ICDAR 2015 test score | from the scratch (ICDAR 2015 train) | pretrained
(no transfer learned) | transfer learned
(ICDAR 2015 train) |
char precision | 0.2925 | 0.9115 | 0.91+ |
char recall | 0.2767 | 0.9180 | 0.91+ |
word acc | 0.0520 | 0.6548 | 0.70+ |
pretrained
11/16 12:58:32 - mmengine - INFO - Epoch(test) [1039/1039] IC15/recog/word_acc: 0.6548 IC15/recog/word_acc_ignore_case: 0.7381 IC15/recog/word_acc_ignore_case_symbol: 0.7607 IC15/recog/char_recall: 0.9180 IC15/recog/char_precision: 0.9115
Bash
복사
transfer learned
•
다음과 같이 pretrained weight 를 추가할 수 있다.
sar_resnet31_parallel-decoder_500e_icdar2015.py
_base_ = [
'_base_sar_resnet31_parallel-decoder.py',
'../_base_/datasets/icdar2015.py',
'../_base_/default_runtime.py',
'../_base_/schedules/schedule_adam_step_500e.py',
]
load_from = './pretrained/sar_resnet31_parallel-decoder_5e_st-sub_mj-sub_sa_real_20220915_171910-04eb4e75.pth'
# dataset settings
train_list = [_base_.ic15_rec_train]
test_list = [_base_.ic15_rec_test]
Bash
복사
•
결과
11/16 14:14:39 - mmengine - INFO - Epoch(val) [39][1039/1039] IC15/recog/word_acc: 0.7102 IC15/recog/word_acc_ignore_case: 0.7366 IC15/recog/word_acc_ignore_case_symbol: 0.7511 IC15/recog/char_recall: 0.9128 IC15/recog/char_precision: 0.9113
Bash
복사