1.
내 End2End F1 Score 계산기가 잘못되었을수도 있지만 물론, 시각화해보니까 그냥 실제로 못 하긴 했다.
2.
시간이 많이 없다. 시간을 아끼려면 끝까지 다 학습하기 전에 중간중간 end to end f1 score 을 확인해볼 수 있어야 한다.
3.
성능을 어떻게 개선할 수 있을까? 많이 실험해볼 시간은 없다.
a.
가장 강력한 후보는 augmentation 이다. augmentation 이 제대로 안 되고 있는 것 같다. 우리 데이터셋들의 텍스트 영역은 세로 길이가 일정하지만, 가로 길이가 뒤죽박죽이라는 특징이 있다. 그래서 세로 길이를 기준으로 텍스트 영역이 뭉개지지 않게 전처리하면 문제가 해결될 것이라는 가설이 있다.
시각화 | 인사이트 |
빨간색 scatter | 빨간색은 바운딩박스 가로 길이와 세로 길이의 분포를 의미한다. 가로로 작성하는 문서인만큼, 세로 길이에는 분명한 경향성이 있음을 확인할 수 있다. |
파란색 scatter | 파란색은 바운딩박스 가로 길이가 증가함에 따라 단어의 길이가 길어짐을 보이고 있다. 물류 데이터셋(하단) 에서는 가로길이는 엄청 긴데 구성하는 단어의 개수가 짧은 데이터가 다수 보인다. |
단어를 구성하는 문자의 수
207,419,207,448,317,419,317,448,타스빌루스
368,600,368,624,474,600,474,624,7,887만원
565,599,565,621,678,599,678,621,9553.6.19
911,1521,911,1549,974,1521,974,1549,매경훈
881,1602,881,1626,947,1602,947,1626,탄이나
Plain Text
복사
금융 데이터셋 발췌
616,146,616,159,677,146,677,159,HG274730
497,909,497,921,513,909,513,921,13
522,908,522,922,546,908,546,922,K/T
170,361,170,376,237,361,237,376,MEBOURNE,
247,361,247,376,271,361,271,376,QLD
147,204,147,217,273,204,273,217,+87-865-7373-4634
418,274,418,286,449,274,449,286,8858
479,290,479,302,540,290,540,302,COLUMBIA
486,255,486,267,562,255,562,267,CONSULTING
551,352,551,364,579,352,579,364,LTD.
418,385,418,398,457,385,457,398,BUSAN
39,260,39,272,78,260,78,272,98-97
287,260,287,274,322,260,322,274,1557,
42,168,42,182,93,168,93,182,71-656,
172,184,172,195,218,184,218,195,TURKEY
107,497,107,511,138,497,138,511,IRAQ
281,496,281,510,320,496,320,510,CY/CY
731,704,731,715,755,704,755,715,CBM
679,599,679,612,723,599,723,612,559.72
529,709,529,718,561,709,561,718,TOTAL
Plain Text
복사
물류 데이터셋 발췌
b.
물류 모델의 경우 영어밖에 사용하지 않기 때문에 당연히 pretrained weight 를 사용한 것이 사용하지 않은 것보다 나을수밖에 없다.
size mismatch for decoder.trg_word_emb.weight: copying a param with shape torch.Size([93, 256]) from checkpoint, the shape in current model is torch.Size([94, 256]).
size mismatch for decoder.classifier.weight: copying a param with shape torch.Size([92, 256]) from checkpoint, the shape in current model is torch.Size([94, 256]).
size mismatch for decoder.classifier.bias: copying a param with shape torch.Size([92]) from checkpoint, the shape in current model is torch.Size([94]).
Plain Text
복사
•
pretrained 모델과 shape 가 약간 달라서 온전한 성능을 뽑아내지 못하고 있을지도 모른다.
◦
물류 데이터에 공백이 포함되어 있을까? 없다면 공백을 제거해 보자. 공백을 포함한 사전 파일 dicts/english_digits_symbols_space.txt 는 총 91개줄로 이루어져 있다. 여기에 unknown class 1 개를 추가하면 92이다. 여기에 무언가를 두 개 추가하면… (왜 추가하는지는 잘 모르겠지만) 94이다.
◦
만약 공백이 없다면 사전파일 dicts/lower_english_digits.txt 을 사용해 보자. 뭔지는 모르겠지만 이 경우 하나가 줄었으니 93이 되지 않을까.
데이터 증강 기법 변경
금융 | 물류 | |
height (고정기준) | 30 | 20 |
min_width | 30 | 10 |
max_width | 250 | 180 |
변경 이전
짧은 길이 문자: after
보통 길이 문자 : before, 다 뭉개져서 알아볼 수 없다.
변경 이후
짧은 문자: after
보통 길이 문자 : after
그런데 이것도 잘 생각해보면 말이 안 된다. 왜냐하면 OCR 문제는 어떻게든 bbox 딱 fitting 시키려고 하는 것 아님? 그런데 이렇게 검정색으로 슥 칠해버리면 모델이 정신 못차리지 않을까? 가로길이를 딱 맞추고 글씨를 가득 채우려면 어떻게 해야 할까.
그런데 학습은 어떻게 그렇게 잘된거지..?
gk 왜 안되지?
•
잘 돌아가는 모델에서 test_pipeline 을 다른 값으로 덮어썼는데 아무 문제 없이 돌아감. 왜지? pipeline 이 바꼈는데 도대체 어떻게.
•
잘 돌아가던 모델에서 test_pipeline 을 진짜 극단적으로 많이 바꾸니까 슬슬 고장나기 시작함. 그럼에도 불구하고 모델은 돌아가긴 함. 어떻게 돌아갈 수 있는지도 모르겠지만 도대체 무슨 일이 일어나는건지 알수가 없다. 그런데 어쨌든 영향을 주긴 준다는 것을 확인했다.
pretrained/satrn_small_20211009-2cf13355.pth
work_dirs/satrn_shallow-small_30e_aihubtransit1of100_pretrained/epoch_6.pth
이렇게 보면 학습이 되고 있는 것 같기는 한데…
=== Aihub 물류 모델 추론 시작 ===[0m
[32;21m2022-12-01 14:02:48,864 - ocr_e2e - INFO - img_file_name: data/e2ef1/aihub_transit/part_100of100/test/imgs/IMG_OCR_6_T_BL_000010.png[0m
[32;21m2022-12-01 14:03:30,651 - ocr_e2e - INFO - img_file_name: data/e2ef1/aihub_transit/part_100of100/test/imgs/IMG_OCR_6_T_BL_000011.png[0m
[32;21m2022-12-01 14:04:10,908 - ocr_e2e - INFO - img_file_name: data/e2ef1/aihub_transit/part_100of100/test/imgs/IMG_OCR_6_T_BL_000020.png[0m
[32;21m2022-12-01 14:04:54,696 - ocr_e2e - INFO - img_file_name: data/e2ef1/aihub_transit/part_100of100/test/imgs/IMG_OCR_6_T_BL_000034.png[0m
[32;21m2022-12-01 14:05:27,413 - ocr_e2e - INFO - img_file_name: data/e2ef1/aihub_transit/part_100of100/test/imgs/IMG_OCR_6_T_BL_000039.png[0m
[32;21m2022-12-01 14:06:08,729 - ocr_e2e - INFO - img_file_name: data/e2ef1/aihub_transit/part_100of100/test/imgs/IMG_OCR_6_T_BL_000041.png[0m
[32;21m2022-12-01 14:06:53,057 - ocr_e2e - INFO - img_file_name: data/e2ef1/aihub_transit/part_100of100/test/imgs/IMG_OCR_6_T_BL_000061.png[0m
[32;21m2022-12-01 14:07:34,045 - ocr_e2e - INFO - img_file_name: data/e2ef1/aihub_transit/part_100of100/test/imgs/IMG_OCR_6_T_BL_000079.png[0m
[32;21m2022-12-01 14:08:10,162 - ocr_e2e - INFO - img_file_name: data/e2ef1/aihub_transit/part_100of100/test/imgs/IMG_OCR_6_T_BL_000098.png[0m
[32;21m2022-12-01 14:08:56,246 - ocr_e2e - INFO - img_file_name: data/e2ef1/aihub_transit/part_100of100/test/imgs/IMG_OCR_6_T_BL_000100.png[0m
[32;21m2022-12-01 14:09:33,946 - ocr_e2e - INFO - img_file_name: data/e2ef1/aihub_transit/part_100of100/test/imgs/IMG_OCR_6_T_BL_000105.png[0m
[32;21m2022-12-01 14:09:34,799 - ocr_e2e - INFO -
=== MMOCR 타입 Aihub 물류데이터 추론값을 ICDAR 타입으로 변경 ===[0m
[32;21m2022-12-01 14:09:50,553 - e2e_f1 - INFO - End-to-End F1 Score 평가 시작[0m
[32;21m2022-12-01 14:09:50,588 - e2e_f1 - INFO - gt_file_name: IMG_OCR_6_T_BL_000010.txt[0m
[32;21m2022-12-01 14:09:50,588 - e2e_f1 - INFO - pred_file_name: IMG_OCR_6_T_BL_000010.txt[0m
[32;21m2022-12-01 14:09:50,596 - e2e_f1 - INFO - gt_file_name: IMG_OCR_6_T_BL_000011.txt[0m
[32;21m2022-12-01 14:09:50,596 - e2e_f1 - INFO - pred_file_name: IMG_OCR_6_T_BL_000011.txt[0m
[32;21m2022-12-01 14:09:50,598 - e2e_f1 - INFO - gt_file_name: IMG_OCR_6_T_BL_000020.txt[0m
[32;21m2022-12-01 14:09:50,598 - e2e_f1 - INFO - pred_file_name: IMG_OCR_6_T_BL_000020.txt[0m
[32;21m2022-12-01 14:09:50,600 - e2e_f1 - INFO - gt_file_name: IMG_OCR_6_T_BL_000034.txt[0m
[32;21m2022-12-01 14:09:50,600 - e2e_f1 - INFO - pred_file_name: IMG_OCR_6_T_BL_000034.txt[0m
[32;21m2022-12-01 14:09:50,602 - e2e_f1 - INFO - gt_file_name: IMG_OCR_6_T_BL_000039.txt[0m
[32;21m2022-12-01 14:09:50,602 - e2e_f1 - INFO - pred_file_name: IMG_OCR_6_T_BL_000039.txt[0m
[32;21m2022-12-01 14:09:50,604 - e2e_f1 - INFO - gt_file_name: IMG_OCR_6_T_BL_000041.txt[0m
[32;21m2022-12-01 14:09:50,604 - e2e_f1 - INFO - pred_file_name: IMG_OCR_6_T_BL_000041.txt[0m
[32;21m2022-12-01 14:09:50,606 - e2e_f1 - INFO - gt_file_name: IMG_OCR_6_T_BL_000061.txt[0m
[32;21m2022-12-01 14:09:50,606 - e2e_f1 - INFO - pred_file_name: IMG_OCR_6_T_BL_000061.txt[0m
[32;21m2022-12-01 14:09:50,607 - e2e_f1 - INFO - gt_file_name: IMG_OCR_6_T_BL_000079.txt[0m
[32;21m2022-12-01 14:09:50,608 - e2e_f1 - INFO - pred_file_name: IMG_OCR_6_T_BL_000079.txt[0m
[32;21m2022-12-01 14:09:50,609 - e2e_f1 - INFO - gt_file_name: IMG_OCR_6_T_BL_000098.txt[0m
[32;21m2022-12-01 14:09:50,609 - e2e_f1 - INFO - pred_file_name: IMG_OCR_6_T_BL_000098.txt[0m
[32;21m2022-12-01 14:09:50,610 - e2e_f1 - INFO - gt_file_name: IMG_OCR_6_T_BL_000100.txt[0m
[32;21m2022-12-01 14:09:50,611 - e2e_f1 - INFO - pred_file_name: IMG_OCR_6_T_BL_000100.txt[0m
[32;21m2022-12-01 14:09:50,612 - e2e_f1 - INFO - gt_file_name: IMG_OCR_6_T_BL_000105.txt[0m
[32;21m2022-12-01 14:09:50,612 - e2e_f1 - INFO - pred_file_name: IMG_OCR_6_T_BL_000105.txt[0m
[32;21m2022-12-01 14:09:51,445 - e2e_f1 - INFO - Found 1213 preds and 1246 gts[0m
[32;21m2022-12-01 14:09:51,446 - e2e_f1 - INFO - Precision: 0.736191[0m
[32;21m2022-12-01 14:09:51,446 - e2e_f1 - INFO - Recall: 0.717496[0m
[32;21m2022-12-01 14:09:51,446 - e2e_f1 - INFO - F1: 0.726723[0m
Bash
복사
2022-12-01 14:02:45,848 - ocr_e2e - INFO -
=== Aihub 물류 모델 로드 ===
2022-12-01 14:02:45,848 - ocr_e2e - INFO - 디버그 모드로 실행합니다. CPU 를 사용합니다. 10개의 이미지만 처리 및 시각화합니다.
[32;21m2022-12-01 14:02:48,830 - ocr_e2e - INFO -
=== Aihub 물류 모델 추론 시작 ===[0m
[32;21m2022-12-01 14:02:48,864 - ocr_e2e - INFO - img_file_name: data/e2ef1/aihub_transit/part_100of100/test/imgs/IMG_OCR_6_T_BL_000010.png[0m
[32;21m2022-12-01 14:03:30,651 - ocr_e2e - INFO - img_file_name: data/e2ef1/aihub_transit/part_100of100/test/imgs/IMG_OCR_6_T_BL_000011.png[0m
[32;21m2022-12-01 14:04:10,908 - ocr_e2e - INFO - img_file_name: data/e2ef1/aihub_transit/part_100of100/test/imgs/IMG_OCR_6_T_BL_000020.png[0m
[32;21m2022-12-01 14:04:54,696 - ocr_e2e - INFO - img_file_name: data/e2ef1/aihub_transit/part_100of100/test/imgs/IMG_OCR_6_T_BL_000034.png[0m
[32;21m2022-12-01 14:05:27,413 - ocr_e2e - INFO - img_file_name: data/e2ef1/aihub_transit/part_100of100/test/imgs/IMG_OCR_6_T_BL_000039.png[0m
[32;21m2022-12-01 14:06:08,729 - ocr_e2e - INFO - img_file_name: data/e2ef1/aihub_transit/part_100of100/test/imgs/IMG_OCR_6_T_BL_000041.png[0m
[32;21m2022-12-01 14:06:53,057 - ocr_e2e - INFO - img_file_name: data/e2ef1/aihub_transit/part_100of100/test/imgs/IMG_OCR_6_T_BL_000061.png[0m
[32;21m2022-12-01 14:07:34,045 - ocr_e2e - INFO - img_file_name: data/e2ef1/aihub_transit/part_100of100/test/imgs/IMG_OCR_6_T_BL_000079.png[0m
[32;21m2022-12-01 14:08:10,162 - ocr_e2e - INFO - img_file_name: data/e2ef1/aihub_transit/part_100of100/test/imgs/IMG_OCR_6_T_BL_000098.png[0m
[32;21m2022-12-01 14:08:56,246 - ocr_e2e - INFO - img_file_name: data/e2ef1/aihub_transit/part_100of100/test/imgs/IMG_OCR_6_T_BL_000100.png[0m
[32;21m2022-12-01 14:09:33,946 - ocr_e2e - INFO - img_file_name: data/e2ef1/aihub_transit/part_100of100/test/imgs/IMG_OCR_6_T_BL_000105.png[0m
[32;21m2022-12-01 14:09:34,799 - ocr_e2e - INFO -
=== MMOCR 타입 Aihub 물류데이터 추론값을 ICDAR 타입으로 변경 ===[0m
[32;21m2022-12-01 14:09:50,553 - e2e_f1 - INFO - End-to-End F1 Score 평가 시작[0m
[32;21m2022-12-01 14:09:50,588 - e2e_f1 - INFO - gt_file_name: IMG_OCR_6_T_BL_000010.txt[0m
[32;21m2022-12-01 14:09:50,588 - e2e_f1 - INFO - pred_file_name: IMG_OCR_6_T_BL_000010.txt[0m
[32;21m2022-12-01 14:09:50,596 - e2e_f1 - INFO - gt_file_name: IMG_OCR_6_T_BL_000011.txt[0m
[32;21m2022-12-01 14:09:50,596 - e2e_f1 - INFO - pred_file_name: IMG_OCR_6_T_BL_000011.txt[0m
[32;21m2022-12-01 14:09:50,598 - e2e_f1 - INFO - gt_file_name: IMG_OCR_6_T_BL_000020.txt[0m
[32;21m2022-12-01 14:09:50,598 - e2e_f1 - INFO - pred_file_name: IMG_OCR_6_T_BL_000020.txt[0m
[32;21m2022-12-01 14:09:50,600 - e2e_f1 - INFO - gt_file_name: IMG_OCR_6_T_BL_000034.txt[0m
[32;21m2022-12-01 14:09:50,600 - e2e_f1 - INFO - pred_file_name: IMG_OCR_6_T_BL_000034.txt[0m
[32;21m2022-12-01 14:09:50,602 - e2e_f1 - INFO - gt_file_name: IMG_OCR_6_T_BL_000039.txt[0m
[32;21m2022-12-01 14:09:50,602 - e2e_f1 - INFO - pred_file_name: IMG_OCR_6_T_BL_000039.txt[0m
[32;21m2022-12-01 14:09:50,604 - e2e_f1 - INFO - gt_file_name: IMG_OCR_6_T_BL_000041.txt[0m
[32;21m2022-12-01 14:09:50,604 - e2e_f1 - INFO - pred_file_name: IMG_OCR_6_T_BL_000041.txt[0m
[32;21m2022-12-01 14:09:50,606 - e2e_f1 - INFO - gt_file_name: IMG_OCR_6_T_BL_000061.txt[0m
[32;21m2022-12-01 14:09:50,606 - e2e_f1 - INFO - pred_file_name: IMG_OCR_6_T_BL_000061.txt[0m
[32;21m2022-12-01 14:09:50,607 - e2e_f1 - INFO - gt_file_name: IMG_OCR_6_T_BL_000079.txt[0m
[32;21m2022-12-01 14:09:50,608 - e2e_f1 - INFO - pred_file_name: IMG_OCR_6_T_BL_000079.txt[0m
[32;21m2022-12-01 14:09:50,609 - e2e_f1 - INFO - gt_file_name: IMG_OCR_6_T_BL_000098.txt[0m
[32;21m2022-12-01 14:09:50,609 - e2e_f1 - INFO - pred_file_name: IMG_OCR_6_T_BL_000098.txt[0m
[32;21m2022-12-01 14:09:50,610 - e2e_f1 - INFO - gt_file_name: IMG_OCR_6_T_BL_000100.txt[0m
[32;21m2022-12-01 14:09:50,611 - e2e_f1 - INFO - pred_file_name: IMG_OCR_6_T_BL_000100.txt[0m
[32;21m2022-12-01 14:09:50,612 - e2e_f1 - INFO - gt_file_name: IMG_OCR_6_T_BL_000105.txt[0m
[32;21m2022-12-01 14:09:50,612 - e2e_f1 - INFO - pred_file_name: IMG_OCR_6_T_BL_000105.txt[0m
[32;21m2022-12-01 14:09:51,445 - e2e_f1 - INFO - Found 1213 preds and 1246 gts[0m
[32;21m2022-12-01 14:09:51,446 - e2e_f1 - INFO - Precision: 0.736191[0m
[32;21m2022-12-01 14:09:51,446 - e2e_f1 - INFO - Recall: 0.717496[0m
[32;21m2022-12-01 14:09:51,446 - e2e_f1 - INFO - F1: 0.726723[0m
2022-12-01 14:27:04,969 - ocr_e2e - INFO -
=== Aihub 물류 모델 로드 ===
2022-12-01 14:27:04,970 - ocr_e2e - INFO - 디버그 모드로 실행합니다. CPU 를 사용합니다. 10개의 이미지만 처리 및 시각화합니다.
2022-12-01 14:27:07,816 - ocr_e2e - INFO -
=== Aihub 물류 모델 추론 시작 ===
2022-12-01 14:27:07,850 - ocr_e2e - INFO - img_file_name: data/e2ef1/aihub_transit/part_100of100/test/imgs/IMG_OCR_6_T_BL_000010.png
2022-12-01 14:27:47,745 - ocr_e2e - INFO - img_file_name: data/e2ef1/aihub_transit/part_100of100/test/imgs/IMG_OCR_6_T_BL_000011.png
2022-12-01 14:28:27,255 - ocr_e2e - INFO - img_file_name: data/e2ef1/aihub_transit/part_100of100/test/imgs/IMG_OCR_6_T_BL_000020.png
2022-12-01 14:29:09,121 - ocr_e2e - INFO - img_file_name: data/e2ef1/aihub_transit/part_100of100/test/imgs/IMG_OCR_6_T_BL_000034.png
2022-12-01 14:29:41,206 - ocr_e2e - INFO - img_file_name: data/e2ef1/aihub_transit/part_100of100/test/imgs/IMG_OCR_6_T_BL_000039.png
2022-12-01 14:30:21,050 - ocr_e2e - INFO - img_file_name: data/e2ef1/aihub_transit/part_100of100/test/imgs/IMG_OCR_6_T_BL_000041.png
2022-12-01 14:31:02,986 - ocr_e2e - INFO - img_file_name: data/e2ef1/aihub_transit/part_100of100/test/imgs/IMG_OCR_6_T_BL_000061.png
2022-12-01 14:31:43,529 - ocr_e2e - INFO - img_file_name: data/e2ef1/aihub_transit/part_100of100/test/imgs/IMG_OCR_6_T_BL_000079.png
2022-12-01 14:32:15,636 - ocr_e2e - INFO - img_file_name: data/e2ef1/aihub_transit/part_100of100/test/imgs/IMG_OCR_6_T_BL_000098.png
2022-12-01 14:33:00,807 - ocr_e2e - INFO - img_file_name: data/e2ef1/aihub_transit/part_100of100/test/imgs/IMG_OCR_6_T_BL_000100.png
2022-12-01 14:33:36,235 - ocr_e2e - INFO - img_file_name: data/e2ef1/aihub_transit/part_100of100/test/imgs/IMG_OCR_6_T_BL_000105.png
2022-12-01 14:33:37,029 - ocr_e2e - INFO -
=== MMOCR 타입 Aihub 물류데이터 추론값을 ICDAR 타입으로 변경 ===
2022-12-01 14:33:52,429 - e2e_f1 - INFO - End-to-End F1 Score 평가 시작
2022-12-01 14:33:52,459 - e2e_f1 - INFO - gt_file_name: IMG_OCR_6_T_BL_000010.txt
2022-12-01 14:33:52,459 - e2e_f1 - INFO - pred_file_name: IMG_OCR_6_T_BL_000010.txt
2022-12-01 14:33:52,463 - e2e_f1 - INFO - gt_file_name: IMG_OCR_6_T_BL_000011.txt
2022-12-01 14:33:52,463 - e2e_f1 - INFO - pred_file_name: IMG_OCR_6_T_BL_000011.txt
2022-12-01 14:33:52,464 - e2e_f1 - INFO - gt_file_name: IMG_OCR_6_T_BL_000020.txt
2022-12-01 14:33:52,464 - e2e_f1 - INFO - pred_file_name: IMG_OCR_6_T_BL_000020.txt
2022-12-01 14:33:52,464 - e2e_f1 - INFO - gt_file_name: IMG_OCR_6_T_BL_000034.txt
2022-12-01 14:33:52,465 - e2e_f1 - INFO - pred_file_name: IMG_OCR_6_T_BL_000034.txt
2022-12-01 14:33:52,465 - e2e_f1 - INFO - gt_file_name: IMG_OCR_6_T_BL_000039.txt
2022-12-01 14:33:52,465 - e2e_f1 - INFO - pred_file_name: IMG_OCR_6_T_BL_000039.txt
2022-12-01 14:33:52,466 - e2e_f1 - INFO - gt_file_name: IMG_OCR_6_T_BL_000041.txt
2022-12-01 14:33:52,466 - e2e_f1 - INFO - pred_file_name: IMG_OCR_6_T_BL_000041.txt
2022-12-01 14:33:52,466 - e2e_f1 - INFO - gt_file_name: IMG_OCR_6_T_BL_000061.txt
2022-12-01 14:33:52,466 - e2e_f1 - INFO - pred_file_name: IMG_OCR_6_T_BL_000061.txt
2022-12-01 14:33:52,467 - e2e_f1 - INFO - gt_file_name: IMG_OCR_6_T_BL_000079.txt
2022-12-01 14:33:52,467 - e2e_f1 - INFO - pred_file_name: IMG_OCR_6_T_BL_000079.txt
2022-12-01 14:33:52,467 - e2e_f1 - INFO - gt_file_name: IMG_OCR_6_T_BL_000098.txt
2022-12-01 14:33:52,468 - e2e_f1 - INFO - pred_file_name: IMG_OCR_6_T_BL_000098.txt
2022-12-01 14:33:52,468 - e2e_f1 - INFO - gt_file_name: IMG_OCR_6_T_BL_000100.txt
2022-12-01 14:33:52,468 - e2e_f1 - INFO - pred_file_name: IMG_OCR_6_T_BL_000100.txt
2022-12-01 14:33:52,469 - e2e_f1 - INFO - gt_file_name: IMG_OCR_6_T_BL_000105.txt
2022-12-01 14:33:52,469 - e2e_f1 - INFO - pred_file_name: IMG_OCR_6_T_BL_000105.txt
2022-12-01 14:33:53,230 - e2e_f1 - INFO - Found 1213 preds and 1246 gts
2022-12-01 14:33:53,231 - e2e_f1 - INFO - Precision: 0.574608
2022-12-01 14:33:53,231 - e2e_f1 - INFO - Recall: 0.560193
2022-12-01 14:33:53,231 - e2e_f1 - INFO - F1: 0.567309
Bash
복사
=== Aihub 물류 모델 로드 ===
2022-12-01 14:37:58,679 - ocr_e2e - INFO - 디버그 모드로 실행합니다. CPU 를 사용합니다. 10개의 이미지만 처리 및 시각화합니다.
2022-12-01 14:38:01,190 - ocr_e2e - INFO -
=== Aihub 물류 모델 추론 시작 ===
2022-12-01 14:38:01,223 - ocr_e2e - INFO - img_file_name: data/e2ef1/aihub_transit/part_100of100/test/imgs/IMG_OCR_6_T_BL_000010.png
2022-12-01 14:38:41,365 - ocr_e2e - INFO - img_file_name: data/e2ef1/aihub_transit/part_100of100/test/imgs/IMG_OCR_6_T_BL_000011.png
2022-12-01 14:39:21,869 - ocr_e2e - INFO - img_file_name: data/e2ef1/aihub_transit/part_100of100/test/imgs/IMG_OCR_6_T_BL_000020.png
2022-12-01 14:40:04,715 - ocr_e2e - INFO - img_file_name: data/e2ef1/aihub_transit/part_100of100/test/imgs/IMG_OCR_6_T_BL_000034.png
2022-12-01 14:40:39,544 - ocr_e2e - INFO - img_file_name: data/e2ef1/aihub_transit/part_100of100/test/imgs/IMG_OCR_6_T_BL_000039.png
2022-12-01 14:41:21,015 - ocr_e2e - INFO - img_file_name: data/e2ef1/aihub_transit/part_100of100/test/imgs/IMG_OCR_6_T_BL_000041.png
2022-12-01 14:42:03,593 - ocr_e2e - INFO - img_file_name: data/e2ef1/aihub_transit/part_100of100/test/imgs/IMG_OCR_6_T_BL_000061.png
2022-12-01 14:42:43,984 - ocr_e2e - INFO - img_file_name: data/e2ef1/aihub_transit/part_100of100/test/imgs/IMG_OCR_6_T_BL_000079.png
2022-12-01 14:43:21,577 - ocr_e2e - INFO - img_file_name: data/e2ef1/aihub_transit/part_100of100/test/imgs/IMG_OCR_6_T_BL_000098.png
2022-12-01 14:44:09,243 - ocr_e2e - INFO - img_file_name: data/e2ef1/aihub_transit/part_100of100/test/imgs/IMG_OCR_6_T_BL_000100.png
2022-12-01 14:44:46,084 - ocr_e2e - INFO - img_file_name: data/e2ef1/aihub_transit/part_100of100/test/imgs/IMG_OCR_6_T_BL_000105.png
2022-12-01 14:44:46,904 - ocr_e2e - INFO -
=== MMOCR 타입 Aihub 물류데이터 추론값을 ICDAR 타입으로 변경 ===
2022-12-01 14:45:02,605 - e2e_f1 - INFO - End-to-End F1 Score 평가 시작
2022-12-01 14:45:02,639 - e2e_f1 - INFO - gt_file_name: IMG_OCR_6_T_BL_000010.txt
2022-12-01 14:45:02,639 - e2e_f1 - INFO - pred_file_name: IMG_OCR_6_T_BL_000010.txt
2022-12-01 14:45:02,643 - e2e_f1 - INFO - gt_file_name: IMG_OCR_6_T_BL_000011.txt
2022-12-01 14:45:02,644 - e2e_f1 - INFO - pred_file_name: IMG_OCR_6_T_BL_000011.txt
2022-12-01 14:45:02,646 - e2e_f1 - INFO - gt_file_name: IMG_OCR_6_T_BL_000020.txt
2022-12-01 14:45:02,646 - e2e_f1 - INFO - pred_file_name: IMG_OCR_6_T_BL_000020.txt
2022-12-01 14:45:02,646 - e2e_f1 - INFO - gt_file_name: IMG_OCR_6_T_BL_000034.txt
2022-12-01 14:45:02,647 - e2e_f1 - INFO - pred_file_name: IMG_OCR_6_T_BL_000034.txt
2022-12-01 14:45:02,647 - e2e_f1 - INFO - gt_file_name: IMG_OCR_6_T_BL_000039.txt
2022-12-01 14:45:02,647 - e2e_f1 - INFO - pred_file_name: IMG_OCR_6_T_BL_000039.txt
2022-12-01 14:45:02,648 - e2e_f1 - INFO - gt_file_name: IMG_OCR_6_T_BL_000041.txt
2022-12-01 14:45:02,648 - e2e_f1 - INFO - pred_file_name: IMG_OCR_6_T_BL_000041.txt
2022-12-01 14:45:02,648 - e2e_f1 - INFO - gt_file_name: IMG_OCR_6_T_BL_000061.txt
2022-12-01 14:45:02,648 - e2e_f1 - INFO - pred_file_name: IMG_OCR_6_T_BL_000061.txt
2022-12-01 14:45:02,649 - e2e_f1 - INFO - gt_file_name: IMG_OCR_6_T_BL_000079.txt
2022-12-01 14:45:02,649 - e2e_f1 - INFO - pred_file_name: IMG_OCR_6_T_BL_000079.txt
2022-12-01 14:45:02,650 - e2e_f1 - INFO - gt_file_name: IMG_OCR_6_T_BL_000098.txt
2022-12-01 14:45:02,650 - e2e_f1 - INFO - pred_file_name: IMG_OCR_6_T_BL_000098.txt
2022-12-01 14:45:02,650 - e2e_f1 - INFO - gt_file_name: IMG_OCR_6_T_BL_000100.txt
2022-12-01 14:45:02,650 - e2e_f1 - INFO - pred_file_name: IMG_OCR_6_T_BL_000100.txt
2022-12-01 14:45:02,651 - e2e_f1 - INFO - gt_file_name: IMG_OCR_6_T_BL_000105.txt
2022-12-01 14:45:02,651 - e2e_f1 - INFO - pred_file_name: IMG_OCR_6_T_BL_000105.txt
2022-12-01 14:45:03,325 - e2e_f1 - INFO - Found 1213 preds and 1246 gts
2022-12-01 14:45:03,325 - e2e_f1 - INFO - Precision: 0.498763
2022-12-01 14:45:03,325 - e2e_f1 - INFO - Recall: 0.485554
2022-12-01 14:45:03,326 - e2e_f1 - INFO - F1: 0.492070
Bash
복사
•
base.aihubtransit10of100_textrecog_train → 데이터 잘못 만들어짐
•
base.aihubtransit_textrecog_sampled4vis2 → 비어 있는 데이터 존재
◦
validation/IMG_OCR_6_T_ET_005268_60.png
◦
test/IMG_OCR_6_T_ET_008201_29.png
최종적으로 선택된 모델의 성능 (이미지 10개 테스트)
2022-12-01 19:07:21,567 - ocr_e2e - INFO -
=== Aihub 금융 모델 로드 ===
2022-12-01 19:07:21,567 - ocr_e2e - INFO - 디버그 모드로 실행합니다. CPU 를 사용합니다. 10개의 이미지만 처리 및 시각화합니다.
2022-12-01 19:07:24,622 - ocr_e2e - INFO -
=== Aihub 금융 모델 추론 시작 ===
2022-12-01 19:07:24,637 - ocr_e2e - INFO - img_file_name: data/e2ef1/aihub_finance/part_100of100/test/imgs/IMG_OCR_6_F_0000007.png
2022-12-01 19:07:31,981 - ocr_e2e - INFO - img_file_name: data/e2ef1/aihub_finance/part_100of100/test/imgs/IMG_OCR_6_F_0000018.png
2022-12-01 19:07:42,607 - ocr_e2e - INFO - img_file_name: data/e2ef1/aihub_finance/part_100of100/test/imgs/IMG_OCR_6_F_0000026.png
2022-12-01 19:07:51,083 - ocr_e2e - INFO - img_file_name: data/e2ef1/aihub_finance/part_100of100/test/imgs/IMG_OCR_6_F_0000118.png
2022-12-01 19:08:02,681 - ocr_e2e - INFO - img_file_name: data/e2ef1/aihub_finance/part_100of100/test/imgs/IMG_OCR_6_F_0000123.png
2022-12-01 19:08:13,324 - ocr_e2e - INFO - img_file_name: data/e2ef1/aihub_finance/part_100of100/test/imgs/IMG_OCR_6_F_0000205.png
2022-12-01 19:08:23,631 - ocr_e2e - INFO - img_file_name: data/e2ef1/aihub_finance/part_100of100/test/imgs/IMG_OCR_6_F_0000219.png
2022-12-01 19:08:33,156 - ocr_e2e - INFO - img_file_name: data/e2ef1/aihub_finance/part_100of100/test/imgs/IMG_OCR_6_F_0000405.png
2022-12-01 19:08:43,480 - ocr_e2e - INFO - img_file_name: data/e2ef1/aihub_finance/part_100of100/test/imgs/IMG_OCR_6_F_0000407.png
2022-12-01 19:08:50,645 - ocr_e2e - INFO - img_file_name: data/e2ef1/aihub_finance/part_100of100/test/imgs/IMG_OCR_6_F_0000417.png
2022-12-01 19:09:00,850 - ocr_e2e - INFO - img_file_name: data/e2ef1/aihub_finance/part_100of100/test/imgs/IMG_OCR_6_F_0000425.png
2022-12-01 19:09:01,524 - ocr_e2e - INFO -
=== MMOCR 타입 Aihub 금융데이터 추론값을 ICDAR 타입으로 변경 ===
2022-12-01 19:09:08,100 - e2e_f1 - INFO - End-to-End F1 Score 평가 시작
2022-12-01 19:09:08,119 - e2e_f1 - INFO - gt_file_name: IMG_OCR_6_F_0000007.txt
2022-12-01 19:09:08,119 - e2e_f1 - INFO - pred_file_name: IMG_OCR_6_F_0000007.txt
2022-12-01 19:09:08,126 - e2e_f1 - INFO - gt_file_name: IMG_OCR_6_F_0000018.txt
2022-12-01 19:09:08,126 - e2e_f1 - INFO - pred_file_name: IMG_OCR_6_F_0000018.txt
2022-12-01 19:09:08,129 - e2e_f1 - INFO - gt_file_name: IMG_OCR_6_F_0000026.txt
2022-12-01 19:09:08,129 - e2e_f1 - INFO - pred_file_name: IMG_OCR_6_F_0000026.txt
2022-12-01 19:09:08,129 - e2e_f1 - INFO - gt_file_name: IMG_OCR_6_F_0000118.txt
2022-12-01 19:09:08,130 - e2e_f1 - INFO - pred_file_name: IMG_OCR_6_F_0000118.txt
2022-12-01 19:09:08,130 - e2e_f1 - INFO - gt_file_name: IMG_OCR_6_F_0000123.txt
2022-12-01 19:09:08,130 - e2e_f1 - INFO - pred_file_name: IMG_OCR_6_F_0000123.txt
2022-12-01 19:09:08,131 - e2e_f1 - INFO - gt_file_name: IMG_OCR_6_F_0000205.txt
2022-12-01 19:09:08,131 - e2e_f1 - INFO - pred_file_name: IMG_OCR_6_F_0000205.txt
2022-12-01 19:09:08,131 - e2e_f1 - INFO - gt_file_name: IMG_OCR_6_F_0000219.txt
2022-12-01 19:09:08,131 - e2e_f1 - INFO - pred_file_name: IMG_OCR_6_F_0000219.txt
2022-12-01 19:09:08,132 - e2e_f1 - INFO - gt_file_name: IMG_OCR_6_F_0000405.txt
2022-12-01 19:09:08,132 - e2e_f1 - INFO - pred_file_name: IMG_OCR_6_F_0000405.txt
2022-12-01 19:09:08,132 - e2e_f1 - INFO - gt_file_name: IMG_OCR_6_F_0000407.txt
2022-12-01 19:09:08,133 - e2e_f1 - INFO - pred_file_name: IMG_OCR_6_F_0000407.txt
2022-12-01 19:09:08,133 - e2e_f1 - INFO - gt_file_name: IMG_OCR_6_F_0000417.txt
2022-12-01 19:09:08,133 - e2e_f1 - INFO - pred_file_name: IMG_OCR_6_F_0000417.txt
2022-12-01 19:09:08,134 - e2e_f1 - INFO - gt_file_name: IMG_OCR_6_F_0000425.txt
2022-12-01 19:09:08,134 - e2e_f1 - INFO - pred_file_name: IMG_OCR_6_F_0000425.txt
2022-12-01 19:09:08,759 - e2e_f1 - INFO - Found 236 preds and 234 gts
2022-12-01 19:09:08,759 - e2e_f1 - INFO - Precision: 0.919492
2022-12-01 19:09:08,759 - e2e_f1 - INFO - Recall: 0.927350
2022-12-01 19:09:08,759 - e2e_f1 - INFO - F1: 0.923404
Bash
복사