Initial commit

.gitignore

@@ -0,0 +1,2 @@
+temps/*
+!temps/.gitkeep

app.py

@@ -0,0 +1,72 @@
+import base64
+import json
+import os, shutil
+import re
+import time
+import uuid
+
+import cv2
+
+import numpy as np
+import streamlit as st
+from PIL import Image
+# from extract_video import extract_method_single_video
+
+import shlex
+import subprocess
+
+def main():
+    st.markdown("###")    
+    uploaded_file = st.file_uploader('Upload a picture', type=['mp4', 'jpg', 'jpeg', 'png'], accept_multiple_files=False)
+    if uploaded_file: 
+        random_id = uuid.uuid1()
+        filename = "{}.{}".format(random_id, uploaded_file.type.split("/")[-1])
+        file_type = uploaded_file.type.split("/")[0]
+
+        if uploaded_file.type == 'video/mp4':
+            with open(f"temps/{filename}", mode='wb') as f:
+                f.write(uploaded_file.read())
+            st.video(uploaded_file)
+            pass
+        else:
+            img = Image.open(uploaded_file).convert('RGB')
+            ext = uploaded_file.type.split("/")[-1]
+            with open(f"temps/{filename}", mode='wb') as f:
+                f.write(uploaded_file.getbuffer())  
+            st.image(img)
+
+        
+
+        with st.spinner(f'Processing {file_type}...'):
+            subprocess.run(shlex.split(f"python.exe extract_video.py --device cuda --max_frames 50 --bs 2 --frame_interval 5 --confidence_threshold 0.997 --data_path temps/{filename}"))
+        st.text(f'1. Processing {file_type} ✅')
+        with st.spinner(f'Analyzing {file_type}...'):
+            pred = subprocess.run(shlex.split(f"python inference.py --weight weights/model_params_ffpp_c23.pickle --device cuda --image_folder temps/images/{filename}"), capture_output=True)
+        st.text(f'2. Analyzing {file_type} ✅')
+        
+        print(pred)
+        try:
+            fake_probability = float(pred.stdout.decode('utf-8').split('Mean prediction: ')[-1])
+            if fake_probability > 0.6:
+                st.error('  FAKE! ', icon="🚨")
+            else:
+                st.success("  REAL FOOTAGE! ", icon="✅")
+            st.text("fake probability {:.2f}".format(fake_probability))
+
+            os.remove(f"temps/{filename}")
+            folder_name = ".".join(filename.split(".")[:-1])
+            shutil.rmtree(f"temps/images/{folder_name}")
+        except:
+            st.text(pred.stdout.decode('utf-8'))
+
+            st.text("")
+            st.text(pred)
+
+
+
+if __name__ == "__main__":
+    st.set_page_config(
+        page_title="Nodeflux Deepfake Detection", page_icon=":pencil2:"
+    )
+    st.title("Deepfake Detection")
+    main()

data/FDDB/img_list.txt

@@ -0,0 +1,2845 @@
+2002/08/11/big/img_591
+2002/08/26/big/img_265
+2002/07/19/big/img_423
+2002/08/24/big/img_490
+2002/08/31/big/img_17676
+2002/07/31/big/img_228
+2002/07/24/big/img_402
+2002/08/04/big/img_769
+2002/07/19/big/img_581
+2002/08/13/big/img_723
+2002/08/12/big/img_821
+2003/01/17/big/img_610
+2002/08/13/big/img_1116
+2002/08/28/big/img_19238
+2002/08/21/big/img_660
+2002/08/14/big/img_607
+2002/08/05/big/img_3708
+2002/08/19/big/img_511
+2002/08/07/big/img_1316
+2002/07/25/big/img_1047
+2002/07/23/big/img_474
+2002/07/27/big/img_970
+2002/09/02/big/img_15752
+2002/09/01/big/img_16378
+2002/09/01/big/img_16189
+2002/08/26/big/img_276
+2002/07/24/big/img_518
+2002/08/14/big/img_1027
+2002/08/24/big/img_733
+2002/08/15/big/img_249
+2003/01/15/big/img_1371
+2002/08/07/big/img_1348
+2003/01/01/big/img_331
+2002/08/23/big/img_536
+2002/07/30/big/img_224
+2002/08/10/big/img_763
+2002/08/21/big/img_293
+2002/08/15/big/img_1211
+2002/08/15/big/img_1194
+2003/01/15/big/img_390
+2002/08/06/big/img_2893
+2002/08/17/big/img_691
+2002/08/07/big/img_1695
+2002/08/16/big/img_829
+2002/07/25/big/img_201
+2002/08/23/big/img_36
+2003/01/15/big/img_763
+2003/01/15/big/img_637
+2002/08/22/big/img_592
+2002/07/25/big/img_817
+2003/01/15/big/img_1219
+2002/08/05/big/img_3508
+2002/08/15/big/img_1108
+2002/07/19/big/img_488
+2003/01/16/big/img_704
+2003/01/13/big/img_1087
+2002/08/10/big/img_670
+2002/07/24/big/img_104
+2002/08/27/big/img_19823
+2002/09/01/big/img_16229
+2003/01/13/big/img_846
+2002/08/04/big/img_412
+2002/07/22/big/img_554
+2002/08/12/big/img_331
+2002/08/02/big/img_533
+2002/08/12/big/img_259
+2002/08/18/big/img_328
+2003/01/14/big/img_630
+2002/08/05/big/img_3541
+2002/08/06/big/img_2390
+2002/08/20/big/img_150
+2002/08/02/big/img_1231
+2002/08/16/big/img_710
+2002/08/19/big/img_591
+2002/07/22/big/img_725
+2002/07/24/big/img_820
+2003/01/13/big/img_568
+2002/08/22/big/img_853
+2002/08/09/big/img_648
+2002/08/23/big/img_528
+2003/01/14/big/img_888
+2002/08/30/big/img_18201
+2002/08/13/big/img_965
+2003/01/14/big/img_660
+2002/07/19/big/img_517
+2003/01/14/big/img_406
+2002/08/30/big/img_18433
+2002/08/07/big/img_1630
+2002/08/06/big/img_2717
+2002/08/21/big/img_470
+2002/07/23/big/img_633
+2002/08/20/big/img_915
+2002/08/16/big/img_893
+2002/07/29/big/img_644
+2002/08/15/big/img_529
+2002/08/16/big/img_668
+2002/08/07/big/img_1871
+2002/07/25/big/img_192
+2002/07/31/big/img_961
+2002/08/19/big/img_738
+2002/07/31/big/img_382
+2002/08/19/big/img_298
+2003/01/17/big/img_608
+2002/08/21/big/img_514
+2002/07/23/big/img_183
+2003/01/17/big/img_536
+2002/07/24/big/img_478
+2002/08/06/big/img_2997
+2002/09/02/big/img_15380
+2002/08/07/big/img_1153
+2002/07/31/big/img_967
+2002/07/31/big/img_711
+2002/08/26/big/img_664
+2003/01/01/big/img_326
+2002/08/24/big/img_775
+2002/08/08/big/img_961
+2002/08/16/big/img_77
+2002/08/12/big/img_296
+2002/07/22/big/img_905
+2003/01/13/big/img_284
+2002/08/13/big/img_887
+2002/08/24/big/img_849
+2002/07/30/big/img_345
+2002/08/18/big/img_419
+2002/08/01/big/img_1347
+2002/08/05/big/img_3670
+2002/07/21/big/img_479
+2002/08/08/big/img_913
+2002/09/02/big/img_15828
+2002/08/30/big/img_18194
+2002/08/08/big/img_471
+2002/08/22/big/img_734
+2002/08/09/big/img_586
+2002/08/09/big/img_454
+2002/07/29/big/img_47
+2002/07/19/big/img_381
+2002/07/29/big/img_733
+2002/08/20/big/img_327
+2002/07/21/big/img_96
+2002/08/06/big/img_2680
+2002/07/25/big/img_919
+2002/07/21/big/img_158
+2002/07/22/big/img_801
+2002/07/22/big/img_567
+2002/07/24/big/img_804
+2002/07/24/big/img_690
+2003/01/15/big/img_576
+2002/08/14/big/img_335
+2003/01/13/big/img_390
+2002/08/11/big/img_258
+2002/07/23/big/img_917
+2002/08/15/big/img_525
+2003/01/15/big/img_505
+2002/07/30/big/img_886
+2003/01/16/big/img_640
+2003/01/14/big/img_642
+2003/01/17/big/img_844
+2002/08/04/big/img_571
+2002/08/29/big/img_18702
+2003/01/15/big/img_240
+2002/07/29/big/img_553
+2002/08/10/big/img_354
+2002/08/18/big/img_17
+2003/01/15/big/img_782
+2002/07/27/big/img_382
+2002/08/14/big/img_970
+2003/01/16/big/img_70
+2003/01/16/big/img_625
+2002/08/18/big/img_341
+2002/08/26/big/img_188
+2002/08/09/big/img_405
+2002/08/02/big/img_37
+2002/08/13/big/img_748
+2002/07/22/big/img_399
+2002/07/25/big/img_844
+2002/08/12/big/img_340
+2003/01/13/big/img_815
+2002/08/26/big/img_5
+2002/08/10/big/img_158
+2002/08/18/big/img_95
+2002/07/29/big/img_1297
+2003/01/13/big/img_508
+2002/09/01/big/img_16680
+2003/01/16/big/img_338
+2002/08/13/big/img_517
+2002/07/22/big/img_626
+2002/08/06/big/img_3024
+2002/07/26/big/img_499
+2003/01/13/big/img_387
+2002/08/31/big/img_18025
+2002/08/13/big/img_520
+2003/01/16/big/img_576
+2002/07/26/big/img_121
+2002/08/25/big/img_703
+2002/08/26/big/img_615
+2002/08/17/big/img_434
+2002/08/02/big/img_677
+2002/08/18/big/img_276
+2002/08/05/big/img_3672
+2002/07/26/big/img_700
+2002/07/31/big/img_277
+2003/01/14/big/img_220
+2002/08/23/big/img_232
+2002/08/31/big/img_17422
+2002/07/22/big/img_508
+2002/08/13/big/img_681
+2003/01/15/big/img_638
+2002/08/30/big/img_18408
+2003/01/14/big/img_533
+2003/01/17/big/img_12
+2002/08/28/big/img_19388
+2002/08/08/big/img_133
+2002/07/26/big/img_885
+2002/08/19/big/img_387
+2002/08/27/big/img_19976
+2002/08/26/big/img_118
+2002/08/28/big/img_19146
+2002/08/05/big/img_3259
+2002/08/15/big/img_536
+2002/07/22/big/img_279
+2002/07/22/big/img_9
+2002/08/13/big/img_301
+2002/08/15/big/img_974
+2002/08/06/big/img_2355
+2002/08/01/big/img_1526
+2002/08/03/big/img_417
+2002/08/04/big/img_407
+2002/08/15/big/img_1029
+2002/07/29/big/img_700
+2002/08/01/big/img_1463
+2002/08/31/big/img_17365
+2002/07/28/big/img_223
+2002/07/19/big/img_827
+2002/07/27/big/img_531
+2002/07/19/big/img_845
+2002/08/20/big/img_382
+2002/07/31/big/img_268
+2002/08/27/big/img_19705
+2002/08/02/big/img_830
+2002/08/23/big/img_250
+2002/07/20/big/img_777
+2002/08/21/big/img_879
+2002/08/26/big/img_20146
+2002/08/23/big/img_789
+2002/08/06/big/img_2683
+2002/08/25/big/img_576
+2002/08/09/big/img_498
+2002/08/08/big/img_384
+2002/08/26/big/img_592
+2002/07/29/big/img_1470
+2002/08/21/big/img_452
+2002/08/30/big/img_18395
+2002/08/15/big/img_215
+2002/07/21/big/img_643
+2002/07/22/big/img_209
+2003/01/17/big/img_346
+2002/08/25/big/img_658
+2002/08/21/big/img_221
+2002/08/14/big/img_60
+2003/01/17/big/img_885
+2003/01/16/big/img_482
+2002/08/19/big/img_593
+2002/08/08/big/img_233
+2002/07/30/big/img_458
+2002/07/23/big/img_384
+2003/01/15/big/img_670
+2003/01/15/big/img_267
+2002/08/26/big/img_540
+2002/07/29/big/img_552
+2002/07/30/big/img_997
+2003/01/17/big/img_377
+2002/08/21/big/img_265
+2002/08/09/big/img_561
+2002/07/31/big/img_945
+2002/09/02/big/img_15252
+2002/08/11/big/img_276
+2002/07/22/big/img_491
+2002/07/26/big/img_517
+2002/08/14/big/img_726
+2002/08/08/big/img_46
+2002/08/28/big/img_19458
+2002/08/06/big/img_2935
+2002/07/29/big/img_1392
+2002/08/13/big/img_776
+2002/08/24/big/img_616
+2002/08/14/big/img_1065
+2002/07/29/big/img_889
+2002/08/18/big/img_188
+2002/08/07/big/img_1453
+2002/08/02/big/img_760
+2002/07/28/big/img_416
+2002/08/07/big/img_1393
+2002/08/26/big/img_292
+2002/08/26/big/img_301
+2003/01/13/big/img_195
+2002/07/26/big/img_532
+2002/08/20/big/img_550
+2002/08/05/big/img_3658
+2002/08/26/big/img_738
+2002/09/02/big/img_15750
+2003/01/17/big/img_451
+2002/07/23/big/img_339
+2002/08/16/big/img_637
+2002/08/14/big/img_748
+2002/08/06/big/img_2739
+2002/07/25/big/img_482
+2002/08/19/big/img_191
+2002/08/26/big/img_537
+2003/01/15/big/img_716
+2003/01/15/big/img_767
+2002/08/02/big/img_452
+2002/08/08/big/img_1011
+2002/08/10/big/img_144
+2003/01/14/big/img_122
+2002/07/24/big/img_586
+2002/07/24/big/img_762
+2002/08/20/big/img_369
+2002/07/30/big/img_146
+2002/08/23/big/img_396
+2003/01/15/big/img_200
+2002/08/15/big/img_1183
+2003/01/14/big/img_698
+2002/08/09/big/img_792
+2002/08/06/big/img_2347
+2002/07/31/big/img_911
+2002/08/26/big/img_722
+2002/08/23/big/img_621
+2002/08/05/big/img_3790
+2003/01/13/big/img_633
+2002/08/09/big/img_224
+2002/07/24/big/img_454
+2002/07/21/big/img_202
+2002/08/02/big/img_630
+2002/08/30/big/img_18315
+2002/07/19/big/img_491
+2002/09/01/big/img_16456
+2002/08/09/big/img_242
+2002/07/25/big/img_595
+2002/07/22/big/img_522
+2002/08/01/big/img_1593
+2002/07/29/big/img_336
+2002/08/15/big/img_448
+2002/08/28/big/img_19281
+2002/07/29/big/img_342
+2002/08/12/big/img_78
+2003/01/14/big/img_525
+2002/07/28/big/img_147
+2002/08/11/big/img_353
+2002/08/22/big/img_513
+2002/08/04/big/img_721
+2002/08/17/big/img_247
+2003/01/14/big/img_891
+2002/08/20/big/img_853
+2002/07/19/big/img_414
+2002/08/01/big/img_1530
+2003/01/14/big/img_924
+2002/08/22/big/img_468
+2002/08/18/big/img_354
+2002/08/30/big/img_18193
+2002/08/23/big/img_492
+2002/08/15/big/img_871
+2002/08/12/big/img_494
+2002/08/06/big/img_2470
+2002/07/23/big/img_923
+2002/08/26/big/img_155
+2002/08/08/big/img_669
+2002/07/23/big/img_404
+2002/08/28/big/img_19421
+2002/08/29/big/img_18993
+2002/08/25/big/img_416
+2003/01/17/big/img_434
+2002/07/29/big/img_1370
+2002/07/28/big/img_483
+2002/08/11/big/img_50
+2002/08/10/big/img_404
+2002/09/02/big/img_15057
+2003/01/14/big/img_911
+2002/09/01/big/img_16697
+2003/01/16/big/img_665
+2002/09/01/big/img_16708
+2002/08/22/big/img_612
+2002/08/28/big/img_19471
+2002/08/02/big/img_198
+2003/01/16/big/img_527
+2002/08/22/big/img_209
+2002/08/30/big/img_18205
+2003/01/14/big/img_114
+2003/01/14/big/img_1028
+2003/01/16/big/img_894
+2003/01/14/big/img_837
+2002/07/30/big/img_9
+2002/08/06/big/img_2821
+2002/08/04/big/img_85
+2003/01/13/big/img_884
+2002/07/22/big/img_570
+2002/08/07/big/img_1773
+2002/07/26/big/img_208
+2003/01/17/big/img_946
+2002/07/19/big/img_930
+2003/01/01/big/img_698
+2003/01/17/big/img_612
+2002/07/19/big/img_372
+2002/07/30/big/img_721
+2003/01/14/big/img_649
+2002/08/19/big/img_4
+2002/07/25/big/img_1024
+2003/01/15/big/img_601
+2002/08/30/big/img_18470
+2002/07/22/big/img_29
+2002/08/07/big/img_1686
+2002/07/20/big/img_294
+2002/08/14/big/img_800
+2002/08/19/big/img_353
+2002/08/19/big/img_350
+2002/08/05/big/img_3392
+2002/08/09/big/img_622
+2003/01/15/big/img_236
+2002/08/11/big/img_643
+2002/08/05/big/img_3458
+2002/08/12/big/img_413
+2002/08/22/big/img_415
+2002/08/13/big/img_635
+2002/08/07/big/img_1198
+2002/08/04/big/img_873
+2002/08/12/big/img_407
+2003/01/15/big/img_346
+2002/08/02/big/img_275
+2002/08/17/big/img_997
+2002/08/21/big/img_958
+2002/08/20/big/img_579
+2002/07/29/big/img_142
+2003/01/14/big/img_1115
+2002/08/16/big/img_365
+2002/07/29/big/img_1414
+2002/08/17/big/img_489
+2002/08/13/big/img_1010
+2002/07/31/big/img_276
+2002/07/25/big/img_1000
+2002/08/23/big/img_524
+2002/08/28/big/img_19147
+2003/01/13/big/img_433
+2002/08/20/big/img_205
+2003/01/01/big/img_458
+2002/07/29/big/img_1449
+2003/01/16/big/img_696
+2002/08/28/big/img_19296
+2002/08/29/big/img_18688
+2002/08/21/big/img_767
+2002/08/20/big/img_532
+2002/08/26/big/img_187
+2002/07/26/big/img_183
+2002/07/27/big/img_890
+2003/01/13/big/img_576
+2002/07/30/big/img_15
+2002/07/31/big/img_889
+2002/08/31/big/img_17759
+2003/01/14/big/img_1114
+2002/07/19/big/img_445
+2002/08/03/big/img_593
+2002/07/24/big/img_750
+2002/07/30/big/img_133
+2002/08/25/big/img_671
+2002/07/20/big/img_351
+2002/08/31/big/img_17276
+2002/08/05/big/img_3231
+2002/09/02/big/img_15882
+2002/08/14/big/img_115
+2002/08/02/big/img_1148
+2002/07/25/big/img_936
+2002/07/31/big/img_639
+2002/08/04/big/img_427
+2002/08/22/big/img_843
+2003/01/17/big/img_17
+2003/01/13/big/img_690
+2002/08/13/big/img_472
+2002/08/09/big/img_425
+2002/08/05/big/img_3450
+2003/01/17/big/img_439
+2002/08/13/big/img_539
+2002/07/28/big/img_35
+2002/08/16/big/img_241
+2002/08/06/big/img_2898
+2003/01/16/big/img_429
+2002/08/05/big/img_3817
+2002/08/27/big/img_19919
+2002/07/19/big/img_422
+2002/08/15/big/img_560
+2002/07/23/big/img_750
+2002/07/30/big/img_353
+2002/08/05/big/img_43
+2002/08/23/big/img_305
+2002/08/01/big/img_2137
+2002/08/30/big/img_18097
+2002/08/01/big/img_1389
+2002/08/02/big/img_308
+2003/01/14/big/img_652
+2002/08/01/big/img_1798
+2003/01/14/big/img_732
+2003/01/16/big/img_294
+2002/08/26/big/img_213
+2002/07/24/big/img_842
+2003/01/13/big/img_630
+2003/01/13/big/img_634
+2002/08/06/big/img_2285
+2002/08/01/big/img_2162
+2002/08/30/big/img_18134
+2002/08/02/big/img_1045
+2002/08/01/big/img_2143
+2002/07/25/big/img_135
+2002/07/20/big/img_645
+2002/08/05/big/img_3666
+2002/08/14/big/img_523
+2002/08/04/big/img_425
+2003/01/14/big/img_137
+2003/01/01/big/img_176
+2002/08/15/big/img_505
+2002/08/24/big/img_386
+2002/08/05/big/img_3187
+2002/08/15/big/img_419
+2003/01/13/big/img_520
+2002/08/04/big/img_444
+2002/08/26/big/img_483
+2002/08/05/big/img_3449
+2002/08/30/big/img_18409
+2002/08/28/big/img_19455
+2002/08/27/big/img_20090
+2002/07/23/big/img_625
+2002/08/24/big/img_205
+2002/08/08/big/img_938
+2003/01/13/big/img_527
+2002/08/07/big/img_1712
+2002/07/24/big/img_801
+2002/08/09/big/img_579
+2003/01/14/big/img_41
+2003/01/15/big/img_1130
+2002/07/21/big/img_672
+2002/08/07/big/img_1590
+2003/01/01/big/img_532
+2002/08/02/big/img_529
+2002/08/05/big/img_3591
+2002/08/23/big/img_5
+2003/01/14/big/img_882
+2002/08/28/big/img_19234
+2002/07/24/big/img_398
+2003/01/14/big/img_592
+2002/08/22/big/img_548
+2002/08/12/big/img_761
+2003/01/16/big/img_497
+2002/08/18/big/img_133
+2002/08/08/big/img_874
+2002/07/19/big/img_247
+2002/08/15/big/img_170
+2002/08/27/big/img_19679
+2002/08/20/big/img_246
+2002/08/24/big/img_358
+2002/07/29/big/img_599
+2002/08/01/big/img_1555
+2002/07/30/big/img_491
+2002/07/30/big/img_371
+2003/01/16/big/img_682
+2002/07/25/big/img_619
+2003/01/15/big/img_587
+2002/08/02/big/img_1212
+2002/08/01/big/img_2152
+2002/07/25/big/img_668
+2003/01/16/big/img_574
+2002/08/28/big/img_19464
+2002/08/11/big/img_536
+2002/07/24/big/img_201
+2002/08/05/big/img_3488
+2002/07/25/big/img_887
+2002/07/22/big/img_789
+2002/07/30/big/img_432
+2002/08/16/big/img_166
+2002/09/01/big/img_16333
+2002/07/26/big/img_1010
+2002/07/21/big/img_793
+2002/07/22/big/img_720
+2002/07/31/big/img_337
+2002/07/27/big/img_185
+2002/08/23/big/img_440
+2002/07/31/big/img_801
+2002/07/25/big/img_478
+2003/01/14/big/img_171
+2002/08/07/big/img_1054
+2002/09/02/big/img_15659
+2002/07/29/big/img_1348
+2002/08/09/big/img_337
+2002/08/26/big/img_684
+2002/07/31/big/img_537
+2002/08/15/big/img_808
+2003/01/13/big/img_740
+2002/08/07/big/img_1667
+2002/08/03/big/img_404
+2002/08/06/big/img_2520
+2002/07/19/big/img_230
+2002/07/19/big/img_356
+2003/01/16/big/img_627
+2002/08/04/big/img_474
+2002/07/29/big/img_833
+2002/07/25/big/img_176
+2002/08/01/big/img_1684
+2002/08/21/big/img_643
+2002/08/27/big/img_19673
+2002/08/02/big/img_838
+2002/08/06/big/img_2378
+2003/01/15/big/img_48
+2002/07/30/big/img_470
+2002/08/15/big/img_963
+2002/08/24/big/img_444
+2002/08/16/big/img_662
+2002/08/15/big/img_1209
+2002/07/24/big/img_25
+2002/08/06/big/img_2740
+2002/07/29/big/img_996
+2002/08/31/big/img_18074
+2002/08/04/big/img_343
+2003/01/17/big/img_509
+2003/01/13/big/img_726
+2002/08/07/big/img_1466
+2002/07/26/big/img_307
+2002/08/10/big/img_598
+2002/08/13/big/img_890
+2002/08/14/big/img_997
+2002/07/19/big/img_392
+2002/08/02/big/img_475
+2002/08/29/big/img_19038
+2002/07/29/big/img_538
+2002/07/29/big/img_502
+2002/08/02/big/img_364
+2002/08/31/big/img_17353
+2002/08/08/big/img_539
+2002/08/01/big/img_1449
+2002/07/22/big/img_363
+2002/08/02/big/img_90
+2002/09/01/big/img_16867
+2002/08/05/big/img_3371
+2002/07/30/big/img_342
+2002/08/07/big/img_1363
+2002/08/22/big/img_790
+2003/01/15/big/img_404
+2002/08/05/big/img_3447
+2002/09/01/big/img_16167
+2003/01/13/big/img_840
+2002/08/22/big/img_1001
+2002/08/09/big/img_431
+2002/07/27/big/img_618
+2002/07/31/big/img_741
+2002/07/30/big/img_964
+2002/07/25/big/img_86
+2002/07/29/big/img_275
+2002/08/21/big/img_921
+2002/07/26/big/img_892
+2002/08/21/big/img_663
+2003/01/13/big/img_567
+2003/01/14/big/img_719
+2002/07/28/big/img_251
+2003/01/15/big/img_1123
+2002/07/29/big/img_260
+2002/08/24/big/img_337
+2002/08/01/big/img_1914
+2002/08/13/big/img_373
+2003/01/15/big/img_589
+2002/08/13/big/img_906
+2002/07/26/big/img_270
+2002/08/26/big/img_313
+2002/08/25/big/img_694
+2003/01/01/big/img_327
+2002/07/23/big/img_261
+2002/08/26/big/img_642
+2002/07/29/big/img_918
+2002/07/23/big/img_455
+2002/07/24/big/img_612
+2002/07/23/big/img_534
+2002/07/19/big/img_534
+2002/07/19/big/img_726
+2002/08/01/big/img_2146
+2002/08/02/big/img_543
+2003/01/16/big/img_777
+2002/07/30/big/img_484
+2002/08/13/big/img_1161
+2002/07/21/big/img_390
+2002/08/06/big/img_2288
+2002/08/21/big/img_677
+2002/08/13/big/img_747
+2002/08/15/big/img_1248
+2002/07/31/big/img_416
+2002/09/02/big/img_15259
+2002/08/16/big/img_781
+2002/08/24/big/img_754
+2002/07/24/big/img_803
+2002/08/20/big/img_609
+2002/08/28/big/img_19571
+2002/09/01/big/img_16140
+2002/08/26/big/img_769
+2002/07/20/big/img_588
+2002/08/02/big/img_898
+2002/07/21/big/img_466
+2002/08/14/big/img_1046
+2002/07/25/big/img_212
+2002/08/26/big/img_353
+2002/08/19/big/img_810
+2002/08/31/big/img_17824
+2002/08/12/big/img_631
+2002/07/19/big/img_828
+2002/07/24/big/img_130
+2002/08/25/big/img_580
+2002/07/31/big/img_699
+2002/07/23/big/img_808
+2002/07/31/big/img_377
+2003/01/16/big/img_570
+2002/09/01/big/img_16254
+2002/07/21/big/img_471
+2002/08/01/big/img_1548
+2002/08/18/big/img_252
+2002/08/19/big/img_576
+2002/08/20/big/img_464
+2002/07/27/big/img_735
+2002/08/21/big/img_589
+2003/01/15/big/img_1192
+2002/08/09/big/img_302
+2002/07/31/big/img_594
+2002/08/23/big/img_19
+2002/08/29/big/img_18819
+2002/08/19/big/img_293
+2002/07/30/big/img_331
+2002/08/23/big/img_607
+2002/07/30/big/img_363
+2002/08/16/big/img_766
+2003/01/13/big/img_481
+2002/08/06/big/img_2515
+2002/09/02/big/img_15913
+2002/09/02/big/img_15827
+2002/09/02/big/img_15053
+2002/08/07/big/img_1576
+2002/07/23/big/img_268
+2002/08/21/big/img_152
+2003/01/15/big/img_578
+2002/07/21/big/img_589
+2002/07/20/big/img_548
+2002/08/27/big/img_19693
+2002/08/31/big/img_17252
+2002/07/31/big/img_138
+2002/07/23/big/img_372
+2002/08/16/big/img_695
+2002/07/27/big/img_287
+2002/08/15/big/img_315
+2002/08/10/big/img_361
+2002/07/29/big/img_899
+2002/08/13/big/img_771
+2002/08/21/big/img_92
+2003/01/15/big/img_425
+2003/01/16/big/img_450
+2002/09/01/big/img_16942
+2002/08/02/big/img_51
+2002/09/02/big/img_15379
+2002/08/24/big/img_147
+2002/08/30/big/img_18122
+2002/07/26/big/img_950
+2002/08/07/big/img_1400
+2002/08/17/big/img_468
+2002/08/15/big/img_470
+2002/07/30/big/img_318
+2002/07/22/big/img_644
+2002/08/27/big/img_19732
+2002/07/23/big/img_601
+2002/08/26/big/img_398
+2002/08/21/big/img_428
+2002/08/06/big/img_2119
+2002/08/29/big/img_19103
+2003/01/14/big/img_933
+2002/08/11/big/img_674
+2002/08/28/big/img_19420
+2002/08/03/big/img_418
+2002/08/17/big/img_312
+2002/07/25/big/img_1044
+2003/01/17/big/img_671
+2002/08/30/big/img_18297
+2002/07/25/big/img_755
+2002/07/23/big/img_471
+2002/08/21/big/img_39
+2002/07/26/big/img_699
+2003/01/14/big/img_33
+2002/07/31/big/img_411
+2002/08/16/big/img_645
+2003/01/17/big/img_116
+2002/09/02/big/img_15903
+2002/08/20/big/img_120
+2002/08/22/big/img_176
+2002/07/29/big/img_1316
+2002/08/27/big/img_19914
+2002/07/22/big/img_719
+2002/08/28/big/img_19239
+2003/01/13/big/img_385
+2002/08/08/big/img_525
+2002/07/19/big/img_782
+2002/08/13/big/img_843
+2002/07/30/big/img_107
+2002/08/11/big/img_752
+2002/07/29/big/img_383
+2002/08/26/big/img_249
+2002/08/29/big/img_18860
+2002/07/30/big/img_70
+2002/07/26/big/img_194
+2002/08/15/big/img_530
+2002/08/08/big/img_816
+2002/07/31/big/img_286
+2003/01/13/big/img_294
+2002/07/31/big/img_251
+2002/07/24/big/img_13
+2002/08/31/big/img_17938
+2002/07/22/big/img_642
+2003/01/14/big/img_728
+2002/08/18/big/img_47
+2002/08/22/big/img_306
+2002/08/20/big/img_348
+2002/08/15/big/img_764
+2002/08/08/big/img_163
+2002/07/23/big/img_531
+2002/07/23/big/img_467
+2003/01/16/big/img_743
+2003/01/13/big/img_535
+2002/08/02/big/img_523
+2002/08/22/big/img_120
+2002/08/11/big/img_496
+2002/08/29/big/img_19075
+2002/08/08/big/img_465
+2002/08/09/big/img_790
+2002/08/19/big/img_588
+2002/08/23/big/img_407
+2003/01/17/big/img_435
+2002/08/24/big/img_398
+2002/08/27/big/img_19899
+2003/01/15/big/img_335
+2002/08/13/big/img_493
+2002/09/02/big/img_15460
+2002/07/31/big/img_470
+2002/08/05/big/img_3550
+2002/07/28/big/img_123
+2002/08/01/big/img_1498
+2002/08/04/big/img_504
+2003/01/17/big/img_427
+2002/08/27/big/img_19708
+2002/07/27/big/img_861
+2002/07/25/big/img_685
+2002/07/31/big/img_207
+2003/01/14/big/img_745
+2002/08/31/big/img_17756
+2002/08/24/big/img_288
+2002/08/18/big/img_181
+2002/08/10/big/img_520
+2002/08/25/big/img_705
+2002/08/23/big/img_226
+2002/08/04/big/img_727
+2002/07/24/big/img_625
+2002/08/28/big/img_19157
+2002/08/23/big/img_586
+2002/07/31/big/img_232
+2003/01/13/big/img_240
+2003/01/14/big/img_321
+2003/01/15/big/img_533
+2002/07/23/big/img_480
+2002/07/24/big/img_371
+2002/08/21/big/img_702
+2002/08/31/big/img_17075
+2002/09/02/big/img_15278
+2002/07/29/big/img_246
+2003/01/15/big/img_829
+2003/01/15/big/img_1213
+2003/01/16/big/img_441
+2002/08/14/big/img_921
+2002/07/23/big/img_425
+2002/08/15/big/img_296
+2002/07/19/big/img_135
+2002/07/26/big/img_402
+2003/01/17/big/img_88
+2002/08/20/big/img_872
+2002/08/13/big/img_1110
+2003/01/16/big/img_1040
+2002/07/23/big/img_9
+2002/08/13/big/img_700
+2002/08/16/big/img_371
+2002/08/27/big/img_19966
+2003/01/17/big/img_391
+2002/08/18/big/img_426
+2002/08/01/big/img_1618
+2002/07/21/big/img_754
+2003/01/14/big/img_1101
+2003/01/16/big/img_1022
+2002/07/22/big/img_275
+2002/08/24/big/img_86
+2002/08/17/big/img_582
+2003/01/15/big/img_765
+2003/01/17/big/img_449
+2002/07/28/big/img_265
+2003/01/13/big/img_552
+2002/07/28/big/img_115
+2003/01/16/big/img_56
+2002/08/02/big/img_1232
+2003/01/17/big/img_925
+2002/07/22/big/img_445
+2002/07/25/big/img_957
+2002/07/20/big/img_589
+2002/08/31/big/img_17107
+2002/07/29/big/img_483
+2002/08/14/big/img_1063
+2002/08/07/big/img_1545
+2002/08/14/big/img_680
+2002/09/01/big/img_16694
+2002/08/14/big/img_257
+2002/08/11/big/img_726
+2002/07/26/big/img_681
+2002/07/25/big/img_481
+2003/01/14/big/img_737
+2002/08/28/big/img_19480
+2003/01/16/big/img_362
+2002/08/27/big/img_19865
+2003/01/01/big/img_547
+2002/09/02/big/img_15074
+2002/08/01/big/img_1453
+2002/08/22/big/img_594
+2002/08/28/big/img_19263
+2002/08/13/big/img_478
+2002/07/29/big/img_1358
+2003/01/14/big/img_1022
+2002/08/16/big/img_450
+2002/08/02/big/img_159
+2002/07/26/big/img_781
+2003/01/13/big/img_601
+2002/08/20/big/img_407
+2002/08/15/big/img_468
+2002/08/31/big/img_17902
+2002/08/16/big/img_81
+2002/07/25/big/img_987
+2002/07/25/big/img_500
+2002/08/02/big/img_31
+2002/08/18/big/img_538
+2002/08/08/big/img_54
+2002/07/23/big/img_686
+2002/07/24/big/img_836
+2003/01/17/big/img_734
+2002/08/16/big/img_1055
+2003/01/16/big/img_521
+2002/07/25/big/img_612
+2002/08/22/big/img_778
+2002/08/03/big/img_251
+2002/08/12/big/img_436
+2002/08/23/big/img_705
+2002/07/28/big/img_243
+2002/07/25/big/img_1029
+2002/08/20/big/img_287
+2002/08/29/big/img_18739
+2002/08/05/big/img_3272
+2002/07/27/big/img_214
+2003/01/14/big/img_5
+2002/08/01/big/img_1380
+2002/08/29/big/img_19097
+2002/07/30/big/img_486
+2002/08/29/big/img_18707
+2002/08/10/big/img_559
+2002/08/15/big/img_365
+2002/08/09/big/img_525
+2002/08/10/big/img_689
+2002/07/25/big/img_502
+2002/08/03/big/img_667
+2002/08/10/big/img_855
+2002/08/10/big/img_706
+2002/08/18/big/img_603
+2003/01/16/big/img_1055
+2002/08/31/big/img_17890
+2002/08/15/big/img_761
+2003/01/15/big/img_489
+2002/08/26/big/img_351
+2002/08/01/big/img_1772
+2002/08/31/big/img_17729
+2002/07/25/big/img_609
+2003/01/13/big/img_539
+2002/07/27/big/img_686
+2002/07/31/big/img_311
+2002/08/22/big/img_799
+2003/01/16/big/img_936
+2002/08/31/big/img_17813
+2002/08/04/big/img_862
+2002/08/09/big/img_332
+2002/07/20/big/img_148
+2002/08/12/big/img_426
+2002/07/24/big/img_69
+2002/07/27/big/img_685
+2002/08/02/big/img_480
+2002/08/26/big/img_154
+2002/07/24/big/img_598
+2002/08/01/big/img_1881
+2002/08/20/big/img_667
+2003/01/14/big/img_495
+2002/07/21/big/img_744
+2002/07/30/big/img_150
+2002/07/23/big/img_924
+2002/08/08/big/img_272
+2002/07/23/big/img_310
+2002/07/25/big/img_1011
+2002/09/02/big/img_15725
+2002/07/19/big/img_814
+2002/08/20/big/img_936
+2002/07/25/big/img_85
+2002/08/24/big/img_662
+2002/08/09/big/img_495
+2003/01/15/big/img_196
+2002/08/16/big/img_707
+2002/08/28/big/img_19370
+2002/08/06/big/img_2366
+2002/08/06/big/img_3012
+2002/08/01/big/img_1452
+2002/07/31/big/img_742
+2002/07/27/big/img_914
+2003/01/13/big/img_290
+2002/07/31/big/img_288
+2002/08/02/big/img_171
+2002/08/22/big/img_191
+2002/07/27/big/img_1066
+2002/08/12/big/img_383
+2003/01/17/big/img_1018
+2002/08/01/big/img_1785
+2002/08/11/big/img_390
+2002/08/27/big/img_20037
+2002/08/12/big/img_38
+2003/01/15/big/img_103
+2002/08/26/big/img_31
+2002/08/18/big/img_660
+2002/07/22/big/img_694
+2002/08/15/big/img_24
+2002/07/27/big/img_1077
+2002/08/01/big/img_1943
+2002/07/22/big/img_292
+2002/09/01/big/img_16857
+2002/07/22/big/img_892
+2003/01/14/big/img_46
+2002/08/09/big/img_469
+2002/08/09/big/img_414
+2003/01/16/big/img_40
+2002/08/28/big/img_19231
+2002/07/27/big/img_978
+2002/07/23/big/img_475
+2002/07/25/big/img_92
+2002/08/09/big/img_799
+2002/07/25/big/img_491
+2002/08/03/big/img_654
+2003/01/15/big/img_687
+2002/08/11/big/img_478
+2002/08/07/big/img_1664
+2002/08/20/big/img_362
+2002/08/01/big/img_1298
+2003/01/13/big/img_500
+2002/08/06/big/img_2896
+2002/08/30/big/img_18529
+2002/08/16/big/img_1020
+2002/07/29/big/img_892
+2002/08/29/big/img_18726
+2002/07/21/big/img_453
+2002/08/17/big/img_437
+2002/07/19/big/img_665
+2002/07/22/big/img_440
+2002/07/19/big/img_582
+2002/07/21/big/img_233
+2003/01/01/big/img_82
+2002/07/25/big/img_341
+2002/07/29/big/img_864
+2002/08/02/big/img_276
+2002/08/29/big/img_18654
+2002/07/27/big/img_1024
+2002/08/19/big/img_373
+2003/01/15/big/img_241
+2002/07/25/big/img_84
+2002/08/13/big/img_834
+2002/08/10/big/img_511
+2002/08/01/big/img_1627
+2002/08/08/big/img_607
+2002/08/06/big/img_2083
+2002/08/01/big/img_1486
+2002/08/08/big/img_700
+2002/08/01/big/img_1954
+2002/08/21/big/img_54
+2002/07/30/big/img_847
+2002/08/28/big/img_19169
+2002/07/21/big/img_549
+2002/08/03/big/img_693
+2002/07/31/big/img_1002
+2003/01/14/big/img_1035
+2003/01/16/big/img_622
+2002/07/30/big/img_1201
+2002/08/10/big/img_444
+2002/07/31/big/img_374
+2002/08/21/big/img_301
+2002/08/13/big/img_1095
+2003/01/13/big/img_288
+2002/07/25/big/img_232
+2003/01/13/big/img_967
+2002/08/26/big/img_360
+2002/08/05/big/img_67
+2002/08/29/big/img_18969
+2002/07/28/big/img_16
+2002/08/16/big/img_515
+2002/07/20/big/img_708
+2002/08/18/big/img_178
+2003/01/15/big/img_509
+2002/07/25/big/img_430
+2002/08/21/big/img_738
+2002/08/16/big/img_886
+2002/09/02/big/img_15605
+2002/09/01/big/img_16242
+2002/08/24/big/img_711
+2002/07/25/big/img_90
+2002/08/09/big/img_491
+2002/07/30/big/img_534
+2003/01/13/big/img_474
+2002/08/25/big/img_510
+2002/08/15/big/img_555
+2002/08/02/big/img_775
+2002/07/23/big/img_975
+2002/08/19/big/img_229
+2003/01/17/big/img_860
+2003/01/02/big/img_10
+2002/07/23/big/img_542
+2002/08/06/big/img_2535
+2002/07/22/big/img_37
+2002/08/06/big/img_2342
+2002/08/25/big/img_515
+2002/08/25/big/img_336
+2002/08/18/big/img_837
+2002/08/21/big/img_616
+2003/01/17/big/img_24
+2002/07/26/big/img_936
+2002/08/14/big/img_896
+2002/07/29/big/img_465
+2002/07/31/big/img_543
+2002/08/01/big/img_1411
+2002/08/02/big/img_423
+2002/08/21/big/img_44
+2002/07/31/big/img_11
+2003/01/15/big/img_628
+2003/01/15/big/img_605
+2002/07/30/big/img_571
+2002/07/23/big/img_428
+2002/08/15/big/img_942
+2002/07/26/big/img_531
+2003/01/16/big/img_59
+2002/08/02/big/img_410
+2002/07/31/big/img_230
+2002/08/19/big/img_806
+2003/01/14/big/img_462
+2002/08/16/big/img_370
+2002/08/13/big/img_380
+2002/08/16/big/img_932
+2002/07/19/big/img_393
+2002/08/20/big/img_764
+2002/08/15/big/img_616
+2002/07/26/big/img_267
+2002/07/27/big/img_1069
+2002/08/14/big/img_1041
+2003/01/13/big/img_594
+2002/09/01/big/img_16845
+2002/08/09/big/img_229
+2003/01/16/big/img_639
+2002/08/19/big/img_398
+2002/08/18/big/img_978
+2002/08/24/big/img_296
+2002/07/29/big/img_415
+2002/07/30/big/img_923
+2002/08/18/big/img_575
+2002/08/22/big/img_182
+2002/07/25/big/img_806
+2002/07/22/big/img_49
+2002/07/29/big/img_989
+2003/01/17/big/img_789
+2003/01/15/big/img_503
+2002/09/01/big/img_16062
+2003/01/17/big/img_794
+2002/08/15/big/img_564
+2003/01/15/big/img_222
+2002/08/01/big/img_1656
+2003/01/13/big/img_432
+2002/07/19/big/img_426
+2002/08/17/big/img_244
+2002/08/13/big/img_805
+2002/09/02/big/img_15067
+2002/08/11/big/img_58
+2002/08/22/big/img_636
+2002/07/22/big/img_416
+2002/08/13/big/img_836
+2002/08/26/big/img_363
+2002/07/30/big/img_917
+2003/01/14/big/img_206
+2002/08/12/big/img_311
+2002/08/31/big/img_17623
+2002/07/29/big/img_661
+2003/01/13/big/img_417
+2002/08/02/big/img_463
+2002/08/02/big/img_669
+2002/08/26/big/img_670
+2002/08/02/big/img_375
+2002/07/19/big/img_209
+2002/08/08/big/img_115
+2002/08/21/big/img_399
+2002/08/20/big/img_911
+2002/08/07/big/img_1212
+2002/08/20/big/img_578
+2002/08/22/big/img_554
+2002/08/21/big/img_484
+2002/07/25/big/img_450
+2002/08/03/big/img_542
+2002/08/15/big/img_561
+2002/07/23/big/img_360
+2002/08/30/big/img_18137
+2002/07/25/big/img_250
+2002/08/03/big/img_647
+2002/08/20/big/img_375
+2002/08/14/big/img_387
+2002/09/01/big/img_16990
+2002/08/28/big/img_19341
+2003/01/15/big/img_239
+2002/08/20/big/img_528
+2002/08/12/big/img_130
+2002/09/02/big/img_15108
+2003/01/15/big/img_372
+2002/08/16/big/img_678
+2002/08/04/big/img_623
+2002/07/23/big/img_477
+2002/08/28/big/img_19590
+2003/01/17/big/img_978
+2002/09/01/big/img_16692
+2002/07/20/big/img_109
+2002/08/06/big/img_2660
+2003/01/14/big/img_464
+2002/08/09/big/img_618
+2002/07/22/big/img_722
+2002/08/25/big/img_419
+2002/08/03/big/img_314
+2002/08/25/big/img_40
+2002/07/27/big/img_430
+2002/08/10/big/img_569
+2002/08/23/big/img_398
+2002/07/23/big/img_893
+2002/08/16/big/img_261
+2002/08/06/big/img_2668
+2002/07/22/big/img_835
+2002/09/02/big/img_15093
+2003/01/16/big/img_65
+2002/08/21/big/img_448
+2003/01/14/big/img_351
+2003/01/17/big/img_133
+2002/07/28/big/img_493
+2003/01/15/big/img_640
+2002/09/01/big/img_16880
+2002/08/15/big/img_350
+2002/08/20/big/img_624
+2002/08/25/big/img_604
+2002/08/06/big/img_2200
+2002/08/23/big/img_290
+2002/08/13/big/img_1152
+2003/01/14/big/img_251
+2002/08/02/big/img_538
+2002/08/22/big/img_613
+2003/01/13/big/img_351
+2002/08/18/big/img_368
+2002/07/23/big/img_392
+2002/07/25/big/img_198
+2002/07/25/big/img_418
+2002/08/26/big/img_614
+2002/07/23/big/img_405
+2003/01/14/big/img_445
+2002/07/25/big/img_326
+2002/08/10/big/img_734
+2003/01/14/big/img_530
+2002/08/08/big/img_561
+2002/08/29/big/img_18990
+2002/08/10/big/img_576
+2002/07/29/big/img_1494
+2002/07/19/big/img_198
+2002/08/10/big/img_562
+2002/07/22/big/img_901
+2003/01/14/big/img_37
+2002/09/02/big/img_15629
+2003/01/14/big/img_58
+2002/08/01/big/img_1364
+2002/07/27/big/img_636
+2003/01/13/big/img_241
+2002/09/01/big/img_16988
+2003/01/13/big/img_560
+2002/08/09/big/img_533
+2002/07/31/big/img_249
+2003/01/17/big/img_1007
+2002/07/21/big/img_64
+2003/01/13/big/img_537
+2003/01/15/big/img_606
+2002/08/18/big/img_651
+2002/08/24/big/img_405
+2002/07/26/big/img_837
+2002/08/09/big/img_562
+2002/08/01/big/img_1983
+2002/08/03/big/img_514
+2002/07/29/big/img_314
+2002/08/12/big/img_493
+2003/01/14/big/img_121
+2003/01/14/big/img_479
+2002/08/04/big/img_410
+2002/07/22/big/img_607
+2003/01/17/big/img_417
+2002/07/20/big/img_547
+2002/08/13/big/img_396
+2002/08/31/big/img_17538
+2002/08/13/big/img_187
+2002/08/12/big/img_328
+2003/01/14/big/img_569
+2002/07/27/big/img_1081
+2002/08/14/big/img_504
+2002/08/23/big/img_785
+2002/07/26/big/img_339
+2002/08/07/big/img_1156
+2002/08/07/big/img_1456
+2002/08/23/big/img_378
+2002/08/27/big/img_19719
+2002/07/31/big/img_39
+2002/07/31/big/img_883
+2003/01/14/big/img_676
+2002/07/29/big/img_214
+2002/07/26/big/img_669
+2002/07/25/big/img_202
+2002/08/08/big/img_259
+2003/01/17/big/img_943
+2003/01/15/big/img_512
+2002/08/05/big/img_3295
+2002/08/27/big/img_19685
+2002/08/08/big/img_277
+2002/08/30/big/img_18154
+2002/07/22/big/img_663
+2002/08/29/big/img_18914
+2002/07/31/big/img_908
+2002/08/27/big/img_19926
+2003/01/13/big/img_791
+2003/01/15/big/img_827
+2002/08/18/big/img_878
+2002/08/14/big/img_670
+2002/07/20/big/img_182
+2002/08/15/big/img_291
+2002/08/06/big/img_2600
+2002/07/23/big/img_587
+2002/08/14/big/img_577
+2003/01/15/big/img_585
+2002/07/30/big/img_310
+2002/08/03/big/img_658
+2002/08/10/big/img_157
+2002/08/19/big/img_811
+2002/07/29/big/img_1318
+2002/08/04/big/img_104
+2002/07/30/big/img_332
+2002/07/24/big/img_789
+2002/07/29/big/img_516
+2002/07/23/big/img_843
+2002/08/01/big/img_1528
+2002/08/13/big/img_798
+2002/08/07/big/img_1729
+2002/08/28/big/img_19448
+2003/01/16/big/img_95
+2002/08/12/big/img_473
+2002/07/27/big/img_269
+2003/01/16/big/img_621
+2002/07/29/big/img_772
+2002/07/24/big/img_171
+2002/07/19/big/img_429
+2002/08/07/big/img_1933
+2002/08/27/big/img_19629
+2002/08/05/big/img_3688
+2002/08/07/big/img_1691
+2002/07/23/big/img_600
+2002/07/29/big/img_666
+2002/08/25/big/img_566
+2002/08/06/big/img_2659
+2002/08/29/big/img_18929
+2002/08/16/big/img_407
+2002/08/18/big/img_774
+2002/08/19/big/img_249
+2002/08/06/big/img_2427
+2002/08/29/big/img_18899
+2002/08/01/big/img_1818
+2002/07/31/big/img_108
+2002/07/29/big/img_500
+2002/08/11/big/img_115
+2002/07/19/big/img_521
+2002/08/02/big/img_1163
+2002/07/22/big/img_62
+2002/08/13/big/img_466
+2002/08/21/big/img_956
+2002/08/23/big/img_602
+2002/08/20/big/img_858
+2002/07/25/big/img_690
+2002/07/19/big/img_130
+2002/08/04/big/img_874
+2002/07/26/big/img_489
+2002/07/22/big/img_548
+2002/08/10/big/img_191
+2002/07/25/big/img_1051
+2002/08/18/big/img_473
+2002/08/12/big/img_755
+2002/08/18/big/img_413
+2002/08/08/big/img_1044
+2002/08/17/big/img_680
+2002/08/26/big/img_235
+2002/08/20/big/img_330
+2002/08/22/big/img_344
+2002/08/09/big/img_593
+2002/07/31/big/img_1006
+2002/08/14/big/img_337
+2002/08/16/big/img_728
+2002/07/24/big/img_834
+2002/08/04/big/img_552
+2002/09/02/big/img_15213
+2002/07/25/big/img_725
+2002/08/30/big/img_18290
+2003/01/01/big/img_475
+2002/07/27/big/img_1083
+2002/08/29/big/img_18955
+2002/08/31/big/img_17232
+2002/08/08/big/img_480
+2002/08/01/big/img_1311
+2002/07/30/big/img_745
+2002/08/03/big/img_649
+2002/08/12/big/img_193
+2002/07/29/big/img_228
+2002/07/25/big/img_836
+2002/08/20/big/img_400
+2002/07/30/big/img_507
+2002/09/02/big/img_15072
+2002/07/26/big/img_658
+2002/07/28/big/img_503
+2002/08/05/big/img_3814
+2002/08/24/big/img_745
+2003/01/13/big/img_817
+2002/08/08/big/img_579
+2002/07/22/big/img_251
+2003/01/13/big/img_689
+2002/07/25/big/img_407
+2002/08/13/big/img_1050
+2002/08/14/big/img_733
+2002/07/24/big/img_82
+2003/01/17/big/img_288
+2003/01/15/big/img_475
+2002/08/14/big/img_620
+2002/08/21/big/img_167
+2002/07/19/big/img_300
+2002/07/26/big/img_219
+2002/08/01/big/img_1468
+2002/07/23/big/img_260
+2002/08/09/big/img_555
+2002/07/19/big/img_160
+2002/08/02/big/img_1060
+2003/01/14/big/img_149
+2002/08/15/big/img_346
+2002/08/24/big/img_597
+2002/08/22/big/img_502
+2002/08/30/big/img_18228
+2002/07/21/big/img_766
+2003/01/15/big/img_841
+2002/07/24/big/img_516
+2002/08/02/big/img_265
+2002/08/15/big/img_1243
+2003/01/15/big/img_223
+2002/08/04/big/img_236
+2002/07/22/big/img_309
+2002/07/20/big/img_656
+2002/07/31/big/img_412
+2002/09/01/big/img_16462
+2003/01/16/big/img_431
+2002/07/22/big/img_793
+2002/08/15/big/img_877
+2002/07/26/big/img_282
+2002/07/25/big/img_529
+2002/08/24/big/img_613
+2003/01/17/big/img_700
+2002/08/06/big/img_2526
+2002/08/24/big/img_394
+2002/08/21/big/img_521
+2002/08/25/big/img_560
+2002/07/29/big/img_966
+2002/07/25/big/img_448
+2003/01/13/big/img_782
+2002/08/21/big/img_296
+2002/09/01/big/img_16755
+2002/08/05/big/img_3552
+2002/09/02/big/img_15823
+2003/01/14/big/img_193
+2002/07/21/big/img_159
+2002/08/02/big/img_564
+2002/08/16/big/img_300
+2002/07/19/big/img_269
+2002/08/13/big/img_676
+2002/07/28/big/img_57
+2002/08/05/big/img_3318
+2002/07/31/big/img_218
+2002/08/21/big/img_898
+2002/07/29/big/img_109
+2002/07/19/big/img_854
+2002/08/23/big/img_311
+2002/08/14/big/img_318
+2002/07/25/big/img_523
+2002/07/21/big/img_678
+2003/01/17/big/img_690
+2002/08/28/big/img_19503
+2002/08/18/big/img_251
+2002/08/22/big/img_672
+2002/08/20/big/img_663
+2002/08/02/big/img_148
+2002/09/02/big/img_15580
+2002/07/25/big/img_778
+2002/08/14/big/img_565
+2002/08/12/big/img_374
+2002/08/13/big/img_1018
+2002/08/20/big/img_474
+2002/08/25/big/img_33
+2002/08/02/big/img_1190
+2002/08/08/big/img_864
+2002/08/14/big/img_1071
+2002/08/30/big/img_18103
+2002/08/18/big/img_533
+2003/01/16/big/img_650
+2002/07/25/big/img_108
+2002/07/26/big/img_81
+2002/07/27/big/img_543
+2002/07/29/big/img_521
+2003/01/13/big/img_434
+2002/08/26/big/img_674
+2002/08/06/big/img_2932
+2002/08/07/big/img_1262
+2003/01/15/big/img_201
+2003/01/16/big/img_673
+2002/09/02/big/img_15988
+2002/07/29/big/img_1306
+2003/01/14/big/img_1072
+2002/08/30/big/img_18232
+2002/08/05/big/img_3711
+2002/07/23/big/img_775
+2002/08/01/big/img_16
+2003/01/16/big/img_630
+2002/08/22/big/img_695
+2002/08/14/big/img_51
+2002/08/14/big/img_782
+2002/08/24/big/img_742
+2003/01/14/big/img_512
+2003/01/15/big/img_1183
+2003/01/15/big/img_714
+2002/08/01/big/img_2078
+2002/07/31/big/img_682
+2002/09/02/big/img_15687
+2002/07/26/big/img_518
+2002/08/27/big/img_19676
+2002/09/02/big/img_15969
+2002/08/02/big/img_931
+2002/08/25/big/img_508
+2002/08/29/big/img_18616
+2002/07/22/big/img_839
+2002/07/28/big/img_313
+2003/01/14/big/img_155
+2002/08/02/big/img_1105
+2002/08/09/big/img_53
+2002/08/16/big/img_469
+2002/08/15/big/img_502
+2002/08/20/big/img_575
+2002/07/25/big/img_138
+2003/01/16/big/img_579
+2002/07/19/big/img_352
+2003/01/14/big/img_762
+2003/01/01/big/img_588
+2002/08/02/big/img_981
+2002/08/21/big/img_447
+2002/09/01/big/img_16151
+2003/01/14/big/img_769
+2002/08/23/big/img_461
+2002/08/17/big/img_240
+2002/09/02/big/img_15220
+2002/07/19/big/img_408
+2002/09/02/big/img_15496
+2002/07/29/big/img_758
+2002/08/28/big/img_19392
+2002/08/06/big/img_2723
+2002/08/31/big/img_17752
+2002/08/23/big/img_469
+2002/08/13/big/img_515
+2002/09/02/big/img_15551
+2002/08/03/big/img_462
+2002/07/24/big/img_613
+2002/07/22/big/img_61
+2002/08/08/big/img_171
+2002/08/21/big/img_177
+2003/01/14/big/img_105
+2002/08/02/big/img_1017
+2002/08/22/big/img_106
+2002/07/27/big/img_542
+2002/07/21/big/img_665
+2002/07/23/big/img_595
+2002/08/04/big/img_657
+2002/08/29/big/img_19002
+2003/01/15/big/img_550
+2002/08/14/big/img_662
+2002/07/20/big/img_425
+2002/08/30/big/img_18528
+2002/07/26/big/img_611
+2002/07/22/big/img_849
+2002/08/07/big/img_1655
+2002/08/21/big/img_638
+2003/01/17/big/img_732
+2003/01/01/big/img_496
+2002/08/18/big/img_713
+2002/08/08/big/img_109
+2002/07/27/big/img_1008
+2002/07/20/big/img_559
+2002/08/16/big/img_699
+2002/08/31/big/img_17702
+2002/07/31/big/img_1013
+2002/08/01/big/img_2027
+2002/08/02/big/img_1001
+2002/08/03/big/img_210
+2002/08/01/big/img_2087
+2003/01/14/big/img_199
+2002/07/29/big/img_48
+2002/07/19/big/img_727
+2002/08/09/big/img_249
+2002/08/04/big/img_632
+2002/08/22/big/img_620
+2003/01/01/big/img_457
+2002/08/05/big/img_3223
+2002/07/27/big/img_240
+2002/07/25/big/img_797
+2002/08/13/big/img_430
+2002/07/25/big/img_615
+2002/08/12/big/img_28
+2002/07/30/big/img_220
+2002/07/24/big/img_89
+2002/08/21/big/img_357
+2002/08/09/big/img_590
+2003/01/13/big/img_525
+2002/08/17/big/img_818
+2003/01/02/big/img_7
+2002/07/26/big/img_636
+2003/01/13/big/img_1122
+2002/07/23/big/img_810
+2002/08/20/big/img_888
+2002/07/27/big/img_3
+2002/08/15/big/img_451
+2002/09/02/big/img_15787
+2002/07/31/big/img_281
+2002/08/05/big/img_3274
+2002/08/07/big/img_1254
+2002/07/31/big/img_27
+2002/08/01/big/img_1366
+2002/07/30/big/img_182
+2002/08/27/big/img_19690
+2002/07/29/big/img_68
+2002/08/23/big/img_754
+2002/07/30/big/img_540
+2002/08/27/big/img_20063
+2002/08/14/big/img_471
+2002/08/02/big/img_615
+2002/07/30/big/img_186
+2002/08/25/big/img_150
+2002/07/27/big/img_626
+2002/07/20/big/img_225
+2003/01/15/big/img_1252
+2002/07/19/big/img_367
+2003/01/15/big/img_582
+2002/08/09/big/img_572
+2002/08/08/big/img_428
+2003/01/15/big/img_639
+2002/08/28/big/img_19245
+2002/07/24/big/img_321
+2002/08/02/big/img_662
+2002/08/08/big/img_1033
+2003/01/17/big/img_867
+2002/07/22/big/img_652
+2003/01/14/big/img_224
+2002/08/18/big/img_49
+2002/07/26/big/img_46
+2002/08/31/big/img_18021
+2002/07/25/big/img_151
+2002/08/23/big/img_540
+2002/08/25/big/img_693
+2002/07/23/big/img_340
+2002/07/28/big/img_117
+2002/09/02/big/img_15768
+2002/08/26/big/img_562
+2002/07/24/big/img_480
+2003/01/15/big/img_341
+2002/08/10/big/img_783
+2002/08/20/big/img_132
+2003/01/14/big/img_370
+2002/07/20/big/img_720
+2002/08/03/big/img_144
+2002/08/20/big/img_538
+2002/08/01/big/img_1745
+2002/08/11/big/img_683
+2002/08/03/big/img_328
+2002/08/10/big/img_793
+2002/08/14/big/img_689
+2002/08/02/big/img_162
+2003/01/17/big/img_411
+2002/07/31/big/img_361
+2002/08/15/big/img_289
+2002/08/08/big/img_254
+2002/08/15/big/img_996
+2002/08/20/big/img_785
+2002/07/24/big/img_511
+2002/08/06/big/img_2614
+2002/08/29/big/img_18733
+2002/08/17/big/img_78
+2002/07/30/big/img_378
+2002/08/31/big/img_17947
+2002/08/26/big/img_88
+2002/07/30/big/img_558
+2002/08/02/big/img_67
+2003/01/14/big/img_325
+2002/07/29/big/img_1357
+2002/07/19/big/img_391
+2002/07/30/big/img_307
+2003/01/13/big/img_219
+2002/07/24/big/img_807
+2002/08/23/big/img_543
+2002/08/29/big/img_18620
+2002/07/22/big/img_769
+2002/08/26/big/img_503
+2002/07/30/big/img_78
+2002/08/14/big/img_1036
+2002/08/09/big/img_58
+2002/07/24/big/img_616
+2002/08/02/big/img_464
+2002/07/26/big/img_576
+2002/07/22/big/img_273
+2003/01/16/big/img_470
+2002/07/29/big/img_329
+2002/07/30/big/img_1086
+2002/07/31/big/img_353
+2002/09/02/big/img_15275
+2003/01/17/big/img_555
+2002/08/26/big/img_212
+2002/08/01/big/img_1692
+2003/01/15/big/img_600
+2002/07/29/big/img_825
+2002/08/08/big/img_68
+2002/08/10/big/img_719
+2002/07/31/big/img_636
+2002/07/29/big/img_325
+2002/07/21/big/img_515
+2002/07/22/big/img_705
+2003/01/13/big/img_818
+2002/08/09/big/img_486
+2002/08/22/big/img_141
+2002/07/22/big/img_303
+2002/08/09/big/img_393
+2002/07/29/big/img_963
+2002/08/02/big/img_1215
+2002/08/19/big/img_674
+2002/08/12/big/img_690
+2002/08/21/big/img_637
+2002/08/21/big/img_841
+2002/08/24/big/img_71
+2002/07/25/big/img_596
+2002/07/24/big/img_864
+2002/08/18/big/img_293
+2003/01/14/big/img_657
+2002/08/15/big/img_411
+2002/08/16/big/img_348
+2002/08/05/big/img_3157
+2002/07/20/big/img_663
+2003/01/13/big/img_654
+2003/01/16/big/img_433
+2002/08/30/big/img_18200
+2002/08/12/big/img_226
+2003/01/16/big/img_491
+2002/08/08/big/img_666
+2002/07/19/big/img_576
+2003/01/15/big/img_776
+2003/01/16/big/img_899
+2002/07/19/big/img_397
+2002/08/14/big/img_44
+2003/01/15/big/img_762
+2002/08/02/big/img_982
+2002/09/02/big/img_15234
+2002/08/17/big/img_556
+2002/08/21/big/img_410
+2002/08/21/big/img_386
+2002/07/19/big/img_690
+2002/08/05/big/img_3052
+2002/08/14/big/img_219
+2002/08/16/big/img_273
+2003/01/15/big/img_752
+2002/08/08/big/img_184
+2002/07/31/big/img_743
+2002/08/23/big/img_338
+2003/01/14/big/img_1055
+2002/08/05/big/img_3405
+2003/01/15/big/img_17
+2002/08/03/big/img_141
+2002/08/14/big/img_549
+2002/07/27/big/img_1034
+2002/07/31/big/img_932
+2002/08/30/big/img_18487
+2002/09/02/big/img_15814
+2002/08/01/big/img_2086
+2002/09/01/big/img_16535
+2002/07/22/big/img_500
+2003/01/13/big/img_400
+2002/08/25/big/img_607
+2002/08/30/big/img_18384
+2003/01/14/big/img_951
+2002/08/13/big/img_1150
+2002/08/08/big/img_1022
+2002/08/10/big/img_428
+2002/08/28/big/img_19242
+2002/08/05/big/img_3098
+2002/07/23/big/img_400
+2002/08/26/big/img_365
+2002/07/20/big/img_318
+2002/08/13/big/img_740
+2003/01/16/big/img_37
+2002/08/26/big/img_274
+2002/08/02/big/img_205
+2002/08/21/big/img_695
+2002/08/06/big/img_2289
+2002/08/20/big/img_794
+2002/08/18/big/img_438
+2002/08/07/big/img_1380
+2002/08/02/big/img_737
+2002/08/07/big/img_1651
+2002/08/15/big/img_1238
+2002/08/01/big/img_1681
+2002/08/06/big/img_3017
+2002/07/23/big/img_706
+2002/07/31/big/img_392
+2002/08/09/big/img_539
+2002/07/29/big/img_835
+2002/08/26/big/img_723
+2002/08/28/big/img_19235
+2003/01/16/big/img_353
+2002/08/10/big/img_150
+2002/08/29/big/img_19025
+2002/08/21/big/img_310
+2002/08/10/big/img_823
+2002/07/26/big/img_981
+2002/08/11/big/img_288
+2002/08/19/big/img_534
+2002/08/21/big/img_300
+2002/07/31/big/img_49
+2002/07/30/big/img_469
+2002/08/28/big/img_19197
+2002/08/25/big/img_205
+2002/08/10/big/img_390
+2002/08/23/big/img_291
+2002/08/26/big/img_230
+2002/08/18/big/img_76
+2002/07/23/big/img_409
+2002/08/14/big/img_1053
+2003/01/14/big/img_291
+2002/08/10/big/img_503
+2002/08/27/big/img_19928
+2002/08/03/big/img_563
+2002/08/17/big/img_250
+2002/08/06/big/img_2381
+2002/08/17/big/img_948
+2002/08/06/big/img_2710
+2002/07/22/big/img_696
+2002/07/31/big/img_670
+2002/08/12/big/img_594
+2002/07/29/big/img_624
+2003/01/17/big/img_934
+2002/08/03/big/img_584
+2002/08/22/big/img_1003
+2002/08/05/big/img_3396
+2003/01/13/big/img_570
+2002/08/02/big/img_219
+2002/09/02/big/img_15774
+2002/08/16/big/img_818
+2002/08/23/big/img_402
+2003/01/14/big/img_552
+2002/07/29/big/img_71
+2002/08/05/big/img_3592
+2002/08/16/big/img_80
+2002/07/27/big/img_672
+2003/01/13/big/img_470
+2003/01/16/big/img_702
+2002/09/01/big/img_16130
+2002/08/08/big/img_240
+2002/09/01/big/img_16338
+2002/07/26/big/img_312
+2003/01/14/big/img_538
+2002/07/20/big/img_695
+2002/08/30/big/img_18098
+2002/08/25/big/img_259
+2002/08/16/big/img_1042
+2002/08/09/big/img_837
+2002/08/31/big/img_17760
+2002/07/31/big/img_14
+2002/08/09/big/img_361
+2003/01/16/big/img_107
+2002/08/14/big/img_124
+2002/07/19/big/img_463
+2003/01/15/big/img_275
+2002/07/25/big/img_1151
+2002/07/29/big/img_1501
+2002/08/27/big/img_19889
+2002/08/29/big/img_18603
+2003/01/17/big/img_601
+2002/08/25/big/img_355
+2002/08/08/big/img_297
+2002/08/20/big/img_290
+2002/07/31/big/img_195
+2003/01/01/big/img_336
+2002/08/18/big/img_369
+2002/07/25/big/img_621
+2002/08/11/big/img_508
+2003/01/14/big/img_458
+2003/01/15/big/img_795
+2002/08/12/big/img_498
+2002/08/01/big/img_1734
+2002/08/02/big/img_246
+2002/08/16/big/img_565
+2002/08/11/big/img_475
+2002/08/22/big/img_408
+2002/07/28/big/img_78
+2002/07/21/big/img_81
+2003/01/14/big/img_697
+2002/08/14/big/img_661
+2002/08/15/big/img_507
+2002/08/19/big/img_55
+2002/07/22/big/img_152
+2003/01/14/big/img_470
+2002/08/03/big/img_379
+2002/08/22/big/img_506
+2003/01/16/big/img_966
+2002/08/18/big/img_698
+2002/08/24/big/img_528
+2002/08/23/big/img_10
+2002/08/01/big/img_1655
+2002/08/22/big/img_953
+2002/07/19/big/img_630
+2002/07/22/big/img_889
+2002/08/16/big/img_351
+2003/01/16/big/img_83
+2002/07/19/big/img_805
+2002/08/14/big/img_704
+2002/07/19/big/img_389
+2002/08/31/big/img_17765
+2002/07/29/big/img_606
+2003/01/17/big/img_939
+2002/09/02/big/img_15081
+2002/08/21/big/img_181
+2002/07/29/big/img_1321
+2002/07/21/big/img_497
+2002/07/20/big/img_539
+2002/08/24/big/img_119
+2002/08/01/big/img_1281
+2002/07/26/big/img_207
+2002/07/26/big/img_432
+2002/07/27/big/img_1006
+2002/08/05/big/img_3087
+2002/08/14/big/img_252
+2002/08/14/big/img_798
+2002/07/24/big/img_538
+2002/09/02/big/img_15507
+2002/08/08/big/img_901
+2003/01/14/big/img_557
+2002/08/07/big/img_1819
+2002/08/04/big/img_470
+2002/08/01/big/img_1504
+2002/08/16/big/img_1070
+2002/08/16/big/img_372
+2002/08/23/big/img_416
+2002/08/30/big/img_18208
+2002/08/01/big/img_2043
+2002/07/22/big/img_385
+2002/08/22/big/img_466
+2002/08/21/big/img_869
+2002/08/28/big/img_19429
+2002/08/02/big/img_770
+2002/07/23/big/img_433
+2003/01/14/big/img_13
+2002/07/27/big/img_953
+2002/09/02/big/img_15728
+2002/08/01/big/img_1361
+2002/08/29/big/img_18897
+2002/08/26/big/img_534
+2002/08/11/big/img_121
+2002/08/26/big/img_20130
+2002/07/31/big/img_363
+2002/08/13/big/img_978
+2002/07/25/big/img_835
+2002/08/02/big/img_906
+2003/01/14/big/img_548
+2002/07/30/big/img_80
+2002/07/26/big/img_982
+2003/01/16/big/img_99
+2002/08/19/big/img_362
+2002/08/24/big/img_376
+2002/08/07/big/img_1264
+2002/07/27/big/img_938
+2003/01/17/big/img_535
+2002/07/26/big/img_457
+2002/08/08/big/img_848
+2003/01/15/big/img_859
+2003/01/15/big/img_622
+2002/07/30/big/img_403
+2002/07/29/big/img_217
+2002/07/26/big/img_891
+2002/07/24/big/img_70
+2002/08/25/big/img_619
+2002/08/05/big/img_3375
+2002/08/01/big/img_2160
+2002/08/06/big/img_2227
+2003/01/14/big/img_117
+2002/08/14/big/img_227
+2002/08/13/big/img_565
+2002/08/19/big/img_625
+2002/08/03/big/img_812
+2002/07/24/big/img_41
+2002/08/16/big/img_235
+2002/07/29/big/img_759
+2002/07/21/big/img_433
+2002/07/29/big/img_190
+2003/01/16/big/img_435
+2003/01/13/big/img_708
+2002/07/30/big/img_57
+2002/08/22/big/img_162
+2003/01/01/big/img_558
+2003/01/15/big/img_604
+2002/08/16/big/img_935
+2002/08/20/big/img_394
+2002/07/28/big/img_465
+2002/09/02/big/img_15534
+2002/08/16/big/img_87
+2002/07/22/big/img_469
+2002/08/12/big/img_245
+2003/01/13/big/img_236
+2002/08/06/big/img_2736
+2002/08/03/big/img_348
+2003/01/14/big/img_218
+2002/07/26/big/img_232
+2003/01/15/big/img_244
+2002/07/25/big/img_1121
+2002/08/01/big/img_1484
+2002/07/26/big/img_541
+2002/08/07/big/img_1244
+2002/07/31/big/img_3
+2002/08/30/big/img_18437
+2002/08/29/big/img_19094
+2002/08/01/big/img_1355
+2002/08/19/big/img_338
+2002/07/19/big/img_255
+2002/07/21/big/img_76
+2002/08/25/big/img_199
+2002/08/12/big/img_740
+2002/07/30/big/img_852
+2002/08/15/big/img_599
+2002/08/23/big/img_254
+2002/08/19/big/img_125
+2002/07/24/big/img_2
+2002/08/04/big/img_145
+2002/08/05/big/img_3137
+2002/07/28/big/img_463
+2003/01/14/big/img_801
+2002/07/23/big/img_366
+2002/08/26/big/img_600
+2002/08/26/big/img_649
+2002/09/02/big/img_15849
+2002/07/26/big/img_248
+2003/01/13/big/img_200
+2002/08/07/big/img_1794
+2002/08/31/big/img_17270
+2002/08/23/big/img_608
+2003/01/13/big/img_837
+2002/08/23/big/img_581
+2002/08/20/big/img_754
+2002/08/18/big/img_183
+2002/08/20/big/img_328
+2002/07/22/big/img_494
+2002/07/29/big/img_399
+2002/08/28/big/img_19284
+2002/08/08/big/img_566
+2002/07/25/big/img_376
+2002/07/23/big/img_138
+2002/07/25/big/img_435
+2002/08/17/big/img_685
+2002/07/19/big/img_90
+2002/07/20/big/img_716
+2002/08/31/big/img_17458
+2002/08/26/big/img_461
+2002/07/25/big/img_355
+2002/08/06/big/img_2152
+2002/07/27/big/img_932
+2002/07/23/big/img_232
+2002/08/08/big/img_1020
+2002/07/31/big/img_366
+2002/08/06/big/img_2667
+2002/08/21/big/img_465
+2002/08/15/big/img_305
+2002/08/02/big/img_247
+2002/07/28/big/img_46
+2002/08/27/big/img_19922
+2002/08/23/big/img_643
+2003/01/13/big/img_624
+2002/08/23/big/img_625
+2002/08/05/big/img_3787
+2003/01/13/big/img_627
+2002/09/01/big/img_16381
+2002/08/05/big/img_3668
+2002/07/21/big/img_535
+2002/08/27/big/img_19680
+2002/07/22/big/img_413
+2002/07/29/big/img_481
+2003/01/15/big/img_496
+2002/07/23/big/img_701
+2002/08/29/big/img_18670
+2002/07/28/big/img_319
+2003/01/14/big/img_517
+2002/07/26/big/img_256
+2003/01/16/big/img_593
+2002/07/30/big/img_956
+2002/07/30/big/img_667
+2002/07/25/big/img_100
+2002/08/11/big/img_570
+2002/07/26/big/img_745
+2002/08/04/big/img_834
+2002/08/25/big/img_521
+2002/08/01/big/img_2148
+2002/09/02/big/img_15183
+2002/08/22/big/img_514
+2002/08/23/big/img_477
+2002/07/23/big/img_336
+2002/07/26/big/img_481
+2002/08/20/big/img_409
+2002/07/23/big/img_918
+2002/08/09/big/img_474
+2002/08/02/big/img_929
+2002/08/31/big/img_17932
+2002/08/19/big/img_161
+2002/08/09/big/img_667
+2002/07/31/big/img_805
+2002/09/02/big/img_15678
+2002/08/31/big/img_17509
+2002/08/29/big/img_18998
+2002/07/23/big/img_301
+2002/08/07/big/img_1612
+2002/08/06/big/img_2472
+2002/07/23/big/img_466
+2002/08/27/big/img_19634
+2003/01/16/big/img_16
+2002/08/14/big/img_193
+2002/08/21/big/img_340
+2002/08/27/big/img_19799
+2002/08/01/big/img_1345
+2002/08/07/big/img_1448
+2002/08/11/big/img_324
+2003/01/16/big/img_754
+2002/08/13/big/img_418
+2003/01/16/big/img_544
+2002/08/19/big/img_135
+2002/08/10/big/img_455
+2002/08/10/big/img_693
+2002/08/31/big/img_17967
+2002/08/28/big/img_19229
+2002/08/04/big/img_811
+2002/09/01/big/img_16225
+2003/01/16/big/img_428
+2002/09/02/big/img_15295
+2002/07/26/big/img_108
+2002/07/21/big/img_477
+2002/08/07/big/img_1354
+2002/08/23/big/img_246
+2002/08/16/big/img_652
+2002/07/27/big/img_553
+2002/07/31/big/img_346
+2002/08/04/big/img_537
+2002/08/08/big/img_498
+2002/08/29/big/img_18956
+2003/01/13/big/img_922
+2002/08/31/big/img_17425
+2002/07/26/big/img_438
+2002/08/19/big/img_185
+2003/01/16/big/img_33
+2002/08/10/big/img_252
+2002/07/29/big/img_598
+2002/08/27/big/img_19820
+2002/08/06/big/img_2664
+2002/08/20/big/img_705
+2003/01/14/big/img_816
+2002/08/03/big/img_552
+2002/07/25/big/img_561
+2002/07/25/big/img_934
+2002/08/01/big/img_1893
+2003/01/14/big/img_746
+2003/01/16/big/img_519
+2002/08/03/big/img_681
+2002/07/24/big/img_808
+2002/08/14/big/img_803
+2002/08/25/big/img_155
+2002/07/30/big/img_1107
+2002/08/29/big/img_18882
+2003/01/15/big/img_598
+2002/08/19/big/img_122
+2002/07/30/big/img_428
+2002/07/24/big/img_684
+2002/08/22/big/img_192
+2002/08/22/big/img_543
+2002/08/07/big/img_1318
+2002/08/18/big/img_25
+2002/07/26/big/img_583
+2002/07/20/big/img_464
+2002/08/19/big/img_664
+2002/08/24/big/img_861
+2002/09/01/big/img_16136
+2002/08/22/big/img_400
+2002/08/12/big/img_445
+2003/01/14/big/img_174
+2002/08/27/big/img_19677
+2002/08/31/big/img_17214
+2002/08/30/big/img_18175
+2003/01/17/big/img_402
+2002/08/06/big/img_2396
+2002/08/18/big/img_448
+2002/08/21/big/img_165
+2002/08/31/big/img_17609
+2003/01/01/big/img_151
+2002/08/26/big/img_372
+2002/09/02/big/img_15994
+2002/07/26/big/img_660
+2002/09/02/big/img_15197
+2002/07/29/big/img_258
+2002/08/30/big/img_18525
+2003/01/13/big/img_368
+2002/07/29/big/img_1538
+2002/07/21/big/img_787
+2002/08/18/big/img_152
+2002/08/06/big/img_2379
+2003/01/17/big/img_864
+2002/08/27/big/img_19998
+2002/08/01/big/img_1634
+2002/07/25/big/img_414
+2002/08/22/big/img_627
+2002/08/07/big/img_1669
+2002/08/16/big/img_1052
+2002/08/31/big/img_17796
+2002/08/18/big/img_199
+2002/09/02/big/img_15147
+2002/08/09/big/img_460
+2002/08/14/big/img_581
+2002/08/30/big/img_18286
+2002/07/26/big/img_337
+2002/08/18/big/img_589
+2003/01/14/big/img_866
+2002/07/20/big/img_624
+2002/08/01/big/img_1801
+2002/07/24/big/img_683
+2002/08/09/big/img_725
+2003/01/14/big/img_34
+2002/07/30/big/img_144
+2002/07/30/big/img_706
+2002/08/08/big/img_394
+2002/08/19/big/img_619
+2002/08/06/big/img_2703
+2002/08/29/big/img_19034
+2002/07/24/big/img_67
+2002/08/27/big/img_19841
+2002/08/19/big/img_427
+2003/01/14/big/img_333
+2002/09/01/big/img_16406
+2002/07/19/big/img_882
+2002/08/17/big/img_238
+2003/01/14/big/img_739
+2002/07/22/big/img_151
+2002/08/21/big/img_743
+2002/07/25/big/img_1048
+2002/07/30/big/img_395
+2003/01/13/big/img_584
+2002/08/13/big/img_742
+2002/08/13/big/img_1168
+2003/01/14/big/img_147
+2002/07/26/big/img_803
+2002/08/05/big/img_3298
+2002/08/07/big/img_1451
+2002/08/16/big/img_424
+2002/07/29/big/img_1069
+2002/09/01/big/img_16735
+2002/07/21/big/img_637
+2003/01/14/big/img_585
+2002/08/02/big/img_358
+2003/01/13/big/img_358
+2002/08/14/big/img_198
+2002/08/17/big/img_935
+2002/08/04/big/img_42
+2002/08/30/big/img_18245
+2002/07/25/big/img_158
+2002/08/22/big/img_744
+2002/08/06/big/img_2291
+2002/08/05/big/img_3044
+2002/07/30/big/img_272
+2002/08/23/big/img_641
+2002/07/24/big/img_797
+2002/07/30/big/img_392
+2003/01/14/big/img_447
+2002/07/31/big/img_898
+2002/08/06/big/img_2812
+2002/08/13/big/img_564
+2002/07/22/big/img_43
+2002/07/26/big/img_634
+2002/07/19/big/img_843
+2002/08/26/big/img_58
+2002/07/21/big/img_375
+2002/08/25/big/img_729
+2002/07/19/big/img_561
+2003/01/15/big/img_884
+2002/07/25/big/img_891
+2002/08/09/big/img_558
+2002/08/26/big/img_587
+2002/08/13/big/img_1146
+2002/09/02/big/img_15153
+2002/07/26/big/img_316
+2002/08/01/big/img_1940
+2002/08/26/big/img_90
+2003/01/13/big/img_347
+2002/07/25/big/img_520
+2002/08/29/big/img_18718
+2002/08/28/big/img_19219
+2002/08/13/big/img_375
+2002/07/20/big/img_719
+2002/08/31/big/img_17431
+2002/07/28/big/img_192
+2002/08/26/big/img_259
+2002/08/18/big/img_484
+2002/07/29/big/img_580
+2002/07/26/big/img_84
+2002/08/02/big/img_302
+2002/08/31/big/img_17007
+2003/01/15/big/img_543
+2002/09/01/big/img_16488
+2002/08/22/big/img_798
+2002/07/30/big/img_383
+2002/08/04/big/img_668
+2002/08/13/big/img_156
+2002/08/07/big/img_1353
+2002/07/25/big/img_281
+2003/01/14/big/img_587
+2003/01/15/big/img_524
+2002/08/19/big/img_726
+2002/08/21/big/img_709
+2002/08/26/big/img_465
+2002/07/31/big/img_658
+2002/08/28/big/img_19148
+2002/07/23/big/img_423
+2002/08/16/big/img_758
+2002/08/22/big/img_523
+2002/08/16/big/img_591
+2002/08/23/big/img_845
+2002/07/26/big/img_678
+2002/08/09/big/img_806
+2002/08/06/big/img_2369
+2002/07/29/big/img_457
+2002/07/19/big/img_278
+2002/08/30/big/img_18107
+2002/07/26/big/img_444
+2002/08/20/big/img_278
+2002/08/26/big/img_92
+2002/08/26/big/img_257
+2002/07/25/big/img_266
+2002/08/05/big/img_3829
+2002/07/26/big/img_757
+2002/07/29/big/img_1536
+2002/08/09/big/img_472
+2003/01/17/big/img_480
+2002/08/28/big/img_19355
+2002/07/26/big/img_97
+2002/08/06/big/img_2503
+2002/07/19/big/img_254
+2002/08/01/big/img_1470
+2002/08/21/big/img_42
+2002/08/20/big/img_217
+2002/08/06/big/img_2459
+2002/07/19/big/img_552
+2002/08/13/big/img_717
+2002/08/12/big/img_586
+2002/08/20/big/img_411
+2003/01/13/big/img_768
+2002/08/07/big/img_1747
+2002/08/15/big/img_385
+2002/08/01/big/img_1648
+2002/08/15/big/img_311
+2002/08/21/big/img_95
+2002/08/09/big/img_108
+2002/08/21/big/img_398
+2002/08/17/big/img_340
+2002/08/14/big/img_474
+2002/08/13/big/img_294
+2002/08/24/big/img_840
+2002/08/09/big/img_808
+2002/08/23/big/img_491
+2002/07/28/big/img_33
+2003/01/13/big/img_664
+2002/08/02/big/img_261
+2002/08/09/big/img_591
+2002/07/26/big/img_309
+2003/01/14/big/img_372
+2002/08/19/big/img_581
+2002/08/19/big/img_168
+2002/08/26/big/img_422
+2002/07/24/big/img_106
+2002/08/01/big/img_1936
+2002/08/05/big/img_3764
+2002/08/21/big/img_266
+2002/08/31/big/img_17968
+2002/08/01/big/img_1941
+2002/08/15/big/img_550
+2002/08/14/big/img_13
+2002/07/30/big/img_171
+2003/01/13/big/img_490
+2002/07/25/big/img_427
+2002/07/19/big/img_770
+2002/08/12/big/img_759
+2003/01/15/big/img_1360
+2002/08/05/big/img_3692
+2003/01/16/big/img_30
+2002/07/25/big/img_1026
+2002/07/22/big/img_288
+2002/08/29/big/img_18801
+2002/07/24/big/img_793
+2002/08/13/big/img_178
+2002/08/06/big/img_2322
+2003/01/14/big/img_560
+2002/08/18/big/img_408
+2003/01/16/big/img_915
+2003/01/16/big/img_679
+2002/08/07/big/img_1552
+2002/08/29/big/img_19050
+2002/08/01/big/img_2172
+2002/07/31/big/img_30
+2002/07/30/big/img_1019
+2002/07/30/big/img_587
+2003/01/13/big/img_773
+2002/07/30/big/img_410
+2002/07/28/big/img_65
+2002/08/05/big/img_3138
+2002/07/23/big/img_541
+2002/08/22/big/img_963
+2002/07/27/big/img_657
+2002/07/30/big/img_1051
+2003/01/16/big/img_150
+2002/07/31/big/img_519
+2002/08/01/big/img_1961
+2002/08/05/big/img_3752
+2002/07/23/big/img_631
+2003/01/14/big/img_237
+2002/07/28/big/img_21
+2002/07/22/big/img_813
+2002/08/05/big/img_3563
+2003/01/17/big/img_620
+2002/07/19/big/img_523
+2002/07/30/big/img_904
+2002/08/29/big/img_18642
+2002/08/11/big/img_492
+2002/08/01/big/img_2130
+2002/07/25/big/img_618
+2002/08/17/big/img_305
+2003/01/16/big/img_520
+2002/07/26/big/img_495
+2002/08/17/big/img_164
+2002/08/03/big/img_440
+2002/07/24/big/img_441
+2002/08/06/big/img_2146
+2002/08/11/big/img_558
+2002/08/02/big/img_545
+2002/08/31/big/img_18090
+2003/01/01/big/img_136
+2002/07/25/big/img_1099
+2003/01/13/big/img_728
+2003/01/16/big/img_197
+2002/07/26/big/img_651
+2002/08/11/big/img_676
+2003/01/15/big/img_10
+2002/08/21/big/img_250
+2002/08/14/big/img_325
+2002/08/04/big/img_390
+2002/07/24/big/img_554
+2003/01/16/big/img_333
+2002/07/31/big/img_922
+2002/09/02/big/img_15586
+2003/01/16/big/img_184
+2002/07/22/big/img_766
+2002/07/21/big/img_608
+2002/08/07/big/img_1578
+2002/08/17/big/img_961
+2002/07/27/big/img_324
+2002/08/05/big/img_3765
+2002/08/23/big/img_462
+2003/01/16/big/img_382
+2002/08/27/big/img_19838
+2002/08/01/big/img_1505
+2002/08/21/big/img_662
+2002/08/14/big/img_605
+2002/08/19/big/img_816
+2002/07/29/big/img_136
+2002/08/20/big/img_719
+2002/08/06/big/img_2826
+2002/08/10/big/img_630
+2003/01/17/big/img_973
+2002/08/14/big/img_116
+2002/08/02/big/img_666
+2002/08/21/big/img_710
+2002/08/05/big/img_55
+2002/07/31/big/img_229
+2002/08/01/big/img_1549
+2002/07/23/big/img_432
+2002/07/21/big/img_430
+2002/08/21/big/img_549
+2002/08/08/big/img_985
+2002/07/20/big/img_610
+2002/07/23/big/img_978
+2002/08/23/big/img_219
+2002/07/25/big/img_175
+2003/01/15/big/img_230
+2002/08/23/big/img_385
+2002/07/31/big/img_879
+2002/08/12/big/img_495
+2002/08/22/big/img_499
+2002/08/30/big/img_18322
+2002/08/15/big/img_795
+2002/08/13/big/img_835
+2003/01/17/big/img_930
+2002/07/30/big/img_873
+2002/08/11/big/img_257
+2002/07/31/big/img_593
+2002/08/21/big/img_916
+2003/01/13/big/img_814
+2002/07/25/big/img_722
+2002/08/16/big/img_379
+2002/07/31/big/img_497
+2002/07/22/big/img_602
+2002/08/21/big/img_642
+2002/08/21/big/img_614
+2002/08/23/big/img_482
+2002/07/29/big/img_603
+2002/08/13/big/img_705
+2002/07/23/big/img_833
+2003/01/14/big/img_511
+2002/07/24/big/img_376
+2002/08/17/big/img_1030
+2002/08/05/big/img_3576
+2002/08/16/big/img_540
+2002/07/22/big/img_630
+2002/08/10/big/img_180
+2002/08/14/big/img_905
+2002/08/29/big/img_18777
+2002/08/22/big/img_693
+2003/01/16/big/img_933
+2002/08/20/big/img_555
+2002/08/15/big/img_549
+2003/01/14/big/img_830
+2003/01/16/big/img_64
+2002/08/27/big/img_19670
+2002/08/22/big/img_729
+2002/07/27/big/img_981
+2002/08/09/big/img_458
+2003/01/17/big/img_884
+2002/07/25/big/img_639
+2002/08/31/big/img_18008
+2002/08/22/big/img_249
+2002/08/17/big/img_971
+2002/08/04/big/img_308
+2002/07/28/big/img_362
+2002/08/12/big/img_142
+2002/08/26/big/img_61
+2002/08/14/big/img_422
+2002/07/19/big/img_607
+2003/01/15/big/img_717
+2002/08/01/big/img_1475
+2002/08/29/big/img_19061
+2003/01/01/big/img_346
+2002/07/20/big/img_315
+2003/01/15/big/img_756
+2002/08/15/big/img_879
+2002/08/08/big/img_615
+2003/01/13/big/img_431
+2002/08/05/big/img_3233
+2002/08/24/big/img_526
+2003/01/13/big/img_717
+2002/09/01/big/img_16408
+2002/07/22/big/img_217
+2002/07/31/big/img_960
+2002/08/21/big/img_610
+2002/08/05/big/img_3753
+2002/08/03/big/img_151
+2002/08/21/big/img_267
+2002/08/01/big/img_2175
+2002/08/04/big/img_556
+2002/08/21/big/img_527
+2002/09/02/big/img_15800
+2002/07/27/big/img_156
+2002/07/20/big/img_590
+2002/08/15/big/img_700
+2002/08/08/big/img_444
+2002/07/25/big/img_94
+2002/07/24/big/img_778
+2002/08/14/big/img_694
+2002/07/20/big/img_666
+2002/08/02/big/img_200
+2002/08/02/big/img_578
+2003/01/17/big/img_332
+2002/09/01/big/img_16352
+2002/08/27/big/img_19668
+2002/07/23/big/img_823
+2002/08/13/big/img_431
+2003/01/16/big/img_463
+2002/08/27/big/img_19711
+2002/08/23/big/img_154
+2002/07/31/big/img_360
+2002/08/23/big/img_555
+2002/08/10/big/img_561
+2003/01/14/big/img_550
+2002/08/07/big/img_1370
+2002/07/30/big/img_1184
+2002/08/01/big/img_1445
+2002/08/23/big/img_22
+2002/07/30/big/img_606
+2003/01/17/big/img_271
+2002/08/31/big/img_17316
+2002/08/16/big/img_973
+2002/07/26/big/img_77
+2002/07/20/big/img_788
+2002/08/06/big/img_2426
+2002/08/07/big/img_1498
+2002/08/16/big/img_358
+2002/08/06/big/img_2851
+2002/08/12/big/img_359
+2002/08/01/big/img_1521
+2002/08/02/big/img_709
+2002/08/20/big/img_935
+2002/08/12/big/img_188
+2002/08/24/big/img_411
+2002/08/22/big/img_680
+2002/08/06/big/img_2480
+2002/07/20/big/img_627
+2002/07/30/big/img_214
+2002/07/25/big/img_354
+2002/08/02/big/img_636
+2003/01/15/big/img_661
+2002/08/07/big/img_1327
+2002/08/01/big/img_2108
+2002/08/31/big/img_17919
+2002/08/29/big/img_18768
+2002/08/05/big/img_3840
+2002/07/26/big/img_242
+2003/01/14/big/img_451
+2002/08/20/big/img_923
+2002/08/27/big/img_19908
+2002/08/16/big/img_282
+2002/08/19/big/img_440
+2003/01/01/big/img_230
+2002/08/08/big/img_212
+2002/07/20/big/img_443
+2002/08/25/big/img_635
+2003/01/13/big/img_1169
+2002/07/26/big/img_998
+2002/08/15/big/img_995
+2002/08/06/big/img_3002
+2002/07/29/big/img_460
+2003/01/14/big/img_925
+2002/07/23/big/img_539
+2002/08/16/big/img_694
+2003/01/13/big/img_459
+2002/07/23/big/img_249
+2002/08/20/big/img_539
+2002/08/04/big/img_186
+2002/08/26/big/img_264
+2002/07/22/big/img_704
+2002/08/25/big/img_277
+2002/08/22/big/img_988
+2002/07/29/big/img_504
+2002/08/05/big/img_3600
+2002/08/30/big/img_18380
+2003/01/14/big/img_937
+2002/08/21/big/img_254
+2002/08/10/big/img_130
+2002/08/20/big/img_339
+2003/01/14/big/img_428
+2002/08/20/big/img_889
+2002/08/31/big/img_17637
+2002/07/26/big/img_644
+2002/09/01/big/img_16776
+2002/08/06/big/img_2239
+2002/08/06/big/img_2646
+2003/01/13/big/img_491
+2002/08/10/big/img_579
+2002/08/21/big/img_713
+2002/08/22/big/img_482
+2002/07/22/big/img_167
+2002/07/24/big/img_539
+2002/08/14/big/img_721
+2002/07/25/big/img_389
+2002/09/01/big/img_16591
+2002/08/13/big/img_543
+2003/01/14/big/img_432
+2002/08/09/big/img_287
+2002/07/26/big/img_126
+2002/08/23/big/img_412
+2002/08/15/big/img_1034
+2002/08/28/big/img_19485
+2002/07/31/big/img_236
+2002/07/30/big/img_523
+2002/07/19/big/img_141
+2003/01/17/big/img_957
+2002/08/04/big/img_81
+2002/07/25/big/img_206
+2002/08/15/big/img_716
+2002/08/13/big/img_403
+2002/08/15/big/img_685
+2002/07/26/big/img_884
+2002/07/19/big/img_499
+2002/07/23/big/img_772
+2002/07/27/big/img_752
+2003/01/14/big/img_493
+2002/08/25/big/img_664
+2002/07/31/big/img_334
+2002/08/26/big/img_678
+2002/09/01/big/img_16541
+2003/01/14/big/img_347
+2002/07/23/big/img_187
+2002/07/30/big/img_1163
+2002/08/05/big/img_35
+2002/08/22/big/img_944
+2002/08/07/big/img_1239
+2002/07/29/big/img_1215
+2002/08/03/big/img_312
+2002/08/05/big/img_3523
+2002/07/29/big/img_218
+2002/08/13/big/img_672
+2002/08/16/big/img_205
+2002/08/17/big/img_594
+2002/07/29/big/img_1411
+2002/07/30/big/img_942
+2003/01/16/big/img_312
+2002/08/08/big/img_312
+2002/07/25/big/img_15
+2002/08/09/big/img_839
+2002/08/01/big/img_2069
+2002/08/31/big/img_17512
+2002/08/01/big/img_3
+2002/07/31/big/img_320
+2003/01/15/big/img_1265
+2002/08/14/big/img_563
+2002/07/31/big/img_167
+2002/08/20/big/img_374
+2002/08/13/big/img_406
+2002/08/08/big/img_625
+2002/08/02/big/img_314
+2002/08/27/big/img_19964
+2002/09/01/big/img_16670
+2002/07/31/big/img_599
+2002/08/29/big/img_18906
+2002/07/24/big/img_373
+2002/07/26/big/img_513
+2002/09/02/big/img_15497
+2002/08/19/big/img_117
+2003/01/01/big/img_158
+2002/08/24/big/img_178
+2003/01/13/big/img_935
+2002/08/13/big/img_609
+2002/08/30/big/img_18341
+2002/08/25/big/img_674
+2003/01/13/big/img_209
+2002/08/13/big/img_258
+2002/08/05/big/img_3543
+2002/08/07/big/img_1970
+2002/08/06/big/img_3004
+2003/01/17/big/img_487
+2002/08/24/big/img_873
+2002/08/29/big/img_18730
+2002/08/09/big/img_375
+2003/01/16/big/img_751
+2002/08/02/big/img_603
+2002/08/19/big/img_325
+2002/09/01/big/img_16420
+2002/08/05/big/img_3633
+2002/08/21/big/img_516
+2002/07/19/big/img_501
+2002/07/26/big/img_688
+2002/07/24/big/img_256
+2002/07/25/big/img_438
+2002/07/31/big/img_1017
+2002/08/22/big/img_512
+2002/07/21/big/img_543
+2002/08/08/big/img_223
+2002/08/19/big/img_189
+2002/08/12/big/img_630
+2002/07/30/big/img_958
+2002/07/28/big/img_208
+2002/08/31/big/img_17691
+2002/07/22/big/img_542
+2002/07/19/big/img_741
+2002/07/19/big/img_158
+2002/08/15/big/img_399
+2002/08/01/big/img_2159
+2002/08/14/big/img_455
+2002/08/17/big/img_1011
+2002/08/26/big/img_744
+2002/08/12/big/img_624
+2003/01/17/big/img_821
+2002/08/16/big/img_980
+2002/07/28/big/img_281
+2002/07/25/big/img_171
+2002/08/03/big/img_116
+2002/07/22/big/img_467
+2002/07/31/big/img_750
+2002/07/26/big/img_435
+2002/07/19/big/img_822
+2002/08/13/big/img_626
+2002/08/11/big/img_344
+2002/08/02/big/img_473
+2002/09/01/big/img_16817
+2002/08/01/big/img_1275
+2002/08/28/big/img_19270
+2002/07/23/big/img_607
+2002/08/09/big/img_316
+2002/07/29/big/img_626
+2002/07/24/big/img_824
+2002/07/22/big/img_342
+2002/08/08/big/img_794
+2002/08/07/big/img_1209
+2002/07/19/big/img_18
+2002/08/25/big/img_634
+2002/07/24/big/img_730
+2003/01/17/big/img_356
+2002/07/23/big/img_305
+2002/07/30/big/img_453
+2003/01/13/big/img_972
+2002/08/06/big/img_2610
+2002/08/29/big/img_18920
+2002/07/31/big/img_123
+2002/07/26/big/img_979
+2002/08/24/big/img_635
+2002/08/05/big/img_3704
+2002/08/07/big/img_1358
+2002/07/22/big/img_306
+2002/08/13/big/img_619
+2002/08/02/big/img_366

data/__init__.py

@@ -0,0 +1,3 @@
+from .wider_face import WiderFaceDetection, detection_collate
+from .data_augment import *
+from .config import *

data/config.py

@@ -0,0 +1,42 @@
+# config.py
+
+cfg_mnet = {
+    'name': 'mobilenet0.25',
+    'min_sizes': [[16, 32], [64, 128], [256, 512]],
+    'steps': [8, 16, 32],
+    'variance': [0.1, 0.2],
+    'clip': False,
+    'loc_weight': 2.0,
+    'gpu_train': True,
+    'batch_size': 32,
+    'ngpu': 1,
+    'epoch': 250,
+    'decay1': 190,
+    'decay2': 220,
+    'image_size': 640,
+    'pretrain': True,
+    'return_layers': {'stage1': 1, 'stage2': 2, 'stage3': 3},
+    'in_channel': 32,
+    'out_channel': 64
+}
+
+cfg_re50 = {
+    'name': 'Resnet50',
+    'min_sizes': [[16, 32], [64, 128], [256, 512]],
+    'steps': [8, 16, 32],
+    'variance': [0.1, 0.2],
+    'clip': False,
+    'loc_weight': 2.0,
+    'gpu_train': True,
+    'batch_size': 24,
+    'ngpu': 4,
+    'epoch': 100,
+    'decay1': 70,
+    'decay2': 90,
+    'image_size': 840,
+    'pretrain': True,
+    'return_layers': {'layer2': 1, 'layer3': 2, 'layer4': 3},
+    'in_channel': 256,
+    'out_channel': 256
+}
+

data/data_augment.py

@@ -0,0 +1,237 @@
+import cv2
+import numpy as np
+import random
+from utils.box_utils import matrix_iof
+
+
+def _crop(image, boxes, labels, landm, img_dim):
+    height, width, _ = image.shape
+    pad_image_flag = True
+
+    for _ in range(250):
+        """
+        if random.uniform(0, 1) <= 0.2:
+            scale = 1.0
+        else:
+            scale = random.uniform(0.3, 1.0)
+        """
+        PRE_SCALES = [0.3, 0.45, 0.6, 0.8, 1.0]
+        scale = random.choice(PRE_SCALES)
+        short_side = min(width, height)
+        w = int(scale * short_side)
+        h = w
+
+        if width == w:
+            l = 0
+        else:
+            l = random.randrange(width - w)
+        if height == h:
+            t = 0
+        else:
+            t = random.randrange(height - h)
+        roi = np.array((l, t, l + w, t + h))
+
+        value = matrix_iof(boxes, roi[np.newaxis])
+        flag = (value >= 1)
+        if not flag.any():
+            continue
+
+        centers = (boxes[:, :2] + boxes[:, 2:]) / 2
+        mask_a = np.logical_and(roi[:2] < centers, centers < roi[2:]).all(axis=1)
+        boxes_t = boxes[mask_a].copy()
+        labels_t = labels[mask_a].copy()
+        landms_t = landm[mask_a].copy()
+        landms_t = landms_t.reshape([-1, 5, 2])
+
+        if boxes_t.shape[0] == 0:
+            continue
+
+        image_t = image[roi[1]:roi[3], roi[0]:roi[2]]
+
+        boxes_t[:, :2] = np.maximum(boxes_t[:, :2], roi[:2])
+        boxes_t[:, :2] -= roi[:2]
+        boxes_t[:, 2:] = np.minimum(boxes_t[:, 2:], roi[2:])
+        boxes_t[:, 2:] -= roi[:2]
+
+        # landm
+        landms_t[:, :, :2] = landms_t[:, :, :2] - roi[:2]
+        landms_t[:, :, :2] = np.maximum(landms_t[:, :, :2], np.array([0, 0]))
+        landms_t[:, :, :2] = np.minimum(landms_t[:, :, :2], roi[2:] - roi[:2])
+        landms_t = landms_t.reshape([-1, 10])
+
+
+	# make sure that the cropped image contains at least one face > 16 pixel at training image scale
+        b_w_t = (boxes_t[:, 2] - boxes_t[:, 0] + 1) / w * img_dim
+        b_h_t = (boxes_t[:, 3] - boxes_t[:, 1] + 1) / h * img_dim
+        mask_b = np.minimum(b_w_t, b_h_t) > 0.0
+        boxes_t = boxes_t[mask_b]
+        labels_t = labels_t[mask_b]
+        landms_t = landms_t[mask_b]
+
+        if boxes_t.shape[0] == 0:
+            continue
+
+        pad_image_flag = False
+
+        return image_t, boxes_t, labels_t, landms_t, pad_image_flag
+    return image, boxes, labels, landm, pad_image_flag
+
+
+def _distort(image):
+
+    def _convert(image, alpha=1, beta=0):
+        tmp = image.astype(float) * alpha + beta
+        tmp[tmp < 0] = 0
+        tmp[tmp > 255] = 255
+        image[:] = tmp
+
+    image = image.copy()
+
+    if random.randrange(2):
+
+        #brightness distortion
+        if random.randrange(2):
+            _convert(image, beta=random.uniform(-32, 32))
+
+        #contrast distortion
+        if random.randrange(2):
+            _convert(image, alpha=random.uniform(0.5, 1.5))
+
+        image = cv2.cvtColor(image, cv2.COLOR_BGR2HSV)
+
+        #saturation distortion
+        if random.randrange(2):
+            _convert(image[:, :, 1], alpha=random.uniform(0.5, 1.5))
+
+        #hue distortion
+        if random.randrange(2):
+            tmp = image[:, :, 0].astype(int) + random.randint(-18, 18)
+            tmp %= 180
+            image[:, :, 0] = tmp
+
+        image = cv2.cvtColor(image, cv2.COLOR_HSV2BGR)
+
+    else:
+
+        #brightness distortion
+        if random.randrange(2):
+            _convert(image, beta=random.uniform(-32, 32))
+
+        image = cv2.cvtColor(image, cv2.COLOR_BGR2HSV)
+
+        #saturation distortion
+        if random.randrange(2):
+            _convert(image[:, :, 1], alpha=random.uniform(0.5, 1.5))
+
+        #hue distortion
+        if random.randrange(2):
+            tmp = image[:, :, 0].astype(int) + random.randint(-18, 18)
+            tmp %= 180
+            image[:, :, 0] = tmp
+
+        image = cv2.cvtColor(image, cv2.COLOR_HSV2BGR)
+
+        #contrast distortion
+        if random.randrange(2):
+            _convert(image, alpha=random.uniform(0.5, 1.5))
+
+    return image
+
+
+def _expand(image, boxes, fill, p):
+    if random.randrange(2):
+        return image, boxes
+
+    height, width, depth = image.shape
+
+    scale = random.uniform(1, p)
+    w = int(scale * width)
+    h = int(scale * height)
+
+    left = random.randint(0, w - width)
+    top = random.randint(0, h - height)
+
+    boxes_t = boxes.copy()
+    boxes_t[:, :2] += (left, top)
+    boxes_t[:, 2:] += (left, top)
+    expand_image = np.empty(
+        (h, w, depth),
+        dtype=image.dtype)
+    expand_image[:, :] = fill
+    expand_image[top:top + height, left:left + width] = image
+    image = expand_image
+
+    return image, boxes_t
+
+
+def _mirror(image, boxes, landms):
+    _, width, _ = image.shape
+    if random.randrange(2):
+        image = image[:, ::-1]
+        boxes = boxes.copy()
+        boxes[:, 0::2] = width - boxes[:, 2::-2]
+
+        # landm
+        landms = landms.copy()
+        landms = landms.reshape([-1, 5, 2])
+        landms[:, :, 0] = width - landms[:, :, 0]
+        tmp = landms[:, 1, :].copy()
+        landms[:, 1, :] = landms[:, 0, :]
+        landms[:, 0, :] = tmp
+        tmp1 = landms[:, 4, :].copy()
+        landms[:, 4, :] = landms[:, 3, :]
+        landms[:, 3, :] = tmp1
+        landms = landms.reshape([-1, 10])
+
+    return image, boxes, landms
+
+
+def _pad_to_square(image, rgb_mean, pad_image_flag):
+    if not pad_image_flag:
+        return image
+    height, width, _ = image.shape
+    long_side = max(width, height)
+    image_t = np.empty((long_side, long_side, 3), dtype=image.dtype)
+    image_t[:, :] = rgb_mean
+    image_t[0:0 + height, 0:0 + width] = image
+    return image_t
+
+
+def _resize_subtract_mean(image, insize, rgb_mean):
+    interp_methods = [cv2.INTER_LINEAR, cv2.INTER_CUBIC, cv2.INTER_AREA, cv2.INTER_NEAREST, cv2.INTER_LANCZOS4]
+    interp_method = interp_methods[random.randrange(5)]
+    image = cv2.resize(image, (insize, insize), interpolation=interp_method)
+    image = image.astype(np.float32)
+    image -= rgb_mean
+    return image.transpose(2, 0, 1)
+
+
+class preproc(object):
+
+    def __init__(self, img_dim, rgb_means):
+        self.img_dim = img_dim
+        self.rgb_means = rgb_means
+
+    def __call__(self, image, targets):
+        assert targets.shape[0] > 0, "this image does not have gt"
+
+        boxes = targets[:, :4].copy()
+        labels = targets[:, -1].copy()
+        landm = targets[:, 4:-1].copy()
+
+        image_t, boxes_t, labels_t, landm_t, pad_image_flag = _crop(image, boxes, labels, landm, self.img_dim)
+        image_t = _distort(image_t)
+        image_t = _pad_to_square(image_t,self.rgb_means, pad_image_flag)
+        image_t, boxes_t, landm_t = _mirror(image_t, boxes_t, landm_t)
+        height, width, _ = image_t.shape
+        image_t = _resize_subtract_mean(image_t, self.img_dim, self.rgb_means)
+        boxes_t[:, 0::2] /= width
+        boxes_t[:, 1::2] /= height
+
+        landm_t[:, 0::2] /= width
+        landm_t[:, 1::2] /= height
+
+        labels_t = np.expand_dims(labels_t, 1)
+        targets_t = np.hstack((boxes_t, landm_t, labels_t))
+
+        return image_t, targets_t

data/wider_face.py

@@ -0,0 +1,101 @@
+import os
+import os.path
+import sys
+import torch
+import torch.utils.data as data
+import cv2
+import numpy as np
+
+class WiderFaceDetection(data.Dataset):
+    def __init__(self, txt_path, preproc=None):
+        self.preproc = preproc
+        self.imgs_path = []
+        self.words = []
+        f = open(txt_path,'r')
+        lines = f.readlines()
+        isFirst = True
+        labels = []
+        for line in lines:
+            line = line.rstrip()
+            if line.startswith('#'):
+                if isFirst is True:
+                    isFirst = False
+                else:
+                    labels_copy = labels.copy()
+                    self.words.append(labels_copy)
+                    labels.clear()
+                path = line[2:]
+                path = txt_path.replace('label.txt','images/') + path
+                self.imgs_path.append(path)
+            else:
+                line = line.split(' ')
+                label = [float(x) for x in line]
+                labels.append(label)
+
+        self.words.append(labels)
+
+    def __len__(self):
+        return len(self.imgs_path)
+
+    def __getitem__(self, index):
+        img = cv2.imread(self.imgs_path[index])
+        height, width, _ = img.shape
+
+        labels = self.words[index]
+        annotations = np.zeros((0, 15))
+        if len(labels) == 0:
+            return annotations
+        for idx, label in enumerate(labels):
+            annotation = np.zeros((1, 15))
+            # bbox
+            annotation[0, 0] = label[0]  # x1
+            annotation[0, 1] = label[1]  # y1
+            annotation[0, 2] = label[0] + label[2]  # x2
+            annotation[0, 3] = label[1] + label[3]  # y2
+
+            # landmarks
+            annotation[0, 4] = label[4]    # l0_x
+            annotation[0, 5] = label[5]    # l0_y
+            annotation[0, 6] = label[7]    # l1_x
+            annotation[0, 7] = label[8]    # l1_y
+            annotation[0, 8] = label[10]   # l2_x
+            annotation[0, 9] = label[11]   # l2_y
+            annotation[0, 10] = label[13]  # l3_x
+            annotation[0, 11] = label[14]  # l3_y
+            annotation[0, 12] = label[16]  # l4_x
+            annotation[0, 13] = label[17]  # l4_y
+            if (annotation[0, 4]<0):
+                annotation[0, 14] = -1
+            else:
+                annotation[0, 14] = 1
+
+            annotations = np.append(annotations, annotation, axis=0)
+        target = np.array(annotations)
+        if self.preproc is not None:
+            img, target = self.preproc(img, target)
+
+        return torch.from_numpy(img), target
+
+def detection_collate(batch):
+    """Custom collate fn for dealing with batches of images that have a different
+    number of associated object annotations (bounding boxes).
+
+    Arguments:
+        batch: (tuple) A tuple of tensor images and lists of annotations
+
+    Return:
+        A tuple containing:
+            1) (tensor) batch of images stacked on their 0 dim
+            2) (list of tensors) annotations for a given image are stacked on 0 dim
+    """
+    targets = []
+    imgs = []
+    for _, sample in enumerate(batch):
+        for _, tup in enumerate(sample):
+            if torch.is_tensor(tup):
+                imgs.append(tup)
+            elif isinstance(tup, type(np.empty(0))):
+                annos = torch.from_numpy(tup).float()
+                targets.append(annos)
+
+    return (torch.stack(imgs, 0), targets)

dataset/__init__.py

@@ -0,0 +1,17 @@
+from .abstract_dataset import AbstractDataset
+from .faceforensics import FaceForensics
+from .wild_deepfake import WildDeepfake
+from .celeb_df import CelebDF
+from .dfdc import DFDC
+
+LOADERS = {
+    "FaceForensics": FaceForensics,
+    "WildDeepfake": WildDeepfake,
+    "CelebDF": CelebDF,
+    "DFDC": DFDC,
+}
+
+
+def load_dataset(name="FaceForensics"):
+    print(f"Loading dataset: '{name}'...")
+    return LOADERS[name]

dataset/abstract_dataset.py

@@ -0,0 +1,41 @@
+import cv2
+import torch
+import numpy as np
+from torchvision.datasets import VisionDataset
+import albumentations
+from albumentations import Compose
+from albumentations.pytorch.transforms import ToTensorV2
+
+
+class AbstractDataset(VisionDataset):
+    def __init__(self, cfg, seed=2022, transforms=None, transform=None, target_transform=None):
+        super(AbstractDataset, self).__init__(cfg['root'], transforms=transforms,
+                                              transform=transform, target_transform=target_transform)
+        # fix for re-production
+        np.random.seed(seed)
+
+        self.images = list()
+        self.targets = list()
+        self.split = cfg['split']
+        if self.transforms is None:
+            self.transforms = Compose(
+                [getattr(albumentations, _['name'])(**_['params']) for _ in cfg['transforms']] +
+                [ToTensorV2()]
+            )
+
+    def __len__(self):
+        return len(self.images)
+
+    def __getitem__(self, index):
+        path = self.images[index]
+        tgt = self.targets[index]
+        return path, tgt
+
+    def load_item(self, items):
+        images = list()
+        for item in items:
+            img = cv2.imread(item)
+            img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
+            image = self.transforms(image=img)['image']
+            images.append(image)
+        return torch.stack(images, dim=0)

dataset/celeb_df.py

@@ -0,0 +1,126 @@
+import numpy as np
+from glob import glob
+from os import listdir
+from os.path import join
+from dataset import AbstractDataset
+
+SPLITS = ["train", "test"]
+
+
+class CelebDF(AbstractDataset):
+    """
+    Celeb-DF v2 Dataset proposed in "Celeb-DF: A Large-scale Challenging Dataset for DeepFake Forensics".
+    """
+
+    def __init__(self, cfg, seed=2022, transforms=None, transform=None, target_transform=None):
+        # pre-check
+        if cfg['split'] not in SPLITS:
+            raise ValueError(f"split should be one of {SPLITS}, but found {cfg['split']}.")
+        super(CelebDF, self).__init__(cfg, seed, transforms, transform, target_transform)
+        print(f"Loading data from 'Celeb-DF' of split '{cfg['split']}'"
+              f"\nPlease wait patiently...")
+        self.categories = ['original', 'fake']
+        self.root = cfg['root']
+        images_ids = self.__get_images_ids()
+        test_ids = self.__get_test_ids()
+        train_ids = [images_ids[0] - test_ids[0],
+                     images_ids[1] - test_ids[1],
+                     images_ids[2] - test_ids[2]]
+        self.images, self.targets = self.__get_images(
+            test_ids if cfg['split'] == "test" else train_ids, cfg['balance'])
+        assert len(self.images) == len(self.targets), "The number of images and targets not consistent."
+        print("Data from 'Celeb-DF' loaded.\n")
+        print(f"Dataset contains {len(self.images)} images.\n")
+
+    def __get_images_ids(self):
+        youtube_real = listdir(join(self.root, 'YouTube-real', 'images'))
+        celeb_real = listdir(join(self.root, 'Celeb-real', 'images'))
+        celeb_fake = listdir(join(self.root, 'Celeb-synthesis', 'images'))
+        return set(youtube_real), set(celeb_real), set(celeb_fake)
+
+    def __get_test_ids(self):
+        youtube_real = set()
+        celeb_real = set()
+        celeb_fake = set()
+        with open(join(self.root, "List_of_testing_videos.txt"), "r", encoding="utf-8") as f:
+            contents = f.readlines()
+            for line in contents:
+                name = line.split(" ")[-1]
+                number = name.split("/")[-1].split(".")[0]
+                if "YouTube-real" in name:
+                    youtube_real.add(number)
+                elif "Celeb-real" in name:
+                    celeb_real.add(number)
+                elif "Celeb-synthesis" in name:
+                    celeb_fake.add(number)
+                else:
+                    raise ValueError("'List_of_testing_videos.txt' file corrupted.")
+        return youtube_real, celeb_real, celeb_fake
+
+    def __get_images(self, ids, balance=False):
+        real = list()
+        fake = list()
+        # YouTube-real
+        for _ in ids[0]:
+            real.extend(glob(join(self.root, 'YouTube-real', 'images', _, '*.png')))
+        # Celeb-real
+        for _ in ids[1]:
+            real.extend(glob(join(self.root, 'Celeb-real', 'images', _, '*.png')))
+        # Celeb-synthesis
+        for _ in ids[2]:
+            fake.extend(glob(join(self.root, 'Celeb-synthesis', 'images', _, '*.png')))
+        print(f"Real: {len(real)}, Fake: {len(fake)}")
+        if balance:
+            fake = np.random.choice(fake, size=len(real), replace=False)
+            print(f"After Balance | Real: {len(real)}, Fake: {len(fake)}")
+        real_tgt = [0] * len(real)
+        fake_tgt = [1] * len(fake)
+        return [*real, *fake], [*real_tgt, *fake_tgt]
+
+
+if __name__ == '__main__':
+    import yaml
+
+    config_path = "../config/dataset/celeb_df.yml"
+    with open(config_path) as config_file:
+        config = yaml.load(config_file, Loader=yaml.FullLoader)
+    config = config["train_cfg"]
+    # config = config["test_cfg"]
+
+    def run_dataset():
+        dataset = CelebDF(config)
+        print(f"dataset: {len(dataset)}")
+        for i, _ in enumerate(dataset):
+            path, target = _
+            print(f"path: {path}, target: {target}")
+            if i >= 9:
+                break
+
+
+    def run_dataloader(display_samples=False):
+        from torch.utils import data
+        import matplotlib.pyplot as plt
+
+        dataset = CelebDF(config)
+        dataloader = data.DataLoader(dataset, batch_size=8, shuffle=True)
+        print(f"dataset: {len(dataset)}")
+        for i, _ in enumerate(dataloader):
+            path, targets = _
+            image = dataloader.dataset.load_item(path)
+            print(f"image: {image.shape}, target: {targets}")
+            if display_samples:
+                plt.figure()
+                img = image[0].permute([1, 2, 0]).numpy()
+                plt.imshow(img)
+                # plt.savefig("./img_" + str(i) + ".png")
+                plt.show()
+            if i >= 9:
+                break
+
+
+    ###########################
+    # run the functions below #
+    ###########################
+
+    # run_dataset()
+    run_dataloader(False)

dataset/dfdc.py

@@ -0,0 +1,124 @@
+import json
+from glob import glob
+from os.path import join
+from dataset import AbstractDataset
+
+SPLIT = ["train", "val", "test"]
+LABEL_MAP = {"REAL": 0, "FAKE": 1}
+
+
+class DFDC(AbstractDataset):
+    """
+    Deepfake Detection Challenge organized by Facebook
+    """
+
+    def __init__(self, cfg, seed=2022, transforms=None, transform=None, target_transform=None):
+        # pre-check
+        if cfg['split'] not in SPLIT:
+            raise ValueError(f"split should be one of {SPLIT}, but found {cfg['split']}.")
+        super(DFDC, self).__init__(cfg, seed, transforms, transform, target_transform)
+        print(f"Loading data from 'DFDC' of split '{cfg['split']}'"
+              f"\nPlease wait patiently...")
+        self.categories = ['original', 'fake']
+        self.root = cfg['root']
+        self.num_real = 0
+        self.num_fake = 0
+        if self.split == "test":
+            self.__load_test_data()
+        elif self.split == "train":
+            self.__load_train_data()
+        assert len(self.images) == len(self.targets), "Length of images and targets not the same!"
+        print(f"Data from 'DFDC' loaded.")
+        print(f"Real: {self.num_real}, Fake: {self.num_fake}.")
+        print(f"Dataset contains {len(self.images)} images\n")
+
+    def __load_test_data(self):
+        label_path = join(self.root, "test", "labels.csv")
+        with open(label_path, encoding="utf-8") as file:
+            content = file.readlines()
+        for _ in content:
+            if ".mp4" in _:
+                key = _.split(".")[0]
+                label = _.split(",")[1].strip()
+                label = int(label)
+                imgs = glob(join(self.root, "test", "images", key, "*.png"))
+                num = len(imgs)
+                self.images.extend(imgs)
+                self.targets.extend([label] * num)
+                if label == 0:
+                    self.num_real += num
+                elif label == 1:
+                    self.num_fake += num
+
+    def __load_train_data(self):
+        train_folds = glob(join(self.root, "dfdc_train_part_*"))
+        for fold in train_folds:
+            fold_imgs = list()
+            fold_tgts = list()
+            metadata_path = join(fold, "metadata.json")
+            try:
+                with open(metadata_path, "r", encoding="utf-8") as file:
+                    metadata = json.loads(file.readline())
+                for k, v in metadata.items():
+                    index = k.split(".")[0]
+                    label = LABEL_MAP[v["label"]]
+                    imgs = glob(join(fold, "images", index, "*.png"))
+                    fold_imgs.extend(imgs)
+                    fold_tgts.extend([label] * len(imgs))
+                    if label == 0:
+                        self.num_real += len(imgs)
+                    elif label == 1:
+                        self.num_fake += len(imgs)
+                self.images.extend(fold_imgs)
+                self.targets.extend(fold_tgts)
+            except FileNotFoundError:
+                continue
+
+
+if __name__ == '__main__':
+    import yaml
+
+    config_path = "../config/dataset/dfdc.yml"
+    with open(config_path) as config_file:
+        config = yaml.load(config_file, Loader=yaml.FullLoader)
+    config = config["train_cfg"]
+    # config = config["test_cfg"]
+
+
+    def run_dataset():
+        dataset = DFDC(config)
+        print(f"dataset: {len(dataset)}")
+        for i, _ in enumerate(dataset):
+            path, target = _
+            print(f"path: {path}, target: {target}")
+            if i >= 9:
+                break
+
+
+    def run_dataloader(display_samples=False):
+        from torch.utils import data
+        import matplotlib.pyplot as plt
+
+        dataset = DFDC(config)
+        dataloader = data.DataLoader(dataset, batch_size=8, shuffle=True)
+        print(f"dataset: {len(dataset)}")
+        for i, _ in enumerate(dataloader):
+            path, targets = _
+            image = dataloader.dataset.load_item(path)
+            print(f"image: {image.shape}, target: {targets}")
+            if display_samples:
+                plt.figure()
+                img = image[0].permute([1, 2, 0]).numpy()
+                plt.imshow(img)
+                # plt.savefig("./img_" + str(i) + ".png")
+                plt.show()
+            if i >= 9:
+                break
+
+
+    ###########################
+    # run the functions below #
+    ###########################
+
+    # run_dataset()
+    run_dataloader(False)

dataset/faceforensics.py

@@ -0,0 +1,107 @@
+import torch
+import numpy as np
+from os.path import join
+from dataset import AbstractDataset
+
+METHOD = ['all', 'Deepfakes', 'Face2Face', 'FaceSwap', 'NeuralTextures']
+SPLIT = ['train', 'val', 'test']
+COMP2NAME = {'c0': 'raw', 'c23': 'c23', 'c40': 'c40'}
+SOURCE_MAP = {'youtube': 2, 'Deepfakes': 3, 'Face2Face': 4, 'FaceSwap': 5, 'NeuralTextures': 6}
+
+
+class FaceForensics(AbstractDataset):
+    """
+    FaceForensics++ Dataset proposed in "FaceForensics++: Learning to Detect Manipulated Facial Images"
+    """
+
+    def __init__(self, cfg, seed=2022, transforms=None, transform=None, target_transform=None):
+        # pre-check
+        if cfg['split'] not in SPLIT:
+            raise ValueError(f"split should be one of {SPLIT}, "
+                             f"but found {cfg['split']}.")
+        if cfg['method'] not in METHOD:
+            raise ValueError(f"method should be one of {METHOD}, "
+                             f"but found {cfg['method']}.")
+        if cfg['compression'] not in COMP2NAME.keys():
+            raise ValueError(f"compression should be one of {COMP2NAME.keys()}, "
+                             f"but found {cfg['compression']}.")
+        super(FaceForensics, self).__init__(
+            cfg, seed, transforms, transform, target_transform)
+        print(f"Loading data from 'FF++ {cfg['method']}' of split '{cfg['split']}' "
+              f"and compression '{cfg['compression']}'\nPlease wait patiently...")
+
+        self.categories = ['original', 'fake']
+        # load the path of dataset images
+        indices = join(self.root, cfg['split'] + "_" + cfg['compression'] + ".pickle")
+        indices = torch.load(indices)
+        if cfg['method'] == "all":
+            # full dataset
+            self.images = [join(cfg['root'], _[0]) for _ in indices]
+            self.targets = [_[1] for _ in indices]
+        else:
+            # specific manipulated method
+            self.images = list()
+            self.targets = list()
+            nums = 0
+            for _ in indices:
+                if cfg['method'] in _[0]:
+                    self.images.append(join(cfg['root'], _[0]))
+                    self.targets.append(_[1])
+                nums = len(self.targets)
+            ori = list()
+            for _ in indices:
+                if "original_sequences" in _[0]:
+                    ori.append(join(cfg['root'], _[0]))
+            choices = np.random.choice(ori, size=nums, replace=False)
+            self.images.extend(choices)
+            self.targets.extend([0] * nums)
+        print("Data from 'FF++' loaded.\n")
+        print(f"Dataset contains {len(self.images)} images.\n")
+
+
+if __name__ == '__main__':
+    import yaml
+
+    config_path = "../config/dataset/faceforensics.yml"
+    with open(config_path) as config_file:
+        config = yaml.load(config_file, Loader=yaml.FullLoader)
+    config = config["train_cfg"]
+    # config = config["test_cfg"]
+
+    def run_dataset():
+        dataset = FaceForensics(config)
+        print(f"dataset: {len(dataset)}")
+        for i, _ in enumerate(dataset):
+            path, target = _
+            print(f"path: {path}, target: {target}")
+            if i >= 9:
+                break
+
+
+    def run_dataloader(display_samples=False):
+        from torch.utils import data
+        import matplotlib.pyplot as plt
+
+        dataset = FaceForensics(config)
+        dataloader = data.DataLoader(dataset, batch_size=8, shuffle=True)
+        print(f"dataset: {len(dataset)}")
+        for i, _ in enumerate(dataloader):
+            path, targets = _
+            image = dataloader.dataset.load_item(path)
+            print(f"image: {image.shape}, target: {targets}")
+            if display_samples:
+                plt.figure()
+                img = image[0].permute([1, 2, 0]).numpy()
+                plt.imshow(img)
+                # plt.savefig("./img_" + str(i) + ".png")
+                plt.show()
+            if i >= 9:
+                break
+
+
+    ###########################
+    # run the functions below #
+    ###########################
+
+    # run_dataset()
+    run_dataloader(False)

dataset/wild_deepfake.py

@@ -0,0 +1,100 @@
+import torch
+import numpy as np
+from os.path import join
+from dataset import AbstractDataset
+
+SPLITS = ["train", "test"]
+
+
+class WildDeepfake(AbstractDataset):
+    """
+    Wild Deepfake Dataset proposed in "WildDeepfake: A Challenging Real-World Dataset for Deepfake Detection"
+    """
+
+    def __init__(self, cfg, seed=2022, transforms=None, transform=None, target_transform=None):
+        # pre-check
+        if cfg['split'] not in SPLITS:
+            raise ValueError(f"split should be one of {SPLITS}, but found {cfg['split']}.")
+        super(WildDeepfake, self).__init__(cfg, seed, transforms, transform, target_transform)
+        print(f"Loading data from 'WildDeepfake' of split '{cfg['split']}'"
+              f"\nPlease wait patiently...")
+        self.categories = ['original', 'fake']
+        self.root = cfg['root']
+        self.num_train = cfg.get('num_image_train', None)
+        self.num_test = cfg.get('num_image_test', None)
+        self.images, self.targets = self.__get_images()
+        print(f"Data from 'WildDeepfake' loaded.")
+        print(f"Dataset contains {len(self.images)} images.\n")
+
+    def __get_images(self):
+        if self.split == 'train':
+            num = self.num_train
+        elif self.split == 'test':
+            num = self.num_test
+        else:
+            num = None
+        real_images = torch.load(join(self.root, self.split, "real.pickle"))
+        if num is not None:
+            real_images = np.random.choice(real_images, num // 3, replace=False)
+        real_tgts = [torch.tensor(0)] * len(real_images)
+        print(f"real: {len(real_tgts)}")
+        fake_images = torch.load(join(self.root, self.split, "fake.pickle"))
+        if num is not None:
+            fake_images = np.random.choice(fake_images, num - num // 3, replace=False)
+        fake_tgts = [torch.tensor(1)] * len(fake_images)
+        print(f"fake: {len(fake_tgts)}")
+        return real_images + fake_images, real_tgts + fake_tgts
+
+    def __getitem__(self, index):
+        path = join(self.root, self.split, self.images[index])
+        tgt = self.targets[index]
+        return path, tgt
+
+
+if __name__ == '__main__':
+    import yaml
+
+    config_path = "../config/dataset/wilddeepfake.yml"
+    with open(config_path) as config_file:
+        config = yaml.load(config_file, Loader=yaml.FullLoader)
+    config = config["train_cfg"]
+    # config = config["test_cfg"]
+
+
+    def run_dataset():
+        dataset = WildDeepfake(config)
+        print(f"dataset: {len(dataset)}")
+        for i, _ in enumerate(dataset):
+            path, target = _
+            print(f"path: {path}, target: {target}")
+            if i >= 9:
+                break
+
+
+    def run_dataloader(display_samples=False):
+        from torch.utils import data
+        import matplotlib.pyplot as plt
+
+        dataset = WildDeepfake(config)
+        dataloader = data.DataLoader(dataset, batch_size=8, shuffle=True)
+        print(f"dataset: {len(dataset)}")
+        for i, _ in enumerate(dataloader):
+            path, targets = _
+            image = dataloader.dataset.load_item(path)
+            print(f"image: {image.shape}, target: {targets}")
+            if display_samples:
+                plt.figure()
+                img = image[0].permute([1, 2, 0]).numpy()
+                plt.imshow(img)
+                # plt.savefig("./img_" + str(i) + ".png")
+                plt.show()
+            if i >= 9:
+                break
+
+
+    ###########################
+    # run the functions below #
+    ###########################
+
+    # run_dataset()
+    run_dataloader(False)

extract_video.py

@@ -0,0 +1,233 @@
+import os
+from os.path import join
+import argparse
+import numpy as np
+import cv2
+import torch
+from tqdm import tqdm
+
+from data import cfg_mnet, cfg_re50
+from layers.functions.prior_box import PriorBox
+from utils.nms.py_cpu_nms import py_cpu_nms
+from models.retinaface import RetinaFace
+from utils.box_utils import decode
+
+np.random.seed(0)
+
+
+def check_keys(model, pretrained_state_dict):
+    ckpt_keys = set(pretrained_state_dict.keys())
+    model_keys = set(model.state_dict().keys())
+    used_pretrained_keys = model_keys & ckpt_keys
+    unused_pretrained_keys = ckpt_keys - model_keys
+    missing_keys = model_keys - ckpt_keys
+    print('Missing keys:{}'.format(len(missing_keys)))
+    print('Unused checkpoint keys:{}'.format(len(unused_pretrained_keys)))
+    print('Used keys:{}'.format(len(used_pretrained_keys)))
+    assert len(used_pretrained_keys) > 0, 'load NONE from pretrained checkpoint'
+    return True
+
+
+def remove_prefix(state_dict, prefix):
+    ''' Old style model is stored with all names of parameters sharing common prefix 'module.' '''
+    print('remove prefix \'{}\''.format(prefix))
+
+    def f(x): return x.split(prefix, 1)[-1] if x.startswith(prefix) else x
+
+    return {f(key): value for key, value in state_dict.items()}
+
+
+def load_model(model, pretrained_path, load_to_cpu):
+    print('Loading pretrained model from {}'.format(pretrained_path))
+    if load_to_cpu:
+        pretrained_dict = torch.load(
+            pretrained_path, map_location=lambda storage, loc: storage)
+    else:
+        pretrained_dict = torch.load(
+            pretrained_path, map_location=lambda storage, loc: storage.cuda(device))
+    if "state_dict" in pretrained_dict.keys():
+        pretrained_dict = remove_prefix(
+            pretrained_dict['state_dict'], 'module.')
+    else:
+        pretrained_dict = remove_prefix(pretrained_dict, 'module.')
+    check_keys(model, pretrained_dict)
+    model.load_state_dict(pretrained_dict, strict=False)
+    model.to(device)
+    return model
+
+
+def detect(img_list, output_path, resize=1):
+    os.makedirs(output_path, exist_ok=True)
+    im_height, im_width, _ = img_list[0].shape
+    scale = torch.Tensor([im_width, im_height, im_width, im_height])
+    img_x = torch.stack(img_list, dim=0).permute([0, 3, 1, 2])
+    scale = scale.to(device)
+
+    # batch size
+    batch_size = args.bs
+    # forward times
+    f_times = img_x.shape[0] // batch_size
+    if img_x.shape[0] % batch_size != 0:
+        f_times += 1
+    locs_list = list()
+    confs_list = list()
+    for _ in range(f_times):
+        if _ != f_times - 1:
+            batch_img_x = img_x[_ * batch_size:(_ + 1) * batch_size]
+        else:
+            batch_img_x = img_x[_ * batch_size:]  # last batch
+        batch_img_x = batch_img_x.to(device).float()
+        l, c, _ = net(batch_img_x)
+        locs_list.append(l)
+        confs_list.append(c)
+    locs = torch.cat(locs_list, dim=0)
+    confs = torch.cat(confs_list, dim=0)
+
+    priorbox = PriorBox(cfg, image_size=(im_height, im_width))
+    priors = priorbox.forward()
+    priors = priors.to(device)
+    prior_data = priors.data
+
+    img_cpu = img_x.permute([0, 2, 3, 1]).cpu().numpy()
+    i = 0
+    for img, loc, conf in zip(img_cpu, locs, confs):
+        boxes = decode(loc.data, prior_data, cfg['variance'])
+        boxes = boxes * scale / resize
+        boxes = boxes.cpu().numpy()
+        scores = conf.data.cpu().numpy()[:, 1]
+
+        # ignore low scores
+        inds = np.where(scores > args.confidence_threshold)[0]
+        boxes = boxes[inds]
+        scores = scores[inds]
+
+        # keep top-K before NMS
+        order = scores.argsort()[::-1][:args.top_k]
+        boxes = boxes[order]
+        scores = scores[order]
+
+        # do NMS
+        dets = np.hstack((boxes, scores[:, np.newaxis])).astype(
+            np.float32, copy=False)
+        keep = py_cpu_nms(dets, args.nms_threshold)
+        # keep = nms(dets, args.nms_threshold,force_cpu=args.cpu)
+        dets = dets[keep, :]
+
+        # keep top-K faster NMS
+        dets = dets[:args.keep_top_k, :]
+
+        if len(dets) == 0:
+            continue
+        det = list(map(int, dets[0]))
+        x, y, size_bb_x, size_bb_y = get_boundingbox(det, img.shape[1], img.shape[0])
+        cropped_img = img[y:y + size_bb_y, x:x + size_bb_x, :] + (104, 117, 123)
+        cv2.imwrite(join(output_path, '{:04d}.png'.format(i)), cropped_img)
+        i += 1
+    pass
+
+
+def extract_frames(data_path, interval=1):
+    """Method to extract frames"""
+    if data_path.split('.')[-1] == "mp4":
+        reader = cv2.VideoCapture(data_path)
+        frame_num = 0
+        frames = list()
+
+        while reader.isOpened():
+            success, image = reader.read()
+            if not success:
+                break
+            cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
+            image = torch.tensor(image) - torch.tensor([104, 117, 123])
+            if frame_num % interval == 0:
+                frames.append(image)
+            frame_num += 1
+            if len(frames) > args.max_frames:
+                break
+        reader.release()
+        if len(frames) > args.max_frames:
+            samples = np.random.choice(
+                np.arange(0, len(frames)), size=args.max_frames, replace=False)
+            return [frames[_] for _ in samples]
+        return frames
+    else:
+        image = cv2.imread(data_path)
+        cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
+        image = torch.tensor(image) - torch.tensor([104, 117, 123])
+        return [image]
+
+
+def get_boundingbox(bbox, width, height, scale=1.8, minsize=None):
+    x1 = bbox[0]
+    y1 = bbox[1]
+    x2 = bbox[2]
+    y2 = bbox[3]
+    size_bb_x = int((x2 - x1) * scale)
+    size_bb_y = int((y2 - y1) * scale)
+    if minsize:
+        if size_bb_x < minsize:
+            size_bb_x = minsize
+        if size_bb_y < minsize:
+            size_bb_y = minsize
+    center_x, center_y = (x1 + x2) // 2, (y1 + y2) // 2
+
+    # Check for out of bounds, x-y top left corner
+    x1 = max(int(center_x - size_bb_x // 2), 0)
+    y1 = max(int(center_y - size_bb_y // 2), 0)
+    # Check for too big bb size for given x, y
+    size_bb_x = min(width - x1, size_bb_x)
+    size_bb_y = min(height - y1, size_bb_y)
+    return x1, y1, size_bb_x, size_bb_y
+
+
+def extract_method_videos(data_path, interval):
+    video = data_path.split('/')[-1]
+    result_path = '/'.join(data_path.split('/')[:-1])
+    images_path = join(result_path, 'images')
+
+    image_folder = video.split('.')[0]
+    try:
+        print(data_path)
+        image_list = extract_frames(data_path, interval)
+        detect(image_list, join(images_path, image_folder))
+    except Exception as ex:
+        f = open("failure.txt", "a", encoding="utf-8")
+        f.writelines(image_folder +
+                        f"  Exception for {image_folder}: {ex}\n")
+        f.close()
+
+
+if __name__ == '__main__':
+    p = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter
+    )
+    p.add_argument('--data_path', '-p', type=str, help='path to the data')
+    p.add_argument('--confidence_threshold', default=0.05,
+                   type=float, help='confidence threshold')
+    p.add_argument('--top_k', default=5, type=int, help='top_k')
+    p.add_argument('--nms_threshold', default=0.4,
+                   type=float, help='nms threshold')
+    p.add_argument('--keep_top_k', default=1, type=int, help='keep_top_k')
+    p.add_argument('--bs', default=32, type=int, help='batch size')
+    p.add_argument('--frame_interval', '-fi', default=1, type=int, help='frame interval')
+    p.add_argument('--device', "-d", default="cuda:0", type=str, help='device')
+    p.add_argument('--max_frames', default=100, type=int, help='maximum frames per video')
+
+    args = p.parse_args()
+
+    torch.set_grad_enabled(False)
+    # use resnet-50
+    cfg = cfg_re50
+    pretrained_weights = './weights/Resnet50_Final.pth'
+
+    torch.backends.cudnn.benchmark = True
+    device = torch.device(args.device)
+    print(device)
+
+    # net and model
+    net = RetinaFace(cfg=cfg, phase='test')
+    net = load_model(net, pretrained_weights, args.device)
+    net.eval()
+    print('Finished loading model!')
+
+    extract_method_videos(args.data_path, args.frame_interval)

inference.py

@@ -0,0 +1,142 @@
+import cv2
+import torch
+import random
+import argparse
+from glob import glob
+from os.path import join
+from model.network import Recce
+from model.common import freeze_weights
+from albumentations import Compose, Normalize, Resize
+from albumentations.pytorch.transforms import ToTensorV2
+
+import os
+
+os.environ['KMP_DUPLICATE_LIB_OK']='True'
+
+# fix random seed
+seed = 0
+random.seed(seed)
+torch.manual_seed(seed)
+torch.cuda.manual_seed(seed)
+torch.cuda.manual_seed_all(seed)
+
+parser = argparse.ArgumentParser(description="This code helps you use a trained model to "
+                                             "do inference.")
+parser.add_argument("--weight", "-w",
+                    type=str,
+                    default=None,
+                    help="Specify the path to the model weight (the state dict file). "
+                         "Do not use this argument when '--bin' is set.")
+parser.add_argument("--bin", "-b",
+                    type=str,
+                    default=None,
+                    help="Specify the path to the model bin which ends up with '.bin' "
+                         "(which is generated by the trainer of this project). "
+                         "Do not use this argument when '--weight' is set.")
+parser.add_argument("--image", "-i",
+                    type=str,
+                    default=None,
+                    help="Specify the path to the input image. "
+                         "Do not use this argument when '--image_folder' is set.")
+parser.add_argument("--image_folder", "-f",
+                    type=str,
+                    default=None,
+                    help="Specify the directory to evaluate all the images. "
+                         "Do not use this argument when '--image' is set.")
+
+parser.add_argument('--device', '-d', type=str,
+                    default="cpu",
+                    help="Specify the device to load the model. Default: 'cpu'.")
+parser.add_argument('--image_size', '-s', type=int,
+                    default=299,
+                    help="Specify the spatial size of the input image(s). Default: 299.")
+parser.add_argument('--visualize', '-v', action="store_true",
+                    default=False, help='Visualize images.')
+
+
+def preprocess(file_path):
+    img = cv2.imread(file_path)
+    img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
+    compose = Compose([Resize(height=args.image_size, width=args.image_size),
+                       Normalize(mean=[0.5] * 3, std=[0.5] * 3),
+                       ToTensorV2()])
+    img = compose(image=img)['image'].unsqueeze(0)
+    return img
+
+
+def prepare_data():
+    paths = list()
+    images = list()
+    # check the console arguments
+    if args.image and args.image_folder:
+        raise ValueError("Only one of '--image' or '--image_folder' can be set.")
+    elif args.image:
+        images.append(preprocess(args.image))
+        paths.append(args.image)
+    elif args.image_folder:
+        image_folder = '.'.join(args.image_folder.split('.')[:-1])
+        image_paths = glob(image_folder + "/*.jpg")
+        image_paths.extend(glob(image_folder + "/*.png"))
+        for _ in image_paths:
+            images.append(preprocess(_))
+            paths.append(_)
+    else:
+        raise ValueError("Neither of '--image' nor '--image_folder' is set. Please specify either "
+                         "one of these two arguments to load input image(s) properly.")
+    return paths, images
+
+
+def inference(model, images, paths, device):
+    mean_pred = 0
+    for img, pt in zip(images, paths):
+        img = img.to(device)
+        prediction = model(img)
+        prediction = torch.sigmoid(prediction).cpu()
+        fake = True if prediction >= 0.5 else False
+
+        mean_pred += prediction.item()
+
+        print(f"path: {pt} \t\t| fake probability: {prediction.item():.4f} \t| "
+              f"prediction: {'fake' if fake else 'real'}")
+        if args.visualize:
+            cvimg = cv2.imread(pt)
+            cvimg = cv2.putText(cvimg, f'p: {prediction.item():.2f}, ' + f"{'fake' if fake else 'real'}",
+                                (5, 50), cv2.FONT_HERSHEY_SIMPLEX, 0.5,
+                                (0, 0, 255) if fake else (255, 0, 0), 2)
+            cv2.imshow("image", cvimg)
+            cv2.waitKey(0)
+            cv2.destroyWindow("image")
+    mean_pred = mean_pred / len(images)
+    return mean_pred
+
+
+def main():
+    print("Arguments:\n", args, end="\n\n")
+    # set device
+    device = torch.device(args.device)
+    # load model
+    model = eval("Recce")(num_classes=1)
+    # check the console arguments
+    if args.weight and args.bin:
+        raise ValueError("Only one of '--weight' or '--bin' can be set.")
+    elif args.weight:
+        weights = torch.load(args.weight, map_location="cpu")
+    elif args.bin:
+        weights = torch.load(args.bin, map_location="cpu")["model"]
+    else:
+        raise ValueError("Neither of '--weight' nor '--bin' is set. Please specify either "
+                         "one of these two arguments to load model's weight properly.")
+    model.load_state_dict(weights)
+    model = model.to(device)
+    freeze_weights(model)
+    model.eval()
+
+    paths, images = prepare_data()
+    print("Inference:")
+    mean_pred = inference(model, images=images, paths=paths, device=device)
+    print("Mean prediction:", mean_pred)
+
+
+if __name__ == '__main__':
+    args = parser.parse_args()
+    main()

layers/__init__.py

@@ -0,0 +1,2 @@
+from .functions import *
+from .modules import *

layers/functions/prior_box.py

@@ -0,0 +1,34 @@
+import torch
+from itertools import product as product
+import numpy as np
+from math import ceil
+
+
+class PriorBox(object):
+    def __init__(self, cfg, image_size=None, phase='train'):
+        super(PriorBox, self).__init__()
+        self.min_sizes = cfg['min_sizes']
+        self.steps = cfg['steps']
+        self.clip = cfg['clip']
+        self.image_size = image_size
+        self.feature_maps = [[ceil(self.image_size[0]/step), ceil(self.image_size[1]/step)] for step in self.steps]
+        self.name = "s"
+
+    def forward(self):
+        anchors = []
+        for k, f in enumerate(self.feature_maps):
+            min_sizes = self.min_sizes[k]
+            for i, j in product(range(f[0]), range(f[1])):
+                for min_size in min_sizes:
+                    s_kx = min_size / self.image_size[1]
+                    s_ky = min_size / self.image_size[0]
+                    dense_cx = [x * self.steps[k] / self.image_size[1] for x in [j + 0.5]]
+                    dense_cy = [y * self.steps[k] / self.image_size[0] for y in [i + 0.5]]
+                    for cy, cx in product(dense_cy, dense_cx):
+                        anchors += [cx, cy, s_kx, s_ky]
+
+        # back to torch land
+        output = torch.Tensor(anchors).view(-1, 4)
+        if self.clip:
+            output.clamp_(max=1, min=0)
+        return output

layers/modules/__init__.py

@@ -0,0 +1,3 @@
+from .multibox_loss import MultiBoxLoss
+
+__all__ = ['MultiBoxLoss']

layers/modules/multibox_loss.py

@@ -0,0 +1,125 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from torch.autograd import Variable
+from utils.box_utils import match, log_sum_exp
+from data import cfg_mnet
+GPU = cfg_mnet['gpu_train']
+
+class MultiBoxLoss(nn.Module):
+    """SSD Weighted Loss Function
+    Compute Targets:
+        1) Produce Confidence Target Indices by matching  ground truth boxes
+           with (default) 'priorboxes' that have jaccard index > threshold parameter
+           (default threshold: 0.5).
+        2) Produce localization target by 'encoding' variance into offsets of ground
+           truth boxes and their matched  'priorboxes'.
+        3) Hard negative mining to filter the excessive number of negative examples
+           that comes with using a large number of default bounding boxes.
+           (default negative:positive ratio 3:1)
+    Objective Loss:
+        L(x,c,l,g) = (Lconf(x, c) + αLloc(x,l,g)) / N
+        Where, Lconf is the CrossEntropy Loss and Lloc is the SmoothL1 Loss
+        weighted by α which is set to 1 by cross val.
+        Args:
+            c: class confidences,
+            l: predicted boxes,
+            g: ground truth boxes
+            N: number of matched default boxes
+        See: https://arxiv.org/pdf/1512.02325.pdf for more details.
+    """
+
+    def __init__(self, num_classes, overlap_thresh, prior_for_matching, bkg_label, neg_mining, neg_pos, neg_overlap, encode_target):
+        super(MultiBoxLoss, self).__init__()
+        self.num_classes = num_classes
+        self.threshold = overlap_thresh
+        self.background_label = bkg_label
+        self.encode_target = encode_target
+        self.use_prior_for_matching = prior_for_matching
+        self.do_neg_mining = neg_mining
+        self.negpos_ratio = neg_pos
+        self.neg_overlap = neg_overlap
+        self.variance = [0.1, 0.2]
+
+    def forward(self, predictions, priors, targets):
+        """Multibox Loss
+        Args:
+            predictions (tuple): A tuple containing loc preds, conf preds,
+            and prior boxes from SSD net.
+                conf shape: torch.size(batch_size,num_priors,num_classes)
+                loc shape: torch.size(batch_size,num_priors,4)
+                priors shape: torch.size(num_priors,4)
+
+            ground_truth (tensor): Ground truth boxes and labels for a batch,
+                shape: [batch_size,num_objs,5] (last idx is the label).
+        """
+
+        loc_data, conf_data, landm_data = predictions
+        priors = priors
+        num = loc_data.size(0)
+        num_priors = (priors.size(0))
+
+        # match priors (default boxes) and ground truth boxes
+        loc_t = torch.Tensor(num, num_priors, 4)
+        landm_t = torch.Tensor(num, num_priors, 10)
+        conf_t = torch.LongTensor(num, num_priors)
+        for idx in range(num):
+            truths = targets[idx][:, :4].data
+            labels = targets[idx][:, -1].data
+            landms = targets[idx][:, 4:14].data
+            defaults = priors.data
+            match(self.threshold, truths, defaults, self.variance, labels, landms, loc_t, conf_t, landm_t, idx)
+        if GPU:
+            loc_t = loc_t.cuda()
+            conf_t = conf_t.cuda()
+            landm_t = landm_t.cuda()
+
+        zeros = torch.tensor(0).cuda()
+        # landm Loss (Smooth L1)
+        # Shape: [batch,num_priors,10]
+        pos1 = conf_t > zeros
+        num_pos_landm = pos1.long().sum(1, keepdim=True)
+        N1 = max(num_pos_landm.data.sum().float(), 1)
+        pos_idx1 = pos1.unsqueeze(pos1.dim()).expand_as(landm_data)
+        landm_p = landm_data[pos_idx1].view(-1, 10)
+        landm_t = landm_t[pos_idx1].view(-1, 10)
+        loss_landm = F.smooth_l1_loss(landm_p, landm_t, reduction='sum')
+
+
+        pos = conf_t != zeros
+        conf_t[pos] = 1
+
+        # Localization Loss (Smooth L1)
+        # Shape: [batch,num_priors,4]
+        pos_idx = pos.unsqueeze(pos.dim()).expand_as(loc_data)
+        loc_p = loc_data[pos_idx].view(-1, 4)
+        loc_t = loc_t[pos_idx].view(-1, 4)
+        loss_l = F.smooth_l1_loss(loc_p, loc_t, reduction='sum')
+
+        # Compute max conf across batch for hard negative mining
+        batch_conf = conf_data.view(-1, self.num_classes)
+        loss_c = log_sum_exp(batch_conf) - batch_conf.gather(1, conf_t.view(-1, 1))
+
+        # Hard Negative Mining
+        loss_c[pos.view(-1, 1)] = 0 # filter out pos boxes for now
+        loss_c = loss_c.view(num, -1)
+        _, loss_idx = loss_c.sort(1, descending=True)
+        _, idx_rank = loss_idx.sort(1)
+        num_pos = pos.long().sum(1, keepdim=True)
+        num_neg = torch.clamp(self.negpos_ratio*num_pos, max=pos.size(1)-1)
+        neg = idx_rank < num_neg.expand_as(idx_rank)
+
+        # Confidence Loss Including Positive and Negative Examples
+        pos_idx = pos.unsqueeze(2).expand_as(conf_data)
+        neg_idx = neg.unsqueeze(2).expand_as(conf_data)
+        conf_p = conf_data[(pos_idx+neg_idx).gt(0)].view(-1,self.num_classes)
+        targets_weighted = conf_t[(pos+neg).gt(0)]
+        loss_c = F.cross_entropy(conf_p, targets_weighted, reduction='sum')
+
+        # Sum of losses: L(x,c,l,g) = (Lconf(x, c) + αLloc(x,l,g)) / N
+        N = max(num_pos.data.sum().float(), 1)
+        loss_l /= N
+        loss_c /= N
+        loss_landm /= N1
+
+        return loss_l, loss_c, loss_landm

loss/__init__.py

@@ -0,0 +1,12 @@
+import torch.nn as nn
+
+
+def get_loss(name="cross_entropy", device="cuda:0"):
+    print(f"Using loss: '{LOSSES[name]}'")
+    return LOSSES[name].to(device)
+
+
+LOSSES = {
+    "binary_ce": nn.BCEWithLogitsLoss(),
+    "cross_entropy": nn.CrossEntropyLoss()
+}

model/__init__.py

@@ -0,0 +1,12 @@
+from .network import *
+from .common import *
+
+MODELS = {
+    "Recce": Recce
+}
+
+
+def load_model(name="Recce"):
+    assert name in MODELS.keys(), f"Model name can only be one of {MODELS.keys()}."
+    print(f"Using model: '{name}'")
+    return MODELS[name]

model/common.py

@@ -0,0 +1,200 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+
+def freeze_weights(module):
+    for param in module.parameters():
+        param.requires_grad = False
+
+
+def l1_regularize(module):
+    reg_loss = 0.
+    for key, param in module.reg_params.items():
+        if "weight" in key and param.requires_grad:
+            reg_loss += torch.sum(torch.abs(param))
+    return reg_loss
+
+
+class SeparableConv2d(nn.Module):
+    def __init__(self, in_channels, out_channels, kernel_size=1, stride=1, padding=0, dilation=1, bias=False):
+        super(SeparableConv2d, self).__init__()
+
+        self.conv1 = nn.Conv2d(in_channels, in_channels, kernel_size, stride, padding, dilation,
+                               groups=in_channels, bias=bias)
+        self.pointwise = nn.Conv2d(in_channels, out_channels, 1, 1, 0, 1, 1, bias=bias)
+
+    def forward(self, x):
+        x = self.conv1(x)
+        x = self.pointwise(x)
+        return x
+
+
+class Block(nn.Module):
+    def __init__(self, in_channels, out_channels, reps, strides=1,
+                 start_with_relu=True, grow_first=True, with_bn=True):
+        super(Block, self).__init__()
+
+        self.with_bn = with_bn
+
+        if out_channels != in_channels or strides != 1:
+            self.skip = nn.Conv2d(in_channels, out_channels, 1, stride=strides, bias=False)
+            if with_bn:
+                self.skipbn = nn.BatchNorm2d(out_channels)
+        else:
+            self.skip = None
+
+        rep = []
+        for i in range(reps):
+            if grow_first:
+                inc = in_channels if i == 0 else out_channels
+                outc = out_channels
+            else:
+                inc = in_channels
+                outc = in_channels if i < (reps - 1) else out_channels
+            rep.append(nn.ReLU(inplace=True))
+            rep.append(SeparableConv2d(inc, outc, 3, stride=1, padding=1))
+            if with_bn:
+                rep.append(nn.BatchNorm2d(outc))
+
+        if not start_with_relu:
+            rep = rep[1:]
+        else:
+            rep[0] = nn.ReLU(inplace=False)
+
+        if strides != 1:
+            rep.append(nn.MaxPool2d(3, strides, 1))
+        self.rep = nn.Sequential(*rep)
+
+    def forward(self, inp):
+        x = self.rep(inp)
+
+        if self.skip is not None:
+            skip = self.skip(inp)
+            if self.with_bn:
+                skip = self.skipbn(skip)
+        else:
+            skip = inp
+
+        x += skip
+        return x
+
+
+class GraphReasoning(nn.Module):
+    """ Graph Reasoning Module for information aggregation. """
+
+    def __init__(self, va_in, va_out, vb_in, vb_out, vc_in, vc_out, spatial_ratio, drop_rate):
+        super(GraphReasoning, self).__init__()
+        self.ratio = spatial_ratio
+        self.va_embedding = nn.Sequential(
+            nn.Conv2d(va_in, va_out, 1, bias=False),
+            nn.ReLU(True),
+            nn.Conv2d(va_out, va_out, 1, bias=False),
+        )
+        self.va_gated_b = nn.Sequential(
+            nn.Conv2d(va_in, va_out, 1, bias=False),
+            nn.Sigmoid()
+        )
+        self.va_gated_c = nn.Sequential(
+            nn.Conv2d(va_in, va_out, 1, bias=False),
+            nn.Sigmoid()
+        )
+        self.vb_embedding = nn.Sequential(
+            nn.Linear(vb_in, vb_out, bias=False),
+            nn.ReLU(True),
+            nn.Linear(vb_out, vb_out, bias=False),
+        )
+        self.vc_embedding = nn.Sequential(
+            nn.Linear(vc_in, vc_out, bias=False),
+            nn.ReLU(True),
+            nn.Linear(vc_out, vc_out, bias=False),
+        )
+        self.unfold_b = nn.Unfold(kernel_size=spatial_ratio[0], stride=spatial_ratio[0])
+        self.unfold_c = nn.Unfold(kernel_size=spatial_ratio[1], stride=spatial_ratio[1])
+        self.reweight_ab = nn.Sequential(
+            nn.Linear(va_out + vb_out, 1, bias=False),
+            nn.ReLU(True),
+            nn.Softmax(dim=1)
+        )
+        self.reweight_ac = nn.Sequential(
+            nn.Linear(va_out + vc_out, 1, bias=False),
+            nn.ReLU(True),
+            nn.Softmax(dim=1)
+        )
+        self.reproject = nn.Sequential(
+            nn.Conv2d(va_out + vb_out + vc_out, va_in, kernel_size=1, bias=False),
+            nn.ReLU(True),
+            nn.Conv2d(va_in, va_in, kernel_size=1, bias=False),
+            nn.Dropout(drop_rate) if drop_rate is not None else nn.Identity(),
+        )
+
+    def forward(self, vert_a, vert_b, vert_c):
+        emb_vert_a = self.va_embedding(vert_a)
+        emb_vert_a = emb_vert_a.reshape([emb_vert_a.shape[0], emb_vert_a.shape[1], -1])
+
+        gate_vert_b = 1 - self.va_gated_b(vert_a)
+        gate_vert_b = gate_vert_b.reshape(*emb_vert_a.shape)
+        gate_vert_c = 1 - self.va_gated_c(vert_a)
+        gate_vert_c = gate_vert_c.reshape(*emb_vert_a.shape)
+
+        vert_b = self.unfold_b(vert_b).reshape(
+            [vert_b.shape[0], vert_b.shape[1], self.ratio[0] * self.ratio[0], -1])
+        vert_b = vert_b.permute([0, 2, 3, 1])
+        emb_vert_b = self.vb_embedding(vert_b)
+
+        vert_c = self.unfold_c(vert_c).reshape(
+            [vert_c.shape[0], vert_c.shape[1], self.ratio[1] * self.ratio[1], -1])
+        vert_c = vert_c.permute([0, 2, 3, 1])
+        emb_vert_c = self.vc_embedding(vert_c)
+
+        agg_vb = list()
+        agg_vc = list()
+        for j in range(emb_vert_a.shape[-1]):
+            # ab propagating
+            emb_v_a = torch.stack([emb_vert_a[:, :, j]] * (self.ratio[0] ** 2), dim=1)
+            emb_v_b = emb_vert_b[:, :, j, :]
+            emb_v_ab = torch.cat([emb_v_a, emb_v_b], dim=-1)
+            w = self.reweight_ab(emb_v_ab)
+            agg_vb.append(torch.bmm(emb_v_b.transpose(1, 2), w).squeeze() * gate_vert_b[:, :, j])
+
+            # ac propagating
+            emb_v_a = torch.stack([emb_vert_a[:, :, j]] * (self.ratio[1] ** 2), dim=1)
+            emb_v_c = emb_vert_c[:, :, j, :]
+            emb_v_ac = torch.cat([emb_v_a, emb_v_c], dim=-1)
+            w = self.reweight_ac(emb_v_ac)
+            agg_vc.append(torch.bmm(emb_v_c.transpose(1, 2), w).squeeze() * gate_vert_c[:, :, j])
+
+        agg_vert_b = torch.stack(agg_vb, dim=-1)
+        agg_vert_c = torch.stack(agg_vc, dim=-1)
+        agg_vert_bc = torch.cat([agg_vert_b, agg_vert_c], dim=1)
+        agg_vert_abc = torch.cat([agg_vert_bc, emb_vert_a], dim=1)
+        agg_vert_abc = torch.sigmoid(agg_vert_abc)
+        agg_vert_abc = agg_vert_abc.reshape(vert_a.shape[0], -1, vert_a.shape[2], vert_a.shape[3])
+        return self.reproject(agg_vert_abc)
+
+
+class GuidedAttention(nn.Module):
+    """ Reconstruction Guided Attention. """
+
+    def __init__(self, depth=728, drop_rate=0.2):
+        super(GuidedAttention, self).__init__()
+        self.depth = depth
+        self.gated = nn.Sequential(
+            nn.Conv2d(3, 3, kernel_size=3, stride=1, padding=1, bias=False),
+            nn.ReLU(True),
+            nn.Conv2d(3, 1, 1, bias=False),
+            nn.Sigmoid()
+        )
+        self.h = nn.Sequential(
+            nn.Conv2d(depth, depth, 1, 1, bias=False),
+            nn.BatchNorm2d(depth),
+            nn.ReLU(True),
+        )
+        self.dropout = nn.Dropout(drop_rate)
+
+    def forward(self, x, pred_x, embedding):
+        residual_full = torch.abs(x - pred_x)
+        residual_x = F.interpolate(residual_full, size=embedding.shape[-2:],
+                                   mode='bilinear', align_corners=True)
+        res_map = self.gated(residual_x)
+        return res_map * self.h(embedding) + self.dropout(embedding)

model/network/Recce.py

@@ -0,0 +1,133 @@
+from functools import partial
+from timm.models import xception
+from model.common import SeparableConv2d, Block
+from model.common import GuidedAttention, GraphReasoning
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+encoder_params = {
+    "xception": {
+        "features": 2048,
+        "init_op": partial(xception, pretrained=True)
+    }
+}
+
+
+class Recce(nn.Module):
+    """ End-to-End Reconstruction-Classification Learning for Face Forgery Detection """
+
+    def __init__(self, num_classes, drop_rate=0.2):
+        super(Recce, self).__init__()
+        self.name = "xception"
+        self.loss_inputs = dict()
+        self.encoder = encoder_params[self.name]["init_op"]()
+        self.global_pool = nn.AdaptiveAvgPool2d((1, 1))
+        self.dropout = nn.Dropout(drop_rate)
+        self.fc = nn.Linear(encoder_params[self.name]["features"], num_classes)
+
+        self.attention = GuidedAttention(depth=728, drop_rate=drop_rate)
+        self.reasoning = GraphReasoning(728, 256, 256, 256, 128, 256, [2, 4], drop_rate)
+
+        self.decoder1 = nn.Sequential(
+            nn.UpsamplingNearest2d(scale_factor=2),
+            SeparableConv2d(728, 256, 3, 1, 1, bias=False),
+            nn.BatchNorm2d(256),
+            nn.ReLU(inplace=True)
+        )
+        self.decoder2 = Block(256, 256, 3, 1)
+        self.decoder3 = nn.Sequential(
+            nn.UpsamplingNearest2d(scale_factor=2),
+            SeparableConv2d(256, 128, 3, 1, 1, bias=False),
+            nn.BatchNorm2d(128),
+            nn.ReLU(inplace=True)
+        )
+        self.decoder4 = Block(128, 128, 3, 1)
+        self.decoder5 = nn.Sequential(
+            nn.UpsamplingNearest2d(scale_factor=2),
+            SeparableConv2d(128, 64, 3, 1, 1, bias=False),
+            nn.BatchNorm2d(64),
+            nn.ReLU(inplace=True)
+        )
+        self.decoder6 = nn.Sequential(
+            nn.Conv2d(64, 3, 1, 1, bias=False),
+            nn.Tanh()
+        )
+
+    def norm_n_corr(self, x):
+        norm_embed = F.normalize(self.global_pool(x), p=2, dim=1)
+        corr = (torch.matmul(norm_embed.squeeze(), norm_embed.squeeze().T) + 1.) / 2.
+        return norm_embed, corr
+
+    @staticmethod
+    def add_white_noise(tensor, mean=0., std=1e-6):
+        rand = torch.rand([tensor.shape[0], 1, 1, 1])
+        rand = torch.where(rand > 0.5, 1., 0.).to(tensor.device)
+        white_noise = torch.normal(mean, std, size=tensor.shape, device=tensor.device)
+        noise_t = tensor + white_noise * rand
+        noise_t = torch.clip(noise_t, -1., 1.)
+        return noise_t
+
+    def forward(self, x):
+        # clear the loss inputs
+        self.loss_inputs = dict(recons=[], contra=[])
+        noise_x = self.add_white_noise(x) if self.training else x
+        out = self.encoder.conv1(noise_x)
+        out = self.encoder.bn1(out)
+        out = self.encoder.act1(out)
+        out = self.encoder.conv2(out)
+        out = self.encoder.bn2(out)
+        out = self.encoder.act2(out)
+        out = self.encoder.block1(out)
+        out = self.encoder.block2(out)
+        out = self.encoder.block3(out)
+        embedding = self.encoder.block4(out)
+
+        norm_embed, corr = self.norm_n_corr(embedding)
+        self.loss_inputs['contra'].append(corr)
+
+        out = self.dropout(embedding)
+        out = self.decoder1(out)
+        out_d2 = self.decoder2(out)
+
+        norm_embed, corr = self.norm_n_corr(out_d2)
+        self.loss_inputs['contra'].append(corr)
+
+        out = self.decoder3(out_d2)
+        out_d4 = self.decoder4(out)
+
+        norm_embed, corr = self.norm_n_corr(out_d4)
+        self.loss_inputs['contra'].append(corr)
+
+        out = self.decoder5(out_d4)
+        pred = self.decoder6(out)
+
+        recons_x = F.interpolate(pred, size=x.shape[-2:], mode='bilinear', align_corners=True)
+        self.loss_inputs['recons'].append(recons_x)
+
+        embedding = self.encoder.block5(embedding)
+        embedding = self.encoder.block6(embedding)
+        embedding = self.encoder.block7(embedding)
+
+        fusion = self.reasoning(embedding, out_d2, out_d4) + embedding
+
+        embedding = self.encoder.block8(fusion)
+        img_att = self.attention(x, recons_x, embedding)
+
+        embedding = self.encoder.block9(img_att)
+        embedding = self.encoder.block10(embedding)
+        embedding = self.encoder.block11(embedding)
+        embedding = self.encoder.block12(embedding)
+
+        embedding = self.encoder.conv3(embedding)
+        embedding = self.encoder.bn3(embedding)
+        embedding = self.encoder.act3(embedding)
+        embedding = self.encoder.conv4(embedding)
+        embedding = self.encoder.bn4(embedding)
+        embedding = self.encoder.act4(embedding)
+
+        embedding = self.global_pool(embedding).squeeze()
+
+        out = self.dropout(embedding)
+        return self.fc(out)

model/network/__init__.py

@@ -0,0 +1 @@
+from .Recce import Recce

models/net.py

@@ -0,0 +1,137 @@
+import time
+import torch
+import torch.nn as nn
+import torchvision.models._utils as _utils
+import torchvision.models as models
+import torch.nn.functional as F
+from torch.autograd import Variable
+
+def conv_bn(inp, oup, stride = 1, leaky = 0):
+    return nn.Sequential(
+        nn.Conv2d(inp, oup, 3, stride, 1, bias=False),
+        nn.BatchNorm2d(oup),
+        nn.LeakyReLU(negative_slope=leaky, inplace=True)
+    )
+
+def conv_bn_no_relu(inp, oup, stride):
+    return nn.Sequential(
+        nn.Conv2d(inp, oup, 3, stride, 1, bias=False),
+        nn.BatchNorm2d(oup),
+    )
+
+def conv_bn1X1(inp, oup, stride, leaky=0):
+    return nn.Sequential(
+        nn.Conv2d(inp, oup, 1, stride, padding=0, bias=False),
+        nn.BatchNorm2d(oup),
+        nn.LeakyReLU(negative_slope=leaky, inplace=True)
+    )
+
+def conv_dw(inp, oup, stride, leaky=0.1):
+    return nn.Sequential(
+        nn.Conv2d(inp, inp, 3, stride, 1, groups=inp, bias=False),
+        nn.BatchNorm2d(inp),
+        nn.LeakyReLU(negative_slope= leaky,inplace=True),
+
+        nn.Conv2d(inp, oup, 1, 1, 0, bias=False),
+        nn.BatchNorm2d(oup),
+        nn.LeakyReLU(negative_slope= leaky,inplace=True),
+    )
+
+class SSH(nn.Module):
+    def __init__(self, in_channel, out_channel):
+        super(SSH, self).__init__()
+        assert out_channel % 4 == 0
+        leaky = 0
+        if (out_channel <= 64):
+            leaky = 0.1
+        self.conv3X3 = conv_bn_no_relu(in_channel, out_channel//2, stride=1)
+
+        self.conv5X5_1 = conv_bn(in_channel, out_channel//4, stride=1, leaky = leaky)
+        self.conv5X5_2 = conv_bn_no_relu(out_channel//4, out_channel//4, stride=1)
+
+        self.conv7X7_2 = conv_bn(out_channel//4, out_channel//4, stride=1, leaky = leaky)
+        self.conv7x7_3 = conv_bn_no_relu(out_channel//4, out_channel//4, stride=1)
+
+    def forward(self, input):
+        conv3X3 = self.conv3X3(input)
+
+        conv5X5_1 = self.conv5X5_1(input)
+        conv5X5 = self.conv5X5_2(conv5X5_1)
+
+        conv7X7_2 = self.conv7X7_2(conv5X5_1)
+        conv7X7 = self.conv7x7_3(conv7X7_2)
+
+        out = torch.cat([conv3X3, conv5X5, conv7X7], dim=1)
+        out = F.relu(out)
+        return out
+
+class FPN(nn.Module):
+    def __init__(self,in_channels_list,out_channels):
+        super(FPN,self).__init__()
+        leaky = 0
+        if (out_channels <= 64):
+            leaky = 0.1
+        self.output1 = conv_bn1X1(in_channels_list[0], out_channels, stride = 1, leaky = leaky)
+        self.output2 = conv_bn1X1(in_channels_list[1], out_channels, stride = 1, leaky = leaky)
+        self.output3 = conv_bn1X1(in_channels_list[2], out_channels, stride = 1, leaky = leaky)
+
+        self.merge1 = conv_bn(out_channels, out_channels, leaky = leaky)
+        self.merge2 = conv_bn(out_channels, out_channels, leaky = leaky)
+
+    def forward(self, input):
+        # names = list(input.keys())
+        input = list(input.values())
+
+        output1 = self.output1(input[0])
+        output2 = self.output2(input[1])
+        output3 = self.output3(input[2])
+
+        up3 = F.interpolate(output3, size=[output2.size(2), output2.size(3)], mode="nearest")
+        output2 = output2 + up3
+        output2 = self.merge2(output2)
+
+        up2 = F.interpolate(output2, size=[output1.size(2), output1.size(3)], mode="nearest")
+        output1 = output1 + up2
+        output1 = self.merge1(output1)
+
+        out = [output1, output2, output3]
+        return out
+
+
+
+class MobileNetV1(nn.Module):
+    def __init__(self):
+        super(MobileNetV1, self).__init__()
+        self.stage1 = nn.Sequential(
+            conv_bn(3, 8, 2, leaky = 0.1),    # 3
+            conv_dw(8, 16, 1),   # 7
+            conv_dw(16, 32, 2),  # 11
+            conv_dw(32, 32, 1),  # 19
+            conv_dw(32, 64, 2),  # 27
+            conv_dw(64, 64, 1),  # 43
+        )
+        self.stage2 = nn.Sequential(
+            conv_dw(64, 128, 2),  # 43 + 16 = 59
+            conv_dw(128, 128, 1), # 59 + 32 = 91
+            conv_dw(128, 128, 1), # 91 + 32 = 123
+            conv_dw(128, 128, 1), # 123 + 32 = 155
+            conv_dw(128, 128, 1), # 155 + 32 = 187
+            conv_dw(128, 128, 1), # 187 + 32 = 219
+        )
+        self.stage3 = nn.Sequential(
+            conv_dw(128, 256, 2), # 219 +3 2 = 241
+            conv_dw(256, 256, 1), # 241 + 64 = 301
+        )
+        self.avg = nn.AdaptiveAvgPool2d((1,1))
+        self.fc = nn.Linear(256, 1000)
+
+    def forward(self, x):
+        x = self.stage1(x)
+        x = self.stage2(x)
+        x = self.stage3(x)
+        x = self.avg(x)
+        # x = self.model(x)
+        x = x.view(-1, 256)
+        x = self.fc(x)
+        return x
+

models/retinaface.py

@@ -0,0 +1,127 @@
+import torch
+import torch.nn as nn
+import torchvision.models.detection.backbone_utils as backbone_utils
+import torchvision.models._utils as _utils
+import torch.nn.functional as F
+from collections import OrderedDict
+
+from models.net import MobileNetV1 as MobileNetV1
+from models.net import FPN as FPN
+from models.net import SSH as SSH
+
+
+
+class ClassHead(nn.Module):
+    def __init__(self,inchannels=512,num_anchors=3):
+        super(ClassHead,self).__init__()
+        self.num_anchors = num_anchors
+        self.conv1x1 = nn.Conv2d(inchannels,self.num_anchors*2,kernel_size=(1,1),stride=1,padding=0)
+
+    def forward(self,x):
+        out = self.conv1x1(x)
+        out = out.permute(0,2,3,1).contiguous()
+        
+        return out.view(out.shape[0], -1, 2)
+
+class BboxHead(nn.Module):
+    def __init__(self,inchannels=512,num_anchors=3):
+        super(BboxHead,self).__init__()
+        self.conv1x1 = nn.Conv2d(inchannels,num_anchors*4,kernel_size=(1,1),stride=1,padding=0)
+
+    def forward(self,x):
+        out = self.conv1x1(x)
+        out = out.permute(0,2,3,1).contiguous()
+
+        return out.view(out.shape[0], -1, 4)
+
+class LandmarkHead(nn.Module):
+    def __init__(self,inchannels=512,num_anchors=3):
+        super(LandmarkHead,self).__init__()
+        self.conv1x1 = nn.Conv2d(inchannels,num_anchors*10,kernel_size=(1,1),stride=1,padding=0)
+
+    def forward(self,x):
+        out = self.conv1x1(x)
+        out = out.permute(0,2,3,1).contiguous()
+
+        return out.view(out.shape[0], -1, 10)
+
+class RetinaFace(nn.Module):
+    def __init__(self, cfg = None, phase = 'train'):
+        """
+        :param cfg:  Network related settings.
+        :param phase: train or test.
+        """
+        super(RetinaFace,self).__init__()
+        self.phase = phase
+        backbone = None
+        if cfg['name'] == 'mobilenet0.25':
+            backbone = MobileNetV1()
+            if cfg['pretrain']:
+                checkpoint = torch.load("./weights/mobilenetV1X0.25_pretrain.tar", map_location=torch.device('cpu'))
+                from collections import OrderedDict
+                new_state_dict = OrderedDict()
+                for k, v in checkpoint['state_dict'].items():
+                    name = k[7:]  # remove module.
+                    new_state_dict[name] = v
+                # load params
+                backbone.load_state_dict(new_state_dict)
+        elif cfg['name'] == 'Resnet50':
+            import torchvision.models as models
+            backbone = models.resnet50(pretrained=cfg['pretrain'])
+
+        self.body = _utils.IntermediateLayerGetter(backbone, cfg['return_layers'])
+        in_channels_stage2 = cfg['in_channel']
+        in_channels_list = [
+            in_channels_stage2 * 2,
+            in_channels_stage2 * 4,
+            in_channels_stage2 * 8,
+        ]
+        out_channels = cfg['out_channel']
+        self.fpn = FPN(in_channels_list,out_channels)
+        self.ssh1 = SSH(out_channels, out_channels)
+        self.ssh2 = SSH(out_channels, out_channels)
+        self.ssh3 = SSH(out_channels, out_channels)
+
+        self.ClassHead = self._make_class_head(fpn_num=3, inchannels=cfg['out_channel'])
+        self.BboxHead = self._make_bbox_head(fpn_num=3, inchannels=cfg['out_channel'])
+        self.LandmarkHead = self._make_landmark_head(fpn_num=3, inchannels=cfg['out_channel'])
+
+    def _make_class_head(self,fpn_num=3,inchannels=64,anchor_num=2):
+        classhead = nn.ModuleList()
+        for i in range(fpn_num):
+            classhead.append(ClassHead(inchannels,anchor_num))
+        return classhead
+    
+    def _make_bbox_head(self,fpn_num=3,inchannels=64,anchor_num=2):
+        bboxhead = nn.ModuleList()
+        for i in range(fpn_num):
+            bboxhead.append(BboxHead(inchannels,anchor_num))
+        return bboxhead
+
+    def _make_landmark_head(self,fpn_num=3,inchannels=64,anchor_num=2):
+        landmarkhead = nn.ModuleList()
+        for i in range(fpn_num):
+            landmarkhead.append(LandmarkHead(inchannels,anchor_num))
+        return landmarkhead
+
+    def forward(self,inputs):
+        out = self.body(inputs)
+
+        # FPN
+        fpn = self.fpn(out)
+
+        # SSH
+        feature1 = self.ssh1(fpn[0])
+        feature2 = self.ssh2(fpn[1])
+        feature3 = self.ssh3(fpn[2])
+        features = [feature1, feature2, feature3]
+
+        bbox_regressions = torch.cat([self.BboxHead[i](feature) for i, feature in enumerate(features)], dim=1)
+        classifications = torch.cat([self.ClassHead[i](feature) for i, feature in enumerate(features)],dim=1)
+        ldm_regressions = torch.cat([self.LandmarkHead[i](feature) for i, feature in enumerate(features)], dim=1)
+
+        if self.phase == 'train':
+            output = (bbox_regressions, classifications, ldm_regressions)
+        else:
+            output = (bbox_regressions, F.softmax(classifications, dim=-1), ldm_regressions)
+        return output

optimizer/__init__.py

@@ -0,0 +1,30 @@
+from torch.optim import SGD
+from torch.optim import Adam
+from torch.optim import ASGD
+from torch.optim import Adamax
+from torch.optim import Adadelta
+from torch.optim import Adagrad
+from torch.optim import RMSprop
+
+key2opt = {
+    'sgd': SGD,
+    'adam': Adam,
+    'asgd': ASGD,
+    'adamax': Adamax,
+    'adadelta': Adadelta,
+    'adagrad': Adagrad,
+    'rmsprop': RMSprop,
+}
+
+
+def get_optimizer(optimizer_name=None):
+    if optimizer_name is None:
+        print("Using default 'SGD' optimizer")
+        return SGD
+
+    else:
+        if optimizer_name not in key2opt:
+            raise NotImplementedError(f"Optimizer '{optimizer_name}' not implemented")
+
+        print(f"Using optimizer: '{optimizer_name}'")
+        return key2opt[optimizer_name]

requirements.txt

@@ -0,0 +1,24 @@
+albumentations==1.3.1
+bbox==0.9.4
+Cython==0.29.35
+ipython==8.14.0
+matplotlib==3.4.3
+numpy==1.25.0
+numpy==1.21.1
+opencv_python==4.5.5.62
+opencv_python_headless==4.7.0.72
+Pillow==9.3.0
+Pillow==9.4.0
+Pillow==9.5.0
+PyYAML==6.0
+PyYAML==6.0
+scikit_learn==1.2.2
+scipy==1.8.0
+streamlit==1.24.0
+tensorboardX==2.6.1
+timm==0.4.12
+torch==1.10.0
+torch==2.0.1+cu117
+torchvision==0.11.1
+torchvision==0.15.2+cu117
+tqdm==4.65.0

scheduler/__init__.py

@@ -0,0 +1,36 @@
+from torch.optim.lr_scheduler import _LRScheduler
+from torch.optim.lr_scheduler import StepLR
+from torch.optim.lr_scheduler import MultiStepLR
+from torch.optim.lr_scheduler import ExponentialLR
+from torch.optim.lr_scheduler import CosineAnnealingLR
+from torch.optim.lr_scheduler import CosineAnnealingWarmRestarts
+from torch.optim.lr_scheduler import ReduceLROnPlateau
+
+
+class ConstantLR(_LRScheduler):
+    def __init__(self, optimizer, last_epoch=-1):
+        super(ConstantLR, self).__init__(optimizer, last_epoch)
+
+    def get_lr(self):
+        return [base_lr for base_lr in self.base_lrs]
+
+
+SCHEDULERS = {
+    'ConstantLR': ConstantLR,
+    "StepLR": StepLR,
+    "MultiStepLR": MultiStepLR,
+    "CosineAnnealingLR": CosineAnnealingLR,
+    "CosineAnnealingWarmRestarts": CosineAnnealingWarmRestarts,
+    "ExponentialLR": ExponentialLR,
+    "ReduceLROnPlateau": ReduceLROnPlateau
+}
+
+
+def get_scheduler(optimizer, kwargs):
+    if kwargs is None:
+        print("No lr scheduler is used.")
+        return ConstantLR(optimizer)
+    name = kwargs["name"]
+    kwargs.pop("name")
+    print("Using scheduler: '%s' with params: %s" % (name, kwargs))
+    return SCHEDULERS[name](optimizer, **kwargs)

trainer/__init__.py

@@ -0,0 +1,5 @@
+from .abstract_trainer import AbstractTrainer, LEGAL_METRIC
+from .exp_mgpu_trainer import ExpMultiGpuTrainer
+from .exp_tester import ExpTester
+from .utils import center_print, reduce_tensor
+from .utils import exp_recons_loss

trainer/abstract_trainer.py

@@ -0,0 +1,100 @@
+import os
+import torch
+import random
+from collections import OrderedDict
+from torchvision.utils import make_grid
+
+LEGAL_METRIC = ['Acc', 'AUC', 'LogLoss']
+
+
+class AbstractTrainer(object):
+    def __init__(self, config, stage="Train"):
+        feasible_stage = ["Train", "Test"]
+        if stage not in feasible_stage:
+            raise ValueError(f"stage should be in {feasible_stage}, but found '{stage}'")
+
+        self.config = config
+        model_cfg = config.get("model", None)
+        data_cfg = config.get("data", None)
+        config_cfg = config.get("config", None)
+
+        self.model_name = model_cfg.pop("name")
+
+        self.gpu = None
+        self.dir = None
+        self.debug = None
+        self.device = None
+        self.resume = None
+        self.local_rank = None
+        self.num_classes = None
+
+        self.best_metric = 0.0
+        self.best_step = 1
+        self.start_step = 1
+
+        self._initiated_settings(model_cfg, data_cfg, config_cfg)
+
+        if stage == 'Train':
+            self._train_settings(model_cfg, data_cfg, config_cfg)
+        if stage == 'Test':
+            self._test_settings(model_cfg, data_cfg, config_cfg)
+
+    def _initiated_settings(self, model_cfg, data_cfg, config_cfg):
+        raise NotImplementedError("Not implemented in abstract class.")
+
+    def _train_settings(self, model_cfg, data_cfg, config_cfg):
+        raise NotImplementedError("Not implemented in abstract class.")
+
+    def _test_settings(self, model_cfg, data_cfg, config_cfg):
+        raise NotImplementedError("Not implemented in abstract class.")
+
+    def _save_ckpt(self, step, best=False):
+        raise NotImplementedError("Not implemented in abstract class.")
+
+    def _load_ckpt(self, best=False, train=False):
+        raise NotImplementedError("Not implemented in abstract class.")
+
+    def to_device(self, items):
+        return [obj.to(self.device) for obj in items]
+
+    @staticmethod
+    def fixed_randomness():
+        random.seed(0)
+        torch.manual_seed(0)
+        torch.cuda.manual_seed(0)
+        torch.cuda.manual_seed_all(0)
+
+    def train(self):
+        raise NotImplementedError("Not implemented in abstract class.")
+
+    def validate(self, epoch, step, timer, writer):
+        raise NotImplementedError("Not implemented in abstract class.")
+
+    def test(self):
+        raise NotImplementedError("Not implemented in abstract class.")
+
+    def plot_figure(self, images, pred, gt, nrow, categories=None, show=True):
+        import matplotlib.pyplot as plt
+        plot = make_grid(
+            images, nrow, padding=4, normalize=True, scale_each=True, pad_value=1)
+        if self.num_classes == 1:
+            pred = (pred >= 0.5).cpu().numpy()
+        else:
+            pred = pred.argmax(1).cpu().numpy()
+        gt = gt.cpu().numpy()
+        if categories is not None:
+            pred = [categories[i] for i in pred]
+            gt = [categories[i] for i in gt]
+        plot = plot.permute([1, 2, 0])
+        plot = plot.cpu().numpy()
+        ret = plt.figure()
+        plt.imshow(plot)
+        plt.title("pred: %s\ngt: %s" % (pred, gt))
+        plt.axis("off")
+        if show:
+            plt.savefig(os.path.join(self.dir, "test_image.png"), dpi=300)
+            plt.show()
+            plt.close()
+        else:
+            plt.close()
+            return ret

trainer/exp_mgpu_trainer.py

@@ -0,0 +1,370 @@
+import os
+import sys
+import time
+import math
+import yaml
+import torch
+import random
+import numpy as np
+
+from tqdm import tqdm
+from pprint import pprint
+from torch.utils import data
+import torch.distributed as dist
+from torch.cuda.amp import autocast, GradScaler
+from tensorboardX import SummaryWriter
+
+from dataset import load_dataset
+from loss import get_loss
+from model import load_model
+from optimizer import get_optimizer
+from scheduler import get_scheduler
+from trainer import AbstractTrainer, LEGAL_METRIC
+from trainer.utils import exp_recons_loss, MLLoss, reduce_tensor, center_print
+from trainer.utils import MODELS_PATH, AccMeter, AUCMeter, AverageMeter, Logger, Timer
+
+
+class ExpMultiGpuTrainer(AbstractTrainer):
+    def __init__(self, config, stage="Train"):
+        super(ExpMultiGpuTrainer, self).__init__(config, stage)
+        np.random.seed(2021)
+
+    def _mprint(self, content=""):
+        if self.local_rank == 0:
+            print(content)
+
+    def _initiated_settings(self, model_cfg=None, data_cfg=None, config_cfg=None):
+        self.local_rank = config_cfg["local_rank"]
+
+    def _train_settings(self, model_cfg, data_cfg, config_cfg):
+        # debug mode: no log dir, no train_val operation.
+        self.debug = config_cfg["debug"]
+        self._mprint(f"Using debug mode: {self.debug}.")
+        self._mprint("*" * 20)
+
+        self.eval_metric = config_cfg["metric"]
+        if self.eval_metric not in LEGAL_METRIC:
+            raise ValueError(f"Evaluation metric must be in {LEGAL_METRIC}, but found "
+                             f"{self.eval_metric}.")
+        if self.eval_metric == LEGAL_METRIC[-1]:
+            self.best_metric = 1.0e8
+
+        # distribution
+        dist.init_process_group(config_cfg["distribute"]["backend"])
+
+        # load training dataset
+        train_dataset = data_cfg["file"]
+        branch = data_cfg["train_branch"]
+        name = data_cfg["name"]
+        with open(train_dataset, "r") as f:
+            options = yaml.load(f, Loader=yaml.FullLoader)
+        train_options = options[branch]
+        self.train_set = load_dataset(name)(train_options)
+        # define training sampler
+        self.train_sampler = data.distributed.DistributedSampler(self.train_set)
+        # wrapped with data loader
+        self.train_loader = data.DataLoader(self.train_set, shuffle=False,
+                                            sampler=self.train_sampler,
+                                            num_workers=data_cfg.get("num_workers", 4),
+                                            batch_size=data_cfg["train_batch_size"])
+
+        if self.local_rank == 0:
+            # load validation dataset
+            val_options = options[data_cfg["val_branch"]]
+            self.val_set = load_dataset(name)(val_options)
+            # wrapped with data loader
+            self.val_loader = data.DataLoader(self.val_set, shuffle=True,
+                                              num_workers=data_cfg.get("num_workers", 4),
+                                              batch_size=data_cfg["val_batch_size"])
+
+        self.resume = config_cfg.get("resume", False)
+
+        if not self.debug:
+            time_format = "%Y-%m-%d...%H.%M.%S"
+            run_id = time.strftime(time_format, time.localtime(time.time()))
+            self.run_id = config_cfg.get("id", run_id)
+            self.dir = os.path.join("runs", self.model_name, self.run_id)
+
+            if self.local_rank == 0:
+                if not self.resume:
+                    if os.path.exists(self.dir):
+                        raise ValueError("Error: given id '%s' already exists." % self.run_id)
+                    os.makedirs(self.dir, exist_ok=True)
+                    print(f"Writing config file to file directory: {self.dir}.")
+                    yaml.dump({"config": self.config,
+                               "train_data": train_options,
+                               "val_data": val_options},
+                              open(os.path.join(self.dir, 'train_config.yml'), 'w'))
+                    # copy the script for the training model
+                    model_file = MODELS_PATH[self.model_name]
+                    os.system("cp " + model_file + " " + self.dir)
+                else:
+                    print(f"Resuming the history in file directory: {self.dir}.")
+
+                print(f"Logging directory: {self.dir}.")
+
+                # redirect the std out stream
+                sys.stdout = Logger(os.path.join(self.dir, 'records.txt'))
+                center_print('Train configurations begins.')
+                pprint(self.config)
+                pprint(train_options)
+                pprint(val_options)
+                center_print('Train configurations ends.')
+
+        # load model
+        self.num_classes = model_cfg["num_classes"]
+        self.device = "cuda:" + str(self.local_rank)
+        self.model = load_model(self.model_name)(**model_cfg)
+        self.model = torch.nn.SyncBatchNorm.convert_sync_batchnorm(self.model).to(self.device)
+        self._mprint(f"Using SyncBatchNorm.")
+        self.model = torch.nn.parallel.DistributedDataParallel(
+            self.model, device_ids=[self.local_rank], find_unused_parameters=True)
+
+        # load optimizer
+        optim_cfg = config_cfg.get("optimizer", None)
+        optim_name = optim_cfg.pop("name")
+        self.optimizer = get_optimizer(optim_name)(self.model.parameters(), **optim_cfg)
+        # load scheduler
+        self.scheduler = get_scheduler(self.optimizer, config_cfg.get("scheduler", None))
+        # load loss
+        self.loss_criterion = get_loss(config_cfg.get("loss", None), device=self.device)
+
+        # total number of steps (or epoch) to train
+        self.num_steps = train_options["num_steps"]
+        self.num_epoch = math.ceil(self.num_steps / len(self.train_loader))
+
+        # the number of steps to write down a log
+        self.log_steps = train_options["log_steps"]
+        # the number of steps to validate on val dataset once
+        self.val_steps = train_options["val_steps"]
+
+        # balance coefficients
+        self.lambda_1 = config_cfg["lambda_1"]
+        self.lambda_2 = config_cfg["lambda_2"]
+        self.warmup_step = config_cfg.get('warmup_step', 0)
+
+        self.contra_loss = MLLoss()
+        self.acc_meter = AccMeter()
+        self.loss_meter = AverageMeter()
+        self.recons_loss_meter = AverageMeter()
+        self.contra_loss_meter = AverageMeter()
+
+        if self.resume and self.local_rank == 0:
+            self._load_ckpt(best=config_cfg.get("resume_best", False), train=True)
+
+    def _test_settings(self, model_cfg, data_cfg, config_cfg):
+        # Not used.
+        raise NotImplementedError("The function is not intended to be used here.")
+
+    def _load_ckpt(self, best=False, train=False):
+        # Not used.
+        raise NotImplementedError("The function is not intended to be used here.")
+
+    def _save_ckpt(self, step, best=False):
+        save_dir = os.path.join(self.dir, f"best_model_{step}.bin" if best else "latest_model.bin")
+        torch.save({
+            "step": step,
+            "best_step": self.best_step,
+            "best_metric": self.best_metric,
+            "eval_metric": self.eval_metric,
+            "model": self.model.module.state_dict(),
+            "optimizer": self.optimizer.state_dict(),
+            "scheduler": self.scheduler.state_dict(),
+        }, save_dir)
+
+    def train(self):
+        try:
+            timer = Timer()
+            grad_scalar = GradScaler(2 ** 10)
+            if self.local_rank == 0:
+                writer = None if self.debug else SummaryWriter(log_dir=self.dir)
+                center_print("Training begins......")
+            else:
+                writer = None
+            start_epoch = self.start_step // len(self.train_loader) + 1
+            for epoch_idx in range(start_epoch, self.num_epoch + 1):
+                # set sampler
+                self.train_sampler.set_epoch(epoch_idx)
+
+                # reset meter
+                self.acc_meter.reset()
+                self.loss_meter.reset()
+                self.recons_loss_meter.reset()
+                self.contra_loss_meter.reset()
+                self.optimizer.step()
+
+                train_generator = enumerate(self.train_loader, 1)
+                # wrap train generator with tqdm for process 0
+                if self.local_rank == 0:
+                    train_generator = tqdm(train_generator, position=0, leave=True)
+
+                for batch_idx, train_data in train_generator:
+                    global_step = (epoch_idx - 1) * len(self.train_loader) + batch_idx
+                    self.model.train()
+                    I, Y = train_data
+                    I = self.train_loader.dataset.load_item(I)
+                    in_I, Y = self.to_device((I, Y))
+
+                    # warm-up lr
+                    if self.warmup_step != 0 and global_step <= self.warmup_step:
+                        lr = self.config['config']['optimizer']['lr'] * float(global_step) / self.warmup_step
+                        for param_group in self.optimizer.param_groups:
+                            param_group['lr'] = lr
+
+                    self.optimizer.zero_grad()
+                    with autocast():
+                        Y_pre = self.model(in_I)
+
+                        # for BCE Setting:
+                        if self.num_classes == 1:
+                            Y_pre = Y_pre.squeeze()
+                            loss = self.loss_criterion(Y_pre, Y.float())
+                            Y_pre = torch.sigmoid(Y_pre)
+                        else:
+                            loss = self.loss_criterion(Y_pre, Y)
+
+                        # flood
+                        loss = (loss - 0.04).abs() + 0.04
+                        recons_loss = exp_recons_loss(self.model.module.loss_inputs['recons'], (in_I, Y))
+                        contra_loss = self.contra_loss(self.model.module.loss_inputs['contra'], Y)
+                        loss += self.lambda_1 * recons_loss + self.lambda_2 * contra_loss
+
+                    grad_scalar.scale(loss).backward()
+                    grad_scalar.step(self.optimizer)
+                    grad_scalar.update()
+                    if self.warmup_step == 0 or global_step > self.warmup_step:
+                        self.scheduler.step()
+
+                    self.acc_meter.update(Y_pre, Y, self.num_classes == 1)
+                    self.loss_meter.update(reduce_tensor(loss).item())
+                    self.recons_loss_meter.update(reduce_tensor(recons_loss).item())
+                    self.contra_loss_meter.update(reduce_tensor(contra_loss).item())
+                    iter_acc = reduce_tensor(self.acc_meter.mean_acc()).item()
+
+                    if self.local_rank == 0:
+                        if global_step % self.log_steps == 0 and writer is not None:
+                            writer.add_scalar("train/Acc", iter_acc, global_step)
+                            writer.add_scalar("train/Loss", self.loss_meter.avg, global_step)
+                            writer.add_scalar("train/Recons_Loss",
+                                              self.recons_loss_meter.avg if self.lambda_1 != 0 else 0.,
+                                              global_step)
+                            writer.add_scalar("train/Contra_Loss", self.contra_loss_meter.avg, global_step)
+                            writer.add_scalar("train/LR", self.scheduler.get_last_lr()[0], global_step)
+
+                        # log training step
+                        train_generator.set_description(
+                            "Train Epoch %d (%d/%d), Global Step %d, Loss %.4f, Recons %.4f, con %.4f, "
+                            "ACC %.4f, LR %.6f" % (
+                                epoch_idx, batch_idx, len(self.train_loader), global_step,
+                                self.loss_meter.avg, self.recons_loss_meter.avg, self.contra_loss_meter.avg,
+                                iter_acc, self.scheduler.get_last_lr()[0])
+                        )
+
+                        # validating process
+                        if global_step % self.val_steps == 0 and not self.debug:
+                            print()
+                            self.validate(epoch_idx, global_step, timer, writer)
+
+                    # when num_steps has been set and the training process will
+                    # be stopped earlier than the specified num_epochs, then stop.
+                    if self.num_steps is not None and global_step == self.num_steps:
+                        if writer is not None:
+                            writer.close()
+                        if self.local_rank == 0:
+                            print()
+                            center_print("Training process ends.")
+                        dist.destroy_process_group()
+                        return
+                # close the tqdm bar when one epoch ends
+                if self.local_rank == 0:
+                    train_generator.close()
+                    print()
+            # training ends with integer epochs
+            if self.local_rank == 0:
+                if writer is not None:
+                    writer.close()
+                center_print("Training process ends.")
+            dist.destroy_process_group()
+        except Exception as e:
+            dist.destroy_process_group()
+            raise e
+
+    def validate(self, epoch, step, timer, writer):
+        v_idx = random.randint(1, len(self.val_loader) + 1)
+        categories = self.val_loader.dataset.categories
+        self.model.eval()
+        with torch.no_grad():
+            acc = AccMeter()
+            auc = AUCMeter()
+            loss_meter = AverageMeter()
+            cur_acc = 0.0  # Higher is better
+            cur_auc = 0.0  # Higher is better
+            cur_loss = 1e8  # Lower is better
+            val_generator = tqdm(enumerate(self.val_loader, 1), position=0, leave=True)
+            for val_idx, val_data in val_generator:
+                I, Y = val_data
+                I = self.val_loader.dataset.load_item(I)
+                in_I, Y = self.to_device((I, Y))
+                Y_pre = self.model(in_I)
+
+                # for BCE Setting:
+                if self.num_classes == 1:
+                    Y_pre = Y_pre.squeeze()
+                    loss = self.loss_criterion(Y_pre, Y.float())
+                    Y_pre = torch.sigmoid(Y_pre)
+                else:
+                    loss = self.loss_criterion(Y_pre, Y)
+
+                acc.update(Y_pre, Y, self.num_classes == 1)
+                auc.update(Y_pre, Y, self.num_classes == 1)
+                loss_meter.update(loss.item())
+
+                cur_acc = acc.mean_acc()
+                cur_loss = loss_meter.avg
+
+                val_generator.set_description(
+                    "Eval Epoch %d (%d/%d), Global Step %d, Loss %.4f, ACC %.4f" % (
+                        epoch, val_idx, len(self.val_loader), step,
+                        cur_loss, cur_acc)
+                )
+
+                if val_idx == v_idx or val_idx == 1:
+                    sample_recons = list()
+                    for _ in self.model.module.loss_inputs['recons']:
+                        sample_recons.append(_[:4].to("cpu"))
+                    # show images
+                    images = I[:4]
+                    images = torch.cat([images, *sample_recons], dim=0)
+                    pred = Y_pre[:4]
+                    gt = Y[:4]
+                    figure = self.plot_figure(images, pred, gt, 4, categories, show=False)
+
+            cur_auc = auc.mean_auc()
+            print("Eval Epoch %d, Loss %.4f, ACC %.4f, AUC %.4f" % (epoch, cur_loss, cur_acc, cur_auc))
+            if writer is not None:
+                writer.add_scalar("val/Loss", cur_loss, step)
+                writer.add_scalar("val/Acc", cur_acc, step)
+                writer.add_scalar("val/AUC", cur_auc, step)
+                writer.add_figure("val/Figures", figure, step)
+            # record the best acc and the corresponding step
+            if self.eval_metric == 'Acc' and cur_acc >= self.best_metric:
+                self.best_metric = cur_acc
+                self.best_step = step
+                self._save_ckpt(step, best=True)
+            elif self.eval_metric == 'AUC' and cur_auc >= self.best_metric:
+                self.best_metric = cur_auc
+                self.best_step = step
+                self._save_ckpt(step, best=True)
+            elif self.eval_metric == 'LogLoss' and cur_loss <= self.best_metric:
+                self.best_metric = cur_loss
+                self.best_step = step
+                self._save_ckpt(step, best=True)
+            print("Best Step %d, Best %s %.4f, Running Time: %s, Estimated Time: %s" % (
+                self.best_step, self.eval_metric, self.best_metric,
+                timer.measure(), timer.measure(step / self.num_steps)
+            ))
+            self._save_ckpt(step, best=False)
+
+    def test(self):
+        # Not used.
+        raise NotImplementedError("The function is not intended to be used here.")

trainer/exp_tester.py

@@ -0,0 +1,144 @@
+import os
+import sys
+import yaml
+import torch
+import random
+
+from tqdm import tqdm
+from pprint import pprint
+from torch.utils import data
+
+from dataset import load_dataset
+from loss import get_loss
+from model import load_model
+from model.common import freeze_weights
+from trainer import AbstractTrainer
+from trainer.utils import AccMeter, AUCMeter, AverageMeter, Logger, center_print
+
+
+class ExpTester(AbstractTrainer):
+    def __init__(self, config, stage="Test"):
+        super(ExpTester, self).__init__(config, stage)
+
+        if torch.cuda.is_available() and self.device is not None:
+            print(f"Using cuda device: {self.device}.")
+            self.gpu = True
+            self.model = self.model.to(self.device)
+        else:
+            print("Using cpu device.")
+            self.device = torch.device("cpu")
+
+    def _initiated_settings(self, model_cfg=None, data_cfg=None, config_cfg=None):
+        self.gpu = False
+        self.device = config_cfg.get("device", None)
+
+    def _train_settings(self, model_cfg=None, data_cfg=None, config_cfg=None):
+        # Not used.
+        raise NotImplementedError("The function is not intended to be used here.")
+
+    def _test_settings(self, model_cfg=None, data_cfg=None, config_cfg=None):
+        # load test dataset
+        test_dataset = data_cfg["file"]
+        branch = data_cfg["test_branch"]
+        name = data_cfg["name"]
+        with open(test_dataset, "r") as f:
+            options = yaml.load(f, Loader=yaml.FullLoader)
+        test_options = options[branch]
+        self.test_set = load_dataset(name)(test_options)
+        # wrapped with data loader
+        self.test_batch_size = data_cfg["test_batch_size"]
+        self.test_loader = data.DataLoader(self.test_set, shuffle=False,
+                                           batch_size=self.test_batch_size)
+        self.run_id = config_cfg["id"]
+        self.ckpt_fold = config_cfg.get("ckpt_fold", "runs")
+        self.dir = os.path.join(self.ckpt_fold, self.model_name, self.run_id)
+
+        # load model
+        self.num_classes = model_cfg["num_classes"]
+        self.model = load_model(self.model_name)(**model_cfg)
+
+        # load loss
+        self.loss_criterion = get_loss(config_cfg.get("loss", None))
+
+        # redirect the std out stream
+        sys.stdout = Logger(os.path.join(self.dir, "test_result.txt"))
+        print('Run dir: {}'.format(self.dir))
+
+        center_print('Test configurations begins')
+        pprint(self.config)
+        pprint(test_options)
+        center_print('Test configurations ends')
+
+        self.ckpt = config_cfg.get("ckpt", "best_model")
+        self._load_ckpt(best=True, train=False)
+
+    def _save_ckpt(self, step, best=False):
+        # Not used.
+        raise NotImplementedError("The function is not intended to be used here.")
+
+    def _load_ckpt(self, best=False, train=False):
+        load_dir = os.path.join(self.dir, self.ckpt + ".bin" if best else "latest_model.bin")
+        load_dict = torch.load(load_dir, map_location=self.device)
+        self.start_step = load_dict["step"]
+        self.best_step = load_dict["best_step"]
+        self.best_metric = load_dict.get("best_metric", None)
+        if self.best_metric is None:
+            self.best_metric = load_dict.get("best_acc")
+        self.eval_metric = load_dict.get("eval_metric", None)
+        if self.eval_metric is None:
+            self.eval_metric = load_dict.get("Acc")
+        self.model.load_state_dict(load_dict["model"])
+        print(f"Loading checkpoint from {load_dir}, best step: {self.best_step}, "
+              f"best {self.eval_metric}: {round(self.best_metric.item(), 4)}.")
+
+    def train(self):
+        # Not used.
+        raise NotImplementedError("The function is not intended to be used here.")
+
+    def validate(self, epoch, step, timer, writer):
+        # Not used.
+        raise NotImplementedError("The function is not intended to be used here.")
+
+    def test(self, display_images=False):
+        freeze_weights(self.model)
+        t_idx = random.randint(1, len(self.test_loader) + 1)
+        self.fixed_randomness()  # for reproduction
+
+        acc = AccMeter()
+        auc = AUCMeter()
+        logloss = AverageMeter()
+        test_generator = tqdm(enumerate(self.test_loader, 1))
+        categories = self.test_loader.dataset.categories
+        for idx, test_data in test_generator:
+            self.model.eval()
+            I, Y = test_data
+            I = self.test_loader.dataset.load_item(I)
+            if self.gpu:
+                in_I, Y = self.to_device((I, Y))
+            else:
+                in_I, Y = (I, Y)
+            Y_pre = self.model(in_I)
+
+            # for BCE Setting:
+            if self.num_classes == 1:
+                Y_pre = Y_pre.squeeze()
+                loss = self.loss_criterion(Y_pre, Y.float())
+                Y_pre = torch.sigmoid(Y_pre)
+            else:
+                loss = self.loss_criterion(Y_pre, Y)
+
+            acc.update(Y_pre, Y, use_bce=self.num_classes == 1)
+            auc.update(Y_pre, Y, use_bce=self.num_classes == 1)
+            logloss.update(loss.item())
+
+            test_generator.set_description("Test %d/%d" % (idx, len(self.test_loader)))
+            if display_images and idx == t_idx:
+                # show images
+                images = I[:4]
+                pred = Y_pre[:4]
+                gt = Y[:4]
+                self.plot_figure(images, pred, gt, 2, categories)
+
+        print("Test, FINAL LOSS %.4f, FINAL ACC %.4f, FINAL AUC %.4f" %
+              (logloss.avg, acc.mean_acc(), auc.mean_auc()))
+        auc.curve(self.dir)

trainer/utils.py

@@ -0,0 +1,183 @@
+import os
+import sys
+import time
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import torch.distributed as dist
+from collections import OrderedDict
+
+import numpy as np
+from sklearn.metrics import roc_auc_score, roc_curve
+from scipy.optimize import brentq
+from scipy.interpolate import interp1d
+
+# Tracking the path to the definition of the model.
+MODELS_PATH = {
+    "Recce": "model/network/Recce.py"
+}
+
+
+def exp_recons_loss(recons, x):
+    x, y = x
+    loss = torch.tensor(0., device=y.device)
+    real_index = torch.where(1 - y)[0]
+    for r in recons:
+        if real_index.numel() > 0:
+            real_x = torch.index_select(x, dim=0, index=real_index)
+            real_rec = torch.index_select(r, dim=0, index=real_index)
+            real_rec = F.interpolate(real_rec, size=x.shape[-2:], mode='bilinear', align_corners=True)
+            loss += torch.mean(torch.abs(real_rec - real_x))
+    return loss
+
+
+def center_print(content, around='*', repeat_around=10):
+    num = repeat_around
+    s = around
+    print(num * s + ' %s ' % content + num * s)
+
+
+def reduce_tensor(t):
+    rt = t.clone()
+    dist.all_reduce(rt)
+    rt /= float(dist.get_world_size())
+    return rt
+
+
+def tensor2image(tensor):
+    image = tensor.permute([1, 2, 0]).cpu().detach().numpy()
+    return (image - np.min(image)) / (np.max(image) - np.min(image))
+
+
+def state_dict(state_dict):
+    """ Remove 'module' keyword in state dictionary. """
+    weights = OrderedDict()
+    for k, v in state_dict.items():
+        weights.update({k.replace("module.", ""): v})
+    return weights
+
+
+class Logger(object):
+    def __init__(self, filename):
+        self.terminal = sys.stdout
+        self.log = open(filename, "a")
+
+    def write(self, message):
+        self.terminal.write(message)
+        self.log.write(message)
+        self.log.flush()
+
+    def flush(self):
+        pass
+
+
+class Timer(object):
+    """The class for timer."""
+
+    def __init__(self):
+        self.o = time.time()
+
+    def measure(self, p=1):
+        x = (time.time() - self.o) / p
+        x = int(x)
+        if x >= 3600:
+            return '{:.1f}h'.format(x / 3600)
+        if x >= 60:
+            return '{}m'.format(round(x / 60))
+        return '{}s'.format(x)
+
+
+class MLLoss(nn.Module):
+    def __init__(self):
+        super(MLLoss, self).__init__()
+
+    def forward(self, input, target, eps=1e-6):
+        # 0 - real; 1 - fake.
+        loss = torch.tensor(0., device=target.device)
+        batch_size = target.shape[0]
+        mat_1 = torch.hstack([target.unsqueeze(-1)] * batch_size)
+        mat_2 = torch.vstack([target] * batch_size)
+        diff_mat = torch.logical_xor(mat_1, mat_2).float()
+        or_mat = torch.logical_or(mat_1, mat_2)
+        eye = torch.eye(batch_size, device=target.device)
+        or_mat = torch.logical_or(or_mat, eye).float()
+        sim_mat = 1. - or_mat
+        for _ in input:
+            diff = torch.sum(_ * diff_mat, dim=[0, 1]) / (torch.sum(diff_mat, dim=[0, 1]) + eps)
+            sim = torch.sum(_ * sim_mat, dim=[0, 1]) / (torch.sum(sim_mat, dim=[0, 1]) + eps)
+            partial_loss = 1. - sim + diff
+            loss += max(partial_loss, torch.zeros_like(partial_loss))
+        return loss
+
+
+class AccMeter(object):
+    def __init__(self):
+        self.nums = 0
+        self.acc = 0
+
+    def reset(self):
+        self.nums = 0
+        self.acc = 0
+
+    def update(self, pred, target, use_bce=False):
+        if use_bce:
+            pred = (pred >= 0.5).int()
+        else:
+            pred = pred.argmax(1)
+        self.nums += target.shape[0]
+        self.acc += torch.sum(pred == target)
+
+    def mean_acc(self):
+        return self.acc / self.nums
+
+
+class AUCMeter(object):
+    def __init__(self):
+        self.score = None
+        self.true = None
+
+    def reset(self):
+        self.score = None
+        self.true = None
+
+    def update(self, score, true, use_bce=False):
+        if use_bce:
+            score = score.detach().cpu().numpy()
+        else:
+            score = torch.softmax(score.detach(), dim=-1)
+            score = torch.select(score, 1, 1).cpu().numpy()
+        true = true.flatten().cpu().numpy()
+        self.score = score if self.score is None else np.concatenate([self.score, score])
+        self.true = true if self.true is None else np.concatenate([self.true, true])
+
+    def mean_auc(self):
+        return roc_auc_score(self.true, self.score)
+
+    def curve(self, prefix):
+        fpr, tpr, thresholds = roc_curve(self.true, self.score, pos_label=1)
+        eer = brentq(lambda x: 1. - x - interp1d(fpr, tpr)(x), 0., 1.)
+        thresh = interp1d(fpr, thresholds)(eer)
+        print(f"# EER: {eer:.4f}(thresh: {thresh:.4f})")
+        torch.save([fpr, tpr, thresholds], os.path.join(prefix, "roc_curve.pickle"))
+
+
+class AverageMeter(object):
+    """Computes and stores the average and current value"""
+
+    def __init__(self):
+        self.val = 0
+        self.avg = 0
+        self.sum = 0
+        self.count = 0
+
+    def reset(self):
+        self.val = 0
+        self.avg = 0
+        self.sum = 0
+        self.count = 0
+
+    def update(self, val, n=1):
+        self.val = val
+        self.sum += val * n
+        self.count += n
+        self.avg = self.sum / self.count