.map) class_id = l.map[class_id];
415 const float class_multiplier = (l.classes_multipliers) ? l.classes_multipliers[class_id] : 1.0f;
416 l.delta[class_index + stride*class_id] = class_multiplier * (iou_multiplier - l.output[class_index + stride*class_id]);
417 }
418 else l.delta[obj_index] = 0;
419 }
420 else if (state.net.adversarial) {
421 int stride = l.w*l.h;
422 float scale = pred.w * pred.h;
423 if (scale > 0) scale = sqrt(scale);
424 l.delta[obj_index] = scale * l.cls_normalizer * (0 - l.output[obj_index]);
425 int cl_id;
426 for (cl_id = 0; cl_id < l.classes; ++cl_id) {
427 if(l.output[class_index + stride*cl_id] * l.output[obj_index] > 0.25)
428 l.delta[class_index + stride*cl_id] = scale * (0 - l.output[class_index + stride*cl_id]);
429 }
430 }
431 if (best_iou > l.truth_thresh) {
432 const float iou_multiplier = best_iou*best_iou;// (best_iou - l.truth_thresh) / (1.0 - l.truth_thresh);
433 if (l.objectness_smooth) l.delta[obj_index] = l.cls_normalizer * (iou_multiplier - l.output[obj_index]);
434 else l.delta[obj_index] = l.cls_normalizer * (1 - l.output[obj_index]);
435 //l.delta[obj_index] = l.cls_normalizer * (1 - l.output[obj_index]);
436
437 int class_id = state.truth[best_t*(4 + 1) + b*l.truths + 4];
438 if (l.map) class_id = l.map[class_id];
439 delta_yolo_class(l.output, l.delta, class_index, class_id, l.classes, l.w*l.h, 0, l.focal_loss, l.label_smooth_eps, l.classes_multipliers);
440 const float class_multiplier = (l.classes_multipliers) ? l.classes_multipliers[class_id] : 1.0f;
441 if (l.objectness_smooth) l.delta[class_index + stride*class_id] = class_multiplier * (iou_multiplier - l.output[class_index + stride*class_id]);
442 box truth = float_to_box_stride(state.truth + best_t*(4 + 1) + b*l.truths, 1);
443 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, l.iou_normalizer * class_multiplier, l.iou_loss, 1, l.max_delta);
444 }
445 }
446 }
447 }
448 for (t = 0; t < l.max_boxes; ++t) {
449 box truth = float_to_box_stride(state.truth + t*(4 + 1) + b*l.truths, 1);
450 if (truth.x < 0 || truth.y < 0 || truth.x > 1 || truth.y > 1 || truth.w < 0 || truth.h < 0) {
451 char buff[256];
452 printf(" Wrong label: truth.x = %f, truth.y = %f, truth.w = %f, truth.h = %f \n", truth.x, truth.y, truth.w, truth.h);
453 sprintf(buff, "echo \"Wrong label: truth.x = %f, truth.y = %f, truth.w = %f, truth.h = %f\" >> bad_label.list",
454 truth.x, truth.y, truth.w, truth.h);
455 system(buff);
456 }
457 int class_id = state.truth[t*(4 + 1) + b*l.truths + 4];
458 if (class_id >= l.classes || class_id < 0) continue; // if label contains class_id more than number of classes in the cfg-file and class_id check garbage value
459
460 if (!truth.x) break; // continue;
461 float best_iou = 0;
462 int best_n = 0;
463 i = (truth.x * l.w);
464 j = (truth.y * l.h);
465 box truth_shift = truth;
466 truth_shift.x = truth_shift.y = 0;
467 for (n = 0; n < l.total; ++n) {
468 box pred = { 0 };
469 pred.w = l.biases[2 * n] / state.net.w;
470 pred.h = l.biases[2 * n + 1] / state.net.h;
471 float iou = box_iou(pred, truth_shift);
472 if (iou > best_iou) {
473 best_iou = iou;
474 best_n = n;
475 }
476 }
477
478 int mask_n = int_index(l.mask, best_n, l.n);
479 if (mask_n >= 0) {
480 int class_id = state.truth[t*(4 + 1) + b*l.truths + 4];
481 if (l.map) class_id = l.map[class_id];
482
483 int box_index = entry_index(l, b, mask_n*l.w*l.h + j*l.w + i, 0);
484 const float class_multiplier = (l.classes_multipliers) ? l.classes_multipliers[class_id] : 1.0f;
485 ious all_ious = delta_yolo_box(truth, l.output, l.biases, best_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, l.iou_normalizer * class_multiplier, l.iou_loss, 1, l.max_delta);
486
487 // range is 0 <= 1
488 tot_iou += all_ious.iou;
489 tot_iou_loss += 1 - all_ious.iou;
490 // range is -1 <= giou <= 1
491 tot_giou += all_ious.giou;
492 tot_giou_loss += 1 - all_ious.giou;
493
494 tot_diou += all_ious.diou;
495 tot_diou_loss += 1 - all_ious.diou;
496
497 tot_ciou += all_ious.ciou;
498 tot_ciou_loss += 1 - all_ious.ciou;
499
500 int obj_index = entry_index(l, b, mask_n*l.w*l.h + j*l.w + i, 4);
501 avg_obj += l.output[obj_index];
502 l.delta[obj_index] = class_multiplier * l.cls_normalizer * (1 - l.output[obj_index]);
503
504 int class_index = entry_index(l, b, mask_n*l.w*l.h + j*l.w + i, 4 + 1);
505 delta_yolo_class(l.output, l.delta, class_index, class_id, l.classes, l.w*l.h, &avg_cat, l.focal_loss, l.label_smooth_eps, l.classes_multipliers);
506
507 //printf(" label: class_id = %d, truth.x = %f, truth.y = %f, truth.w = %f, truth.h = %f \n", class_id, truth.x, truth.y, truth.w, truth.h);
508 //printf(" mask_n = %d, l.output[obj_index] = %f, l.output[class_index + class_id] = %f \n\n", mask_n, l.output[obj_index], l.output[class_index + class_id]);
509
510 ++count;
511 ++class_count;
512 if (all_ious.iou > .5) recall += 1;
513 if (all_ious.iou > .75) recall75 += 1;
514 }
515
516 // iou_thresh
517 for (n = 0; n < l.total; ++n) {
518 int mask_n = int_index(l.mask, n, l.n);
519 if (mask_n >= 0 && n != best_n && l.iou_thresh < 1.0f) {
520 box pred = { 0 };
521 pred.w = l.biases[2 * n] / state.net.w;
522 pred.h = l.biases[2 * n + 1] / state.net.h;
523 float iou = box_iou_kind(pred, truth_shift, l.iou_thresh_kind); // IOU, GIOU, MSE, DIOU, CIOU
524 // iou, n
525
526 if (iou > l.iou_thresh) {
527 int class_id = state.truth[t*(4 + 1) + b*l.truths + 4];
528 if (l.map) class_id = l.map[class_id];
529
530 int box_index = entry_index(l, b, mask_n*l.w*l.h + j*l.w + i, 0);
531 const float class_multiplier = (l.classes_multipliers) ? l.classes_multipliers[class_id] : 1.0f;
532 ious all_ious = delta_yolo_box(truth, l.output, l.biases, 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, l.iou_normalizer * class_multiplier, l.iou_loss, 1, l.max_delta);
533
534 // range is 0 <= 1
535 tot_iou += all_ious.iou;
536 tot_iou_loss += 1 - all_ious.iou;
537 // range is -1 <= giou <= 1
538 tot_giou += all_ious.giou;
539 tot_giou_loss += 1 - all_ious.giou;
540
541 tot_diou += all_ious.diou;
542 tot_diou_loss += 1 - all_ious.diou;
543
544 tot_ciou += all_ious.ciou;
545 tot_ciou_loss += 1 - all_ious.ciou;
546
547 int obj_index = entry_index(l, b, mask_n*l.w*l.h + j*l.w + i, 4);
548 avg_obj += l.output[obj_index];
549 l.delta[obj_index] = class_multiplier * l.cls_normalizer * (1 - l.output[obj_index]);
550
551 int class_index = entry_index(l, b, mask_n*l.w*l.h + j*l.w + i, 4 + 1);
552 delta_yolo_class(l.output, l.delta, class_index, class_id, l.classes, l.w*l.h, &avg_cat, l.focal_loss, l.label_smooth_eps, l.classes_multipliers);
553
554 ++count;
555 ++class_count;
556 if (all_ious.iou > .5) recall += 1;
557 if (all_ious.iou > .75) recall75 += 1;
558 }
559 }
560 }
561 }
562
563 // averages the deltas obtained by the function: delta_yolo_box()_accumulate
564 for (j = 0; j < l.h; ++j) {
565 for (i = 0; i < l.w; ++i) {
566 for (n = 0; n < l.n; ++n) {
567 int box_index = entry_index(l, b, n*l.w*l.h + j*l.w + i, 0);
568 int class_index = entry_index(l, b, n*l.w*l.h + j*l.w + i, 4 + 1);
569 const int stride = l.w*l.h;
570
571 averages_yolo_deltas(class_index, box_index, stride, l.classes, l.delta);
572 }
573 }
574 }
575 }
576
577 if (count == 0) count = 1;
578 if (class_count == 0) class_count = 1;
579
580 //*(l.cost) = pow(mag_array(l.delta, l.outputs * l.batch), 2);
581 //printf("Region %d Avg IOU: %f, Class: %f, Obj: %f, No Obj: %f, .5R: %f, .75R: %f, count: %d\n", state.index, avg_iou / count, avg_cat / class_count, avg_obj / count, avg_anyobj / (l.w*l.h*l.n*l.batch), recall / count, recall75 / count, count);
582
583 int stride = l.w*l.h;
584 float* no_iou_loss_delta = (float *)calloc(l.batch * l.outputs, sizeof(float));
585 memcpy(no_iou_loss_delta, l.delta, l.batch * l.outputs * sizeof(float));
586 for (b = 0; b < l.batch; ++b) {
587 for (j = 0; j < l.h; ++j) {
588 for (i = 0; i < l.w; ++i) {
589 for (n = 0; n < l.n; ++n) {
590 int index = entry_index(l, b, n*l.w*l.h + j*l.w + i, 0);
591 no_iou_loss_delta[index + 0 * stride] = 0;
592 no_iou_loss_delta[index + 1 * stride] = 0;
593 no_iou_loss_delta[index + 2 * stride] = 0;
594 no_iou_loss_delta[index + 3 * stride] = 0;
595 }
596 }
597 }
598 }
599 float classification_loss = l.cls_normalizer * pow(mag_array(no_iou_loss_delta, l.outputs * l.batch), 2);
600 free(no_iou_loss_delta);
601 float loss = pow(mag_array(l.delta, l.outputs * l.batch), 2);
602 float iou_loss = loss - classification_loss;
603
604 float avg_iou_loss = 0;
605 // gIOU loss + MSE (objectness) loss
606 if (l.iou_loss == MSE) {
607 *(l.cost) = pow(mag_array(l.delta, l.outputs * l.batch), 2);
608 }
609 else {
610 // Always compute classification loss both for iou + cls loss and for logging with mse loss
611 // TODO: remove IOU loss fields before computing MSE on class
612 // probably split into two arrays
613 if (l.iou_loss == GIOU) {
614 avg_iou_loss = count > 0 ? l.iou_normalizer * (tot_giou_loss / count) : 0;
615 }
616 else {
617 avg_iou_loss = count > 0 ? l.iou_normalizer * (tot_iou_loss / count) : 0;
618 }
619 *(l.cost) = avg_iou_loss + classification_loss;
620 }
621
622 loss /= l.batch;
623 classification_loss /= l.batch;
624 iou_loss /= l.batch;
625
626 fprintf(stderr, "v3 (%s loss, Normalizer: (iou: %.2f, cls: %.2f) Region %d Avg (IOU: %f, GIOU: %f), Class: %f, Obj: %f, No Obj: %f, .5R: %f, .75R: %f, count: %d, class_loss = %f, iou_loss = %f, total_loss = %f \n",
627 (l.iou_loss == MSE ? "mse" : (l.iou_loss == GIOU ? "giou" : "iou")), l.iou_normalizer, l.cls_normalizer, state.index, tot_iou / count, tot_giou / count, avg_cat / class_count, avg_obj / count, avg_anyobj / (l.w*l.h*l.n*l.batch), recall / count, recall75 / count, count,
628 classification_loss, iou_loss, loss);
629 }
630
backward_yolo_layer(const layer l,network_state state)631 void backward_yolo_layer(const layer l, network_state state)
632 {
633 axpy_cpu(l.batch*l.inputs, 1, l.delta, 1, state.delta, 1);
634 }
635
636 // Converts output of the network to detection boxes
637 // w,h: image width,height
638 // netw,neth: network width,height
639 // relative: 1 (all callers seems to pass TRUE)
correct_yolo_boxes(detection * dets,int n,int w,int h,int netw,int neth,int relative,int letter)640 void correct_yolo_boxes(detection *dets, int n, int w, int h, int netw, int neth, int relative, int letter)
641 {
642 int i;
643 // network height (or width)
644 int new_w = 0;
645 // network height (or width)
646 int new_h = 0;
647 // Compute scale given image w,h vs network w,h
648 // I think this "rotates" the image to match network to input image w/h ratio
649 // new_h and new_w are really just network width and height
650 if (letter) {
651 if (((float)netw / w) < ((float)neth / h)) {
652 new_w = netw;
653 new_h = (h * netw) / w;
654 }
655 else {
656 new_h = neth;
657 new_w = (w * neth) / h;
658 }
659 }
660 else {
661 new_w = netw;
662 new_h = neth;
663 }
664 // difference between network width and "rotated" width
665 float deltaw = netw - new_w;
666 // difference between network height and "rotated" height
667 float deltah = neth - new_h;
668 // ratio between rotated network width and network width
669 float ratiow = (float)new_w / netw;
670 // ratio between rotated network width and network width
671 float ratioh = (float)new_h / neth;
672 for (i = 0; i < n; ++i) {
673
674 box b = dets[i].bbox;
675 // x = ( x - (deltaw/2)/netw ) / ratiow;
676 // x - [(1/2 the difference of the network width and rotated width) / (network width)]
677 b.x = (b.x - deltaw / 2. / netw) / ratiow;
678 b.y = (b.y - deltah / 2. / neth) / ratioh;
679 // scale to match rotation of incoming image
680 b.w *= 1 / ratiow;
681 b.h *= 1 / ratioh;
682
683 // relative seems to always be == 1, I don't think we hit this condition, ever.
684 if (!relative) {
685 b.x *= w;
686 b.w *= w;
687 b.y *= h;
688 b.h *= h;
689 }
690
691 dets[i].bbox = b;
692 }
693 }
694
695 /*
696 void correct_yolo_boxes(detection *dets, int n, int w, int h, int netw, int neth, int relative, int letter)
697 {
698 int i;
699 int new_w=0;
700 int new_h=0;
701 if (letter) {
702 if (((float)netw / w) < ((float)neth / h)) {
703 new_w = netw;
704 new_h = (h * netw) / w;
705 }
706 else {
707 new_h = neth;
708 new_w = (w * neth) / h;
709 }
710 }
711 else {
712 new_w = netw;
713 new_h = neth;
714 }
715 for (i = 0; i < n; ++i){
716 box b = dets[i].bbox;
717 b.x = (b.x - (netw - new_w)/2./netw) / ((float)new_w/netw);
718 b.y = (b.y - (neth - new_h)/2./neth) / ((float)new_h/neth);
719 b.w *= (float)netw/new_w;
720 b.h *= (float)neth/new_h;
721 if(!relative){
722 b.x *= w;
723 b.w *= w;
724 b.y *= h;
725 b.h *= h;
726 }
727 dets[i].bbox = b;
728 }
729 }
730 */
731
yolo_num_detections(layer l,float thresh)732 int yolo_num_detections(layer l, float thresh)
733 {
734 int i, n;
735 int count = 0;
736 for(n = 0; n < l.n; ++n){
737 for (i = 0; i < l.w*l.h; ++i) {
738 int obj_index = entry_index(l, 0, n*l.w*l.h + i, 4);
739 if(l.output[obj_index] > thresh){
740 ++count;
741 }
742 }
743 }
744 return count;
745 }
746
yolo_num_detections_batch(layer l,float thresh,int batch)747 int yolo_num_detections_batch(layer l, float thresh, int batch)
748 {
749 int i, n;
750 int count = 0;
751 for (i = 0; i < l.w*l.h; ++i){
752 for(n = 0; n < l.n; ++n){
753 int obj_index = entry_index(l, batch, n*l.w*l.h + i, 4);
754 if(l.output[obj_index] > thresh){
755 ++count;
756 }
757 }
758 }
759 return count;
760 }
761
avg_flipped_yolo(layer l)762 void avg_flipped_yolo(layer l)
763 {
764 int i,j,n,z;
765 float *flip = l.output + l.outputs;
766 for (j = 0; j < l.h; ++j) {
767 for (i = 0; i < l.w/2; ++i) {
768 for (n = 0; n < l.n; ++n) {
769 for(z = 0; z < l.classes + 4 + 1; ++z){
770 int i1 = z*l.w*l.h*l.n + n*l.w*l.h + j*l.w + i;
771 int i2 = z*l.w*l.h*l.n + n*l.w*l.h + j*l.w + (l.w - i - 1);
772 float swap = flip[i1];
773 flip[i1] = flip[i2];
774 flip[i2] = swap;
775 if(z == 0){
776 flip[i1] = -flip[i1];
777 flip[i2] = -flip[i2];
778 }
779 }
780 }
781 }
782 }
783 for(i = 0; i < l.outputs; ++i){
784 l.output[i] = (l.output[i] + flip[i])/2.;
785 }
786 }
787
get_yolo_detections(layer l,int w,int h,int netw,int neth,float thresh,int * map,int relative,detection * dets,int letter)788 int get_yolo_detections(layer l, int w, int h, int netw, int neth, float thresh, int *map, int relative, detection *dets, int letter)
789 {
790 //printf("\n l.batch = %d, l.w = %d, l.h = %d, l.n = %d \n", l.batch, l.w, l.h, l.n);
791 int i,j,n;
792 float *predictions = l.output;
793 // This snippet below is not necessary
794 // Need to comment it in order to batch processing >= 2 images
795 //if (l.batch == 2) avg_flipped_yolo(l);
796 int count = 0;
797 for (i = 0; i < l.w*l.h; ++i){
798 int row = i / l.w;
799 int col = i % l.w;
800 for(n = 0; n < l.n; ++n){
801 int obj_index = entry_index(l, 0, n*l.w*l.h + i, 4);
802 float objectness = predictions[obj_index];
803 //if(objectness <= thresh) continue; // incorrect behavior for Nan values
804 if (objectness > thresh) {
805 //printf("\n objectness = %f, thresh = %f, i = %d, n = %d \n", objectness, thresh, i, n);
806 int box_index = entry_index(l, 0, n*l.w*l.h + i, 0);
807 dets[count].bbox = get_yolo_box(predictions, l.biases, l.mask[n], box_index, col, row, l.w, l.h, netw, neth, l.w*l.h);
808 dets[count].objectness = objectness;
809 dets[count].classes = l.classes;
810 for (j = 0; j < l.classes; ++j) {
811 int class_index = entry_index(l, 0, n*l.w*l.h + i, 4 + 1 + j);
812 float prob = objectness*predictions[class_index];
813 dets[count].prob[j] = (prob > thresh) ? prob : 0;
814 }
815 ++count;
816 }
817 }
818 }
819 correct_yolo_boxes(dets, count, w, h, netw, neth, relative, letter);
820 return count;
821 }
822
get_yolo_detections_batch(layer l,int w,int h,int netw,int neth,float thresh,int * map,int relative,detection * dets,int letter,int batch)823 int get_yolo_detections_batch(layer l, int w, int h, int netw, int neth, float thresh, int *map, int relative, detection *dets, int letter, int batch)
824 {
825 int i,j,n;
826 float *predictions = l.output;
827 //if (l.batch == 2) avg_flipped_yolo(l);
828 int count = 0;
829 for (i = 0; i < l.w*l.h; ++i){
830 int row = i / l.w;
831 int col = i % l.w;
832 for(n = 0; n < l.n; ++n){
833 int obj_index = entry_index(l, batch, n*l.w*l.h + i, 4);
834 float objectness = predictions[obj_index];
835 //if(objectness <= thresh) continue; // incorrect behavior for Nan values
836 if (objectness > thresh) {
837 //printf("\n objectness = %f, thresh = %f, i = %d, n = %d \n", objectness, thresh, i, n);
838 int box_index = entry_index(l, batch, n*l.w*l.h + i, 0);
839 dets[count].bbox = get_yolo_box(predictions, l.biases, l.mask[n], box_index, col, row, l.w, l.h, netw, neth, l.w*l.h);
840 dets[count].objectness = objectness;
841 dets[count].classes = l.classes;
842 for (j = 0; j < l.classes; ++j) {
843 int class_index = entry_index(l, batch, n*l.w*l.h + i, 4 + 1 + j);
844 float prob = objectness*predictions[class_index];
845 dets[count].prob[j] = (prob > thresh) ? prob : 0;
846 }
847 ++count;
848 }
849 }
850 }
851 correct_yolo_boxes(dets, count, w, h, netw, neth, relative, letter);
852 return count;
853 }
854
855 #ifdef GPU
856
forward_yolo_layer_gpu(const layer l,network_state state)857 void forward_yolo_layer_gpu(const layer l, network_state state)
858 {
859 //copy_ongpu(l.batch*l.inputs, state.input, 1, l.output_gpu, 1);
860 simple_copy_ongpu(l.batch*l.inputs, state.input, l.output_gpu);
861 int b, n;
862 for (b = 0; b < l.batch; ++b){
863 for(n = 0; n < l.n; ++n){
864 int index = entry_index(l, b, n*l.w*l.h, 0);
865 // y = 1./(1. + exp(-x))
866 // x = ln(y/(1-y)) // ln - natural logarithm (base = e)
867 // if(y->1) x -> inf
868 // if(y->0) x -> -inf
869 activate_array_ongpu(l.output_gpu + index, 2*l.w*l.h, LOGISTIC); // x,y
870 if (l.scale_x_y != 1) scal_add_ongpu(2 * l.w*l.h, l.scale_x_y, -0.5*(l.scale_x_y - 1), l.output_gpu + index, 1); // scale x,y
871 index = entry_index(l, b, n*l.w*l.h, 4);
872 activate_array_ongpu(l.output_gpu + index, (1+l.classes)*l.w*l.h, LOGISTIC); // classes and objectness
873 }
874 }
875 if(!state.train || l.onlyforward){
876 //cuda_pull_array(l.output_gpu, l.output, l.batch*l.outputs);
877 if (l.mean_alpha && l.output_avg_gpu) mean_array_gpu(l.output_gpu, l.batch*l.outputs, l.mean_alpha, l.output_avg_gpu);
878 cuda_pull_array_async(l.output_gpu, l.output, l.batch*l.outputs);
879 CHECK_CUDA(cudaPeekAtLastError());
880 return;
881 }
882
883 float *in_cpu = (float *)xcalloc(l.batch*l.inputs, sizeof(float));
884 cuda_pull_array(l.output_gpu, l.output, l.batch*l.outputs);
885 memcpy(in_cpu, l.output, l.batch*l.outputs*sizeof(float));
886 float *truth_cpu = 0;
887 if (state.truth) {
888 int num_truth = l.batch*l.truths;
889 truth_cpu = (float *)xcalloc(num_truth, sizeof(float));
890 cuda_pull_array(state.truth, truth_cpu, num_truth);
891 }
892 network_state cpu_state = state;
893 cpu_state.net = state.net;
894 cpu_state.index = state.index;
895 cpu_state.train = state.train;
896 cpu_state.truth = truth_cpu;
897 cpu_state.input = in_cpu;
898 forward_yolo_layer(l, cpu_state);
899 //forward_yolo_layer(l, state);
900 cuda_push_array(l.delta_gpu, l.delta, l.batch*l.outputs);
901 free(in_cpu);
902 if (cpu_state.truth) free(cpu_state.truth);
903 }
904
backward_yolo_layer_gpu(const layer l,network_state state)905 void backward_yolo_layer_gpu(const layer l, network_state state)
906 {
907 axpy_ongpu(l.batch*l.inputs, state.net.loss_scale, l.delta_gpu, 1, state.delta, 1);
908 }
909 #endif
910