-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathmodel_inference.py
1182 lines (1046 loc) · 65.3 KB
/
model_inference.py
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
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
# pmp_inter test
import torch
import argparse
import time
import os
import re
import math
import numpy as np
from torch.utils.data import DataLoader, TensorDataset, Dataset
from tqdm import tqdm
import torch.nn.functional as F
import gc
from einops import rearrange
import matplotlib.pyplot as plt
from tqdm import tqdm
from models.backbone import MTT_mask_net, MTT_Net_HLG
import options.options as option
from utils.Map2Partition_qtmtt import map_to_parititon_qtmtt
import concurrent.futures
import matplotlib.pyplot as plt
raster2zscan4 = np.array([0, 1, 4, 5, 2, 3, 6, 7, 8, 9, 12, 13, 10, 11, 14, 15], dtype=np.int8)
work_on_PC = False
def str2bool(v):
if isinstance(v, bool):
return v
if v.lower() in ('yes', 'true', 't', 'y', '1'):
return True
elif v.lower() in ('no', 'false', 'f', 'n', '0'):
return False
else:
raise argparse.ArgumentTypeError('Boolean value expected.')
def batch_index_select(x, idx):
if len(x.size()) == 3:
B, N, C = x.size()
N_new = idx.size(1)
offset = torch.arange(B, dtype=torch.long, device=x.device).view(B, 1) * N
idx = idx + offset
out = x.reshape(B * N, C)[idx.reshape(-1)].reshape(B, N_new, C)
return out
elif len(x.size()) == 2:
B, N = x.size()
N_new = idx.size(1)
offset = torch.arange(B, dtype=torch.long, device=x.device).view(B, 1) * N
idx = idx + offset
out = x.reshape(B * N)[idx.reshape(-1)].reshape(B, N_new)
return out
else:
raise NotImplementedError
def batch_index_fill(x, x1, x2, idx1, idx2):
B, N, C = x.size()
B, N1, C = x1.size()
B, N2, C = x2.size()
offset = torch.arange(B, dtype=torch.long, device=x.device).view(B, 1)
idx1 = idx1 + offset * N
idx2 = idx2 + offset * N
x = x.reshape(B * N, C)
x[idx1.reshape(-1)] = x1.reshape(B * N1, C)
x[idx2.reshape(-1)] = x2.reshape(B * N2, C)
x = x.reshape(B, N, C)
return x
def load_ifo_from_cfg(cfg_path):
fp = open(cfg_path)
input_path = None
bit_depth = None
width = None
height = None
for line in fp:
if "InputFile" in line:
line = line.rstrip('\n').replace(' ', '').split('#')[0]
loc = line.find(':')
input_path = line[loc + 1:]
elif "InputBitDepth" in line:
bit_depth = int(line.rstrip('\n').replace(' ', '').split('#')[0].split(':')[-1])
elif "SourceWidth" in line:
width = int(line.rstrip('\n').replace(' ', '').split('#')[0].split(':')[-1])
elif "SourceHeight" in line:
height = int(line.rstrip('\n').replace(' ', '').split('#')[0].split(':')[-1])
elif "FramesToBeEncoded" in line:
frame_num = int(line.rstrip('\n').replace(' ', '').split('#')[0].split(':')[-1])
elif "IntraPeriod" in line:
intraperiod = int(line.rstrip('\n').replace(' ', '').split('#')[0].split(':')[-1])
if (input_path is None) or (bit_depth is None) or (width is None) or (height is None) or (intraperiod is None):
print("Format of CFG error !!!!!!!!")
return
return input_path, bit_depth, width, height, frame_num, intraperiod
def import_yuv420(file_path, width, height, frm_num, sub_sample_ratio=1, is10bit=False, gop_size=None):
fp = open(file_path, 'rb')
pix_num = width * height
if gop_size:
I_frm_num = frm_num // gop_size + 1
sub_frm_num = frm_num - I_frm_num
else:
sub_frm_num = (frm_num + sub_sample_ratio - 1) // sub_sample_ratio # actual frame number after downsampling
if is10bit:
data_type = np.uint16
else:
data_type = np.uint8
y_temp = np.zeros(pix_num * sub_frm_num, dtype=data_type)
u_temp = np.zeros(pix_num*sub_frm_num // 4, dtype=data_type)
v_temp = np.zeros(pix_num*sub_frm_num // 4, dtype=data_type)
for i in range(0, frm_num, sub_sample_ratio):
if gop_size and i % gop_size == 0:
continue
if is10bit:
fp.seek(i * pix_num * 3, 0)
else:
fp.seek(i * pix_num * 3 // 2, 0)
if gop_size:
subi = max(i - i // gop_size - 1, 0)
else:
subi = i // sub_sample_ratio
y_temp[subi * pix_num: (subi + 1) * pix_num] = np.fromfile(fp, dtype=data_type, count=pix_num,
sep='') # u_temp[subi*pix_num//4 : (subi+1)*pix_num//4] = np.fromfile(fp, dtype=data_type, count=pix_num//4, sep='') # v_temp[subi*pix_num//4 : (subi+1)*pix_num//4] = np.fromfile(fp, dtype=data_type, count=pix_num//4, sep='')
fp.close()
y = y_temp.reshape((sub_frm_num, height, width))
u = u_temp.reshape((sub_frm_num, height//2, width//2))
v = v_temp.reshape((sub_frm_num, height//2, width//2))
# print(y.max(), u.max(), 'is_10bit ', is10bit)
if is10bit:
y = np.clip((y + 2) / 4, 0, 255).astype(np.uint8)
u = np.clip((u + 2) / 4, 0, 255).astype(np.uint8)
v = np.clip((v + 2) / 4, 0, 255).astype(np.uint8)
if opt['datasets']['yuv']:
return y, u, v # return frm_num * H * W
else:
return y
def RA_recurrent(cur_id, a=0, b=32):
if cur_id == a + (b - a) // 2:
return a, b
elif cur_id < a + (b - a) // 2:
return RA_recurrent(cur_id, a=a, b=a + (b - a) // 2)
elif cur_id > a + (b - a) // 2:
return RA_recurrent(cur_id, a=a + (b - a) // 2, b=b)
def get_cand_id_list(cur_id, mode='RA', gop_size=32, ref_len=2, frm_num=32):
"""get candidate reference frame id list"""
if mode == 'LDP' or mode == 'LDB':
cand_id_list = set([max(0, (cur_id // gop_size - i) * gop_size) for i in range(4)])
tmp_id = max(0, cur_id - 1)
while len(cand_id_list) < 4:
cand_id_list.add(tmp_id)
tmp_id -= 1
if tmp_id <= 0:
break
cand_id_list = list(cand_id_list)
cand_id_list.sort(reverse=True)
while len(cand_id_list) < 4:
cand_id_list.append(cand_id_list[-1])
elif mode == 'RA':
base_id, remainder_id = cur_id // 32, cur_id % 32
if base_id * 32 + 32 >= frm_num:
f0_id, f1_id = RA_recurrent(cur_id=remainder_id, a=0, b=min(32, frm_num - base_id * 32))
else:
f0_id, f1_id = RA_recurrent(cur_id=remainder_id, a=0, b=32)
cand_id_list = [base_id * 32 + f0_id, base_id * 32 + f0_id if (base_id * 32 + f1_id) >= frm_num else base_id * 32 + f1_id] # 前向参考列表和后向参考列表的第一个frame id
return cand_id_list[:ref_len]
def enc_dec_test(seq_name, qp, frm_num=64):
cur_path = os.getcwd()
# print(os.path.join(cur_path, "codec"))
# exe_path = os.path.join(cur_path, 'codec')
exe_path = os.path.join(cur_path)
cfg_path = os.path.join(cur_path, "cfg")
seq_path = os.path.join(opt['path']['test_seq_dir'], seq_name + '.yuv')
profile_list = ['dpfast']
for profile in profile_list: # ["dpfast", "anchor"]:
out_path = os.path.join(cur_path, "output", "RA", "QP" + str(qp), profile)
log_path = os.path.join(cur_path, "log", "RA", "QP" + str(qp), profile)
if not os.path.exists(log_path):
os.makedirs(log_path)
if not os.path.exists(out_path):
os.makedirs(out_path)
bin_name = seq_name + ".bin"
enc_log_name = "enc_" + seq_name + ".log"
dec_log_name = "dec_" + seq_name + ".log"
if opt['enc_mode'] == 'LDP':
encoder_type = '_LDP'
elif opt['enc_mode'] == 'LDB':
encoder_type = '_LDB'
elif opt['enc_mode'] == 'RA':
encoder_type = '_RA'
cur_exe_path = os.path.join(exe_path, "TAppEncoder_" + profile + encoder_type + ".exe")
mode_cfg_path = os.path.join(cfg_path, "encoder_randomaccess_vtm.cfg") # RA mode
cur_cfg_path = os.path.join(cfg_path, 'per-sequence', seq_name.split('_')[0] + ".cfg")
cur_out_path = os.path.join(out_path, bin_name)
cur_log_path = os.path.join(log_path, enc_log_name)
enc_order = cur_exe_path + " -c " + mode_cfg_path + " -c " + cur_cfg_path + " -i " + seq_path + " -q " + str(qp) + " -b " + cur_out_path + " -f " + str(frm_num) + " > " + cur_log_path
print(enc_order)
# print(dec_order)
os.system(enc_order) # os.system(dec_order)
def depth2flag(depth_map, depth):
cu_size = 64 >> depth # 64 32 16
cu_map_size = 8 >> depth # 8 4 2
offset = 1 << depth # 1 2 4
block_num = depth_map.shape[0]
# print(block_num)
flag_list = []
for k in range(block_num):
temp = []
for i in range(0, 8, cu_map_size):
for j in range(0, 8, cu_map_size):
if depth_map[k, i, j] > depth:
temp.append(1)
else:
temp.append(0)
for i in range(len(temp)):
if depth == 2:
flag_list.append(temp[raster2zscan4[i]])
else:
flag_list.append(temp[i])
return np.array(flag_list, dtype=np.int8)
def get_qt_flag(depth_map, save_name, qp):
flag_map64 = depth2flag(depth_map, 0)
flag_map32 = depth2flag(depth_map, 1)
flag_map16 = depth2flag(depth_map, 2)
save_name = os.path.join('./DepthFlag', save_name + '_Q' + str(qp))
# print('save name:', save_name)
out64_file = open(save_name + '_64.txt', 'w')
out32_file = open(save_name + '_32.txt', 'w')
out16_file = open(save_name + '_16.txt', 'w')
for i in range(flag_map64.size):
out64_file.write(str(flag_map64[i]) + '\n')
for i in range(flag_map32.size):
out32_file.write(str(flag_map32[i]) + '\n')
for i in range(flag_map16.size):
out16_file.write(str(flag_map16[i]) + '\n')
out64_file.close()
out32_file.close()
out16_file.close()
def valid_qt_func(qt_map, depth):
# input: 16x16
if depth >= 4:
return qt_map
cu_size = 16 // (2 ** depth)
map_size = qt_map.shape[-1]
ratio_list = [0.2, 0.9, 0.9, 0.9]
ratio = ratio_list[depth]
depth_base = 1e-1 # 防止1e-6也变成
for x in range(0, map_size, cu_size):
for y in range(0, map_size, cu_size):
if (qt_map[x:x+cu_size, y:y+cu_size] > depth+depth_base).sum() >= ratio * cu_size * cu_size:
qt_map[x:x+cu_size, y:y+cu_size] = qt_map[x:x+cu_size, y:y+cu_size].clip(min=depth+1)
qt_map[x:x+cu_size, y:y+cu_size] = valid_qt_func(qt_map[x:x+cu_size, y:y+cu_size], depth+1)
else:
qt_map[x:x+cu_size, y:y+cu_size] = qt_map[x:x+cu_size, y:y+cu_size].clip(max=depth)
return qt_map
def get_sequence_partition_for_VTM(qt_map, bt_map, dire_map, is_luma, save_path, frm_num, frm_width, frm_height, qp, block_size=128, p_frame_id_list=None, label_qt=None, label_mtt=None, label_dire=None, ori_mtt_mask=None, m_label_depth_list=None, ):
"""qt_map shape(N, 3, 16, 16) / None bt_map shape(N, 4, 2, 32, 32)"""
block_ratio = block_size // 64
executor = concurrent.futures.ThreadPoolExecutor(max_workers=opt['thread_num'])
future_results = []
# out_file = open(save_path, 'w')
block_num_in_height = frm_height // block_size
block_num_in_width = frm_width // block_size
seq_partition_hor_mat = np.zeros((frm_num, frm_height // 4, frm_width // 4)) # store whether is partition edge or not (1 or 0) for edges of all the basic unit (4*4)
seq_partition_ver_mat = np.zeros((frm_num, frm_height // 4, frm_width // 4))
seq_qt_map = np.zeros((frm_num, frm_height // 8, frm_width // 8))
seq_dire_map = np.zeros((frm_num, 3, frm_height // 4, frm_width // 4))
seq_mtt_mask = np.zeros((frm_num, frm_height // 128, frm_width // 128)).astype(np.int8)
assert p_frame_id_list is not None
for total_id, frm_id in tqdm(enumerate(p_frame_id_list)):
# print("Frame ", frm_id)
frm_block_id = total_id * block_num_in_height * block_num_in_width
for block_x in range(block_num_in_height):
for block_y in range(block_num_in_width):
block_id = frm_block_id + block_x * block_num_in_width + block_y
if ori_mtt_mask is not None:
seq_mtt_mask[frm_id, block_x, block_y] = int(ori_mtt_mask[block_id] * 100)
if bt_map is None:
bt_block = np.zeros((3,32,32))
else:
bt_block = bt_map[block_id]
if dire_map is None:
dire_block = np.zeros((3,32,32))
else:
dire_block = dire_map[block_id]
future = executor.submit(
map_to_parititon_qtmtt,
qt_map[block_id].clip(min=0, max=4), bt_block, dire_block, qp, chroma_factor=1, block_size=128, debug_mode=False,
no_dir=False, acc_level=opt['acc_level'],
frm_id=frm_id, block_x=block_x, block_y=block_y
)
future_results.append(future)
concurrent.futures.wait(future_results)
results = []
for future in future_results:
result = future.result()
results.append(result)
for result in results:
hor_mat, ver_mat, out_dire_map, valid_qt_map, frm_id, block_x, block_y = result
seq_partition_hor_mat[frm_id, block_x * 16 * block_ratio:(block_x + 1) * 16 * block_ratio, block_y * 16 * block_ratio:(block_y + 1) * 16 * block_ratio] = hor_mat
seq_partition_ver_mat[frm_id, block_x * 16 * block_ratio:(block_x + 1) * 16 * block_ratio, block_y * 16 * block_ratio:(block_y + 1) * 16 * block_ratio] = ver_mat
seq_dire_map[frm_id, :, block_x * 16 * block_ratio:(block_x + 1) * 16 * block_ratio, block_y * 16 * block_ratio:(block_y + 1) * 16 * block_ratio] = out_dire_map
seq_qt_map[frm_id, block_x * 8 * block_ratio:(block_x + 1) * 8 * block_ratio, block_y * 8 * block_ratio:(block_y + 1) * 8 * block_ratio] = valid_qt_map
if save_path is not None:
if os.path.exists(save_path):
os.remove(save_path)
out_file = open(save_path, 'w')
# print(p_frame_id_list)
for total_id, frm_id in tqdm(enumerate(p_frame_id_list)):
hor_vec = seq_partition_hor_mat[frm_id].reshape(-1).astype(np.uint8)
ver_vec = seq_partition_ver_mat[frm_id].reshape(-1).astype(np.uint8)
qtdepth_vec = seq_qt_map[frm_id].reshape(-1).astype(np.uint8)
dire_vec = seq_dire_map[frm_id].reshape(-1).astype(np.int8)
if ori_mtt_mask is not None:
mask_vec = seq_mtt_mask[frm_id].reshape(-1).astype(np.int8)
for i in range(hor_vec.size): # horizontal edge vector
out_file.write(str(hor_vec[i]) + '\n')
for i in range(ver_vec.size): # vertical edge vector
out_file.write(str(ver_vec[i]) + '\n')
for i in range(qtdepth_vec.size): # qt depth vector
out_file.write(str(qtdepth_vec[i]) + '\n')
for i in range(dire_vec.size):
out_file.write(str(dire_vec[i]) + '\n') # print(hor_vec.size) # print(qtdepth_vec.size) # print(dire_vec.size)
if ori_mtt_mask is not None:
for i in range(mask_vec.size):
out_file.write(str(mask_vec[i]) + '\n')
out_file.close()
return seq_partition_hor_mat, seq_partition_ver_mat
def get_sequence_partition_for_VTM_debug(qt_map, bt_map, dire_map, is_luma, save_path, frm_num, frm_width, frm_height, qp, block_size=128, p_frame_id_list=None, label_qt=None, label_mtt=None, label_dire=None, ori_mtt_mask=None, m_label_depth_list=None, gop_size=None):
"""qt_map shape(N, 3, 16, 16) / None bt_map shape(N, 4, 2, 32, 32)"""
valid_qt_map = None if qt_map is None else np.zeros_like(np.asarray(qt_map))
valid_bt_map, valid_dire_map = np.zeros_like(bt_map), np.zeros_like(dire_map)
block_num = bt_map.shape[0]
qt_map = qt_map.clip(min=0, max=3)
block_num_in_height = frm_height // block_size
block_num_in_width = frm_width // block_size
p_frame_id_list = list(set(range(frm_num)) - set(range(0, frm_num, gop_size)))
seq_mtt_mask = np.zeros((frm_num, frm_height // 128, frm_width // 128)).astype(np.int8)
for total_id, frm_id in tqdm(enumerate(p_frame_id_list)):
# print("Frame ", frm_id)
frm_block_id = total_id * block_num_in_height * block_num_in_width
for block_x in range(block_num_in_height):
for block_y in range(block_num_in_width):
block_id = frm_block_id + block_x * block_num_in_width + block_y
seq_mtt_mask[frm_id, block_x, block_y] = int(ori_mtt_mask[block_id] * 100)
if bt_map is None:
bt_block = np.zeros((3,32,32))
else:
bt_block = bt_map[block_id]
if dire_map is None:
dire_block = np.zeros((3,32,32))
else:
dire_block = dire_map[block_id]
valid_qt_block, valid_bt_block, valid_dire_block = map_to_parititon_qtmtt(qt_map[block_id].clip(min=0, max=4), bt_block, dire_block, qp, chroma_factor=1, block_size=128, debug_mode=True, no_dir=False, acc_level=opt['acc_level'])
valid_bt_map[block_id] = valid_bt_block
valid_dire_map[block_id] = valid_dire_block
valid_qt_map[block_id] = valid_qt_block
# vis
block_ratio = block_size // 64
chroma_factor = 2
if is_luma:
chroma_factor = 1
# assert save_path is not None
block_num_in_height = frm_height // block_size
block_num_in_width = frm_width // block_size
inconsist_error = np.sum(valid_qt_map != qt_map.round()) / qt_map.size
seq_bt_map = np.zeros((frm_num, 3, frm_height // 4, frm_width // 4))
seq_qt_map = np.zeros((frm_num, frm_height // 8, frm_width // 8))
seq_dire_map = np.zeros((frm_num, 3, frm_height // 4, frm_width // 4))
# get p-frame id
# p_frame_id_list = list(set(range(frm_num)) - set(range(0, frm_num, gop_size)))
for total_id, frm_id in (enumerate(p_frame_id_list)):
# print("Frame ", frm_id)
frm_block_id = total_id * block_num_in_height * block_num_in_width
for block_x in range(block_num_in_height):
for block_y in range(block_num_in_width):
block_id = frm_block_id + block_x * block_num_in_width + block_y
seq_qt_map[frm_id, block_x * 8 * block_ratio:(block_x + 1) * 8 * block_ratio, block_y * 8 * block_ratio:(block_y + 1) * 8 * block_ratio] = valid_qt_map[block_id]
seq_bt_map[frm_id, :, block_x * 16 * block_ratio:(block_x + 1) * 16 * block_ratio, block_y * 16 * block_ratio:(block_y + 1) * 16 * block_ratio] = valid_bt_map[block_id]
seq_dire_map[frm_id, :, block_x * 16 * block_ratio:(block_x + 1) * 16 * block_ratio, block_y * 16 * block_ratio:(block_y + 1) * 16 * block_ratio] = valid_dire_map[block_id]
print("AFTER post-processing")
seq_qt_map = seq_qt_map[p_frame_id_list]
seq_bt_map = seq_bt_map[p_frame_id_list]
seq_dire_map = seq_dire_map[p_frame_id_list]
accu_list, accu_mask = print_metric(m_label_depth_list=m_label_depth_list, out_pred_qt=seq_qt_map, out_pred_mtt=seq_bt_map, out_pred_dir=seq_dire_map, out_pred_mask=None)
return seq_qt_map, seq_bt_map, seq_dire_map, accu_list, accu_mask, inconsist_error
def reblock_array(input_array, frame_num, frame_height, frame_width, origin_block_size=128, target_block_size=64, patch_size=16):
"""origin_block_size: origin luma block size, target_block_size: target luma block size, patch_size: map's patch size"""
"""input shape (N, C, 64, 64) output shape (N*4*4, C, 16, 16) 保持target_block的raster-scan"""
block_ratio = origin_block_size // target_block_size
target_array = np.zeros((input_array.shape[0] * block_ratio * block_ratio, input_array.shape[1], input_array.shape[2] // block_ratio, input_array.shape[3] // block_ratio))
block_num_in_height, block_num_in_width = frame_height // origin_block_size, frame_width // origin_block_size
for frm_id in range(frame_num):
frm_block_id = frm_id * block_num_in_height * block_num_in_width
for block_x in range(block_num_in_height):
for block_y in range(block_num_in_width):
block_id = frm_block_id + block_x * block_num_in_width + block_y
target_parent_upper_id = (frm_block_id + block_x * block_num_in_width) * block_ratio * block_ratio
for sub_height_id in range(block_ratio):
for sub_width_id in range(block_ratio):
target_block_id = target_parent_upper_id + sub_height_id * (frame_width // target_block_size) + block_y * block_ratio + sub_width_id # 当前块上边的parent block行，当前块上边target_block行，当前块在当前行的index
target_array[target_block_id] = input_array[block_id, :, sub_height_id * patch_size: (sub_height_id + 1) * patch_size, sub_width_id * patch_size: (sub_width_id + 1) * patch_size]
return target_array
def get_order(file_path):
poc_list = []
p_list = []
reorder_list = []
refer_id_f = {}
refer_id_b = {}
qp_dict = {}
tid_dict = {}
with open(file_path, 'r') as f:
content = f.readlines()
for line in content:
if line[:3] == 'POC':
pattern = r'POC\s+(\d+)'
match = re.search(pattern, line)
poc = int(match.group(1))
poc_list.append(poc)
if 'I-SLICE' not in line:
p_list.append(poc)
# qp
qp_pattern = r"QP (\d+)"
qp_match = re.search(qp_pattern, line)
qp = int(qp_match.group(1))
qp_dict[poc] = qp
# tid
tid_pattern = r'TId:\s*(\d+)'
tid_match = re.search(tid_pattern, line)
tid = int(tid_match.group(1))
tid_dict[poc] = tid
pattern = r'\[L0 ([\d\sc]+)\] \[L1 ([\d\sc]+)\]'
matches = re.search(pattern, line)
if matches:
L0 = matches.group(1)
L1 = matches.group(2)
L0 = [int(item[:-1]) if item[-1] == "c" else int(item) for item in L0.split()]
L1 = [int(item[:-1]) if item[-1] == "c" else int(item) for item in L1.split()]
refer_id_f[poc] = L0
refer_id_b[poc] = L1
else:
refer_id_b[poc] = 0
refer_id_f[poc] = 0
for i in range(len(poc_list)):
reorder_list.append(poc_list.index(i))
return refer_id_f, refer_id_b, reorder_list, qp_dict, tid_dict, sorted(p_list)
class MyDataset(Dataset):
def __init__(self, file_path, width, height, frm_num, sub_sample_ratio, is10bit, mode, ref_len, enc_mode='RA', pad128=False, qp=None, opt=None, intra_period=None):
super().__init__()
assert intra_period is not None
self.intra_period = intra_period
self.enc_mode = enc_mode
self.qp = qp
self.gop_size = int(opt['encode_cfg'].split('gop')[-1])
self.__load_dataset_y__( file_path, width, height, frm_num, sub_sample_ratio, is10bit, mode, ref_len, gop_size=self.gop_size, pad128=pad128)
if opt['network']['qml'] or opt['network']['tml']:
if work_on_PC:
self.refer_id_f, self.refer_id_b, self.reorder_list, self.qp_dict, self.tid_dict, self.p_list = get_order(r"D:\VTM\VVCLAST\log\qp%d_frame64.log" % self.qp)
else:
self.refer_id_f, self.refer_id_b, self.reorder_list, self.qp_dict, self.tid_dict, self.p_list = get_order(f"/data/fengxm/pmp_dataset/log/{opt['encode_cfg']}/{opt['encode_cfg']}_ip{self.intra_period}_q{self.qp}.log")
self.opt = opt
def __load_dataset_y__(self, file_path, width, height, frm_num, sub_sample_ratio, is10bit, mode, ref_len, gop_size, pad128):
if opt['datasets']['yuv']:
y,u,v = import_yuv420(file_path, width, height, frm_num, sub_sample_ratio, is10bit)
u,v = u.repeat(2, axis=-1).repeat(2, axis=-2), v.repeat(2, axis=-1).repeat(2, axis=-2)
self.content = np.concatenate([y[:,None],u[:,None],v[:,None]], axis=1)
if self.content.shape[-1]%128!=0 or self.content.shape[-2]%128!=0:
if pad128:
content = np.zeros((self.content.shape[0], self.content.shape[1], math.ceil(self.content.shape[2] / 128) * 128, math.ceil(self.content.shape[3] / 128) * 128))
content[:, :, :self.content.shape[2], :self.content.shape[3]] = self.content
self.content = content
else:
h_cropped, w_cropped = self.content.shape[-2]//128*128, self.content.shape[-1]//128*128
self.content = self.content[:,:,:h_cropped, :w_cropped]
else:
y = import_yuv420(file_path, width, height, frm_num, sub_sample_ratio, is10bit)
self.content = y[:,None]
if self.content.shape[-1]%128!=0 or self.content.shape[-2]%128!=0:
if pad128:
content = np.zeros((self.content.shape[0], self.content.shape[1], math.ceil(self.content.shape[2] / 128) * 128, math.ceil(self.content.shape[3] / 128) * 128))
content[:, :, :self.content.shape[2], :self.content.shape[3]] = self.content
self.content = content
else:
h_cropped, w_cropped = self.content.shape[-2]//128*128, self.content.shape[-1]//128*128
self.content = self.content[:,:,:h_cropped, :w_cropped]
def __getitem__(self, index):
p_idx = index
abs_idx = self.p_list[p_idx]
cand_frmid = [self.refer_id_f[abs_idx][0], self.refer_id_b[abs_idx][0]]
i_frame = self.content[abs_idx]
p0_frame = self.content[cand_frmid[0]]
p1_frame = self.content[cand_frmid[1]]
if self.opt['network']['qml'] or self.opt['network']['tml']:
qp_list = [self.qp_dict[abs_idx], self.qp_dict[cand_frmid[0]], self.qp_dict[cand_frmid[1]]]
tid_list = [self.tid_dict[abs_idx], self.tid_dict[cand_frmid[0]], self.tid_dict[cand_frmid[1]]]
output = {'i_frame': i_frame, 'p0_frame': p0_frame, 'p1_frame':p1_frame, "qp_list": torch.tensor(qp_list), "tid_list": torch.tensor(tid_list)}
else:
output = {'i_frame': i_frame, 'p0_frame': p0_frame, 'p1_frame':p1_frame}
return output
def __len__(self):
return len(self.p_list)
def YUV2RGB(Y,U,V, isYUV420 = True):
"""Y: (frame_num, height, width), U/V: (frame_num, height//2, width//2)"""
FRAME_NUM, IMG_HEIGHT, IMG_WIDTH = Y.shape[-3], Y.shape[-2], Y.shape[-1]
bgr_data = torch.zeros(FRAME_NUM, 3, IMG_HEIGHT, IMG_WIDTH, dtype=torch.uint8).to(Y.device)
if (isYUV420):
U = F.interpolate(U, scale_factor=2)
V = F.interpolate(V, scale_factor=2)
c = (Y-16) * 298
d = U - 128
e = V - 128
r = torch.floor((c + 409 * e + 128) / 256).long()
g = torch.floor((c - 100 * d - 208 * e + 128)/256).long()
b = torch.floor((c + 516 * d + 128)/256).long()
r = torch.where(r < 0, 0, r)
r = torch.where(r > 255,255,r)
g = torch.where(g < 0, 0, g)
g = torch.where(g > 255,255,g)
b = torch.where(b < 0, 0, b)
b = torch.where(b > 255,255,b)
bgr_data[:, 2, :, :] = r
bgr_data[:, 1, :, :] = g
bgr_data[:, 0, :, :] = b
# return (n,3,h,w)
return bgr_data
@torch.no_grad()
def flow_norm(tensorFlow):
Backward_tensorGrid_cpu = {}
tensorHorizontal = torch.linspace(-1.0, 1.0, tensorFlow.size(3)).view(
1, 1, 1, tensorFlow.size(3)).expand(tensorFlow.size(0), -1, tensorFlow.size(2), -1)
tensorVertical = torch.linspace(-1.0, 1.0, tensorFlow.size(2)).view(
1, 1, tensorFlow.size(2), 1).expand(tensorFlow.size(0), -1, -1, tensorFlow.size(3))
Backward_tensorGrid_cpu[str(tensorFlow.size())] = torch.cat(
[tensorHorizontal, tensorVertical], 1).cpu()
tensorFlow = torch.cat([tensorFlow[:, 0:1, :, :] / ((tensorFlow.size(3) - 1.0) / 2.0),
tensorFlow[:, 1:2, :, :] / ((tensorFlow.size(2) - 1.0) / 2.0)], 1)
return tensorFlow
@torch.no_grad()
def dataset2dataset(i_frame, p0_frame, p1_frame, flow_net, ds=1, label_batch_list=None, out_type='res', flag=None, p0_flow=None, p1_flow=None):
"""
Returns:
- input_batch: A batch containing the current frame's YUV texture, the residuals from warping the previous and current frames,
the residuals from warping the next and current frames, or the p_frame_aligned after 64x64 alignment.
- flow: The optical flow and MDF from the previous frame, and the optical flow and MDF from the next frame.
- When mtt-mask is set to true, the initial optical flow is reused from p0_flow and p1_flow.
"""
if flag == 'mtt':
Lm = 0
elif flag == 'mtt_mask':
Lm = 1
elif flag == 'qt':
Lm = 2
if not opt['network']['light_spynet']:
I_frame_YYY = F.interpolate(i_frame[:, 0:1], scale_factor=1 / ds, mode='nearest').repeat(1, 3, 1, 1) / 255.0
P0_frame_YYY = F.interpolate(p0_frame[:, 0:1], scale_factor=1 / ds, mode='nearest').repeat(1, 3, 1, 1) / 255.0
P1_frame_YYY = F.interpolate(p1_frame[:, 0:1], scale_factor=1 / ds, mode='nearest').repeat(1, 3, 1, 1) / 255.0
P0_flow_list = flow_net(im1=I_frame_YYY, im2=P0_frame_YYY, Lm=Lm, init_flow=p0_flow)
P1_flow_list = flow_net(im1=I_frame_YYY, im2=P1_frame_YYY, Lm=Lm, init_flow=p1_flow)
if flag == 'qt':
return P0_flow_list[-3], P1_flow_list[-3]
elif flag == 'mtt_mask':
return [P0_flow_list[-2]], [P1_flow_list[-2]]
elif flag == 'mtt':
return [P0_flow_list[-3], P0_flow_list[-2], P0_flow_list[-1]], [P1_flow_list[-3], P1_flow_list[-2], P1_flow_list[-1]]
else:
raise Exception('invalid flag')
else:
I_frame_YYY = F.interpolate(i_frame[:, 0:1], scale_factor=1 / ds, mode='nearest').repeat(1, 1, 1, 1) / 255.0
P0_frame_YYY = F.interpolate(p0_frame[:, 0:1], scale_factor=1 / ds, mode='nearest').repeat(1, 1, 1, 1) / 255.0
P1_frame_YYY = F.interpolate(p1_frame[:, 0:1], scale_factor=1 / ds, mode='nearest').repeat(1, 1, 1, 1) / 255.0
p0_flow = flow_net(im1=I_frame_YYY, im2=P0_frame_YYY)
p1_flow = flow_net(im1=I_frame_YYY, im2=P1_frame_YYY)
return p0_flow, p1_flow
def metric_qt(out_qt, label_qt):
accu, error = 0, 0
accu = np.sum((out_qt.round() == label_qt)) / label_qt.size
return accu, error
def print_metric(m_label_depth_list, out_pred_qt, out_pred_mtt=None, out_pred_dir=None, out_pred_mask=None):
if out_pred_mask is not None:
out_pred_mask = out_pred_mask.round()
accu_list = [0, 0, 0, 0, 0, 0, 0]
recall_list = [0, 0, 0, 0, 0, 0, 0]
zero_rate_list = [0,0,0,0,0,0, 0, 0]
zero_rate_list[0] = np.sum(out_pred_qt[:,::2,::2].round() == 0) / m_label_depth_list[0].size * 100
recall_list[0] = np.sum((out_pred_qt[:,::2,::2].round() == m_label_depth_list[0].round()) * ( m_label_depth_list[0].round() != 0)) / np.sum( m_label_depth_list[0].round() != 0) * 100
accu_list[0] = np.sum(out_pred_qt[:,::2,::2].round() == m_label_depth_list[0].round()) / out_pred_qt[:,::2,::2].size * 100
accu_label = 0
for pred_depth in range(3):
mtt_depth = pred_depth + 1
accu_label += m_label_depth_list[mtt_depth][:, 0]
dir_label = m_label_depth_list[mtt_depth][:, 1]
if out_pred_mtt is None:
zero_rate_list[mtt_depth] = 0
recall_list[mtt_depth] = 0
accu_list[mtt_depth] = 0
else:
zero_rate_list[mtt_depth] = np.sum(out_pred_mtt[:, pred_depth].round() == 0) / accu_label.size * 100
recall_list[mtt_depth] = np.sum((out_pred_mtt[:, pred_depth].round() == accu_label.round()) * ( out_pred_mtt[:, pred_depth].round() != 0)) / np.sum( out_pred_mtt[:, pred_depth].round() != 0) * 100
accu_list[mtt_depth] = np.sum(out_pred_mtt[:, pred_depth].round() == accu_label.round()) / out_pred_mtt[:, pred_depth].size * 100
if out_pred_dir is None:
zero_rate_list[mtt_depth + 3] = 0
recall_list[mtt_depth + 3] = 0
accu_list[mtt_depth + 3] = 0
else:
zero_rate_list[mtt_depth + 3] = np.sum(out_pred_dir[:, pred_depth].round() == 0) / dir_label.size * 100
recall_list[mtt_depth + 3] = np.sum((out_pred_dir[:, pred_depth].round() == dir_label.round()) * ( out_pred_dir[:, pred_depth].round() != 0)) / np.sum( out_pred_dir[:, pred_depth].round() != 0) * 100
accu_list[mtt_depth + 3] = np.sum(out_pred_dir[:, pred_depth].round() == dir_label.round()) / out_pred_dir[:, pred_depth].size * 100
# mt mask
sum_mtt_weight = m_label_depth_list[1][:, 0] + m_label_depth_list[2][:, 0] + m_label_depth_list[3][:, 0]
sum_mtt_weight = ((F.max_pool2d(torch.from_numpy(sum_mtt_weight).float(), kernel_size=4) + m_label_depth_list[0]) > torch.from_numpy(out_pred_qt[:,::2,::2]).round())
mtt_mask_label = rearrange(rearrange(sum_mtt_weight, 'b (hi h) (wi w) -> b hi h wi w', h=8, w=8), 'b hi h wi w -> b hi wi h w').sum(dim=(3, 4)).float()
mtt_mask_label = torch.clip(mtt_mask_label, min=0, max=1)
if out_pred_mask is None:
accu_mask = 0
zero_mask = 0
else:
accu_mask = torch.sum(torch.from_numpy(out_pred_mask).flatten() == mtt_mask_label.flatten()) / mtt_mask_label.numel() * 100
zero_mask = np.sum(out_pred_mask == 0) / mtt_mask_label.numel() * 100
print('TEST_METRIC of DEPTH ')
for pred_depth in range(7):
if pred_depth < 4:
print("depth%d\t accuracy: " % pred_depth, '%.4f%%' % accu_list[pred_depth], "recall: ", \
'%.4f%%' % recall_list[pred_depth], "\t zero_rate: ", '%.4f%%' % zero_rate_list[pred_depth])
else:
print("dire%d\t accuracy: " % (pred_depth-4), '%.4f%%' % accu_list[pred_depth], "recall: ", \
'%.4f%%' % recall_list[pred_depth], "\t zero_rate: ", '%.4f%%' % zero_rate_list[pred_depth])
if pred_depth == 3:
print("--------------------------------")
print("--------------------------------")
print("mtt mask:\t accuracy: %.2f \t zero_rate: %.2f"%(accu_mask, zero_mask) )
return accu_list, accu_mask
@torch.no_grad()
def inference_VVC(qp):
print("QP:", qp)
target_qp = qp
if qp <= 24:
target_qp = 22
elif qp >= 25 and qp <= 29:
target_qp = 27
elif qp >= 30 and qp <= 34:
target_qp = 32
elif qp >= 35:
target_qp = 37
else:
raise Exception('invalid qp')
inference_time, post_time = {}, {}
mask_ratio = {}
print('stage: ', opt['stage'])
qp_dict = {22:0, 27:1, 32:2, 37:3}
qp_id = qp_dict[target_qp]
stage = opt['stage']
if target_qp == 22:
text_fe = True
else:
text_fe = False
if opt['network']['large_model'][qp_id]:
if target_qp == 37:
guide = False
else:
guide = True
from models.backbone import QT_Net_HLG
qt_net = QT_Net_HLG(spp=False, qml=True, tml=True, guide=guide)
elif 'backbone' in opt['network'] and opt['network']['backbone'] != 'HourglassBlock':
from models.backbone import QT_Net_HLG
qt_net = QT_Net_HLG(qml=opt['network']['qml'], guide=True, backbone=opt['network']['backbone'])
else:
from models.backbone import QT_Net_HLG
qt_net = QT_Net_HLG(qml=opt['network']['qml'], guide=True)
from spynet.Spy_net import ME_Spynet
flow_net = ME_Spynet(me_model_dir=opt['path']['me_model_dir'])
mtt_mask_net = MTT_mask_net(qml=opt['network']['qml'], dlm=opt['network']['mtt_mask_dlm'])
mtt_depth_net = MTT_Net_HLG(qml=opt['network']['qml'], residual_type=opt['network']['residual_type'], max_depth=3)
mtt_dire_net = MTT_Net_HLG(qml=opt['network']['qml'], residual_type=opt['network']['residual_type'], max_depth=3)
if torch.cuda.is_available() and opt['gpu_num'] > 0:
print("Testing on GPU!")
qt_net = qt_net.cuda()
mtt_mask_net = mtt_mask_net.cuda()
flow_net = flow_net.cuda()
mtt_dire_net = mtt_dire_net.cuda()
mtt_depth_net = mtt_depth_net.cuda()
if opt['network']['large_model'][qp_id]:
qt_net_path = os.path.join(opt['path']['qt_large_model_path'], 'model_qp%d.pth'%target_qp) if opt['path']['qt_model_path'][-4:] != ".pth" else opt['path']['qt_model_path']
else:
qt_net_path = os.path.join(opt['path']['qt_model_path'], 'model_qp%d.pth'%target_qp) if opt['path']['qt_model_path'][-4:] != ".pth" else opt['path']['qt_model_path']
if stage >= 2:
mtt_mask_net_path = os.path.join(opt['path']['mtt_mask_model_path'], 'model_qp%d.pth'%target_qp)
if stage >= 3:
mtt_depth_net_path = os.path.join(opt['path']['mtt_model_path'], 'model_qp%d.pth'%target_qp)
mtt_dire_net_path = os.path.join(opt['path']['dire_model_path'], 'model_qp%d.pth'%target_qp)
if torch.cuda.is_available() and opt['gpu_num'] > 0:
if torch.cuda.device_count() > 1:
qt_net.load_state_dict(torch.load(qt_net_path)['qt_net'])
if stage >= 2:
mtt_mask_net.load_state_dict(torch.load(mtt_mask_net_path)['mtt_mask_net'])
if stage >= 3:
mtt_depth_net.load_state_dict(torch.load(mtt_depth_net_path)['mtt_mask_net'])
mtt_dire_net.load_state_dict(torch.load(mtt_dire_net_path)['mtt_mask_net'])
else:
qt_net.load_state_dict({k.replace('module.', ''): v for k, v in torch.load(qt_net_path)['qt_net'].items()})
if stage >= 2:
mtt_mask_net.load_state_dict({k.replace('module.', ''): v for k, v in torch.load(mtt_mask_net_path)['mtt_mask_net'].items()})
if stage >= 3:
mtt_depth_net.load_state_dict({k.replace('module.', ''): v for k, v in torch.load(mtt_depth_net_path)['mtt_depth_model'].items()})
mtt_dire_net.load_state_dict({k.replace('module.', ''): v for k, v in torch.load(mtt_dire_net_path)['mtt_dire_model'].items()})
else:
qt_net.load_state_dict({k.replace('module.', ''): v for k, v in torch.load(qt_net_path, map_location=torch.device('cpu'))['qt_net'].items()})
if stage >= 2:
mtt_mask_net.load_state_dict({k.replace('module.', ''): v for k, v in torch.load(mtt_mask_net_path, map_location=torch.device('cpu'))['mtt_mask_net'].items()})
if stage >= 3:
mtt_depth_net.load_state_dict({k.replace('module.', ''): v for k, v in torch.load(mtt_depth_net_path, map_location=torch.device('cpu'))['mtt_depth_model'].items()})
mtt_dire_net.load_state_dict({k.replace('module.', ''): v for k, v in torch.load(mtt_dire_net_path, map_location=torch.device('cpu'))['mtt_dire_model'].items()})
flow_net = flow_net.eval()
qt_net = qt_net.eval()
mtt_mask_net = mtt_mask_net.eval()
mtt_depth_net = mtt_depth_net.eval()
mtt_dire_net = mtt_dire_net.eval()
if torch.cuda.is_available() and opt['gpu_num'] > 0:
pass
else:
flow_net = flow_net.cpu()
qt_net = qt_net.cpu()
mtt_mask_net = mtt_mask_net.cpu()
mtt_depth_net = mtt_depth_net.cpu()
mtt_dire_net = mtt_dire_net.cpu()
# load video
seqs_info_path = "VVC_RA_Test_Sequences.txt"
# seqs_info_fp = open(seqs_info_path, 'r')
seqs = []
with open(seqs_info_path, 'r') as seqs_info_fp:
for line in seqs_info_fp:
if line is None:
break
seqs.append(line.rstrip('\n').split(',')[0])
# hyper parameters
seq_accu_list = []
for seq_id, seq_name in enumerate(seqs):
if seq_id not in opt['datasets']['test_ids']:
continue
start_time = time.time()
if 'RaceHorses' in seq_name and '832x480' in seq_name:
cfg_path = '/data/fengxm/pmp_dataset/log/cfg/RaceHorsesC.cfg'
else:
cfg_path = os.path.join("/data/fengxm/pmp_dataset/log/cfg" if opt['path']['cfg_path'] is None else opt['path']['cfg_path'], seq_name.split('_')[0] + ".cfg")
print("SEQ:", seq_name)
seq_path, bit_depth, width, height, frame_num, intra_period = load_ifo_from_cfg(cfg_path)
# continue
# frame_num = opt['datasets']['frm_num']
frame_num, gop_size = int(opt['encode_cfg'].split('_')[0].strip('f')), int(opt['encode_cfg'].split('_')[1].strip('gop'))
if gop_size == 32 and intra_period == 48:
intra_period = 64
width, height = int(seq_name.split('_')[1].split('x')[0]), int(seq_name.split('_')[1].split('x')[1]) # to make Horsers and HorsersC different
seq_path = seq_name + '.yuv'
if bit_depth != 10:
is10bit = False
else:
is10bit = True
input_path = os.path.join(opt['path']['test_seq_dir'], seq_path)
if opt['boundary_handling']:
width_round, height_round = math.ceil(width / 128) * 128, math.ceil(height / 128) * 128
block_width_num, block_height_num = width_round // 128, height_round // 128
else:
block_width_num, block_height_num = width // 128, height // 128
width_round, height_round = block_width_num * 128, block_height_num * 128
if opt['module']['metric']:
dataset_dir = opt['module']['metric_path']
qt_label = np.load(os.path.join(dataset_dir, opt['enc_mode'] , 'QP%d'%target_qp, 'qt_label', '%s.npy'%seq_id))
mtt_label = np.load(os.path.join(dataset_dir, opt['enc_mode'] , 'QP%d'%target_qp, 'mtt_label', '%s.npy'%seq_id))
m_label_depth_list = [qt_label, mtt_label[:,0], mtt_label[:,1], mtt_label[:,2]]
else:
m_label_depth_list = None
train_dataset = MyDataset(input_path, width, height, frm_num=frame_num, sub_sample_ratio=1, is10bit=is10bit, \
ref_len=2 if opt['enc_mode'] == 'RA' else 8, mode=opt['enc_mode'], qp=qp, pad128=opt['boundary_handling'], opt=opt, intra_period=intra_period)
train_data = DataLoader(train_dataset, batch_size=opt['datasets']['batchSize'], shuffle=False, num_workers=opt['datasets']['num_workers'], pin_memory=True, multiprocessing_context='spawn')
p_frame_num = len(train_dataset.p_list)
out_pred_qt = np.zeros((p_frame_num, height_round // 8, width_round // 8))
out_pred_mt_mask = np.zeros((p_frame_num, height_round // 128, width_round // 128))
out_pred_mtt = np.zeros((p_frame_num, 3, height_round // 4, width_round // 4))
out_pred_dir = np.zeros((p_frame_num, 3, height_round // 4, width_round // 4))
frm_id = 0
for step, batch in tqdm(enumerate(train_data)):
if torch.cuda.is_available() and opt['gpu_num'] > 0:
i_frame, p0_frame, p1_frame = batch['i_frame'].cuda().float(), batch['p0_frame'].cuda().float(), batch['p1_frame'].cuda().float() # input_batch shape(B, ref_frm_num+1, 384, 384), shape(B, ref_frm_num+1, 2, 192, 192)
temp_qp_list = batch['qp_list'].cuda().float()
else:
i_frame, p0_frame, p1_frame = batch['i_frame'].float(), batch['p0_frame'].float(), batch['p1_frame'].float()
temp_qp_list = batch['qp_list'].float()
if stage == 1:
if opt['network']['large_model'][qp_id]:
if opt['network']['light_spynet']:
p0_flow, p1_flow = dataset2dataset(i_frame, p0_frame, p1_frame, flow_net, ds=1, flag='mtt')
else:
p0_flow, p1_flow = dataset2dataset(i_frame, p0_frame, p1_frame, flow_net, ds=1, flag='mtt')
p0_flow, p1_flow = p0_flow[-1], p1_flow[-1]
qt_pred_list, _, _ = qt_net(i_frame[:, 0:1] / 255.0, torch.stack([p0_flow, p1_flow], dim=1), qp=temp_qp_list[:, 0:1],\
trans_flow_DAM=True, make_res=True, p0_frame=p0_frame[:, 0:1] / 255.0, p1_frame=p1_frame[:, 0:1] / 255.0, out_medium_feat=True, upsample=0)
else:
# new architecture
i_frame, p0_frame, p1_frame = i_frame[:, :, ::4, ::4], p0_frame[:, :, ::4, ::4], p1_frame[:, :, ::4, ::4]
p0_flow, p1_flow = dataset2dataset(i_frame, p0_frame, p1_frame, flow_net, ds=1, flag='qt')
qt_pred_list, qt_feature = qt_net(i_frame[:, 0:1] / 255.0, torch.stack([p0_flow, p1_flow], dim=1), qp=temp_qp_list[:, 0:1], trans_flow_DAM=True, p0_frame=p0_frame[:, 0:1] / 255.0, p1_frame=p1_frame[:, 0:1] / 255.0, out_medium_feat=True, upsample=4)
qt_pred_list = [ele * 4 - 0.5 for ele in qt_pred_list]
qt_pred = qt_pred_list[-1]
out_pred_qt[frm_id:frm_id+i_frame.shape[0]] = F.interpolate(qt_pred, scale_factor=2)[:,0].cpu()
max_pred_depth = 1
elif stage == 2:
# mtt mask
if opt['network']['large_model'][qp_id]:
p0_flow, p1_flow = dataset2dataset(i_frame, p0_frame, p1_frame, flow_net, ds=1, flag='mtt')
qt_pred_list, _, _ = qt_net(i_frame[:, 0:1] / 255.0, torch.stack([p0_flow[-1], p1_flow[-1]], dim=1), qp=temp_qp_list[:, 0:1],\
trans_flow_DAM=True, make_res=True, p0_frame=p0_frame[:, 0:1] / 255.0, p1_frame=p1_frame[:, 0:1] / 255.0, out_medium_feat=True, upsample=0)
p0_flow, p1_flow = p0_flow[-2], p1_flow[-2]
else:
p0_flow, p1_flow = dataset2dataset(F.interpolate(i_frame, scale_factor=1/4, mode='bilinear'), F.interpolate(p0_frame, scale_factor=1/4, mode='bilinear'), F.interpolate(p1_frame, scale_factor=1/4, mode='bilinear'), flow_net, ds=1, flag='mtt')
qt_pred_list, _ = qt_net(F.interpolate(i_frame, scale_factor=1/4, mode='bilinear') / 255.0, torch.stack([p0_flow[-1], p1_flow[-1]], dim=1), qp=temp_qp_list[:, 0:1], trans_flow_DAM=True, p0_frame=F.interpolate(p0_frame, scale_factor=1/4, mode='bilinear') / 255.0, p1_frame=F.interpolate(p1_frame, scale_factor=1/4, mode='bilinear') / 255.0, out_medium_feat=True, upsample=4)
p0_flow, p1_flow = dataset2dataset(F.interpolate(i_frame, scale_factor=1/2, mode='bilinear'), F.interpolate(p0_frame, scale_factor=1/2, mode='bilinear'), F.interpolate(p1_frame, scale_factor=1/2, mode='bilinear'),\
flow_net, ds=1, flag='mtt_mask', p0_flow=p0_flow[-1], p1_flow=p1_flow[-1])
p0_flow, p1_flow = p0_flow[-1], p1_flow[-1]
qt_pred_list = [ele * 4 - 0.5 for ele in qt_pred_list]
qt_pred = qt_pred_list[-1]
out_pred_qt[frm_id:frm_id+i_frame.shape[0]] = F.interpolate(qt_pred, scale_factor=2)[:,0].cpu()
mtt_mask_list = mtt_mask_net(F.interpolate(i_frame, scale_factor=1/2, mode='bilinear'), p0_flow=p0_flow, p1_flow=p1_flow, qt_pred=qt_pred, qp=temp_qp_list[:, 0:1], trans_flow_DAM=True, p0_frame=F.interpolate(p0_frame, scale_factor=1/2, mode='bilinear'), p1_frame=F.interpolate(p1_frame, scale_factor=1/2, mode='bilinear'))
mtt_mask_pred = mtt_mask_list[-1]
out_pred_qt[frm_id:frm_id+i_frame.shape[0]] = F.interpolate(qt_pred, scale_factor=2)[:,0].cpu()
out_pred_mt_mask[frm_id:frm_id+i_frame.shape[0]] = mtt_mask_pred.reshape(out_pred_mt_mask.shape[-2], out_pred_mt_mask.shape[-1]).cpu().numpy()
max_pred_depth = 2
elif stage >= 3:
temp_qp_list = torch.ones_like(temp_qp_list) * target_qp
if opt['network']['large_model'][qp_id]:
p0_flow, p1_flow = dataset2dataset(i_frame, p0_frame, p1_frame, flow_net, ds=1, flag='mtt')
qt_pred_list, _, _ = qt_net(i_frame[:, 0:1] / 255.0, torch.stack([p0_flow[-1], p1_flow[-1]], dim=1), qp=temp_qp_list[:, 0:1],\
trans_flow_DAM=True, make_res=True, p0_frame=p0_frame[:, 0:1] / 255.0, p1_frame=p1_frame[:, 0:1] / 255.0, out_medium_feat=True, upsample=0)
else:
p0_flow, p1_flow = dataset2dataset(F.interpolate(i_frame, scale_factor=1/4, mode='bilinear'), F.interpolate(p0_frame, scale_factor=1/4, mode='bilinear'), F.interpolate(p1_frame, scale_factor=1/4, mode='bilinear'), flow_net, ds=1, flag='mtt')
qt_pred_list, _ = qt_net(F.interpolate(i_frame, scale_factor=1/4, mode='bilinear') / 255.0, torch.stack([p0_flow[-1], p1_flow[-1]], dim=1), qp=temp_qp_list[:, 0:1], trans_flow_DAM=True, p0_frame=F.interpolate(p0_frame, scale_factor=1/4, mode='bilinear') / 255.0, p1_frame=F.interpolate(p1_frame, scale_factor=1/4, mode='bilinear') / 255.0, out_medium_feat=True, upsample=4)
p0_flow, p1_flow = dataset2dataset(i_frame, p0_frame, p1_frame, flow_net, ds=1, flag='mtt', p0_flow=p0_flow[-1], p1_flow=p1_flow[-1])
qt_pred_list = [ele * 4 - 0.5 for ele in qt_pred_list]
qt_pred = qt_pred_list[-1]
out_pred_qt[frm_id:frm_id+i_frame.shape[0]] = F.interpolate(qt_pred, scale_factor=2)[:,0].cpu()
if target_qp == 27:
if seq_id == 10:
mtt_mask_net_path_tmp = os.path.join(opt['path']['mtt_mask_model_path'], 'model_qp22.pth')
if torch.cuda.is_available() and opt['gpu_num'] > 0:
if torch.cuda.device_count() > 1:
mtt_mask_net.load_state_dict(torch.load(mtt_mask_net_path_tmp)['mtt_mask_net'])
else:
mtt_mask_net.load_state_dict({k.replace('module.', ''): v for k, v in torch.load(mtt_mask_net_path_tmp)['mtt_mask_net'].items()})
else:
mtt_mask_net.load_state_dict({k.replace('module.', ''): v for k, v in torch.load(mtt_mask_net_path_tmp, map_location=torch.device('cpu'))['mtt_mask_net'].items()})
else:
mtt_mask_net_path_tmp = os.path.join(opt['path']['mtt_mask_model_path'], 'model_qp27.pth')
if torch.cuda.is_available() and opt['gpu_num'] > 0:
if torch.cuda.device_count() > 1:
mtt_mask_net.load_state_dict(torch.load(mtt_mask_net_path_tmp)['mtt_mask_net'])
else:
mtt_mask_net.load_state_dict({k.replace('module.', ''): v for k, v in torch.load(mtt_mask_net_path_tmp)['mtt_mask_net'].items()})
else:
mtt_mask_net.load_state_dict({k.replace('module.', ''): v for k, v in torch.load(mtt_mask_net_path_tmp, map_location=torch.device('cpu'))['mtt_mask_net'].items()})
mtt_mask_list = mtt_mask_net(F.interpolate(i_frame, scale_factor=1/2, mode='bilinear'), p0_flow=p0_flow[-2], p1_flow=p1_flow[-2], qt_pred=qt_pred, qp=temp_qp_list[:, 0:1], trans_flow_DAM=True, p0_frame=F.interpolate(p0_frame, scale_factor=1/2, mode='bilinear'), p1_frame=F.interpolate(p1_frame, scale_factor=1/2, mode='bilinear'))
mtt_mask_pred = mtt_mask_list[-1]
out_pred_mt_mask[frm_id:frm_id+i_frame.shape[0]] = mtt_mask_pred.reshape(out_pred_mt_mask.shape[-2], out_pred_mt_mask.shape[-1]).cpu().numpy()
mtt_depth_map_list, ctu_decision, drop_decision = mtt_depth_net(luma=i_frame[:, 0:1] / 255.0, p0_flow=p0_flow[-1], p1_flow=p1_flow[-1], qt_pred=qt_pred, qp=temp_qp_list[:, 0:1], trans_flow_DAM=True, p0_frame=p0_frame / 255.0, p1_frame=p1_frame / 255.0, ctu_decision= mtt_mask_list[-1].view(i_frame.shape[0], -1), mask_ratio=0.3)
mtt_dire_map_list, _, _ = mtt_dire_net(luma=i_frame[:, 0:1] / 255.0, p0_flow=p0_flow[-1], p1_flow=p1_flow[-1], qt_pred=qt_pred, qp=temp_qp_list[:, 0:1], trans_flow_DAM=True, p0_frame=p0_frame / 255.0, p1_frame=p1_frame / 255.0, ctu_decision= mtt_mask_list[-1].view(i_frame.shape[0], -1), mask_ratio=0.3)
mtt_pred, mtt_dire_pred = [], []
for layer_depth in range(3):
if opt['open_offset']:
offset = torch.zeros_like(mtt_depth_map_list[layer_depth])
offset[:,:,:,2] = opt['offset'][qp_id][layer_depth]
save_mtt_depth = torch.argmax(torch.softmax(mtt_depth_map_list[layer_depth], dim=-1) + offset, dim=-1)
else:
save_mtt_depth = torch.argmax(mtt_depth_map_list[layer_depth], dim=-1)
mtt_pred_single = torch.zeros(i_frame.shape[0], i_frame.shape[-1] // 128 * i_frame.shape[-2] // 128, 32, 32)
drop_mtt_depth = torch.zeros(i_frame.shape[0], i_frame.shape[-1] // 128 * i_frame.shape[-2] // 128 - save_mtt_depth.shape[0], 32, 32)
if torch.cuda.is_available() and opt['gpu_num'] > 0:
mtt_pred_single, drop_mtt_depth = mtt_pred_single.cuda(), drop_mtt_depth.cuda()
mtt_pred_single = rearrange(mtt_pred_single, 'b n h w -> b n (h w)')
drop_mtt_depth = rearrange(drop_mtt_depth, 'b n h w -> b n (h w)')
save_mtt_depth = rearrange(save_mtt_depth, '(b n) h w -> b n (h w)', b=i_frame.shape[0]).float()
mtt_pred_single = batch_index_fill(mtt_pred_single, save_mtt_depth, drop_mtt_depth, ctu_decision, drop_decision)
mtt_pred.append(rearrange(mtt_pred_single, 'b (hi wi) (h w) -> b (hi h) (wi w)', hi=i_frame.shape[-2] // 128, wi=i_frame.shape[-1] // 128, h=32, w=32))
# dire
save_mtt_depth = torch.argmax(mtt_dire_map_list[layer_depth], dim=-1) - 1
mtt_pred_single = torch.zeros(i_frame.shape[0], i_frame.shape[-1] // 128 * i_frame.shape[-2] // 128, 32, 32)
drop_mtt_depth = torch.zeros(i_frame.shape[0], i_frame.shape[-1] // 128 * i_frame.shape[-2] // 128 - save_mtt_depth.shape[0], 32, 32)
if torch.cuda.is_available() and opt['gpu_num'] > 0:
save_mtt_depth, mtt_pred_single, drop_mtt_depth = save_mtt_depth.cuda(), mtt_pred_single.cuda(), drop_mtt_depth.cuda()
mtt_pred_single = rearrange(mtt_pred_single, 'b n h w -> b n (h w)')
drop_mtt_depth = rearrange(drop_mtt_depth, 'b n h w -> b n (h w)')
save_mtt_depth = rearrange(save_mtt_depth, '(b n) h w -> b n (h w)', b=i_frame.shape[0]).float()
mtt_pred_single = batch_index_fill(mtt_pred_single, save_mtt_depth, drop_mtt_depth, ctu_decision, drop_decision)
mtt_dire_pred.append(rearrange(mtt_pred_single, 'b (hi wi) (h w) -> b (hi h) (wi w)', hi=i_frame.shape[-2] // 128, wi=i_frame.shape[-1] // 128, h=32, w=32))