Skip to content

GitLab

Unverified Commit a8a9b2e5 authored by xiaoting's avatar xiaoting Committed by GitHub
Browse files

Merge branch 'dygraph' into fix_prepare

parents 49e42a16 0f5a5d96
Hide whitespace changes
Inline Side-by-side
Showing with 1497 additions and 147 deletions
paddleocr.py
View file @ a8a9b2e5
...@@ -33,104 +33,141 @@ from tools.infer.utility import draw_ocr, str2bool ...@@ -33,104 +33,141 @@ from tools.infer.utility import draw_ocr, str2bool
from ppstructure.utility import init_args, draw_structure_result from ppstructure.utility import init_args, draw_structure_result
from ppstructure.predict_system import OCRSystem, save_structure_res from ppstructure.predict_system import OCRSystem, save_structure_res
__all__ = ['PaddleOCR', 'PPStructure', 'draw_ocr', 'draw_structure_result', 'save_structure_res','download_with_progressbar'] __all__ = [
'PaddleOCR', 'PPStructure', 'draw_ocr', 'draw_structure_result',
model_urls = { 'save_structure_res', 'download_with_progressbar'
'det': { ]
'ch':
'https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_mobile_v2.0_det_infer.tar', SUPPORT_DET_MODEL = ['DB']
'en': VERSION = '2.2.1'
'https://paddleocr.bj.bcebos.com/dygraph_v2.0/multilingual/en_ppocr_mobile_v2.0_det_infer.tar', SUPPORT_REC_MODEL = ['CRNN']
'structure': 'https://paddleocr.bj.bcebos.com/dygraph_v2.0/table/en_ppocr_mobile_v2.0_table_det_infer.tar' BASE_DIR = os.path.expanduser("~/.paddleocr/")
DEFAULT_MODEL_VERSION = '2.0'
MODEL_URLS = {
'2.1': {
'det': {
'ch': {
'url':
'https://paddleocr.bj.bcebos.com/PP-OCRv2/chinese/ch_PP-OCRv2_det_infer.tar',
},
},
'rec': {
'ch': {
'url':
'https://paddleocr.bj.bcebos.com/PP-OCRv2/chinese/ch_PP-OCRv2_rec_infer.tar',
'dict_path': './ppocr/utils/ppocr_keys_v1.txt'
}
}
}, },
'rec': { '2.0': {
'ch': { 'det': {
'url': 'ch': {
'https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_mobile_v2.0_rec_infer.tar', 'url':
'dict_path': './ppocr/utils/ppocr_keys_v1.txt' 'https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_mobile_v2.0_det_infer.tar',
},
'en': {
'url':
'https://paddleocr.bj.bcebos.com/dygraph_v2.0/multilingual/en_ppocr_mobile_v2.0_det_infer.tar',
},
'structure': {
'url':
'https://paddleocr.bj.bcebos.com/dygraph_v2.0/table/en_ppocr_mobile_v2.0_table_det_infer.tar'
}
}, },
'en': { 'rec': {
'url': 'ch': {
'url':
'https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_mobile_v2.0_rec_infer.tar',
'dict_path': './ppocr/utils/ppocr_keys_v1.txt'
},
'en': {
'url':
'https://paddleocr.bj.bcebos.com/dygraph_v2.0/multilingual/en_number_mobile_v2.0_rec_infer.tar', 'https://paddleocr.bj.bcebos.com/dygraph_v2.0/multilingual/en_number_mobile_v2.0_rec_infer.tar',
'dict_path': './ppocr/utils/en_dict.txt' 'dict_path': './ppocr/utils/en_dict.txt'
}, },
'french': { 'french': {
'url': 'url':
'https://paddleocr.bj.bcebos.com/dygraph_v2.0/multilingual/french_mobile_v2.0_rec_infer.tar', 'https://paddleocr.bj.bcebos.com/dygraph_v2.0/multilingual/french_mobile_v2.0_rec_infer.tar',
'dict_path': './ppocr/utils/dict/french_dict.txt' 'dict_path': './ppocr/utils/dict/french_dict.txt'
}, },
'german': { 'german': {
'url': 'url':
'https://paddleocr.bj.bcebos.com/dygraph_v2.0/multilingual/german_mobile_v2.0_rec_infer.tar', 'https://paddleocr.bj.bcebos.com/dygraph_v2.0/multilingual/german_mobile_v2.0_rec_infer.tar',
'dict_path': './ppocr/utils/dict/german_dict.txt' 'dict_path': './ppocr/utils/dict/german_dict.txt'
}, },
'korean': { 'korean': {
'url': 'url':
'https://paddleocr.bj.bcebos.com/dygraph_v2.0/multilingual/korean_mobile_v2.0_rec_infer.tar', 'https://paddleocr.bj.bcebos.com/dygraph_v2.0/multilingual/korean_mobile_v2.0_rec_infer.tar',
'dict_path': './ppocr/utils/dict/korean_dict.txt' 'dict_path': './ppocr/utils/dict/korean_dict.txt'
}, },
'japan': { 'japan': {
'url': 'url':
'https://paddleocr.bj.bcebos.com/dygraph_v2.0/multilingual/japan_mobile_v2.0_rec_infer.tar', 'https://paddleocr.bj.bcebos.com/dygraph_v2.0/multilingual/japan_mobile_v2.0_rec_infer.tar',
'dict_path': './ppocr/utils/dict/japan_dict.txt' 'dict_path': './ppocr/utils/dict/japan_dict.txt'
}, },
'chinese_cht': { 'chinese_cht': {
'url': 'url':
'https://paddleocr.bj.bcebos.com/dygraph_v2.0/multilingual/chinese_cht_mobile_v2.0_rec_infer.tar', 'https://paddleocr.bj.bcebos.com/dygraph_v2.0/multilingual/chinese_cht_mobile_v2.0_rec_infer.tar',
'dict_path': './ppocr/utils/dict/chinese_cht_dict.txt' 'dict_path': './ppocr/utils/dict/chinese_cht_dict.txt'
}, },
'ta': { 'ta': {
'url': 'url':
'https://paddleocr.bj.bcebos.com/dygraph_v2.0/multilingual/ta_mobile_v2.0_rec_infer.tar', 'https://paddleocr.bj.bcebos.com/dygraph_v2.0/multilingual/ta_mobile_v2.0_rec_infer.tar',
'dict_path': './ppocr/utils/dict/ta_dict.txt' 'dict_path': './ppocr/utils/dict/ta_dict.txt'
}, },
'te': { 'te': {
'url': 'url':
'https://paddleocr.bj.bcebos.com/dygraph_v2.0/multilingual/te_mobile_v2.0_rec_infer.tar', 'https://paddleocr.bj.bcebos.com/dygraph_v2.0/multilingual/te_mobile_v2.0_rec_infer.tar',
'dict_path': './ppocr/utils/dict/te_dict.txt' 'dict_path': './ppocr/utils/dict/te_dict.txt'
}, },
'ka': { 'ka': {
'url': 'url':
'https://paddleocr.bj.bcebos.com/dygraph_v2.0/multilingual/ka_mobile_v2.0_rec_infer.tar', 'https://paddleocr.bj.bcebos.com/dygraph_v2.0/multilingual/ka_mobile_v2.0_rec_infer.tar',
'dict_path': './ppocr/utils/dict/ka_dict.txt' 'dict_path': './ppocr/utils/dict/ka_dict.txt'
}, },
'latin': { 'latin': {
'url': 'url':
'https://paddleocr.bj.bcebos.com/dygraph_v2.0/multilingual/latin_ppocr_mobile_v2.0_rec_infer.tar', 'https://paddleocr.bj.bcebos.com/dygraph_v2.0/multilingual/latin_ppocr_mobile_v2.0_rec_infer.tar',
'dict_path': './ppocr/utils/dict/latin_dict.txt' 'dict_path': './ppocr/utils/dict/latin_dict.txt'
}, },
'arabic': { 'arabic': {
'url': 'url':
'https://paddleocr.bj.bcebos.com/dygraph_v2.0/multilingual/arabic_ppocr_mobile_v2.0_rec_infer.tar', 'https://paddleocr.bj.bcebos.com/dygraph_v2.0/multilingual/arabic_ppocr_mobile_v2.0_rec_infer.tar',
'dict_path': './ppocr/utils/dict/arabic_dict.txt' 'dict_path': './ppocr/utils/dict/arabic_dict.txt'
}, },
'cyrillic': { 'cyrillic': {
'url': 'url':
'https://paddleocr.bj.bcebos.com/dygraph_v2.0/multilingual/cyrillic_ppocr_mobile_v2.0_rec_infer.tar', 'https://paddleocr.bj.bcebos.com/dygraph_v2.0/multilingual/cyrillic_ppocr_mobile_v2.0_rec_infer.tar',
'dict_path': './ppocr/utils/dict/cyrillic_dict.txt' 'dict_path': './ppocr/utils/dict/cyrillic_dict.txt'
}, },
'devanagari': { 'devanagari': {
'url': 'url':
'https://paddleocr.bj.bcebos.com/dygraph_v2.0/multilingual/devanagari_ppocr_mobile_v2.0_rec_infer.tar', 'https://paddleocr.bj.bcebos.com/dygraph_v2.0/multilingual/devanagari_ppocr_mobile_v2.0_rec_infer.tar',
'dict_path': './ppocr/utils/dict/devanagari_dict.txt' 'dict_path': './ppocr/utils/dict/devanagari_dict.txt'
},
'structure': {
'url':
'https://paddleocr.bj.bcebos.com/dygraph_v2.0/table/en_ppocr_mobile_v2.0_table_rec_infer.tar',
'dict_path': 'ppocr/utils/dict/table_dict.txt'
}
},
'cls': {
'ch': {
'url':
'https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_mobile_v2.0_cls_infer.tar',
}
}, },
'structure': { 'table': {
'url': 'https://paddleocr.bj.bcebos.com/dygraph_v2.0/table/en_ppocr_mobile_v2.0_table_rec_infer.tar', 'en': {
'dict_path': 'ppocr/utils/dict/table_dict.txt' 'url':
'https://paddleocr.bj.bcebos.com/dygraph_v2.0/table/en_ppocr_mobile_v2.0_table_structure_infer.tar',
'dict_path': 'ppocr/utils/dict/table_structure_dict.txt'
}
} }
},
'cls': 'https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_mobile_v2.0_cls_infer.tar',
'table': {
'url': 'https://paddleocr.bj.bcebos.com/dygraph_v2.0/table/en_ppocr_mobile_v2.0_table_structure_infer.tar',
'dict_path': 'ppocr/utils/dict/table_structure_dict.txt'
} }
} }
SUPPORT_DET_MODEL = ['DB']
VERSION = '2.2'
SUPPORT_REC_MODEL = ['CRNN']
BASE_DIR = os.path.expanduser("~/.paddleocr/")
def parse_args(mMain=True): def parse_args(mMain=True):
import argparse import argparse
...@@ -140,6 +177,7 @@ def parse_args(mMain=True): ...@@ -140,6 +177,7 @@ def parse_args(mMain=True):
parser.add_argument("--det", type=str2bool, default=True) parser.add_argument("--det", type=str2bool, default=True)
parser.add_argument("--rec", type=str2bool, default=True) parser.add_argument("--rec", type=str2bool, default=True)
parser.add_argument("--type", type=str, default='ocr') parser.add_argument("--type", type=str, default='ocr')
parser.add_argument("--version", type=str, default='2.1')
for action in parser._actions: for action in parser._actions:
if action.dest in ['rec_char_dict_path', 'table_char_dict_path']: if action.dest in ['rec_char_dict_path', 'table_char_dict_path']:
...@@ -155,19 +193,19 @@ def parse_args(mMain=True): ...@@ -155,19 +193,19 @@ def parse_args(mMain=True):
def parse_lang(lang): def parse_lang(lang):
latin_lang = [ latin_lang = [
'af', 'az', 'bs', 'cs', 'cy', 'da', 'de', 'es', 'et', 'fr', 'ga', 'af', 'az', 'bs', 'cs', 'cy', 'da', 'de', 'es', 'et', 'fr', 'ga', 'hr',
'hr', 'hu', 'id', 'is', 'it', 'ku', 'la', 'lt', 'lv', 'mi', 'ms', 'hu', 'id', 'is', 'it', 'ku', 'la', 'lt', 'lv', 'mi', 'ms', 'mt', 'nl',
'mt', 'nl', 'no', 'oc', 'pi', 'pl', 'pt', 'ro', 'rs_latin', 'sk', 'no', 'oc', 'pi', 'pl', 'pt', 'ro', 'rs_latin', 'sk', 'sl', 'sq', 'sv',
'sl', 'sq', 'sv', 'sw', 'tl', 'tr', 'uz', 'vi' 'sw', 'tl', 'tr', 'uz', 'vi'
] ]
arabic_lang = ['ar', 'fa', 'ug', 'ur'] arabic_lang = ['ar', 'fa', 'ug', 'ur']
cyrillic_lang = [ cyrillic_lang = [
'ru', 'rs_cyrillic', 'be', 'bg', 'uk', 'mn', 'abq', 'ady', 'kbd', 'ru', 'rs_cyrillic', 'be', 'bg', 'uk', 'mn', 'abq', 'ady', 'kbd', 'ava',
'ava', 'dar', 'inh', 'che', 'lbe', 'lez', 'tab' 'dar', 'inh', 'che', 'lbe', 'lez', 'tab'
] ]
devanagari_lang = [ devanagari_lang = [
'hi', 'mr', 'ne', 'bh', 'mai', 'ang', 'bho', 'mah', 'sck', 'new', 'hi', 'mr', 'ne', 'bh', 'mai', 'ang', 'bho', 'mah', 'sck', 'new', 'gom',
'gom', 'sa', 'bgc' 'sa', 'bgc'
] ]
if lang in latin_lang: if lang in latin_lang:
lang = "latin" lang = "latin"
...@@ -177,9 +215,9 @@ def parse_lang(lang): ...@@ -177,9 +215,9 @@ def parse_lang(lang):
lang = "cyrillic" lang = "cyrillic"
elif lang in devanagari_lang: elif lang in devanagari_lang:
lang = "devanagari" lang = "devanagari"
assert lang in model_urls[ assert lang in MODEL_URLS[DEFAULT_MODEL_VERSION][
'rec'], 'param lang must in {}, but got {}'.format( 'rec'], 'param lang must in {}, but got {}'.format(
model_urls['rec'].keys(), lang) MODEL_URLS[DEFAULT_MODEL_VERSION]['rec'].keys(), lang)
if lang == "ch": if lang == "ch":
det_lang = "ch" det_lang = "ch"
elif lang == 'structure': elif lang == 'structure':
...@@ -189,6 +227,35 @@ def parse_lang(lang): ...@@ -189,6 +227,35 @@ def parse_lang(lang):
return lang, det_lang return lang, det_lang
def get_model_config(version, model_type, lang):
if version not in MODEL_URLS:
logger.warning('version {} not in {}, use version {} instead'.format(
version, MODEL_URLS.keys(), DEFAULT_MODEL_VERSION))
version = DEFAULT_MODEL_VERSION
if model_type not in MODEL_URLS[version]:
if model_type in MODEL_URLS[DEFAULT_MODEL_VERSION]:
logger.warning(
'version {} not support {} models, use version {} instead'.
format(version, model_type, DEFAULT_MODEL_VERSION))
version = DEFAULT_MODEL_VERSION
else:
logger.error('{} models is not support, we only support {}'.format(
model_type, MODEL_URLS[DEFAULT_MODEL_VERSION].keys()))
sys.exit(-1)
if lang not in MODEL_URLS[version][model_type]:
if lang in MODEL_URLS[DEFAULT_MODEL_VERSION][model_type]:
logger.warning('lang {} is not support in {}, use {} instead'.
format(lang, version, DEFAULT_MODEL_VERSION))
version = DEFAULT_MODEL_VERSION
else:
logger.error(
'lang {} is not support, we only support {} for {} models'.
format(lang, MODEL_URLS[DEFAULT_MODEL_VERSION][model_type].keys(
), model_type))
sys.exit(-1)
return MODEL_URLS[version][model_type][lang]
class PaddleOCR(predict_system.TextSystem): class PaddleOCR(predict_system.TextSystem):
def __init__(self, **kwargs): def __init__(self, **kwargs):
""" """
...@@ -204,15 +271,21 @@ class PaddleOCR(predict_system.TextSystem): ...@@ -204,15 +271,21 @@ class PaddleOCR(predict_system.TextSystem):
lang, det_lang = parse_lang(params.lang) lang, det_lang = parse_lang(params.lang)
# init model dir # init model dir
params.det_model_dir, det_url = confirm_model_dir_url(params.det_model_dir, det_model_config = get_model_config(params.version, 'det', det_lang)
os.path.join(BASE_DIR, VERSION, 'ocr', 'det', det_lang), params.det_model_dir, det_url = confirm_model_dir_url(
model_urls['det'][det_lang]) params.det_model_dir,
params.rec_model_dir, rec_url = confirm_model_dir_url(params.rec_model_dir, os.path.join(BASE_DIR, VERSION, 'ocr', 'det', det_lang),
os.path.join(BASE_DIR, VERSION, 'ocr', 'rec', lang), det_model_config['url'])
model_urls['rec'][lang]['url']) rec_model_config = get_model_config(params.version, 'rec', lang)
params.cls_model_dir, cls_url = confirm_model_dir_url(params.cls_model_dir, params.rec_model_dir, rec_url = confirm_model_dir_url(
os.path.join(BASE_DIR, VERSION, 'ocr', 'cls'), params.rec_model_dir,
model_urls['cls']) os.path.join(BASE_DIR, VERSION, 'ocr', 'rec', lang),
rec_model_config['url'])
cls_model_config = get_model_config(params.version, 'cls', 'ch')
params.cls_model_dir, cls_url = confirm_model_dir_url(
params.cls_model_dir,
os.path.join(BASE_DIR, VERSION, 'ocr', 'cls'),
cls_model_config['url'])
# download model # download model
maybe_download(params.det_model_dir, det_url) maybe_download(params.det_model_dir, det_url)
maybe_download(params.rec_model_dir, rec_url) maybe_download(params.rec_model_dir, rec_url)
...@@ -226,7 +299,8 @@ class PaddleOCR(predict_system.TextSystem): ...@@ -226,7 +299,8 @@ class PaddleOCR(predict_system.TextSystem):
sys.exit(0) sys.exit(0)
if params.rec_char_dict_path is None: if params.rec_char_dict_path is None:
params.rec_char_dict_path = str(Path(__file__).parent / model_urls['rec'][lang]['dict_path']) params.rec_char_dict_path = str(
Path(__file__).parent / rec_model_config['dict_path'])
print(params) print(params)
# init det_model and rec_model # init det_model and rec_model
...@@ -293,24 +367,32 @@ class PPStructure(OCRSystem): ...@@ -293,24 +367,32 @@ class PPStructure(OCRSystem):
lang, det_lang = parse_lang(params.lang) lang, det_lang = parse_lang(params.lang)
# init model dir # init model dir
params.det_model_dir, det_url = confirm_model_dir_url(params.det_model_dir, det_model_config = get_model_config(params.version, 'det', det_lang)
os.path.join(BASE_DIR, VERSION, 'ocr', 'det', det_lang), params.det_model_dir, det_url = confirm_model_dir_url(
model_urls['det'][det_lang]) params.det_model_dir,
params.rec_model_dir, rec_url = confirm_model_dir_url(params.rec_model_dir, os.path.join(BASE_DIR, VERSION, 'ocr', 'det', det_lang),
os.path.join(BASE_DIR, VERSION, 'ocr', 'rec', lang), det_model_config['url'])
model_urls['rec'][lang]['url']) rec_model_config = get_model_config(params.version, 'rec', lang)
params.table_model_dir, table_url = confirm_model_dir_url(params.table_model_dir, params.rec_model_dir, rec_url = confirm_model_dir_url(
os.path.join(BASE_DIR, VERSION, 'ocr', 'table'), params.rec_model_dir,
model_urls['table']['url']) os.path.join(BASE_DIR, VERSION, 'ocr', 'rec', lang),
rec_model_config['url'])
table_model_config = get_model_config(params.version, 'table', 'en')
params.table_model_dir, table_url = confirm_model_dir_url(
params.table_model_dir,
os.path.join(BASE_DIR, VERSION, 'ocr', 'table'),
table_model_config['url'])
# download model # download model
maybe_download(params.det_model_dir, det_url) maybe_download(params.det_model_dir, det_url)
maybe_download(params.rec_model_dir, rec_url) maybe_download(params.rec_model_dir, rec_url)
maybe_download(params.table_model_dir, table_url) maybe_download(params.table_model_dir, table_url)
if params.rec_char_dict_path is None: if params.rec_char_dict_path is None:
params.rec_char_dict_path = str(Path(__file__).parent / model_urls['rec'][lang]['dict_path']) params.rec_char_dict_path = str(
Path(__file__).parent / rec_model_config['dict_path'])
if params.table_char_dict_path is None: if params.table_char_dict_path is None:
params.table_char_dict_path = str(Path(__file__).parent / model_urls['table']['dict_path']) params.table_char_dict_path = str(
Path(__file__).parent / table_model_config['dict_path'])
print(params) print(params)
super().__init__(params) super().__init__(params)
...@@ -374,4 +456,3 @@ def main(): ...@@ -374,4 +456,3 @@ def main():
for item in result: for item in result:
item.pop('img') item.pop('img')
logger.info(item) logger.info(item)
ppocr/data/__init__.py
View file @ a8a9b2e5
...@@ -49,14 +49,12 @@ def term_mp(sig_num, frame): ...@@ -49,14 +49,12 @@ def term_mp(sig_num, frame):
os.killpg(pgid, signal.SIGKILL) os.killpg(pgid, signal.SIGKILL)
signal.signal(signal.SIGINT, term_mp)
signal.signal(signal.SIGTERM, term_mp)
def build_dataloader(config, mode, device, logger, seed=None): def build_dataloader(config, mode, device, logger, seed=None):
config = copy.deepcopy(config) config = copy.deepcopy(config)
support_dict = ['SimpleDataSet', 'LMDBDataSet', 'PGDataSet', 'PubTabDataSet'] support_dict = [
'SimpleDataSet', 'LMDBDataSet', 'PGDataSet', 'PubTabDataSet'
]
module_name = config[mode]['dataset']['name'] module_name = config[mode]['dataset']['name']
assert module_name in support_dict, Exception( assert module_name in support_dict, Exception(
'DataSet only support {}'.format(support_dict)) 'DataSet only support {}'.format(support_dict))
...@@ -96,4 +94,8 @@ def build_dataloader(config, mode, device, logger, seed=None): ...@@ -96,4 +94,8 @@ def build_dataloader(config, mode, device, logger, seed=None):
return_list=True, return_list=True,
use_shared_memory=use_shared_memory) use_shared_memory=use_shared_memory)
# support exit using ctrl+c
signal.signal(signal.SIGINT, term_mp)
signal.signal(signal.SIGTERM, term_mp)
return data_loader return data_loader
ppocr/data/imaug/__init__.py
View file @ a8a9b2e5
...@@ -21,7 +21,7 @@ from .make_border_map import MakeBorderMap ...@@ -21,7 +21,7 @@ from .make_border_map import MakeBorderMap
from .make_shrink_map import MakeShrinkMap from .make_shrink_map import MakeShrinkMap
from .random_crop_data import EastRandomCropData, PSERandomCrop from .random_crop_data import EastRandomCropData, PSERandomCrop
from .rec_img_aug import RecAug, RecResizeImg, ClsResizeImg, SRNRecResizeImg from .rec_img_aug import RecAug, RecResizeImg, ClsResizeImg, SRNRecResizeImg, NRTRRecResizeImg
from .randaugment import RandAugment from .randaugment import RandAugment
from .copy_paste import CopyPaste from .copy_paste import CopyPaste
from .operators import * from .operators import *
......
ppocr/data/imaug/label_ops.py
View file @ a8a9b2e5
...@@ -161,6 +161,34 @@ class BaseRecLabelEncode(object): ...@@ -161,6 +161,34 @@ class BaseRecLabelEncode(object):
return text_list return text_list
class NRTRLabelEncode(BaseRecLabelEncode):
""" Convert between text-label and text-index """
def __init__(self,
max_text_length,
character_dict_path=None,
character_type='EN_symbol',
use_space_char=False,
**kwargs):
super(NRTRLabelEncode,
self).__init__(max_text_length, character_dict_path,
character_type, use_space_char)
def __call__(self, data):
text = data['label']
text = self.encode(text)
if text is None:
return None
data['length'] = np.array(len(text))
text.insert(0, 2)
text.append(3)
text = text + [0] * (self.max_text_len - len(text))
data['label'] = np.array(text)
return data
def add_special_char(self, dict_character):
dict_character = ['blank','<unk>','<s>','</s>'] + dict_character
return dict_character
class CTCLabelEncode(BaseRecLabelEncode): class CTCLabelEncode(BaseRecLabelEncode):
""" Convert between text-label and text-index """ """ Convert between text-label and text-index """
......
ppocr/data/imaug/operators.py
View file @ a8a9b2e5
...@@ -57,6 +57,38 @@ class DecodeImage(object): ...@@ -57,6 +57,38 @@ class DecodeImage(object):
return data return data
class NRTRDecodeImage(object):
""" decode image """
def __init__(self, img_mode='RGB', channel_first=False, **kwargs):
self.img_mode = img_mode
self.channel_first = channel_first
def __call__(self, data):
img = data['image']
if six.PY2:
assert type(img) is str and len(
img) > 0, "invalid input 'img' in DecodeImage"
else:
assert type(img) is bytes and len(
img) > 0, "invalid input 'img' in DecodeImage"
img = np.frombuffer(img, dtype='uint8')
img = cv2.imdecode(img, 1)
if img is None:
return None
if self.img_mode == 'GRAY':
img = cv2.cvtColor(img, cv2.COLOR_GRAY2BGR)
elif self.img_mode == 'RGB':
assert img.shape[2] == 3, 'invalid shape of image[%s]' % (img.shape)
img = img[:, :, ::-1]
img = cv2.cvtColor(img,cv2.COLOR_BGR2GRAY)
if self.channel_first:
img = img.transpose((2, 0, 1))
data['image'] = img
return data
class NormalizeImage(object): class NormalizeImage(object):
""" normalize image such as substract mean, divide std """ normalize image such as substract mean, divide std
""" """
...@@ -81,7 +113,7 @@ class NormalizeImage(object): ...@@ -81,7 +113,7 @@ class NormalizeImage(object):
assert isinstance(img, assert isinstance(img,
np.ndarray), "invalid input 'img' in NormalizeImage" np.ndarray), "invalid input 'img' in NormalizeImage"
data['image'] = ( data['image'] = (
img.astype('float32') * self.scale - self.mean) / self.std img.astype('float32') * self.scale - self.mean) / self.std
return data return data
...@@ -112,6 +144,34 @@ class KeepKeys(object): ...@@ -112,6 +144,34 @@ class KeepKeys(object):
return data_list return data_list
class Resize(object):
def __init__(self, size=(640, 640), **kwargs):
self.size = size
def resize_image(self, img):
resize_h, resize_w = self.size
ori_h, ori_w = img.shape[:2] # (h, w, c)
ratio_h = float(resize_h) / ori_h
ratio_w = float(resize_w) / ori_w
img = cv2.resize(img, (int(resize_w), int(resize_h)))
return img, [ratio_h, ratio_w]
def __call__(self, data):
img = data['image']
text_polys = data['polys']
img_resize, [ratio_h, ratio_w] = self.resize_image(img)
new_boxes = []
for box in text_polys:
new_box = []
for cord in box:
new_box.append([cord[0] * ratio_w, cord[1] * ratio_h])
new_boxes.append(new_box)
data['image'] = img_resize
data['polys'] = np.array(new_boxes, dtype=np.float32)
return data
class DetResizeForTest(object): class DetResizeForTest(object):
def __init__(self, **kwargs): def __init__(self, **kwargs):
super(DetResizeForTest, self).__init__() super(DetResizeForTest, self).__init__()
...@@ -183,7 +243,7 @@ class DetResizeForTest(object): ...@@ -183,7 +243,7 @@ class DetResizeForTest(object):
else: else:
ratio = 1. ratio = 1.
elif self.limit_type == 'resize_long': elif self.limit_type == 'resize_long':
ratio = float(limit_side_len) / max(h,w) ratio = float(limit_side_len) / max(h, w)
else: else:
raise Exception('not support limit type, image ') raise Exception('not support limit type, image ')
resize_h = int(h * ratio) resize_h = int(h * ratio)
......
ppocr/data/imaug/rec_img_aug.py
View file @ a8a9b2e5
...@@ -16,7 +16,7 @@ import math ...@@ -16,7 +16,7 @@ import math
import cv2 import cv2
import numpy as np import numpy as np
import random import random
from PIL import Image
from .text_image_aug import tia_perspective, tia_stretch, tia_distort from .text_image_aug import tia_perspective, tia_stretch, tia_distort
...@@ -43,6 +43,25 @@ class ClsResizeImg(object): ...@@ -43,6 +43,25 @@ class ClsResizeImg(object):
return data return data
class NRTRRecResizeImg(object):
def __init__(self, image_shape, resize_type, **kwargs):
self.image_shape = image_shape
self.resize_type = resize_type
def __call__(self, data):
img = data['image']
if self.resize_type == 'PIL':
image_pil = Image.fromarray(np.uint8(img))
img = image_pil.resize(self.image_shape, Image.ANTIALIAS)
img = np.array(img)
if self.resize_type == 'OpenCV':
img = cv2.resize(img, self.image_shape)
norm_img = np.expand_dims(img, -1)
norm_img = norm_img.transpose((2, 0, 1))
data['image'] = norm_img.astype(np.float32) / 128. - 1.
return data
class RecResizeImg(object): class RecResizeImg(object):
def __init__(self, def __init__(self,
image_shape, image_shape,
......
ppocr/losses/__init__.py
View file @ a8a9b2e5
...@@ -25,7 +25,7 @@ from .det_sast_loss import SASTLoss ...@@ -25,7 +25,7 @@ from .det_sast_loss import SASTLoss
from .rec_ctc_loss import CTCLoss from .rec_ctc_loss import CTCLoss
from .rec_att_loss import AttentionLoss from .rec_att_loss import AttentionLoss
from .rec_srn_loss import SRNLoss from .rec_srn_loss import SRNLoss
from .rec_nrtr_loss import NRTRLoss
# cls loss # cls loss
from .cls_loss import ClsLoss from .cls_loss import ClsLoss
...@@ -44,8 +44,9 @@ from .table_att_loss import TableAttentionLoss ...@@ -44,8 +44,9 @@ from .table_att_loss import TableAttentionLoss
def build_loss(config): def build_loss(config):
support_dict = [ support_dict = [
'DBLoss', 'EASTLoss', 'SASTLoss', 'CTCLoss', 'ClsLoss', 'AttentionLoss', 'DBLoss', 'EASTLoss', 'SASTLoss', 'CTCLoss', 'ClsLoss', 'AttentionLoss',
'SRNLoss', 'PGLoss', 'CombinedLoss', 'TableAttentionLoss' 'SRNLoss', 'PGLoss', 'CombinedLoss', 'NRTRLoss', 'TableAttentionLoss'
] ]
config = copy.deepcopy(config) config = copy.deepcopy(config)
module_name = config.pop('name') module_name = config.pop('name')
assert module_name in support_dict, Exception('loss only support {}'.format( assert module_name in support_dict, Exception('loss only support {}'.format(
......
ppocr/losses/cls_loss.py
View file @ a8a9b2e5
...@@ -25,6 +25,6 @@ class ClsLoss(nn.Layer): ...@@ -25,6 +25,6 @@ class ClsLoss(nn.Layer):
self.loss_func = nn.CrossEntropyLoss(reduction='mean') self.loss_func = nn.CrossEntropyLoss(reduction='mean')
def forward(self, predicts, batch): def forward(self, predicts, batch):
label = batch[1] label = batch[1].astype("int64")
loss = self.loss_func(input=predicts, label=label) loss = self.loss_func(input=predicts, label=label)
return {'loss': loss} return {'loss': loss}
ppocr/losses/rec_nrtr_loss.py 0 → 100644
View file @ a8a9b2e5
import paddle
from paddle import nn
import paddle.nn.functional as F
class NRTRLoss(nn.Layer):
def __init__(self, smoothing=True, **kwargs):
super(NRTRLoss, self).__init__()
self.loss_func = nn.CrossEntropyLoss(reduction='mean', ignore_index=0)
self.smoothing = smoothing
def forward(self, pred, batch):
pred = pred.reshape([-1, pred.shape[2]])
max_len = batch[2].max()
tgt = batch[1][:, 1:2 + max_len]
tgt = tgt.reshape([-1])
if self.smoothing:
eps = 0.1
n_class = pred.shape[1]
one_hot = F.one_hot(tgt, pred.shape[1])
one_hot = one_hot * (1 - eps) + (1 - one_hot) * eps / (n_class - 1)
log_prb = F.log_softmax(pred, axis=1)
non_pad_mask = paddle.not_equal(
tgt, paddle.zeros(
tgt.shape, dtype='int64'))
loss = -(one_hot * log_prb).sum(axis=1)
loss = loss.masked_select(non_pad_mask).mean()
else:
loss = self.loss_func(pred, tgt)
return {'loss': loss}
ppocr/metrics/rec_metric.py
View file @ a8a9b2e5
...@@ -57,3 +57,4 @@ class RecMetric(object): ...@@ -57,3 +57,4 @@ class RecMetric(object):
self.correct_num = 0 self.correct_num = 0
self.all_num = 0 self.all_num = 0
self.norm_edit_dis = 0 self.norm_edit_dis = 0
ppocr/modeling/architectures/base_model.py
View file @ a8a9b2e5
...@@ -14,7 +14,6 @@ ...@@ -14,7 +14,6 @@
from __future__ import absolute_import from __future__ import absolute_import
from __future__ import division from __future__ import division
from __future__ import print_function from __future__ import print_function
from paddle import nn from paddle import nn
from ppocr.modeling.transforms import build_transform from ppocr.modeling.transforms import build_transform
from ppocr.modeling.backbones import build_backbone from ppocr.modeling.backbones import build_backbone
......
ppocr/modeling/backbones/__init__.py
View file @ a8a9b2e5
...@@ -26,8 +26,9 @@ def build_backbone(config, model_type): ...@@ -26,8 +26,9 @@ def build_backbone(config, model_type):
from .rec_resnet_vd import ResNet from .rec_resnet_vd import ResNet
from .rec_resnet_fpn import ResNetFPN from .rec_resnet_fpn import ResNetFPN
from .rec_mv1_enhance import MobileNetV1Enhance from .rec_mv1_enhance import MobileNetV1Enhance
from .rec_nrtr_mtb import MTB
support_dict = [ support_dict = [
"MobileNetV1Enhance", "MobileNetV3", "ResNet", "ResNetFPN" 'MobileNetV1Enhance', 'MobileNetV3', 'ResNet', 'ResNetFPN', 'MTB'
] ]
elif model_type == "e2e": elif model_type == "e2e":
from .e2e_resnet_vd_pg import ResNet from .e2e_resnet_vd_pg import ResNet
......
ppocr/modeling/backbones/rec_nrtr_mtb.py 0 → 100644
View file @ a8a9b2e5
# copyright (c) 2020 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from paddle import nn
class MTB(nn.Layer):
def __init__(self, cnn_num, in_channels):
super(MTB, self).__init__()
self.block = nn.Sequential()
self.out_channels = in_channels
self.cnn_num = cnn_num
if self.cnn_num == 2:
for i in range(self.cnn_num):
self.block.add_sublayer(
'conv_{}'.format(i),
nn.Conv2D(
in_channels=in_channels
if i == 0 else 32 * (2**(i - 1)),
out_channels=32 * (2**i),
kernel_size=3,
stride=2,
padding=1))
self.block.add_sublayer('relu_{}'.format(i), nn.ReLU())
self.block.add_sublayer('bn_{}'.format(i),
nn.BatchNorm2D(32 * (2**i)))
def forward(self, images):
x = self.block(images)
if self.cnn_num == 2:
# (b, w, h, c)
x = x.transpose([0, 3, 2, 1])
x_shape = x.shape
x = x.reshape([x_shape[0], x_shape[1], x_shape[2] * x_shape[3]])
return x
ppocr/modeling/heads/__init__.py
View file @ a8a9b2e5
...@@ -26,12 +26,14 @@ def build_head(config): ...@@ -26,12 +26,14 @@ def build_head(config):
from .rec_ctc_head import CTCHead from .rec_ctc_head import CTCHead
from .rec_att_head import AttentionHead from .rec_att_head import AttentionHead
from .rec_srn_head import SRNHead from .rec_srn_head import SRNHead
from .rec_nrtr_head import Transformer
# cls head # cls head
from .cls_head import ClsHead from .cls_head import ClsHead
support_dict = [ support_dict = [
'DBHead', 'EASTHead', 'SASTHead', 'CTCHead', 'ClsHead', 'AttentionHead', 'DBHead', 'EASTHead', 'SASTHead', 'CTCHead', 'ClsHead', 'AttentionHead',
'SRNHead', 'PGHead', 'TableAttentionHead'] 'SRNHead', 'PGHead', 'Transformer', 'TableAttentionHead'
]
#table head #table head
from .table_att_head import TableAttentionHead from .table_att_head import TableAttentionHead
......
ppocr/modeling/heads/multiheadAttention.py 0 → 100755
View file @ a8a9b2e5
# copyright (c) 2021 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import paddle
from paddle import nn
import paddle.nn.functional as F
from paddle.nn import Linear
from paddle.nn.initializer import XavierUniform as xavier_uniform_
from paddle.nn.initializer import Constant as constant_
from paddle.nn.initializer import XavierNormal as xavier_normal_
zeros_ = constant_(value=0.)
ones_ = constant_(value=1.)
class MultiheadAttention(nn.Layer):
"""Allows the model to jointly attend to information
from different representation subspaces.
See reference: Attention Is All You Need
.. math::
\text{MultiHead}(Q, K, V) = \text{Concat}(head_1,\dots,head_h)W^O
\text{where} head_i = \text{Attention}(QW_i^Q, KW_i^K, VW_i^V)
Args:
embed_dim: total dimension of the model
num_heads: parallel attention layers, or heads
"""
def __init__(self,
embed_dim,
num_heads,
dropout=0.,
bias=True,
add_bias_kv=False,
add_zero_attn=False):
super(MultiheadAttention, self).__init__()
self.embed_dim = embed_dim
self.num_heads = num_heads
self.dropout = dropout
self.head_dim = embed_dim // num_heads
assert self.head_dim * num_heads == self.embed_dim, "embed_dim must be divisible by num_heads"
self.scaling = self.head_dim**-0.5
self.out_proj = Linear(embed_dim, embed_dim, bias_attr=bias)
self._reset_parameters()
self.conv1 = paddle.nn.Conv2D(
in_channels=embed_dim, out_channels=embed_dim, kernel_size=(1, 1))
self.conv2 = paddle.nn.Conv2D(
in_channels=embed_dim, out_channels=embed_dim, kernel_size=(1, 1))
self.conv3 = paddle.nn.Conv2D(
in_channels=embed_dim, out_channels=embed_dim, kernel_size=(1, 1))
def _reset_parameters(self):
xavier_uniform_(self.out_proj.weight)
def forward(self,
query,
key,
value,
key_padding_mask=None,
incremental_state=None,
need_weights=True,
static_kv=False,
attn_mask=None):
"""
Inputs of forward function
query: [target length, batch size, embed dim]
key: [sequence length, batch size, embed dim]
value: [sequence length, batch size, embed dim]
key_padding_mask: if True, mask padding based on batch size
incremental_state: if provided, previous time steps are cashed
need_weights: output attn_output_weights
static_kv: key and value are static
Outputs of forward function
attn_output: [target length, batch size, embed dim]
attn_output_weights: [batch size, target length, sequence length]
"""
tgt_len, bsz, embed_dim = query.shape
assert embed_dim == self.embed_dim
assert list(query.shape) == [tgt_len, bsz, embed_dim]
assert key.shape == value.shape
q = self._in_proj_q(query)
k = self._in_proj_k(key)
v = self._in_proj_v(value)
q *= self.scaling
q = q.reshape([tgt_len, bsz * self.num_heads, self.head_dim]).transpose(
[1, 0, 2])
k = k.reshape([-1, bsz * self.num_heads, self.head_dim]).transpose(
[1, 0, 2])
v = v.reshape([-1, bsz * self.num_heads, self.head_dim]).transpose(
[1, 0, 2])
src_len = k.shape[1]
if key_padding_mask is not None:
assert key_padding_mask.shape[0] == bsz
assert key_padding_mask.shape[1] == src_len
attn_output_weights = paddle.bmm(q, k.transpose([0, 2, 1]))
assert list(attn_output_weights.
shape) == [bsz * self.num_heads, tgt_len, src_len]
if attn_mask is not None:
attn_mask = attn_mask.unsqueeze(0)
attn_output_weights += attn_mask
if key_padding_mask is not None:
attn_output_weights = attn_output_weights.reshape(
[bsz, self.num_heads, tgt_len, src_len])
key = key_padding_mask.unsqueeze(1).unsqueeze(2).astype('float32')
y = paddle.full(shape=key.shape, dtype='float32', fill_value='-inf')
y = paddle.where(key == 0., key, y)
attn_output_weights += y
attn_output_weights = attn_output_weights.reshape(
[bsz * self.num_heads, tgt_len, src_len])
attn_output_weights = F.softmax(
attn_output_weights.astype('float32'),
axis=-1,
dtype=paddle.float32 if attn_output_weights.dtype == paddle.float16
else attn_output_weights.dtype)
attn_output_weights = F.dropout(
attn_output_weights, p=self.dropout, training=self.training)
attn_output = paddle.bmm(attn_output_weights, v)
assert list(attn_output.
shape) == [bsz * self.num_heads, tgt_len, self.head_dim]
attn_output = attn_output.transpose([1, 0, 2]).reshape(
[tgt_len, bsz, embed_dim])
attn_output = self.out_proj(attn_output)
if need_weights:
# average attention weights over heads
attn_output_weights = attn_output_weights.reshape(
[bsz, self.num_heads, tgt_len, src_len])
attn_output_weights = attn_output_weights.sum(
axis=1) / self.num_heads
else:
attn_output_weights = None
return attn_output, attn_output_weights
def _in_proj_q(self, query):
query = query.transpose([1, 2, 0])
query = paddle.unsqueeze(query, axis=2)
res = self.conv1(query)
res = paddle.squeeze(res, axis=2)
res = res.transpose([2, 0, 1])
return res
def _in_proj_k(self, key):
key = key.transpose([1, 2, 0])
key = paddle.unsqueeze(key, axis=2)
res = self.conv2(key)
res = paddle.squeeze(res, axis=2)
res = res.transpose([2, 0, 1])
return res
def _in_proj_v(self, value):
value = value.transpose([1, 2, 0]) #(1, 2, 0)
value = paddle.unsqueeze(value, axis=2)
res = self.conv3(value)
res = paddle.squeeze(res, axis=2)
res = res.transpose([2, 0, 1])
return res
ppocr/modeling/heads/rec_nrtr_head.py 0 → 100644
View file @ a8a9b2e5
# copyright (c) 2021 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import math
import paddle
import copy
from paddle import nn
import paddle.nn.functional as F
from paddle.nn import LayerList
from paddle.nn.initializer import XavierNormal as xavier_uniform_
from paddle.nn import Dropout, Linear, LayerNorm, Conv2D
import numpy as np
from ppocr.modeling.heads.multiheadAttention import MultiheadAttention
from paddle.nn.initializer import Constant as constant_
from paddle.nn.initializer import XavierNormal as xavier_normal_
zeros_ = constant_(value=0.)
ones_ = constant_(value=1.)
class Transformer(nn.Layer):
"""A transformer model. User is able to modify the attributes as needed. The architechture
is based on the paper "Attention Is All You Need". Ashish Vaswani, Noam Shazeer,
Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and
Illia Polosukhin. 2017. Attention is all you need. In Advances in Neural Information
Processing Systems, pages 6000-6010.
Args:
d_model: the number of expected features in the encoder/decoder inputs (default=512).
nhead: the number of heads in the multiheadattention models (default=8).
num_encoder_layers: the number of sub-encoder-layers in the encoder (default=6).
num_decoder_layers: the number of sub-decoder-layers in the decoder (default=6).
dim_feedforward: the dimension of the feedforward network model (default=2048).
dropout: the dropout value (default=0.1).
custom_encoder: custom encoder (default=None).
custom_decoder: custom decoder (default=None).
"""
def __init__(self,
d_model=512,
nhead=8,
num_encoder_layers=6,
beam_size=0,
num_decoder_layers=6,
dim_feedforward=1024,
attention_dropout_rate=0.0,
residual_dropout_rate=0.1,
custom_encoder=None,
custom_decoder=None,
in_channels=0,
out_channels=0,
dst_vocab_size=99,
scale_embedding=True):
super(Transformer, self).__init__()
self.embedding = Embeddings(
d_model=d_model,
vocab=dst_vocab_size,
padding_idx=0,
scale_embedding=scale_embedding)
self.positional_encoding = PositionalEncoding(
dropout=residual_dropout_rate,
dim=d_model, )
if custom_encoder is not None:
self.encoder = custom_encoder
else:
if num_encoder_layers > 0:
encoder_layer = TransformerEncoderLayer(
d_model, nhead, dim_feedforward, attention_dropout_rate,
residual_dropout_rate)
self.encoder = TransformerEncoder(encoder_layer,
num_encoder_layers)
else:
self.encoder = None
if custom_decoder is not None:
self.decoder = custom_decoder
else:
decoder_layer = TransformerDecoderLayer(
d_model, nhead, dim_feedforward, attention_dropout_rate,
residual_dropout_rate)
self.decoder = TransformerDecoder(decoder_layer, num_decoder_layers)
self._reset_parameters()
self.beam_size = beam_size
self.d_model = d_model
self.nhead = nhead
self.tgt_word_prj = nn.Linear(d_model, dst_vocab_size, bias_attr=False)
w0 = np.random.normal(0.0, d_model**-0.5,
(d_model, dst_vocab_size)).astype(np.float32)
self.tgt_word_prj.weight.set_value(w0)
self.apply(self._init_weights)
def _init_weights(self, m):
if isinstance(m, nn.Conv2D):
xavier_normal_(m.weight)
if m.bias is not None:
zeros_(m.bias)
def forward_train(self, src, tgt):
tgt = tgt[:, :-1]
tgt_key_padding_mask = self.generate_padding_mask(tgt)
tgt = self.embedding(tgt).transpose([1, 0, 2])
tgt = self.positional_encoding(tgt)
tgt_mask = self.generate_square_subsequent_mask(tgt.shape[0])
if self.encoder is not None:
src = self.positional_encoding(src.transpose([1, 0, 2]))
memory = self.encoder(src)
else:
memory = src.squeeze(2).transpose([2, 0, 1])
output = self.decoder(
tgt,
memory,
tgt_mask=tgt_mask,
memory_mask=None,
tgt_key_padding_mask=tgt_key_padding_mask,
memory_key_padding_mask=None)
output = output.transpose([1, 0, 2])
logit = self.tgt_word_prj(output)
return logit
def forward(self, src, targets=None):
"""Take in and process masked source/target sequences.
Args:
src: the sequence to the encoder (required).
tgt: the sequence to the decoder (required).
Shape:
- src: :math:`(S, N, E)`.
- tgt: :math:`(T, N, E)`.
Examples:
>>> output = transformer_model(src, tgt)
"""
if self.training:
max_len = targets[1].max()
tgt = targets[0][:, :2 + max_len]
return self.forward_train(src, tgt)
else:
if self.beam_size > 0:
return self.forward_beam(src)
else:
return self.forward_test(src)
def forward_test(self, src):
bs = src.shape[0]
if self.encoder is not None:
src = self.positional_encoding(src.transpose([1, 0, 2]))
memory = self.encoder(src)
else:
memory = src.squeeze(2).transpose([2, 0, 1])
dec_seq = paddle.full((bs, 1), 2, dtype=paddle.int64)
for len_dec_seq in range(1, 25):
src_enc = memory.clone()
tgt_key_padding_mask = self.generate_padding_mask(dec_seq)
dec_seq_embed = self.embedding(dec_seq).transpose([1, 0, 2])
dec_seq_embed = self.positional_encoding(dec_seq_embed)
tgt_mask = self.generate_square_subsequent_mask(dec_seq_embed.shape[
0])
output = self.decoder(
dec_seq_embed,
src_enc,
tgt_mask=tgt_mask,
memory_mask=None,
tgt_key_padding_mask=tgt_key_padding_mask,
memory_key_padding_mask=None)
dec_output = output.transpose([1, 0, 2])
dec_output = dec_output[:,
-1, :] # Pick the last step: (bh * bm) * d_h
word_prob = F.log_softmax(self.tgt_word_prj(dec_output), axis=1)
word_prob = word_prob.reshape([1, bs, -1])
preds_idx = word_prob.argmax(axis=2)
if paddle.equal_all(
preds_idx[-1],
paddle.full(
preds_idx[-1].shape, 3, dtype='int64')):
break
preds_prob = word_prob.max(axis=2)
dec_seq = paddle.concat(
[dec_seq, preds_idx.reshape([-1, 1])], axis=1)
return dec_seq
def forward_beam(self, images):
''' Translation work in one batch '''
def get_inst_idx_to_tensor_position_map(inst_idx_list):
''' Indicate the position of an instance in a tensor. '''
return {
inst_idx: tensor_position
for tensor_position, inst_idx in enumerate(inst_idx_list)
}
def collect_active_part(beamed_tensor, curr_active_inst_idx,
n_prev_active_inst, n_bm):
''' Collect tensor parts associated to active instances. '''
_, *d_hs = beamed_tensor.shape
n_curr_active_inst = len(curr_active_inst_idx)
new_shape = (n_curr_active_inst * n_bm, *d_hs)
beamed_tensor = beamed_tensor.reshape([n_prev_active_inst, -1])
beamed_tensor = beamed_tensor.index_select(
paddle.to_tensor(curr_active_inst_idx), axis=0)
beamed_tensor = beamed_tensor.reshape([*new_shape])
return beamed_tensor
def collate_active_info(src_enc, inst_idx_to_position_map,
active_inst_idx_list):
# Sentences which are still active are collected,
# so the decoder will not run on completed sentences.
n_prev_active_inst = len(inst_idx_to_position_map)
active_inst_idx = [
inst_idx_to_position_map[k] for k in active_inst_idx_list
]
active_inst_idx = paddle.to_tensor(active_inst_idx, dtype='int64')
active_src_enc = collect_active_part(
src_enc.transpose([1, 0, 2]), active_inst_idx,
n_prev_active_inst, n_bm).transpose([1, 0, 2])
active_inst_idx_to_position_map = get_inst_idx_to_tensor_position_map(
active_inst_idx_list)
return active_src_enc, active_inst_idx_to_position_map
def beam_decode_step(inst_dec_beams, len_dec_seq, enc_output,
inst_idx_to_position_map, n_bm,
memory_key_padding_mask):
''' Decode and update beam status, and then return active beam idx '''
def prepare_beam_dec_seq(inst_dec_beams, len_dec_seq):
dec_partial_seq = [
b.get_current_state() for b in inst_dec_beams if not b.done
]
dec_partial_seq = paddle.stack(dec_partial_seq)
dec_partial_seq = dec_partial_seq.reshape([-1, len_dec_seq])
return dec_partial_seq
def prepare_beam_memory_key_padding_mask(
inst_dec_beams, memory_key_padding_mask, n_bm):
keep = []
for idx in (memory_key_padding_mask):
if not inst_dec_beams[idx].done:
keep.append(idx)
memory_key_padding_mask = memory_key_padding_mask[
paddle.to_tensor(keep)]
len_s = memory_key_padding_mask.shape[-1]
n_inst = memory_key_padding_mask.shape[0]
memory_key_padding_mask = paddle.concat(
[memory_key_padding_mask for i in range(n_bm)], axis=1)
memory_key_padding_mask = memory_key_padding_mask.reshape(
[n_inst * n_bm, len_s]) #repeat(1, n_bm)
return memory_key_padding_mask
def predict_word(dec_seq, enc_output, n_active_inst, n_bm,
memory_key_padding_mask):
tgt_key_padding_mask = self.generate_padding_mask(dec_seq)
dec_seq = self.embedding(dec_seq).transpose([1, 0, 2])
dec_seq = self.positional_encoding(dec_seq)
tgt_mask = self.generate_square_subsequent_mask(dec_seq.shape[
0])
dec_output = self.decoder(
dec_seq,
enc_output,
tgt_mask=tgt_mask,
tgt_key_padding_mask=tgt_key_padding_mask,
memory_key_padding_mask=memory_key_padding_mask,
).transpose([1, 0, 2])
dec_output = dec_output[:,
-1, :] # Pick the last step: (bh * bm) * d_h
word_prob = F.log_softmax(self.tgt_word_prj(dec_output), axis=1)
word_prob = word_prob.reshape([n_active_inst, n_bm, -1])
return word_prob
def collect_active_inst_idx_list(inst_beams, word_prob,
inst_idx_to_position_map):
active_inst_idx_list = []
for inst_idx, inst_position in inst_idx_to_position_map.items():
is_inst_complete = inst_beams[inst_idx].advance(word_prob[
inst_position])
if not is_inst_complete:
active_inst_idx_list += [inst_idx]
return active_inst_idx_list
n_active_inst = len(inst_idx_to_position_map)
dec_seq = prepare_beam_dec_seq(inst_dec_beams, len_dec_seq)
memory_key_padding_mask = None
word_prob = predict_word(dec_seq, enc_output, n_active_inst, n_bm,
memory_key_padding_mask)
# Update the beam with predicted word prob information and collect incomplete instances
active_inst_idx_list = collect_active_inst_idx_list(
inst_dec_beams, word_prob, inst_idx_to_position_map)
return active_inst_idx_list
def collect_hypothesis_and_scores(inst_dec_beams, n_best):
all_hyp, all_scores = [], []
for inst_idx in range(len(inst_dec_beams)):
scores, tail_idxs = inst_dec_beams[inst_idx].sort_scores()
all_scores += [scores[:n_best]]
hyps = [
inst_dec_beams[inst_idx].get_hypothesis(i)
for i in tail_idxs[:n_best]
]
all_hyp += [hyps]
return all_hyp, all_scores
with paddle.no_grad():
#-- Encode
if self.encoder is not None:
src = self.positional_encoding(images.transpose([1, 0, 2]))
src_enc = self.encoder(src).transpose([1, 0, 2])
else:
src_enc = images.squeeze(2).transpose([0, 2, 1])
#-- Repeat data for beam search
n_bm = self.beam_size
n_inst, len_s, d_h = src_enc.shape
src_enc = paddle.concat([src_enc for i in range(n_bm)], axis=1)
src_enc = src_enc.reshape([n_inst * n_bm, len_s, d_h]).transpose(
[1, 0, 2])
#-- Prepare beams
inst_dec_beams = [Beam(n_bm) for _ in range(n_inst)]
#-- Bookkeeping for active or not
active_inst_idx_list = list(range(n_inst))
inst_idx_to_position_map = get_inst_idx_to_tensor_position_map(
active_inst_idx_list)
#-- Decode
for len_dec_seq in range(1, 25):
src_enc_copy = src_enc.clone()
active_inst_idx_list = beam_decode_step(
inst_dec_beams, len_dec_seq, src_enc_copy,
inst_idx_to_position_map, n_bm, None)
if not active_inst_idx_list:
break # all instances have finished their path to <EOS>
src_enc, inst_idx_to_position_map = collate_active_info(
src_enc_copy, inst_idx_to_position_map,
active_inst_idx_list)
batch_hyp, batch_scores = collect_hypothesis_and_scores(inst_dec_beams,
1)
result_hyp = []
for bs_hyp in batch_hyp:
bs_hyp_pad = bs_hyp[0] + [3] * (25 - len(bs_hyp[0]))
result_hyp.append(bs_hyp_pad)
return paddle.to_tensor(np.array(result_hyp), dtype=paddle.int64)
def generate_square_subsequent_mask(self, sz):
"""Generate a square mask for the sequence. The masked positions are filled with float('-inf').
Unmasked positions are filled with float(0.0).
"""
mask = paddle.zeros([sz, sz], dtype='float32')
mask_inf = paddle.triu(
paddle.full(
shape=[sz, sz], dtype='float32', fill_value='-inf'),
diagonal=1)
mask = mask + mask_inf
return mask
def generate_padding_mask(self, x):
padding_mask = x.equal(paddle.to_tensor(0, dtype=x.dtype))
return padding_mask
def _reset_parameters(self):
"""Initiate parameters in the transformer model."""
for p in self.parameters():
if p.dim() > 1:
xavier_uniform_(p)
class TransformerEncoder(nn.Layer):
"""TransformerEncoder is a stack of N encoder layers
Args:
encoder_layer: an instance of the TransformerEncoderLayer() class (required).
num_layers: the number of sub-encoder-layers in the encoder (required).
norm: the layer normalization component (optional).
"""
def __init__(self, encoder_layer, num_layers):
super(TransformerEncoder, self).__init__()
self.layers = _get_clones(encoder_layer, num_layers)
self.num_layers = num_layers
def forward(self, src):
"""Pass the input through the endocder layers in turn.
Args:
src: the sequnce to the encoder (required).
mask: the mask for the src sequence (optional).
src_key_padding_mask: the mask for the src keys per batch (optional).
"""
output = src
for i in range(self.num_layers):
output = self.layers[i](output,
src_mask=None,
src_key_padding_mask=None)
return output
class TransformerDecoder(nn.Layer):
"""TransformerDecoder is a stack of N decoder layers
Args:
decoder_layer: an instance of the TransformerDecoderLayer() class (required).
num_layers: the number of sub-decoder-layers in the decoder (required).
norm: the layer normalization component (optional).
"""
def __init__(self, decoder_layer, num_layers):
super(TransformerDecoder, self).__init__()
self.layers = _get_clones(decoder_layer, num_layers)
self.num_layers = num_layers
def forward(self,
tgt,
memory,
tgt_mask=None,
memory_mask=None,
tgt_key_padding_mask=None,
memory_key_padding_mask=None):
"""Pass the inputs (and mask) through the decoder layer in turn.
Args:
tgt: the sequence to the decoder (required).
memory: the sequnce from the last layer of the encoder (required).
tgt_mask: the mask for the tgt sequence (optional).
memory_mask: the mask for the memory sequence (optional).
tgt_key_padding_mask: the mask for the tgt keys per batch (optional).
memory_key_padding_mask: the mask for the memory keys per batch (optional).
"""
output = tgt
for i in range(self.num_layers):
output = self.layers[i](
output,
memory,
tgt_mask=tgt_mask,
memory_mask=memory_mask,
tgt_key_padding_mask=tgt_key_padding_mask,
memory_key_padding_mask=memory_key_padding_mask)
return output
class TransformerEncoderLayer(nn.Layer):
"""TransformerEncoderLayer is made up of self-attn and feedforward network.
This standard encoder layer is based on the paper "Attention Is All You Need".
Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez,
Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In Advances in
Neural Information Processing Systems, pages 6000-6010. Users may modify or implement
in a different way during application.
Args:
d_model: the number of expected features in the input (required).
nhead: the number of heads in the multiheadattention models (required).
dim_feedforward: the dimension of the feedforward network model (default=2048).
dropout: the dropout value (default=0.1).
"""
def __init__(self,
d_model,
nhead,
dim_feedforward=2048,
attention_dropout_rate=0.0,
residual_dropout_rate=0.1):
super(TransformerEncoderLayer, self).__init__()
self.self_attn = MultiheadAttention(
d_model, nhead, dropout=attention_dropout_rate)
self.conv1 = Conv2D(
in_channels=d_model,
out_channels=dim_feedforward,
kernel_size=(1, 1))
self.conv2 = Conv2D(
in_channels=dim_feedforward,
out_channels=d_model,
kernel_size=(1, 1))
self.norm1 = LayerNorm(d_model)
self.norm2 = LayerNorm(d_model)
self.dropout1 = Dropout(residual_dropout_rate)
self.dropout2 = Dropout(residual_dropout_rate)
def forward(self, src, src_mask=None, src_key_padding_mask=None):
"""Pass the input through the endocder layer.
Args:
src: the sequnce to the encoder layer (required).
src_mask: the mask for the src sequence (optional).
src_key_padding_mask: the mask for the src keys per batch (optional).
"""
src2 = self.self_attn(
src,
src,
src,
attn_mask=src_mask,
key_padding_mask=src_key_padding_mask)[0]
src = src + self.dropout1(src2)
src = self.norm1(src)
src = src.transpose([1, 2, 0])
src = paddle.unsqueeze(src, 2)
src2 = self.conv2(F.relu(self.conv1(src)))
src2 = paddle.squeeze(src2, 2)
src2 = src2.transpose([2, 0, 1])
src = paddle.squeeze(src, 2)
src = src.transpose([2, 0, 1])
src = src + self.dropout2(src2)
src = self.norm2(src)
return src
class TransformerDecoderLayer(nn.Layer):
"""TransformerDecoderLayer is made up of self-attn, multi-head-attn and feedforward network.
This standard decoder layer is based on the paper "Attention Is All You Need".
Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez,
Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In Advances in
Neural Information Processing Systems, pages 6000-6010. Users may modify or implement
in a different way during application.
Args:
d_model: the number of expected features in the input (required).
nhead: the number of heads in the multiheadattention models (required).
dim_feedforward: the dimension of the feedforward network model (default=2048).
dropout: the dropout value (default=0.1).
"""
def __init__(self,
d_model,
nhead,
dim_feedforward=2048,
attention_dropout_rate=0.0,
residual_dropout_rate=0.1):
super(TransformerDecoderLayer, self).__init__()
self.self_attn = MultiheadAttention(
d_model, nhead, dropout=attention_dropout_rate)
self.multihead_attn = MultiheadAttention(
d_model, nhead, dropout=attention_dropout_rate)
self.conv1 = Conv2D(
in_channels=d_model,
out_channels=dim_feedforward,
kernel_size=(1, 1))
self.conv2 = Conv2D(
in_channels=dim_feedforward,
out_channels=d_model,
kernel_size=(1, 1))
self.norm1 = LayerNorm(d_model)
self.norm2 = LayerNorm(d_model)
self.norm3 = LayerNorm(d_model)
self.dropout1 = Dropout(residual_dropout_rate)
self.dropout2 = Dropout(residual_dropout_rate)
self.dropout3 = Dropout(residual_dropout_rate)
def forward(self,
tgt,
memory,
tgt_mask=None,
memory_mask=None,
tgt_key_padding_mask=None,
memory_key_padding_mask=None):
"""Pass the inputs (and mask) through the decoder layer.
Args:
tgt: the sequence to the decoder layer (required).
memory: the sequnce from the last layer of the encoder (required).
tgt_mask: the mask for the tgt sequence (optional).
memory_mask: the mask for the memory sequence (optional).
tgt_key_padding_mask: the mask for the tgt keys per batch (optional).
memory_key_padding_mask: the mask for the memory keys per batch (optional).
"""
tgt2 = self.self_attn(
tgt,
tgt,
tgt,
attn_mask=tgt_mask,
key_padding_mask=tgt_key_padding_mask)[0]
tgt = tgt + self.dropout1(tgt2)
tgt = self.norm1(tgt)
tgt2 = self.multihead_attn(
tgt,
memory,
memory,
attn_mask=memory_mask,
key_padding_mask=memory_key_padding_mask)[0]
tgt = tgt + self.dropout2(tgt2)
tgt = self.norm2(tgt)
# default
tgt = tgt.transpose([1, 2, 0])
tgt = paddle.unsqueeze(tgt, 2)
tgt2 = self.conv2(F.relu(self.conv1(tgt)))
tgt2 = paddle.squeeze(tgt2, 2)
tgt2 = tgt2.transpose([2, 0, 1])
tgt = paddle.squeeze(tgt, 2)
tgt = tgt.transpose([2, 0, 1])
tgt = tgt + self.dropout3(tgt2)
tgt = self.norm3(tgt)
return tgt
def _get_clones(module, N):
return LayerList([copy.deepcopy(module) for i in range(N)])
class PositionalEncoding(nn.Layer):
"""Inject some information about the relative or absolute position of the tokens
in the sequence. The positional encodings have the same dimension as
the embeddings, so that the two can be summed. Here, we use sine and cosine
functions of different frequencies.
.. math::
\text{PosEncoder}(pos, 2i) = sin(pos/10000^(2i/d_model))
\text{PosEncoder}(pos, 2i+1) = cos(pos/10000^(2i/d_model))
\text{where pos is the word position and i is the embed idx)
Args:
d_model: the embed dim (required).
dropout: the dropout value (default=0.1).
max_len: the max. length of the incoming sequence (default=5000).
Examples:
>>> pos_encoder = PositionalEncoding(d_model)
"""
def __init__(self, dropout, dim, max_len=5000):
super(PositionalEncoding, self).__init__()
self.dropout = nn.Dropout(p=dropout)
pe = paddle.zeros([max_len, dim])
position = paddle.arange(0, max_len, dtype=paddle.float32).unsqueeze(1)
div_term = paddle.exp(
paddle.arange(0, dim, 2).astype('float32') *
(-math.log(10000.0) / dim))
pe[:, 0::2] = paddle.sin(position * div_term)
pe[:, 1::2] = paddle.cos(position * div_term)
pe = pe.unsqueeze(0)
pe = pe.transpose([1, 0, 2])
self.register_buffer('pe', pe)
def forward(self, x):
"""Inputs of forward function
Args:
x: the sequence fed to the positional encoder model (required).
Shape:
x: [sequence length, batch size, embed dim]
output: [sequence length, batch size, embed dim]
Examples:
>>> output = pos_encoder(x)
"""
x = x + self.pe[:x.shape[0], :]
return self.dropout(x)
class PositionalEncoding_2d(nn.Layer):
"""Inject some information about the relative or absolute position of the tokens
in the sequence. The positional encodings have the same dimension as
the embeddings, so that the two can be summed. Here, we use sine and cosine
functions of different frequencies.
.. math::
\text{PosEncoder}(pos, 2i) = sin(pos/10000^(2i/d_model))
\text{PosEncoder}(pos, 2i+1) = cos(pos/10000^(2i/d_model))
\text{where pos is the word position and i is the embed idx)
Args:
d_model: the embed dim (required).
dropout: the dropout value (default=0.1).
max_len: the max. length of the incoming sequence (default=5000).
Examples:
>>> pos_encoder = PositionalEncoding(d_model)
"""
def __init__(self, dropout, dim, max_len=5000):
super(PositionalEncoding_2d, self).__init__()
self.dropout = nn.Dropout(p=dropout)
pe = paddle.zeros([max_len, dim])
position = paddle.arange(0, max_len, dtype=paddle.float32).unsqueeze(1)
div_term = paddle.exp(
paddle.arange(0, dim, 2).astype('float32') *
(-math.log(10000.0) / dim))
pe[:, 0::2] = paddle.sin(position * div_term)
pe[:, 1::2] = paddle.cos(position * div_term)
pe = pe.unsqueeze(0).transpose([1, 0, 2])
self.register_buffer('pe', pe)
self.avg_pool_1 = nn.AdaptiveAvgPool2D((1, 1))
self.linear1 = nn.Linear(dim, dim)
self.linear1.weight.data.fill_(1.)
self.avg_pool_2 = nn.AdaptiveAvgPool2D((1, 1))
self.linear2 = nn.Linear(dim, dim)
self.linear2.weight.data.fill_(1.)
def forward(self, x):
"""Inputs of forward function
Args:
x: the sequence fed to the positional encoder model (required).
Shape:
x: [sequence length, batch size, embed dim]
output: [sequence length, batch size, embed dim]
Examples:
>>> output = pos_encoder(x)
"""
w_pe = self.pe[:x.shape[-1], :]
w1 = self.linear1(self.avg_pool_1(x).squeeze()).unsqueeze(0)
w_pe = w_pe * w1
w_pe = w_pe.transpose([1, 2, 0])
w_pe = w_pe.unsqueeze(2)
h_pe = self.pe[:x.shape[-2], :]
w2 = self.linear2(self.avg_pool_2(x).squeeze()).unsqueeze(0)
h_pe = h_pe * w2
h_pe = h_pe.transpose([1, 2, 0])
h_pe = h_pe.unsqueeze(3)
x = x + w_pe + h_pe
x = x.reshape(
[x.shape[0], x.shape[1], x.shape[2] * x.shape[3]]).transpose(
[2, 0, 1])
return self.dropout(x)
class Embeddings(nn.Layer):
def __init__(self, d_model, vocab, padding_idx, scale_embedding):
super(Embeddings, self).__init__()
self.embedding = nn.Embedding(vocab, d_model, padding_idx=padding_idx)
w0 = np.random.normal(0.0, d_model**-0.5,
(vocab, d_model)).astype(np.float32)
self.embedding.weight.set_value(w0)
self.d_model = d_model
self.scale_embedding = scale_embedding
def forward(self, x):
if self.scale_embedding:
x = self.embedding(x)
return x * math.sqrt(self.d_model)
return self.embedding(x)
class Beam():
''' Beam search '''
def __init__(self, size, device=False):
self.size = size
self._done = False
# The score for each translation on the beam.
self.scores = paddle.zeros((size, ), dtype=paddle.float32)
self.all_scores = []
# The backpointers at each time-step.
self.prev_ks = []
# The outputs at each time-step.
self.next_ys = [paddle.full((size, ), 0, dtype=paddle.int64)]
self.next_ys[0][0] = 2
def get_current_state(self):
"Get the outputs for the current timestep."
return self.get_tentative_hypothesis()
def get_current_origin(self):
"Get the backpointers for the current timestep."
return self.prev_ks[-1]
@property
def done(self):
return self._done
def advance(self, word_prob):
"Update beam status and check if finished or not."
num_words = word_prob.shape[1]
# Sum the previous scores.
if len(self.prev_ks) > 0:
beam_lk = word_prob + self.scores.unsqueeze(1).expand_as(word_prob)
else:
beam_lk = word_prob[0]
flat_beam_lk = beam_lk.reshape([-1])
best_scores, best_scores_id = flat_beam_lk.topk(self.size, 0, True,
True) # 1st sort
self.all_scores.append(self.scores)
self.scores = best_scores
# bestScoresId is flattened as a (beam x word) array,
# so we need to calculate which word and beam each score came from
prev_k = best_scores_id // num_words
self.prev_ks.append(prev_k)
self.next_ys.append(best_scores_id - prev_k * num_words)
# End condition is when top-of-beam is EOS.
if self.next_ys[-1][0] == 3:
self._done = True
self.all_scores.append(self.scores)
return self._done
def sort_scores(self):
"Sort the scores."
return self.scores, paddle.to_tensor(
[i for i in range(self.scores.shape[0])], dtype='int32')
def get_the_best_score_and_idx(self):
"Get the score of the best in the beam."
scores, ids = self.sort_scores()
return scores[1], ids[1]
def get_tentative_hypothesis(self):
"Get the decoded sequence for the current timestep."
if len(self.next_ys) == 1:
dec_seq = self.next_ys[0].unsqueeze(1)
else:
_, keys = self.sort_scores()
hyps = [self.get_hypothesis(k) for k in keys]
hyps = [[2] + h for h in hyps]
dec_seq = paddle.to_tensor(hyps, dtype='int64')
return dec_seq
def get_hypothesis(self, k):
""" Walk back to construct the full hypothesis. """
hyp = []
for j in range(len(self.prev_ks) - 1, -1, -1):
hyp.append(self.next_ys[j + 1][k])
k = self.prev_ks[j][k]
return list(map(lambda x: x.item(), hyp[::-1]))
ppocr/postprocess/__init__.py
View file @ a8a9b2e5
...@@ -24,18 +24,16 @@ __all__ = ['build_post_process'] ...@@ -24,18 +24,16 @@ __all__ = ['build_post_process']
from .db_postprocess import DBPostProcess, DistillationDBPostProcess from .db_postprocess import DBPostProcess, DistillationDBPostProcess
from .east_postprocess import EASTPostProcess from .east_postprocess import EASTPostProcess
from .sast_postprocess import SASTPostProcess from .sast_postprocess import SASTPostProcess
from .rec_postprocess import CTCLabelDecode, AttnLabelDecode, SRNLabelDecode, DistillationCTCLabelDecode, \ from .rec_postprocess import CTCLabelDecode, AttnLabelDecode, SRNLabelDecode, DistillationCTCLabelDecode, NRTRLabelDecode, \
TableLabelDecode TableLabelDecode
from .cls_postprocess import ClsPostProcess from .cls_postprocess import ClsPostProcess
from .pg_postprocess import PGPostProcess from .pg_postprocess import PGPostProcess
def build_post_process(config, global_config=None): def build_post_process(config, global_config=None):
support_dict = [ support_dict = [
'DBPostProcess', 'EASTPostProcess', 'SASTPostProcess', 'CTCLabelDecode', 'DBPostProcess', 'EASTPostProcess', 'SASTPostProcess', 'CTCLabelDecode',
'AttnLabelDecode', 'ClsPostProcess', 'SRNLabelDecode', 'PGPostProcess', 'AttnLabelDecode', 'ClsPostProcess', 'SRNLabelDecode', 'PGPostProcess',
'DistillationCTCLabelDecode', 'TableLabelDecode', 'DistillationCTCLabelDecode', 'NRTRLabelDecode', 'TableLabelDecode', 'DistillationDBPostProcess'
'DistillationDBPostProcess'
] ]
config = copy.deepcopy(config) config = copy.deepcopy(config)
......
ppocr/postprocess/rec_postprocess.py
View file @ a8a9b2e5
...@@ -156,6 +156,69 @@ class DistillationCTCLabelDecode(CTCLabelDecode): ...@@ -156,6 +156,69 @@ class DistillationCTCLabelDecode(CTCLabelDecode):
return output return output
class NRTRLabelDecode(BaseRecLabelDecode):
""" Convert between text-label and text-index """
def __init__(self,
character_dict_path=None,
character_type='EN_symbol',
use_space_char=True,
**kwargs):
super(NRTRLabelDecode, self).__init__(character_dict_path,
character_type, use_space_char)
def __call__(self, preds, label=None, *args, **kwargs):
if preds.dtype == paddle.int64:
if isinstance(preds, paddle.Tensor):
preds = preds.numpy()
if preds[0][0]==2:
preds_idx = preds[:,1:]
else:
preds_idx = preds
text = self.decode(preds_idx)
if label is None:
return text
label = self.decode(label[:,1:])
else:
if isinstance(preds, paddle.Tensor):
preds = preds.numpy()
preds_idx = preds.argmax(axis=2)
preds_prob = preds.max(axis=2)
text = self.decode(preds_idx, preds_prob, is_remove_duplicate=False)
if label is None:
return text
label = self.decode(label[:,1:])
return text, label
def add_special_char(self, dict_character):
dict_character = ['blank','<unk>','<s>','</s>'] + dict_character
return dict_character
def decode(self, text_index, text_prob=None, is_remove_duplicate=False):
""" convert text-index into text-label. """
result_list = []
batch_size = len(text_index)
for batch_idx in range(batch_size):
char_list = []
conf_list = []
for idx in range(len(text_index[batch_idx])):
if text_index[batch_idx][idx] == 3: # end
break
try:
char_list.append(self.character[int(text_index[batch_idx][idx])])
except:
continue
if text_prob is not None:
conf_list.append(text_prob[batch_idx][idx])
else:
conf_list.append(1)
text = ''.join(char_list)
result_list.append((text.lower(), np.mean(conf_list)))
return result_list
class AttnLabelDecode(BaseRecLabelDecode): class AttnLabelDecode(BaseRecLabelDecode):
""" Convert between text-label and text-index """ """ Convert between text-label and text-index """
...@@ -193,8 +256,7 @@ class AttnLabelDecode(BaseRecLabelDecode): ...@@ -193,8 +256,7 @@ class AttnLabelDecode(BaseRecLabelDecode):
if idx > 0 and text_index[batch_idx][idx - 1] == text_index[ if idx > 0 and text_index[batch_idx][idx - 1] == text_index[
batch_idx][idx]: batch_idx][idx]:
continue continue
char_list.append(self.character[int(text_index[batch_idx][ char_list.append(self.character[int(text_index[batch_idx][idx])])
idx])])
if text_prob is not None: if text_prob is not None:
conf_list.append(text_prob[batch_idx][idx]) conf_list.append(text_prob[batch_idx][idx])
else: else:
......
ppstructure/README.md
View file @ a8a9b2e5
...@@ -30,13 +30,13 @@ python3 -m pip install paddlepaddle-gpu==2.1.1 -i https://mirror.baidu.com/pypi/ ...@@ -30,13 +30,13 @@ python3 -m pip install paddlepaddle-gpu==2.1.1 -i https://mirror.baidu.com/pypi/
# CPU # CPU
python3 -m pip install paddlepaddle==2.1.1 -i https://mirror.baidu.com/pypi/simple python3 -m pip install paddlepaddle==2.1.1 -i https://mirror.baidu.com/pypi/simple
# For more，refer[Installation](https://www.paddlepaddle.org.cn/install/quick)。
``` ```
For more，refer [Installation](https://www.paddlepaddle.org.cn/install/quick) .
- **(2) Install Layout-Parser** - **(2) Install Layout-Parser**
```bash ```bash
pip3 install -U premailer paddleocr https://paddleocr.bj.bcebos.com/whl/layoutparser-0.0.0-py3-none-any.whl pip3 install -U https://paddleocr.bj.bcebos.com/whl/layoutparser-0.0.0-py3-none-any.whl
``` ```
### 2.2 Install PaddleOCR（including PP-OCR and PP-Structure） ### 2.2 Install PaddleOCR（including PP-OCR and PP-Structure）
...@@ -124,8 +124,6 @@ Most of the parameters are consistent with the paddleocr whl package, see [doc o ...@@ -124,8 +124,6 @@ Most of the parameters are consistent with the paddleocr whl package, see [doc o
After running, each image will have a directory with the same name under the directory specified in the output field. Each table in the picture will be stored as an excel and figure area will be cropped and saved, the excel and image file name will be the coordinates of the table in the image. After running, each image will have a directory with the same name under the directory specified in the output field. Each table in the picture will be stored as an excel and figure area will be cropped and saved, the excel and image file name will be the coordinates of the table in the image.
## 4. PP-Structure Pipeline ## 4. PP-Structure Pipeline
the process is as follows
![pipeline](../doc/table/pipeline_en.jpg) ![pipeline](../doc/table/pipeline_en.jpg)
In PP-Structure, the image will be analyzed by layoutparser first. In the layout analysis, the area in the image will be classified, including **text, title, image, list and table** 5 categories. For the first 4 types of areas, directly use the PP-OCR to complete the text detection and recognition. The table area will be converted to an excel file of the same table style via Table OCR. In PP-Structure, the image will be analyzed by layoutparser first. In the layout analysis, the area in the image will be classified, including **text, title, image, list and table** 5 categories. For the first 4 types of areas, directly use the PP-OCR to complete the text detection and recognition. The table area will be converted to an excel file of the same table style via Table OCR.
...@@ -180,10 +178,10 @@ OCR and table recognition model ...@@ -180,10 +178,10 @@ OCR and table recognition model
|model name|description|model size|download| |model name|description|model size|download|
| --- | --- | --- | --- | | --- | --- | --- | --- |
|ch_ppocr_mobile_slim_v2.0_det|Slim pruned lightweight model, supporting Chinese, English, multilingual text detection|2.6M|[inference model](https://paddleocr.bj.bcebos.com/dygraph_v2.0/slim/ch_ppocr_mobile_v2.0_det_prune_infer.tar) | |ch_ppocr_mobile_slim_v2.0_det|Slim pruned lightweight model, supporting Chinese, English, multilingual text detection|2.6M|[inference model](https://paddleocr.bj.bcebos.com/dygraph_v2.0/slim/ch_ppocr_mobile_v2.0_det_prune_infer.tar) / [trained model](https://paddleocr.bj.bcebos.com/dygraph_v2.0/slim/ch_ppocr_mobile_v2.0_det_prune_infer.tar) |
|ch_ppocr_mobile_slim_v2.0_rec|Slim pruned and quantized lightweight model, supporting Chinese, English and number recognition|6M|[inference model](https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_mobile_v2.0_rec_slim_infer.tar) | |ch_ppocr_mobile_slim_v2.0_rec|Slim pruned and quantized lightweight model, supporting Chinese, English and number recognition|6M|[inference model](https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_mobile_v2.0_rec_slim_infer.tar) / [trained model](https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_mobile_v2.0_rec_slim_train.tar) |
|en_ppocr_mobile_v2.0_table_det|Text detection of English table scenes trained on PubLayNet dataset|4.7M|[inference model](https://paddleocr.bj.bcebos.com/dygraph_v2.0/table/en_ppocr_mobile_v2.0_table_det_infer.tar) | |en_ppocr_mobile_v2.0_table_det|Text detection of English table scenes trained on PubLayNet dataset|4.7M|[inference model](https://paddleocr.bj.bcebos.com/dygraph_v2.0/table/en_ppocr_mobile_v2.0_table_det_infer.tar) / [trained model](https://paddleocr.bj.bcebos.com/dygraph_v2.1/table/en_ppocr_mobile_v2.0_table_det_train.tar) |
|en_ppocr_mobile_v2.0_table_rec|Text recognition of English table scene trained on PubLayNet dataset|6.9M|[inference model](https://paddleocr.bj.bcebos.com/dygraph_v2.0/table/en_ppocr_mobile_v2.0_table_rec_infer.tar) | |en_ppocr_mobile_v2.0_table_rec|Text recognition of English table scene trained on PubLayNet dataset|6.9M|[inference model](https://paddleocr.bj.bcebos.com/dygraph_v2.0/table/en_ppocr_mobile_v2.0_table_rec_infer.tar) [trained model](https://paddleocr.bj.bcebos.com/dygraph_v2.1/table/en_ppocr_mobile_v2.0_table_rec_train.tar) |
|en_ppocr_mobile_v2.0_table_structure|Table structure prediction of English table scene trained on PubLayNet dataset|18.6M|[inference model](https://paddleocr.bj.bcebos.com/dygraph_v2.0/table/en_ppocr_mobile_v2.0_table_structure_infer.tar) | |en_ppocr_mobile_v2.0_table_structure|Table structure prediction of English table scene trained on PubLayNet dataset|18.6M|[inference model](https://paddleocr.bj.bcebos.com/dygraph_v2.0/table/en_ppocr_mobile_v2.0_table_structure_infer.tar) / [trained model](https://paddleocr.bj.bcebos.com/dygraph_v2.1/table/en_ppocr_mobile_v2.0_table_structure_train.tar) |
If you need to use other models, you can download the model in [model_list](../doc/doc_en/models_list_en.md) or use your own trained model to configure it to the three fields of `det_model_dir`, `rec_model_dir`, `table_model_dir` . If you need to use other models, you can download the model in [model_list](../doc/doc_en/models_list_en.md) or use your own trained model to configure it to the three fields of `det_model_dir`, `rec_model_dir`, `table_model_dir` .
ppstructure/README_ch.md
View file @ a8a9b2e5
...@@ -30,13 +30,13 @@ python3 -m pip install paddlepaddle-gpu==2.1.1 -i https://mirror.baidu.com/pypi/ ...@@ -30,13 +30,13 @@ python3 -m pip install paddlepaddle-gpu==2.1.1 -i https://mirror.baidu.com/pypi/
# CPU安装 # CPU安装
python3 -m pip install paddlepaddle==2.1.1 -i https://mirror.baidu.com/pypi/simple python3 -m pip install paddlepaddle==2.1.1 -i https://mirror.baidu.com/pypi/simple
# 更多需求，请参照[安装文档](https://www.paddlepaddle.org.cn/install/quick)中的说明进行操作。
``` ```
更多需求，请参照[安装文档](https://www.paddlepaddle.org.cn/install/quick)中的说明进行操作。
- **(2) 安装 Layout-Parser** - **(2) 安装 Layout-Parser**
```bash ```bash
pip3 install -U premailer paddleocr https://paddleocr.bj.bcebos.com/whl/layoutparser-0.0.0-py3-none-any.whl pip3 install -U https://paddleocr.bj.bcebos.com/whl/layoutparser-0.0.0-py3-none-any.whl
``` ```
### 2.2 安装PaddleOCR（包含PP-OCR和PP-Structure） ### 2.2 安装PaddleOCR（包含PP-OCR和PP-Structure）
...@@ -179,10 +179,10 @@ OCR和表格识别模型 ...@@ -179,10 +179,10 @@ OCR和表格识别模型
|模型名称|模型简介|推理模型大小|下载地址| |模型名称|模型简介|推理模型大小|下载地址|
| --- | --- | --- | --- | | --- | --- | --- | --- |
|ch_ppocr_mobile_slim_v2.0_det|slim裁剪版超轻量模型，支持中英文、多语种文本检测|2.6M|[推理模型](https://paddleocr.bj.bcebos.com/dygraph_v2.0/slim/ch_ppocr_mobile_v2.0_det_prune_infer.tar) | |ch_ppocr_mobile_slim_v2.0_det|slim裁剪版超轻量模型，支持中英文、多语种文本检测|2.6M|[推理模型](https://paddleocr.bj.bcebos.com/dygraph_v2.0/slim/ch_ppocr_mobile_v2.0_det_prune_infer.tar) / [训练模型](https://paddleocr.bj.bcebos.com/dygraph_v2.0/slim/ch_ppocr_mobile_v2.0_det_prune_infer.tar) |
|ch_ppocr_mobile_slim_v2.0_rec|slim裁剪量化版超轻量模型，支持中英文、数字识别|6M|[推理模型](https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_mobile_v2.0_rec_slim_infer.tar) | |ch_ppocr_mobile_slim_v2.0_rec|slim裁剪量化版超轻量模型，支持中英文、数字识别|6M|[推理模型](https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_mobile_v2.0_rec_slim_infer.tar) / [训练模型](https://paddleocr.bj.bcebos.com/dygraph_v2.0/ch/ch_ppocr_mobile_v2.0_rec_slim_train.tar) |
|en_ppocr_mobile_v2.0_table_det|PubLayNet数据集训练的英文表格场景的文字检测|4.7M|[推理模型](https://paddleocr.bj.bcebos.com/dygraph_v2.0/table/en_ppocr_mobile_v2.0_table_det_infer.tar) | |en_ppocr_mobile_v2.0_table_det|PubLayNet数据集训练的英文表格场景的文字检测|4.7M|[推理模型](https://paddleocr.bj.bcebos.com/dygraph_v2.0/table/en_ppocr_mobile_v2.0_table_det_infer.tar) / [训练模型](https://paddleocr.bj.bcebos.com/dygraph_v2.1/table/en_ppocr_mobile_v2.0_table_det_train.tar) |
|en_ppocr_mobile_v2.0_table_rec|PubLayNet数据集训练的英文表格场景的文字识别|6.9M|[推理模型](https://paddleocr.bj.bcebos.com/dygraph_v2.0/table/en_ppocr_mobile_v2.0_table_rec_infer.tar) | |en_ppocr_mobile_v2.0_table_rec|PubLayNet数据集训练的英文表格场景的文字识别|6.9M|[推理模型](https://paddleocr.bj.bcebos.com/dygraph_v2.0/table/en_ppocr_mobile_v2.0_table_rec_infer.tar) / [训练模型](https://paddleocr.bj.bcebos.com/dygraph_v2.1/table/en_ppocr_mobile_v2.0_table_rec_train.tar) |
|en_ppocr_mobile_v2.0_table_structure|PubLayNet数据集训练的英文表格场景的表格结构预测|18.6M|[推理模型](https://paddleocr.bj.bcebos.com/dygraph_v2.0/table/en_ppocr_mobile_v2.0_table_structure_infer.tar) | |en_ppocr_mobile_v2.0_table_structure|PubLayNet数据集训练的英文表格场景的表格结构预测|18.6M|[推理模型](https://paddleocr.bj.bcebos.com/dygraph_v2.0/table/en_ppocr_mobile_v2.0_table_structure_infer.tar) / [训练模型](https://paddleocr.bj.bcebos.com/dygraph_v2.1/table/en_ppocr_mobile_v2.0_table_structure_train.tar) |
如需要使用其他模型，可以在 [model_list](../doc/doc_ch/models_list.md) 下载模型或者使用自己训练好的模型配置到`det_model_dir`,`rec_model_dir`,`table_model_dir`三个字段即可。 如需要使用其他模型，可以在 [model_list](../doc/doc_ch/models_list.md) 下载模型或者使用自己训练好的模型配置到`det_model_dir`,`rec_model_dir`,`table_model_dir`三个字段即可。
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment