Search
🔨

[1_3.2] 금융 recognition 데이터의 유효성을 확인한다.

상태
Done
담당
마감일
2022/11/24
요약
1. 랜덤하게 선택한 200장을 시각화해볼 때 문제가 없어야 하며, 2. 1% 모델로 10% 데이터, 10% 데이터로 100% 데이터가 문제 없이 추론가능해야 함.
선행 태스크
선행 태스크 상태
관련 마일스톤
2 more properties
1% 의 금융 데이터로 학습시킨 모델로 10% 테스트 데이터에 대해 추론을 시키는 경우
from the scratch (Aihub 금융 train)
transfer learned (Aihub 금융 train)
IC2015 word acc
할필요없음
0.0303
1% word acc
할필요없음
0.9354
10% word acc
할필요없음
0.9474
100% word acc
할필요없음
0.9369
•
1% transfer learned
nohup tools/dist_train.sh \ configs/textrecog/sar/sar_resnet31_parallel-decoder_500e_aihubfinance1of100_pretrained.py \ 2 > nohup.out &
Bash
복사
train
ai-solution-lab
Weights & Biases, developer tools for machine learning
https://wandb.ai/ai-solution-lab/mmocr/runs/332i7nao?workspace=user-deer
cp work_dirs/sar_resnet31_parallel-decoder_500e_aihubfinance1of100_pretrained/epoch_500.pth \ pretrained/sar_resnet31_parallel-decoder_500e_aihubfinance1of100_pretrained_zesty-sun-97.pth nohup tools/dist_test.sh \ configs/textrecog/sar/sar_resnet31_parallel-decoder_500e_aihubfinance1of100_pretrained.py \ pretrained/sar_resnet31_parallel-decoder_500e_aihubfinance1of100_pretrained_zesty-sun-97.pth \ 2 > nohup.out &
Bash
복사
eval
11/28 11:44:16 - mmengine - INFO - Epoch(test) [66916/66916] AihubFinance1of100/recog/word_acc: 0.9354 AihubFinance1of100/recog/word_acc_ignore_case: 0.9354 AihubFinance1of100/recog/word_acc_ignore_case_symbol: 0.9699 AihubFinance10of100/recog/word_acc: 0.9474 AihubFinance10of100/recog/word_acc_ignore_case: 0.9474 AihubFinance10of100/recog/word_acc_ignore_case_symbol: 0.9755 AihubFinance100of100/recog/word_acc: 0.9369 AihubFinance100of100/recog/word_acc_ignore_case: 0.9373 AihubFinance100of100/recog/word_acc_ignore_case_symbol: 0.9706 IC15/recog/word_acc: 0.0303 IC15/recog/word_acc_ignore_case: 0.0327 IC15/recog/word_acc_ignore_case_symbol: 0.0342 AihubFinance1of100/recog/char_recall: 0.9897 AihubFinance1of100/recog/char_precision: 0.9892 AihubFinance10of100/recog/char_recall: 0.9912 AihubFinance10of100/recog/char_precision: 0.9921 AihubFinance100of100/recog/char_recall: 0.9900 AihubFinance100of100/recog/char_precision: 0.9909 IC15/recog/char_recall: 0.1938 IC15/recog/char_precision: 0.3663
Bash
복사
eval res
python3 -m mmocr.ocr \ --det-ckpt pretrained/dbnet_resnet18_fpnc_20e_aihubfinance10of100_sparkling-cloud-104.pth \ --det-config configs/textdet/dbnet/dbnet_resnet18_fpnc_20e_aihubfinance10of100.py \ --recog-ckpt pretrained/sar_resnet31_parallel-decoder_500e_aihubfinance1of100_pretrained_zesty-sun-97.pth \ --recog-config configs/textrecog/sar/sar_resnet31_parallel-decoder_500e_aihubfinance1of100_pretrained.py \ data/det/aihub_finance/part_10of100/imgs/IMG_OCR_6_F_00964.png \ --img-out-dir work_dirs/sar_resnet31_parallel-decoder_500e_aihubfinance1of100_pretrained \ --pred-out-file work_dirs/sar_resnet31_parallel-decoder_500e_aihubfinance1of100_pretrained/output.pkl \ --device cpu
Bash
복사
infer
◦
detection model 은 다음을 사용한다.
(할필요없음) 1%, from the scratch
10% 의 금융 데이터로 학습시킨 모델로 100% 테스트 데이터에 대해 추론을 시키는 경우
할 필요 없음
200장 랜덤 샘플링
•
템플릿이라 그런지 문제 없어 보임. 사람이 레이블한 것이 아닌듯? 선직님이 발견하신 real data 제외하고.