forked from terryky/tflite_gles_app
-
Notifications
You must be signed in to change notification settings - Fork 0
/
trt_posenet.cpp
593 lines (494 loc) · 17.7 KB
/
trt_posenet.cpp
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
/* ------------------------------------------------ *
* The MIT License (MIT)
* Copyright (c) 2020 terryky1220@gmail.com
* ------------------------------------------------ */
#include "util_trt.h"
#include "trt_posenet.h"
#include <unistd.h>
#include <float.h>
#define UFF_MODEL_PATH "./models/posenet_mobilenet_v1_100_257x257_multi_kpt_stripped.uff"
#define PLAN_MODEL_PATH "./models/posenet_mobilenet_v1_100_257x257_multi_kpt_stripped.plan"
static IExecutionContext *s_trt_context;
static trt_tensor_t s_tensor_input;
static trt_tensor_t s_tensor_heatmap;
static trt_tensor_t s_tensor_offsets;
static trt_tensor_t s_tensor_fw_disp;
static trt_tensor_t s_tensor_bw_disp;
static std::vector<void *> s_gpu_buffers;
static int s_img_w = 0;
static int s_img_h = 0;
static int s_hmp_w = 0;
static int s_hmp_h = 0;
static int s_edge_num = 0;
typedef struct part_score_t {
float score;
int idx_x;
int idx_y;
int key_id;
} part_score_t;
typedef struct keypoint_t {
float pos_x;
float pos_y;
float score;
int valid;
} keypoint_t;
static int pose_edges[][2] =
{
/* parent, child */
{ kNose, kLeftEye }, // 0
{ kLeftEye, kLeftEar }, // 1
{ kNose, kRightEye }, // 2
{ kRightEye, kRightEar }, // 3
{ kNose, kLeftShoulder }, // 4
{ kLeftShoulder, kLeftElbow }, // 5
{ kLeftElbow, kLeftWrist }, // 6
{ kLeftShoulder, kLeftHip }, // 7
{ kLeftHip, kLeftKnee }, // 8
{ kLeftKnee, kLeftAnkle }, // 9
{ kNose, kRightShoulder}, // 10
{ kRightShoulder, kRightElbow }, // 11
{ kRightElbow, kRightWrist }, // 12
{ kRightShoulder, kRightHip }, // 13
{ kRightHip, kRightKnee }, // 14
{ kRightKnee, kRightAnkle }, // 15
};
/* -------------------------------------------------- *
* create TensorRT engine.
* -------------------------------------------------- */
static int
convert_uff_to_plan (const std::string &plan_file_name, const std::string &uff_file_name)
{
std::vector<trt_uff_inputdef_t> uff_input_array;
trt_uff_inputdef_t uff_input;
uff_input.name = "sub_2";
uff_input.dims = nvinfer1::Dims3(257, 257, 3);
uff_input.order = nvuffparser::UffInputOrder::kNHWC;
uff_input_array.push_back (uff_input);
std::vector<trt_uff_outputdef_t> uff_output_array;
trt_uff_outputdef_t uff_output[4];
uff_output[0].name = "MobilenetV1/heatmap_2/BiasAdd";
uff_output[1].name = "MobilenetV1/offset_2/BiasAdd";
uff_output[2].name = "MobilenetV1/displacement_bwd_2/BiasAdd";
uff_output[3].name = "MobilenetV1/displacement_fwd_2/BiasAdd";
uff_output_array.push_back (uff_output[0]);
uff_output_array.push_back (uff_output[1]);
uff_output_array.push_back (uff_output[2]);
uff_output_array.push_back (uff_output[3]);
ICudaEngine *engine;
engine = trt_create_engine_from_uff (uff_file_name, uff_input_array, uff_output_array);
if (!engine)
{
fprintf (stderr, "ERR:%s(%d): Failed to load graph from file.\n", __FILE__, __LINE__);
return -1;
}
trt_emit_plan_file (engine, plan_file_name);
engine->destroy();
return 0;
}
/* -------------------------------------------------- *
* Create TensorRT Interpreter
* -------------------------------------------------- */
int
init_trt_posenet ()
{
ICudaEngine *engine = NULL;
trt_initialize ();
/* Try to load Prebuilt TensorRT Engine */
fprintf (stderr, "loading prebuilt TensorRT engine...\n");
engine = trt_load_plan_file (PLAN_MODEL_PATH);
/* Build TensorRT Engine */
if (engine == NULL)
{
convert_uff_to_plan (PLAN_MODEL_PATH, UFF_MODEL_PATH);
engine = trt_load_plan_file (PLAN_MODEL_PATH);
if (engine == NULL)
{
fprintf (stderr, "%s(%d)\n", __FILE__, __LINE__);
return -1;
}
}
s_trt_context = engine->createExecutionContext();
/* Allocate IO tensors */
trt_get_tensor_by_name (engine, "sub_2", &s_tensor_input);
trt_get_tensor_by_name (engine, "MobilenetV1/heatmap_2/BiasAdd", &s_tensor_heatmap);
trt_get_tensor_by_name (engine, "MobilenetV1/offset_2/BiasAdd", &s_tensor_offsets);
trt_get_tensor_by_name (engine, "MobilenetV1/displacement_fwd_2/BiasAdd", &s_tensor_fw_disp);
trt_get_tensor_by_name (engine, "MobilenetV1/displacement_bwd_2/BiasAdd", &s_tensor_bw_disp);
int num_bindings = engine->getNbBindings();
s_gpu_buffers.resize (num_bindings);
s_gpu_buffers[s_tensor_input .bind_idx] = s_tensor_input .gpu_mem;
s_gpu_buffers[s_tensor_heatmap.bind_idx] = s_tensor_heatmap.gpu_mem;
s_gpu_buffers[s_tensor_offsets.bind_idx] = s_tensor_offsets.gpu_mem;
s_gpu_buffers[s_tensor_fw_disp.bind_idx] = s_tensor_fw_disp.gpu_mem;
s_gpu_buffers[s_tensor_bw_disp.bind_idx] = s_tensor_bw_disp.gpu_mem;
/* input image dimention */
s_img_w = s_tensor_input.dims.d[1];
s_img_h = s_tensor_input.dims.d[0];
fprintf (stderr, "input image size: (%d, %d)\n", s_img_w, s_img_h);
/* heatmap dimention */
s_hmp_w = s_tensor_heatmap.dims.d[1];
s_hmp_h = s_tensor_heatmap.dims.d[0];
fprintf (stderr, "heatmap size: (%d, %d)\n", s_hmp_w, s_hmp_h);
/* displacement forward vector dimention */
s_edge_num = s_tensor_fw_disp.dims.d[2] / 2;
return 0;
}
void *
get_posenet_input_buf (int *w, int *h)
{
*w = s_tensor_input.dims.d[1];
*h = s_tensor_input.dims.d[0];
return s_tensor_input.cpu_mem;
}
static float
get_heatmap_score (int idx_y, int idx_x, int key_id)
{
int idx = (idx_y * s_hmp_w * kPoseKeyNum) + (idx_x * kPoseKeyNum) + key_id;
float *heatmap_ptr = (float *)s_tensor_heatmap.cpu_mem;
return heatmap_ptr[idx];
}
static void
get_displacement_vector (void *disp_buf, float *dis_x, float *dis_y, int idx_y, int idx_x, int edge_id)
{
int idx0 = (idx_y * s_hmp_w * s_edge_num*2) + (idx_x * s_edge_num*2) + (edge_id + s_edge_num);
int idx1 = (idx_y * s_hmp_w * s_edge_num*2) + (idx_x * s_edge_num*2) + (edge_id);
float *disp_buf_fp = (float *)disp_buf;
*dis_x = disp_buf_fp[idx0];
*dis_y = disp_buf_fp[idx1];
}
static void
get_offset_vector (float *ofst_x, float *ofst_y, int idx_y, int idx_x, int pose_id)
{
int idx0 = (idx_y * s_hmp_w * kPoseKeyNum*2) + (idx_x * kPoseKeyNum*2) + (pose_id + kPoseKeyNum);
int idx1 = (idx_y * s_hmp_w * kPoseKeyNum*2) + (idx_x * kPoseKeyNum*2) + (pose_id);
float *offsets_ptr = (float *)s_tensor_offsets.cpu_mem;
*ofst_x = offsets_ptr[idx0];
*ofst_y = offsets_ptr[idx1];
}
/* enqueue an item in descending order. */
static void
enqueue_score (std::list<part_score_t> &queue, int x, int y, int key, float score)
{
std::list<part_score_t>::iterator itr;
for (itr = queue.begin(); itr != queue.end(); itr++)
{
if (itr->score < score)
break;
}
part_score_t item;
item.score = score;
item.idx_x = x;
item.idx_y = y;
item.key_id= key;
queue.insert(itr, item);
}
/*
* If the score is the highest in local window, return true.
*
* xs xe
* +--+--+--+
* | | | | ys
* +--+--+--+
* | |##| | ##: (idx_x, idx_y)
* +--+--+--+
* | | | | ye
* +--+--+--+
*/
static bool
score_is_max_in_local_window (int key, float score, int idx_y, int idx_x, int max_rad)
{
int xs = std::max (idx_x - max_rad, 0);
int ys = std::max (idx_y - max_rad, 0);
int xe = std::min (idx_x + max_rad + 1, s_hmp_w);
int ye = std::min (idx_y + max_rad + 1, s_hmp_h);
for (int y = ys; y < ye; y ++)
{
for (int x = xs; x < xe; x ++)
{
/* if a higher score is found, return false */
if (get_heatmap_score (y, x, key) > score)
return false;
}
}
return true;
}
static void
build_score_queue (std::list<part_score_t> &queue, float thresh, int max_rad)
{
for (int y = 0; y < s_hmp_h; y ++)
{
for (int x = 0; x < s_hmp_w; x ++)
{
for (int key = 0; key < kPoseKeyNum; key ++)
{
float score = get_heatmap_score (y, x, key);
/* if this score is lower than thresh, skip this pixel. */
if (score < thresh)
continue;
/* if there is a higher score near this pixel, skip this pixel. */
if (!score_is_max_in_local_window (key, score, y, x, max_rad))
continue;
enqueue_score (queue, x, y, key, score);
}
}
}
}
/*
* 0 28.5 57.1 85.6 114.2 142.7 171.3 199.9 228.4 257 [pos_x]
* |---+---|---+---|---+---|---+---|---+---|---+---|---+---|---+---|---+---|
* 0.0 1.0 2.0 3.0 4.0 5.0 6.0 7.0 8.0 [hmp_pos_x]
*/
static void
get_pos_to_near_index (float pos_x, float pos_y, int *idx_x, int *idx_y)
{
float ratio_x = pos_x / (float)s_img_w;
float ratio_y = pos_y / (float)s_img_h;
float hmp_pos_x = ratio_x * (s_hmp_w - 1);
float hmp_pos_y = ratio_y * (s_hmp_h - 1);
int hmp_idx_x = roundf (hmp_pos_x);
int hmp_idx_y = roundf (hmp_pos_y);
hmp_idx_x = std::min (hmp_idx_x, s_hmp_w -1);
hmp_idx_y = std::min (hmp_idx_y, s_hmp_h -1);
hmp_idx_x = std::max (hmp_idx_x, 0);
hmp_idx_y = std::max (hmp_idx_y, 0);
*idx_x = hmp_idx_x;
*idx_y = hmp_idx_y;
}
static void
get_index_to_pos (int idx_x, int idx_y, int key_id, float *pos_x, float *pos_y)
{
float ofst_x, ofst_y;
get_offset_vector (&ofst_x, &ofst_y, idx_y, idx_x, key_id);
float rel_x = (float)idx_x / (float)(s_hmp_w -1);
float rel_y = (float)idx_y / (float)(s_hmp_h -1);
float pos0_x = rel_x * s_img_w;
float pos0_y = rel_y * s_img_h;
*pos_x = pos0_x + ofst_x;
*pos_y = pos0_y + ofst_y;
}
static keypoint_t
traverse_to_tgt_key(int edge, keypoint_t src_key, int tgt_key_id, void *disp)
{
float src_pos_x = src_key.pos_x;
float src_pos_y = src_key.pos_y;
int src_idx_x, src_idx_y;
get_pos_to_near_index (src_pos_x, src_pos_y, &src_idx_x, &src_idx_y);
/* get displacement vector from source to target */
float disp_x, disp_y;
get_displacement_vector (disp, &disp_x, &disp_y, src_idx_y, src_idx_x, edge);
/* calculate target position */
float tgt_pos_x = src_pos_x + disp_x;
float tgt_pos_y = src_pos_y + disp_y;
int tgt_idx_x, tgt_idx_y;
int offset_refine_step = 2;
for (int i = 0; i < offset_refine_step; i ++)
{
get_pos_to_near_index (tgt_pos_x, tgt_pos_y, &tgt_idx_x, &tgt_idx_y);
get_index_to_pos (tgt_idx_x, tgt_idx_y, tgt_key_id, &tgt_pos_x, &tgt_pos_y);
}
keypoint_t tgt_key = {0};
tgt_key.pos_x = tgt_pos_x;
tgt_key.pos_y = tgt_pos_y;
tgt_key.score = get_heatmap_score (tgt_idx_y, tgt_idx_x, tgt_key_id);
tgt_key.valid = 1;
return tgt_key;
}
static void
decode_pose (part_score_t &root, keypoint_t *keys)
{
/* calculate root key position. */
int idx_x = root.idx_x;
int idx_y = root.idx_y;
int keyid = root.key_id;
float *fw_disp_ptr = (float *)s_tensor_fw_disp.cpu_mem;
float *bw_disp_ptr = (float *)s_tensor_bw_disp.cpu_mem;
float pos_x, pos_y;
get_index_to_pos (idx_x, idx_y, keyid, &pos_x, &pos_y);
keys[keyid].pos_x = pos_x;
keys[keyid].pos_y = pos_y;
keys[keyid].score = root.score;
keys[keyid].valid = 1;
for (int edge = s_edge_num - 1; edge >= 0; edge --)
{
int src_key_id = pose_edges[edge][1];
int tgt_key_id = pose_edges[edge][0];
if ( keys[src_key_id].valid &&
!keys[tgt_key_id].valid)
{
keys[tgt_key_id] = traverse_to_tgt_key(edge, keys[src_key_id], tgt_key_id, bw_disp_ptr);
}
}
for (int edge = 0; edge < s_edge_num; edge ++)
{
int src_key_id = pose_edges[edge][0];
int tgt_key_id = pose_edges[edge][1];
if ( keys[src_key_id].valid &&
!keys[tgt_key_id].valid)
{
keys[tgt_key_id] = traverse_to_tgt_key(edge, keys[src_key_id], tgt_key_id, fw_disp_ptr);
}
}
}
static bool
within_nms_of_corresponding_point (posenet_result_t *pose_result,
float pos_x, float pos_y, int key_id, float nms_rad)
{
for (int i = 0; i < pose_result->num; i ++)
{
pose_t *pose = &pose_result->pose[i];
float prev_pos_x = pose->key[key_id].x * s_img_w;
float prev_pos_y = pose->key[key_id].y * s_img_h;
float dx = pos_x - prev_pos_x;
float dy = pos_y - prev_pos_y;
float len = (dx * dx) + (dy * dy);
if (len <= (nms_rad * nms_rad))
return true;
}
return false;
}
static float
get_instance_score (posenet_result_t *pose_result, keypoint_t *keys, float nms_rad)
{
float score_total = 0.0f;
for (int i = 0; i < kPoseKeyNum; i ++)
{
float pos_x = keys[i].pos_x;
float pos_y = keys[i].pos_y;
if (within_nms_of_corresponding_point (pose_result, pos_x, pos_y, i, nms_rad))
continue;
score_total += keys[i].score;
}
return score_total / (float)kPoseKeyNum;
}
static int
regist_detected_pose (posenet_result_t *pose_result, keypoint_t *keys, float score)
{
int pose_id = pose_result->num;
if (pose_id >= MAX_POSE_NUM)
{
fprintf (stderr, "ERR: %s(%d): pose_num overflow.\n", __FILE__, __LINE__);
return -1;
}
for (int i = 0; i < kPoseKeyNum; i++)
{
pose_result->pose[pose_id].key[i].x = keys[i].pos_x / (float)s_img_w;
pose_result->pose[pose_id].key[i].y = keys[i].pos_y / (float)s_img_h;
pose_result->pose[pose_id].key[i].score = keys[i].score;
}
pose_result->pose[pose_id].pose_score = score;
pose_result->num ++;
return 0;
}
static void
decode_multiple_poses (posenet_result_t *pose_result)
{
std::list<part_score_t> queue;
float score_thresh = 0.5f;
int local_max_rad = 1;
build_score_queue (queue, score_thresh, local_max_rad);
memset (pose_result, 0, sizeof (posenet_result_t));
while (pose_result->num < MAX_POSE_NUM && !queue.empty())
{
part_score_t &root = queue.front();
float pos_x, pos_y;
get_index_to_pos (root.idx_x, root.idx_y, root.key_id, &pos_x, &pos_y);
float nms_rad = 20.0f;
if (within_nms_of_corresponding_point (pose_result, pos_x, pos_y, root.key_id, nms_rad))
{
queue.pop_front();
continue;
}
keypoint_t key_points[kPoseKeyNum] = {0};
decode_pose (root, key_points);
float score = get_instance_score (pose_result, key_points, nms_rad);
regist_detected_pose (pose_result, key_points, score);
queue.pop_front();
}
}
static void
decode_single_pose (posenet_result_t *pose_result)
{
int max_block_idx[kPoseKeyNum][2] = {0};
float max_block_cnf[kPoseKeyNum] = {0};
/* find the highest heatmap block for each key */
for (int i = 0; i < kPoseKeyNum; i ++)
{
float max_confidence = -FLT_MAX;
for (int y = 0; y < s_hmp_h; y ++)
{
for (int x = 0; x < s_hmp_w; x ++)
{
float confidence = get_heatmap_score (y, x, i);
if (confidence > max_confidence)
{
max_confidence = confidence;
max_block_cnf[i] = confidence;
max_block_idx[i][0] = x;
max_block_idx[i][1] = y;
}
}
}
}
#if 0
for (int i = 0; i < kPoseKeyNum; i ++)
{
fprintf (stderr, "---------[%d] --------\n", i);
for (int y = 0; y < s_hmp_h; y ++)
{
fprintf (stderr, "[%d] ", y);
for (int x = 0; x < s_hmp_w; x ++)
{
float confidence = get_heatmap_score (y, x, i);
fprintf (stderr, "%6.3f ", confidence);
if (x == max_block_idx[i][0] && y == max_block_idx[i][1])
fprintf (stderr, "#");
else
fprintf (stderr, " ");
}
fprintf (stderr, "\n");
}
}
#endif
/* find the offset vector and calculate the keypoint coordinates. */
for (int i = 0; i < kPoseKeyNum;i ++ )
{
int idx_x = max_block_idx[i][0];
int idx_y = max_block_idx[i][1];
float key_posex, key_posey;
get_index_to_pos (idx_x, idx_y, i, &key_posex, &key_posey);
pose_result->pose[0].key[i].x = key_posex / (float)s_img_w;
pose_result->pose[0].key[i].y = key_posey / (float)s_img_h;
pose_result->pose[0].key[i].score = max_block_cnf[i];
}
pose_result->num = 1;
pose_result->pose[0].pose_score = 1.0f;
}
/* -------------------------------------------------- *
* Invoke TensorRT
* -------------------------------------------------- */
int
invoke_posenet (posenet_result_t *pose_result)
{
/* copy to CUDA buffer */
trt_copy_tensor_to_gpu (s_tensor_input);
/* invoke inference */
int batchSize = 1;
s_trt_context->execute (batchSize, &s_gpu_buffers[0]);
/* copy from CUDA buffer */
trt_copy_tensor_from_gpu (s_tensor_heatmap);
trt_copy_tensor_from_gpu (s_tensor_offsets);
trt_copy_tensor_from_gpu (s_tensor_fw_disp);
trt_copy_tensor_from_gpu (s_tensor_bw_disp);
/*
* decode algorithm is from:
* https://github.com/tensorflow/tfjs-models/tree/master/posenet/src/multi_pose
*/
if (1)
decode_multiple_poses (pose_result);
else
decode_single_pose (pose_result);
pose_result->pose[0].heatmap = s_tensor_heatmap.cpu_mem;
pose_result->pose[0].heatmap_dims[0] = s_hmp_w;
pose_result->pose[0].heatmap_dims[1] = s_hmp_h;
return 0;
}