Skip to content

GitLab

Commit 19eb7eb8 authored by Leif's avatar Leif
Browse files

Merge remote-tracking branch 'origin/dygraph' into dy1

parents 0afe6c32 03b7daa5
Hide whitespace changes
Inline Side-by-side
Showing with 2484 additions and 31 deletions
ppocr/data/imaug/label_ops.py
View file @ 19eb7eb8
......@@ -19,6 +19,7 @@ from __future__ import unicode_literals
import numpy as np
import string
import json
class ClsLabelEncode(object):
......@@ -39,7 +40,6 @@ class DetLabelEncode(object):
pass
def __call__(self, data):
import json
label = data['label']
label = json.loads(label)
nBox = len(label)
......@@ -53,6 +53,8 @@ class DetLabelEncode(object):
txt_tags.append(True)
else:
txt_tags.append(False)
if len(boxes) == 0:
return None
boxes = self.expand_points_num(boxes)
boxes = np.array(boxes, dtype=np.float32)
txt_tags = np.array(txt_tags, dtype=np.bool)
......@@ -96,7 +98,7 @@ class BaseRecLabelEncode(object):
'ch', 'en', 'EN_symbol', 'french', 'german', 'japan', 'korean',
'EN', 'it', 'xi', 'pu', 'ru', 'ar', 'ta', 'ug', 'fa', 'ur', 'rs',
'oc', 'rsc', 'bg', 'uk', 'be', 'te', 'ka', 'chinese_cht', 'hi',
'mr', 'ne'
'mr', 'ne', 'latin', 'arabic', 'cyrillic', 'devanagari'
]
assert character_type in support_character_type, "Only {} are supported now but get {}".format(
support_character_type, character_type)
......@@ -159,6 +161,34 @@ class BaseRecLabelEncode(object):
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):
""" Convert between text-label and text-index """
......@@ -187,6 +217,78 @@ class CTCLabelEncode(BaseRecLabelEncode):
return dict_character
class E2ELabelEncodeTest(BaseRecLabelEncode):
def __init__(self,
max_text_length,
character_dict_path=None,
character_type='EN',
use_space_char=False,
**kwargs):
super(E2ELabelEncodeTest,
self).__init__(max_text_length, character_dict_path,
character_type, use_space_char)
def __call__(self, data):
import json
padnum = len(self.dict)
label = data['label']
label = json.loads(label)
nBox = len(label)
boxes, txts, txt_tags = [], [], []
for bno in range(0, nBox):
box = label[bno]['points']
txt = label[bno]['transcription']
boxes.append(box)
txts.append(txt)
if txt in ['*', '###']:
txt_tags.append(True)
else:
txt_tags.append(False)
boxes = np.array(boxes, dtype=np.float32)
txt_tags = np.array(txt_tags, dtype=np.bool)
data['polys'] = boxes
data['ignore_tags'] = txt_tags
temp_texts = []
for text in txts:
text = text.lower()
text = self.encode(text)
if text is None:
return None
text = text + [padnum] * (self.max_text_len - len(text)
) # use 36 to pad
temp_texts.append(text)
data['texts'] = np.array(temp_texts)
return data
class E2ELabelEncodeTrain(object):
def __init__(self, **kwargs):
pass
def __call__(self, data):
import json
label = data['label']
label = json.loads(label)
nBox = len(label)
boxes, txts, txt_tags = [], [], []
for bno in range(0, nBox):
box = label[bno]['points']
txt = label[bno]['transcription']
boxes.append(box)
txts.append(txt)
if txt in ['*', '###']:
txt_tags.append(True)
else:
txt_tags.append(False)
boxes = np.array(boxes, dtype=np.float32)
txt_tags = np.array(txt_tags, dtype=np.bool)
data['polys'] = boxes
data['texts'] = txts
data['ignore_tags'] = txt_tags
return data
class AttnLabelEncode(BaseRecLabelEncode):
""" Convert between text-label and text-index """
......@@ -279,3 +381,171 @@ class SRNLabelEncode(BaseRecLabelEncode):
assert False, "Unsupport type %s in get_beg_end_flag_idx" \
% beg_or_end
return idx
class TableLabelEncode(object):
""" Convert between text-label and text-index """
def __init__(self,
max_text_length,
max_elem_length,
max_cell_num,
character_dict_path,
span_weight=1.0,
**kwargs):
self.max_text_length = max_text_length
self.max_elem_length = max_elem_length
self.max_cell_num = max_cell_num
list_character, list_elem = self.load_char_elem_dict(
character_dict_path)
list_character = self.add_special_char(list_character)
list_elem = self.add_special_char(list_elem)
self.dict_character = {}
for i, char in enumerate(list_character):
self.dict_character[char] = i
self.dict_elem = {}
for i, elem in enumerate(list_elem):
self.dict_elem[elem] = i
self.span_weight = span_weight
def load_char_elem_dict(self, character_dict_path):
list_character = []
list_elem = []
with open(character_dict_path, "rb") as fin:
lines = fin.readlines()
substr = lines[0].decode('utf-8').strip("\r\n").split("\t")
character_num = int(substr[0])
elem_num = int(substr[1])
for cno in range(1, 1 + character_num):
character = lines[cno].decode('utf-8').strip("\r\n")
list_character.append(character)
for eno in range(1 + character_num, 1 + character_num + elem_num):
elem = lines[eno].decode('utf-8').strip("\r\n")
list_elem.append(elem)
return list_character, list_elem
def add_special_char(self, list_character):
self.beg_str = "sos"
self.end_str = "eos"
list_character = [self.beg_str] + list_character + [self.end_str]
return list_character
def get_span_idx_list(self):
span_idx_list = []
for elem in self.dict_elem:
if 'span' in elem:
span_idx_list.append(self.dict_elem[elem])
return span_idx_list
def __call__(self, data):
cells = data['cells']
structure = data['structure']['tokens']
structure = self.encode(structure, 'elem')
if structure is None:
return None
elem_num = len(structure)
structure = [0] + structure + [len(self.dict_elem) - 1]
structure = structure + [0] * (self.max_elem_length + 2 - len(structure)
)
structure = np.array(structure)
data['structure'] = structure
elem_char_idx1 = self.dict_elem['<td>']
elem_char_idx2 = self.dict_elem['<td']
span_idx_list = self.get_span_idx_list()
td_idx_list = np.logical_or(structure == elem_char_idx1,
structure == elem_char_idx2)
td_idx_list = np.where(td_idx_list)[0]
structure_mask = np.ones(
(self.max_elem_length + 2, 1), dtype=np.float32)
bbox_list = np.zeros((self.max_elem_length + 2, 4), dtype=np.float32)
bbox_list_mask = np.zeros(
(self.max_elem_length + 2, 1), dtype=np.float32)
img_height, img_width, img_ch = data['image'].shape
if len(span_idx_list) > 0:
span_weight = len(td_idx_list) * 1.0 / len(span_idx_list)
span_weight = min(max(span_weight, 1.0), self.span_weight)
for cno in range(len(cells)):
if 'bbox' in cells[cno]:
bbox = cells[cno]['bbox'].copy()
bbox[0] = bbox[0] * 1.0 / img_width
bbox[1] = bbox[1] * 1.0 / img_height
bbox[2] = bbox[2] * 1.0 / img_width
bbox[3] = bbox[3] * 1.0 / img_height
td_idx = td_idx_list[cno]
bbox_list[td_idx] = bbox
bbox_list_mask[td_idx] = 1.0
cand_span_idx = td_idx + 1
if cand_span_idx < (self.max_elem_length + 2):
if structure[cand_span_idx] in span_idx_list:
structure_mask[cand_span_idx] = span_weight
data['bbox_list'] = bbox_list
data['bbox_list_mask'] = bbox_list_mask
data['structure_mask'] = structure_mask
char_beg_idx = self.get_beg_end_flag_idx('beg', 'char')
char_end_idx = self.get_beg_end_flag_idx('end', 'char')
elem_beg_idx = self.get_beg_end_flag_idx('beg', 'elem')
elem_end_idx = self.get_beg_end_flag_idx('end', 'elem')
data['sp_tokens'] = np.array([
char_beg_idx, char_end_idx, elem_beg_idx, elem_end_idx,
elem_char_idx1, elem_char_idx2, self.max_text_length,
self.max_elem_length, self.max_cell_num, elem_num
])
return data
def encode(self, text, char_or_elem):
"""convert text-label into text-index.
"""
if char_or_elem == "char":
max_len = self.max_text_length
current_dict = self.dict_character
else:
max_len = self.max_elem_length
current_dict = self.dict_elem
if len(text) > max_len:
return None
if len(text) == 0:
if char_or_elem == "char":
return [self.dict_character['space']]
else:
return None
text_list = []
for char in text:
if char not in current_dict:
return None
text_list.append(current_dict[char])
if len(text_list) == 0:
if char_or_elem == "char":
return [self.dict_character['space']]
else:
return None
return text_list
def get_ignored_tokens(self, char_or_elem):
beg_idx = self.get_beg_end_flag_idx("beg", char_or_elem)
end_idx = self.get_beg_end_flag_idx("end", char_or_elem)
return [beg_idx, end_idx]
def get_beg_end_flag_idx(self, beg_or_end, char_or_elem):
if char_or_elem == "char":
if beg_or_end == "beg":
idx = np.array(self.dict_character[self.beg_str])
elif beg_or_end == "end":
idx = np.array(self.dict_character[self.end_str])
else:
assert False, "Unsupport type %s in get_beg_end_flag_idx of char" \
% beg_or_end
elif char_or_elem == "elem":
if beg_or_end == "beg":
idx = np.array(self.dict_elem[self.beg_str])
elif beg_or_end == "end":
idx = np.array(self.dict_elem[self.end_str])
else:
assert False, "Unsupport type %s in get_beg_end_flag_idx of elem" \
% beg_or_end
else:
assert False, "Unsupport type %s in char_or_elem" \
% char_or_elem
return idx
ppocr/data/imaug/operators.py
View file @ 19eb7eb8
......@@ -57,6 +57,38 @@ class DecodeImage(object):
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):
""" normalize image such as substract mean, divide std
"""
......@@ -81,7 +113,7 @@ class NormalizeImage(object):
assert isinstance(img,
np.ndarray), "invalid input 'img' in NormalizeImage"
data['image'] = (
img.astype('float32') * self.scale - self.mean) / self.std
img.astype('float32') * self.scale - self.mean) / self.std
return data
......@@ -163,7 +195,7 @@ class DetResizeForTest(object):
img, (ratio_h, ratio_w)
"""
limit_side_len = self.limit_side_len
h, w, _ = img.shape
h, w, c = img.shape
# limit the max side
if self.limit_type == 'max':
......@@ -174,7 +206,7 @@ class DetResizeForTest(object):
ratio = float(limit_side_len) / w
else:
ratio = 1.
else:
elif self.limit_type == 'min':
if min(h, w) < limit_side_len:
if h < w:
ratio = float(limit_side_len) / h
......@@ -182,6 +214,10 @@ class DetResizeForTest(object):
ratio = float(limit_side_len) / w
else:
ratio = 1.
elif self.limit_type == 'resize_long':
ratio = float(limit_side_len) / max(h,w)
else:
raise Exception('not support limit type, image ')
resize_h = int(h * ratio)
resize_w = int(w * ratio)
......@@ -197,7 +233,6 @@ class DetResizeForTest(object):
sys.exit(0)
ratio_h = resize_h / float(h)
ratio_w = resize_w / float(w)
# return img, np.array([h, w])
return img, [ratio_h, ratio_w]
def resize_image_type2(self, img):
......@@ -206,7 +241,6 @@ class DetResizeForTest(object):
resize_w = w
resize_h = h
# Fix the longer side
if resize_h > resize_w:
ratio = float(self.resize_long) / resize_h
else:
......@@ -223,3 +257,72 @@ class DetResizeForTest(object):
ratio_w = resize_w / float(w)
return img, [ratio_h, ratio_w]
class E2EResizeForTest(object):
def __init__(self, **kwargs):
super(E2EResizeForTest, self).__init__()
self.max_side_len = kwargs['max_side_len']
self.valid_set = kwargs['valid_set']
def __call__(self, data):
img = data['image']
src_h, src_w, _ = img.shape
if self.valid_set == 'totaltext':
im_resized, [ratio_h, ratio_w] = self.resize_image_for_totaltext(
img, max_side_len=self.max_side_len)
else:
im_resized, (ratio_h, ratio_w) = self.resize_image(
img, max_side_len=self.max_side_len)
data['image'] = im_resized
data['shape'] = np.array([src_h, src_w, ratio_h, ratio_w])
return data
def resize_image_for_totaltext(self, im, max_side_len=512):
h, w, _ = im.shape
resize_w = w
resize_h = h
ratio = 1.25
if h * ratio > max_side_len:
ratio = float(max_side_len) / resize_h
resize_h = int(resize_h * ratio)
resize_w = int(resize_w * ratio)
max_stride = 128
resize_h = (resize_h + max_stride - 1) // max_stride * max_stride
resize_w = (resize_w + max_stride - 1) // max_stride * max_stride
im = cv2.resize(im, (int(resize_w), int(resize_h)))
ratio_h = resize_h / float(h)
ratio_w = resize_w / float(w)
return im, (ratio_h, ratio_w)
def resize_image(self, im, max_side_len=512):
"""
resize image to a size multiple of max_stride which is required by the network
:param im: the resized image
:param max_side_len: limit of max image size to avoid out of memory in gpu
:return: the resized image and the resize ratio
"""
h, w, _ = im.shape
resize_w = w
resize_h = h
# Fix the longer side
if resize_h > resize_w:
ratio = float(max_side_len) / resize_h
else:
ratio = float(max_side_len) / resize_w
resize_h = int(resize_h * ratio)
resize_w = int(resize_w * ratio)
max_stride = 128
resize_h = (resize_h + max_stride - 1) // max_stride * max_stride
resize_w = (resize_w + max_stride - 1) // max_stride * max_stride
im = cv2.resize(im, (int(resize_w), int(resize_h)))
ratio_h = resize_h / float(h)
ratio_w = resize_w / float(w)
return im, (ratio_h, ratio_w)
ppocr/data/imaug/pg_process.py 0 → 100644
View file @ 19eb7eb8
# 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 cv2
import numpy as np
__all__ = ['PGProcessTrain']
class PGProcessTrain(object):
def __init__(self,
character_dict_path,
max_text_length,
max_text_nums,
tcl_len,
batch_size=14,
min_crop_size=24,
min_text_size=4,
max_text_size=512,
**kwargs):
self.tcl_len = tcl_len
self.max_text_length = max_text_length
self.max_text_nums = max_text_nums
self.batch_size = batch_size
self.min_crop_size = min_crop_size
self.min_text_size = min_text_size
self.max_text_size = max_text_size
self.Lexicon_Table = self.get_dict(character_dict_path)
self.pad_num = len(self.Lexicon_Table)
self.img_id = 0
def get_dict(self, character_dict_path):
character_str = ""
with open(character_dict_path, "rb") as fin:
lines = fin.readlines()
for line in lines:
line = line.decode('utf-8').strip("\n").strip("\r\n")
character_str += line
dict_character = list(character_str)
return dict_character
def quad_area(self, poly):
"""
compute area of a polygon
:param poly:
:return:
"""
edge = [(poly[1][0] - poly[0][0]) * (poly[1][1] + poly[0][1]),
(poly[2][0] - poly[1][0]) * (poly[2][1] + poly[1][1]),
(poly[3][0] - poly[2][0]) * (poly[3][1] + poly[2][1]),
(poly[0][0] - poly[3][0]) * (poly[0][1] + poly[3][1])]
return np.sum(edge) / 2.
def gen_quad_from_poly(self, poly):
"""
Generate min area quad from poly.
"""
point_num = poly.shape[0]
min_area_quad = np.zeros((4, 2), dtype=np.float32)
rect = cv2.minAreaRect(poly.astype(
np.int32)) # (center (x,y), (width, height), angle of rotation)
box = np.array(cv2.boxPoints(rect))
first_point_idx = 0
min_dist = 1e4
for i in range(4):
dist = np.linalg.norm(box[(i + 0) % 4] - poly[0]) + \
np.linalg.norm(box[(i + 1) % 4] - poly[point_num // 2 - 1]) + \
np.linalg.norm(box[(i + 2) % 4] - poly[point_num // 2]) + \
np.linalg.norm(box[(i + 3) % 4] - poly[-1])
if dist < min_dist:
min_dist = dist
first_point_idx = i
for i in range(4):
min_area_quad[i] = box[(first_point_idx + i) % 4]
return min_area_quad
def check_and_validate_polys(self, polys, tags, im_size):
"""
check so that the text poly is in the same direction,
and also filter some invalid polygons
:param polys:
:param tags:
:return:
"""
(h, w) = im_size
if polys.shape[0] == 0:
return polys, np.array([]), np.array([])
polys[:, :, 0] = np.clip(polys[:, :, 0], 0, w - 1)
polys[:, :, 1] = np.clip(polys[:, :, 1], 0, h - 1)
validated_polys = []
validated_tags = []
hv_tags = []
for poly, tag in zip(polys, tags):
quad = self.gen_quad_from_poly(poly)
p_area = self.quad_area(quad)
if abs(p_area) < 1:
print('invalid poly')
continue
if p_area > 0:
if tag == False:
print('poly in wrong direction')
tag = True # reversed cases should be ignore
poly = poly[(0, 15, 14, 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2,
1), :]
quad = quad[(0, 3, 2, 1), :]
len_w = np.linalg.norm(quad[0] - quad[1]) + np.linalg.norm(quad[3] -
quad[2])
len_h = np.linalg.norm(quad[0] - quad[3]) + np.linalg.norm(quad[1] -
quad[2])
hv_tag = 1
if len_w * 2.0 < len_h:
hv_tag = 0
validated_polys.append(poly)
validated_tags.append(tag)
hv_tags.append(hv_tag)
return np.array(validated_polys), np.array(validated_tags), np.array(
hv_tags)
def crop_area(self,
im,
polys,
tags,
hv_tags,
txts,
crop_background=False,
max_tries=25):
"""
make random crop from the input image
:param im:
:param polys: [b,4,2]
:param tags:
:param crop_background:
:param max_tries: 50 -> 25
:return:
"""
h, w, _ = im.shape
pad_h = h // 10
pad_w = w // 10
h_array = np.zeros((h + pad_h * 2), dtype=np.int32)
w_array = np.zeros((w + pad_w * 2), dtype=np.int32)
for poly in polys:
poly = np.round(poly, decimals=0).astype(np.int32)
minx = np.min(poly[:, 0])
maxx = np.max(poly[:, 0])
w_array[minx + pad_w:maxx + pad_w] = 1
miny = np.min(poly[:, 1])
maxy = np.max(poly[:, 1])
h_array[miny + pad_h:maxy + pad_h] = 1
# ensure the cropped area not across a text
h_axis = np.where(h_array == 0)[0]
w_axis = np.where(w_array == 0)[0]
if len(h_axis) == 0 or len(w_axis) == 0:
return im, polys, tags, hv_tags, txts
for i in range(max_tries):
xx = np.random.choice(w_axis, size=2)
xmin = np.min(xx) - pad_w
xmax = np.max(xx) - pad_w
xmin = np.clip(xmin, 0, w - 1)
xmax = np.clip(xmax, 0, w - 1)
yy = np.random.choice(h_axis, size=2)
ymin = np.min(yy) - pad_h
ymax = np.max(yy) - pad_h
ymin = np.clip(ymin, 0, h - 1)
ymax = np.clip(ymax, 0, h - 1)
if xmax - xmin < self.min_crop_size or \
ymax - ymin < self.min_crop_size:
continue
if polys.shape[0] != 0:
poly_axis_in_area = (polys[:, :, 0] >= xmin) & (polys[:, :, 0] <= xmax) \
& (polys[:, :, 1] >= ymin) & (polys[:, :, 1] <= ymax)
selected_polys = np.where(
np.sum(poly_axis_in_area, axis=1) == 4)[0]
else:
selected_polys = []
if len(selected_polys) == 0:
# no text in this area
if crop_background:
txts_tmp = []
for selected_poly in selected_polys:
txts_tmp.append(txts[selected_poly])
txts = txts_tmp
return im[ymin: ymax + 1, xmin: xmax + 1, :], \
polys[selected_polys], tags[selected_polys], hv_tags[selected_polys], txts
else:
continue
im = im[ymin:ymax + 1, xmin:xmax + 1, :]
polys = polys[selected_polys]
tags = tags[selected_polys]
hv_tags = hv_tags[selected_polys]
txts_tmp = []
for selected_poly in selected_polys:
txts_tmp.append(txts[selected_poly])
txts = txts_tmp
polys[:, :, 0] -= xmin
polys[:, :, 1] -= ymin
return im, polys, tags, hv_tags, txts
return im, polys, tags, hv_tags, txts
def fit_and_gather_tcl_points_v2(self,
min_area_quad,
poly,
max_h,
max_w,
fixed_point_num=64,
img_id=0,
reference_height=3):
"""
Find the center point of poly as key_points, then fit and gather.
"""
key_point_xys = []
point_num = poly.shape[0]
for idx in range(point_num // 2):
center_point = (poly[idx] + poly[point_num - 1 - idx]) / 2.0
key_point_xys.append(center_point)
tmp_image = np.zeros(
shape=(
max_h,
max_w, ), dtype='float32')
cv2.polylines(tmp_image, [np.array(key_point_xys).astype('int32')],
False, 1.0)
ys, xs = np.where(tmp_image > 0)
xy_text = np.array(list(zip(xs, ys)), dtype='float32')
left_center_pt = (
(min_area_quad[0] - min_area_quad[1]) / 2.0).reshape(1, 2)
right_center_pt = (
(min_area_quad[1] - min_area_quad[2]) / 2.0).reshape(1, 2)
proj_unit_vec = (right_center_pt - left_center_pt) / (
np.linalg.norm(right_center_pt - left_center_pt) + 1e-6)
proj_unit_vec_tile = np.tile(proj_unit_vec,
(xy_text.shape[0], 1)) # (n, 2)
left_center_pt_tile = np.tile(left_center_pt,
(xy_text.shape[0], 1)) # (n, 2)
xy_text_to_left_center = xy_text - left_center_pt_tile
proj_value = np.sum(xy_text_to_left_center * proj_unit_vec_tile, axis=1)
xy_text = xy_text[np.argsort(proj_value)]
# convert to np and keep the num of point not greater then fixed_point_num
pos_info = np.array(xy_text).reshape(-1, 2)[:, ::-1] # xy-> yx
point_num = len(pos_info)
if point_num > fixed_point_num:
keep_ids = [
int((point_num * 1.0 / fixed_point_num) * x)
for x in range(fixed_point_num)
]
pos_info = pos_info[keep_ids, :]
keep = int(min(len(pos_info), fixed_point_num))
if np.random.rand() < 0.2 and reference_height >= 3:
dl = (np.random.rand(keep) - 0.5) * reference_height * 0.3
random_float = np.array([1, 0]).reshape([1, 2]) * dl.reshape(
[keep, 1])
pos_info += random_float
pos_info[:, 0] = np.clip(pos_info[:, 0], 0, max_h - 1)
pos_info[:, 1] = np.clip(pos_info[:, 1], 0, max_w - 1)
# padding to fixed length
pos_l = np.zeros((self.tcl_len, 3), dtype=np.int32)
pos_l[:, 0] = np.ones((self.tcl_len, )) * img_id
pos_m = np.zeros((self.tcl_len, 1), dtype=np.float32)
pos_l[:keep, 1:] = np.round(pos_info).astype(np.int32)
pos_m[:keep] = 1.0
return pos_l, pos_m
def generate_direction_map(self, poly_quads, n_char, direction_map):
"""
"""
width_list = []
height_list = []
for quad in poly_quads:
quad_w = (np.linalg.norm(quad[0] - quad[1]) +
np.linalg.norm(quad[2] - quad[3])) / 2.0
quad_h = (np.linalg.norm(quad[0] - quad[3]) +
np.linalg.norm(quad[2] - quad[1])) / 2.0
width_list.append(quad_w)
height_list.append(quad_h)
norm_width = max(sum(width_list) / n_char, 1.0)
average_height = max(sum(height_list) / len(height_list), 1.0)
k = 1
for quad in poly_quads:
direct_vector_full = (
(quad[1] + quad[2]) - (quad[0] + quad[3])) / 2.0
direct_vector = direct_vector_full / (
np.linalg.norm(direct_vector_full) + 1e-6) * norm_width
direction_label = tuple(
map(float,
[direct_vector[0], direct_vector[1], 1.0 / average_height]))
cv2.fillPoly(direction_map,
quad.round().astype(np.int32)[np.newaxis, :, :],
direction_label)
k += 1
return direction_map
def calculate_average_height(self, poly_quads):
"""
"""
height_list = []
for quad in poly_quads:
quad_h = (np.linalg.norm(quad[0] - quad[3]) +
np.linalg.norm(quad[2] - quad[1])) / 2.0
height_list.append(quad_h)
average_height = max(sum(height_list) / len(height_list), 1.0)
return average_height
def generate_tcl_ctc_label(self,
h,
w,
polys,
tags,
text_strs,
ds_ratio,
tcl_ratio=0.3,
shrink_ratio_of_width=0.15):
"""
Generate polygon.
"""
score_map_big = np.zeros(
(
h,
w, ), dtype=np.float32)
h, w = int(h * ds_ratio), int(w * ds_ratio)
polys = polys * ds_ratio
score_map = np.zeros(
(
h,
w, ), dtype=np.float32)
score_label_map = np.zeros(
(
h,
w, ), dtype=np.float32)
tbo_map = np.zeros((h, w, 5), dtype=np.float32)
training_mask = np.ones(
(
h,
w, ), dtype=np.float32)
direction_map = np.ones((h, w, 3)) * np.array([0, 0, 1]).reshape(
[1, 1, 3]).astype(np.float32)
label_idx = 0
score_label_map_text_label_list = []
pos_list, pos_mask, label_list = [], [], []
for poly_idx, poly_tag in enumerate(zip(polys, tags)):
poly = poly_tag[0]
tag = poly_tag[1]
# generate min_area_quad
min_area_quad, center_point = self.gen_min_area_quad_from_poly(poly)
min_area_quad_h = 0.5 * (
np.linalg.norm(min_area_quad[0] - min_area_quad[3]) +
np.linalg.norm(min_area_quad[1] - min_area_quad[2]))
min_area_quad_w = 0.5 * (
np.linalg.norm(min_area_quad[0] - min_area_quad[1]) +
np.linalg.norm(min_area_quad[2] - min_area_quad[3]))
if min(min_area_quad_h, min_area_quad_w) < self.min_text_size * ds_ratio \
or min(min_area_quad_h, min_area_quad_w) > self.max_text_size * ds_ratio:
continue
if tag:
cv2.fillPoly(training_mask,
poly.astype(np.int32)[np.newaxis, :, :], 0.15)
else:
text_label = text_strs[poly_idx]
text_label = self.prepare_text_label(text_label,
self.Lexicon_Table)
text_label_index_list = [[self.Lexicon_Table.index(c_)]
for c_ in text_label
if c_ in self.Lexicon_Table]
if len(text_label_index_list) < 1:
continue
tcl_poly = self.poly2tcl(poly, tcl_ratio)
tcl_quads = self.poly2quads(tcl_poly)
poly_quads = self.poly2quads(poly)
stcl_quads, quad_index = self.shrink_poly_along_width(
tcl_quads,
shrink_ratio_of_width=shrink_ratio_of_width,
expand_height_ratio=1.0 / tcl_ratio)
cv2.fillPoly(score_map,
np.round(stcl_quads).astype(np.int32), 1.0)
cv2.fillPoly(score_map_big,
np.round(stcl_quads / ds_ratio).astype(np.int32),
1.0)
for idx, quad in enumerate(stcl_quads):
quad_mask = np.zeros((h, w), dtype=np.float32)
quad_mask = cv2.fillPoly(
quad_mask,
np.round(quad[np.newaxis, :, :]).astype(np.int32), 1.0)
tbo_map = self.gen_quad_tbo(poly_quads[quad_index[idx]],
quad_mask, tbo_map)
# score label map and score_label_map_text_label_list for refine
if label_idx == 0:
text_pos_list_ = [[len(self.Lexicon_Table)], ]
score_label_map_text_label_list.append(text_pos_list_)
label_idx += 1
cv2.fillPoly(score_label_map,
np.round(poly_quads).astype(np.int32), label_idx)
score_label_map_text_label_list.append(text_label_index_list)
# direction info, fix-me
n_char = len(text_label_index_list)
direction_map = self.generate_direction_map(poly_quads, n_char,
direction_map)
# pos info
average_shrink_height = self.calculate_average_height(
stcl_quads)
pos_l, pos_m = self.fit_and_gather_tcl_points_v2(
min_area_quad,
poly,
max_h=h,
max_w=w,
fixed_point_num=64,
img_id=self.img_id,
reference_height=average_shrink_height)
label_l = text_label_index_list
if len(text_label_index_list) < 2:
continue
pos_list.append(pos_l)
pos_mask.append(pos_m)
label_list.append(label_l)
# use big score_map for smooth tcl lines
score_map_big_resized = cv2.resize(
score_map_big, dsize=None, fx=ds_ratio, fy=ds_ratio)
score_map = np.array(score_map_big_resized > 1e-3, dtype='float32')
return score_map, score_label_map, tbo_map, direction_map, training_mask, \
pos_list, pos_mask, label_list, score_label_map_text_label_list
def adjust_point(self, poly):
"""
adjust point order.
"""
point_num = poly.shape[0]
if point_num == 4:
len_1 = np.linalg.norm(poly[0] - poly[1])
len_2 = np.linalg.norm(poly[1] - poly[2])
len_3 = np.linalg.norm(poly[2] - poly[3])
len_4 = np.linalg.norm(poly[3] - poly[0])
if (len_1 + len_3) * 1.5 < (len_2 + len_4):
poly = poly[[1, 2, 3, 0], :]
elif point_num > 4:
vector_1 = poly[0] - poly[1]
vector_2 = poly[1] - poly[2]
cos_theta = np.dot(vector_1, vector_2) / (
np.linalg.norm(vector_1) * np.linalg.norm(vector_2) + 1e-6)
theta = np.arccos(np.round(cos_theta, decimals=4))
if abs(theta) > (70 / 180 * math.pi):
index = list(range(1, point_num)) + [0]
poly = poly[np.array(index), :]
return poly
def gen_min_area_quad_from_poly(self, poly):
"""
Generate min area quad from poly.
"""
point_num = poly.shape[0]
min_area_quad = np.zeros((4, 2), dtype=np.float32)
if point_num == 4:
min_area_quad = poly
center_point = np.sum(poly, axis=0) / 4
else:
rect = cv2.minAreaRect(poly.astype(
np.int32)) # (center (x,y), (width, height), angle of rotation)
center_point = rect[0]
box = np.array(cv2.boxPoints(rect))
first_point_idx = 0
min_dist = 1e4
for i in range(4):
dist = np.linalg.norm(box[(i + 0) % 4] - poly[0]) + \
np.linalg.norm(box[(i + 1) % 4] - poly[point_num // 2 - 1]) + \
np.linalg.norm(box[(i + 2) % 4] - poly[point_num // 2]) + \
np.linalg.norm(box[(i + 3) % 4] - poly[-1])
if dist < min_dist:
min_dist = dist
first_point_idx = i
for i in range(4):
min_area_quad[i] = box[(first_point_idx + i) % 4]
return min_area_quad, center_point
def shrink_quad_along_width(self,
quad,
begin_width_ratio=0.,
end_width_ratio=1.):
"""
Generate shrink_quad_along_width.
"""
ratio_pair = np.array(
[[begin_width_ratio], [end_width_ratio]], dtype=np.float32)
p0_1 = quad[0] + (quad[1] - quad[0]) * ratio_pair
p3_2 = quad[3] + (quad[2] - quad[3]) * ratio_pair
return np.array([p0_1[0], p0_1[1], p3_2[1], p3_2[0]])
def shrink_poly_along_width(self,
quads,
shrink_ratio_of_width,
expand_height_ratio=1.0):
"""
shrink poly with given length.
"""
upper_edge_list = []
def get_cut_info(edge_len_list, cut_len):
for idx, edge_len in enumerate(edge_len_list):
cut_len -= edge_len
if cut_len <= 0.000001:
ratio = (cut_len + edge_len_list[idx]) / edge_len_list[idx]
return idx, ratio
for quad in quads:
upper_edge_len = np.linalg.norm(quad[0] - quad[1])
upper_edge_list.append(upper_edge_len)
# length of left edge and right edge.
left_length = np.linalg.norm(quads[0][0] - quads[0][
3]) * expand_height_ratio
right_length = np.linalg.norm(quads[-1][1] - quads[-1][
2]) * expand_height_ratio
shrink_length = min(left_length, right_length,
sum(upper_edge_list)) * shrink_ratio_of_width
# shrinking length
upper_len_left = shrink_length
upper_len_right = sum(upper_edge_list) - shrink_length
left_idx, left_ratio = get_cut_info(upper_edge_list, upper_len_left)
left_quad = self.shrink_quad_along_width(
quads[left_idx], begin_width_ratio=left_ratio, end_width_ratio=1)
right_idx, right_ratio = get_cut_info(upper_edge_list, upper_len_right)
right_quad = self.shrink_quad_along_width(
quads[right_idx], begin_width_ratio=0, end_width_ratio=right_ratio)
out_quad_list = []
if left_idx == right_idx:
out_quad_list.append(
[left_quad[0], right_quad[1], right_quad[2], left_quad[3]])
else:
out_quad_list.append(left_quad)
for idx in range(left_idx + 1, right_idx):
out_quad_list.append(quads[idx])
out_quad_list.append(right_quad)
return np.array(out_quad_list), list(range(left_idx, right_idx + 1))
def prepare_text_label(self, label_str, Lexicon_Table):
"""
Prepare text lablel by given Lexicon_Table.
"""
if len(Lexicon_Table) == 36:
return label_str.lower()
else:
return label_str
def vector_angle(self, A, B):
"""
Calculate the angle between vector AB and x-axis positive direction.
"""
AB = np.array([B[1] - A[1], B[0] - A[0]])
return np.arctan2(*AB)
def theta_line_cross_point(self, theta, point):
"""
Calculate the line through given point and angle in ax + by + c =0 form.
"""
x, y = point
cos = np.cos(theta)
sin = np.sin(theta)
return [sin, -cos, cos * y - sin * x]
def line_cross_two_point(self, A, B):
"""
Calculate the line through given point A and B in ax + by + c =0 form.
"""
angle = self.vector_angle(A, B)
return self.theta_line_cross_point(angle, A)
def average_angle(self, poly):
"""
Calculate the average angle between left and right edge in given poly.
"""
p0, p1, p2, p3 = poly
angle30 = self.vector_angle(p3, p0)
angle21 = self.vector_angle(p2, p1)
return (angle30 + angle21) / 2
def line_cross_point(self, line1, line2):
"""
line1 and line2 in 0=ax+by+c form, compute the cross point of line1 and line2
"""
a1, b1, c1 = line1
a2, b2, c2 = line2
d = a1 * b2 - a2 * b1
if d == 0:
print('Cross point does not exist')
return np.array([0, 0], dtype=np.float32)
else:
x = (b1 * c2 - b2 * c1) / d
y = (a2 * c1 - a1 * c2) / d
return np.array([x, y], dtype=np.float32)
def quad2tcl(self, poly, ratio):
"""
Generate center line by poly clock-wise point. (4, 2)
"""
ratio_pair = np.array(
[[0.5 - ratio / 2], [0.5 + ratio / 2]], dtype=np.float32)
p0_3 = poly[0] + (poly[3] - poly[0]) * ratio_pair
p1_2 = poly[1] + (poly[2] - poly[1]) * ratio_pair
return np.array([p0_3[0], p1_2[0], p1_2[1], p0_3[1]])
def poly2tcl(self, poly, ratio):
"""
Generate center line by poly clock-wise point.
"""
ratio_pair = np.array(
[[0.5 - ratio / 2], [0.5 + ratio / 2]], dtype=np.float32)
tcl_poly = np.zeros_like(poly)
point_num = poly.shape[0]
for idx in range(point_num // 2):
point_pair = poly[idx] + (poly[point_num - 1 - idx] - poly[idx]
) * ratio_pair
tcl_poly[idx] = point_pair[0]
tcl_poly[point_num - 1 - idx] = point_pair[1]
return tcl_poly
def gen_quad_tbo(self, quad, tcl_mask, tbo_map):
"""
Generate tbo_map for give quad.
"""
# upper and lower line function: ax + by + c = 0;
up_line = self.line_cross_two_point(quad[0], quad[1])
lower_line = self.line_cross_two_point(quad[3], quad[2])
quad_h = 0.5 * (np.linalg.norm(quad[0] - quad[3]) +
np.linalg.norm(quad[1] - quad[2]))
quad_w = 0.5 * (np.linalg.norm(quad[0] - quad[1]) +
np.linalg.norm(quad[2] - quad[3]))
# average angle of left and right line.
angle = self.average_angle(quad)
xy_in_poly = np.argwhere(tcl_mask == 1)
for y, x in xy_in_poly:
point = (x, y)
line = self.theta_line_cross_point(angle, point)
cross_point_upper = self.line_cross_point(up_line, line)
cross_point_lower = self.line_cross_point(lower_line, line)
##FIX, offset reverse
upper_offset_x, upper_offset_y = cross_point_upper - point
lower_offset_x, lower_offset_y = cross_point_lower - point
tbo_map[y, x, 0] = upper_offset_y
tbo_map[y, x, 1] = upper_offset_x
tbo_map[y, x, 2] = lower_offset_y
tbo_map[y, x, 3] = lower_offset_x
tbo_map[y, x, 4] = 1.0 / max(min(quad_h, quad_w), 1.0) * 2
return tbo_map
def poly2quads(self, poly):
"""
Split poly into quads.
"""
quad_list = []
point_num = poly.shape[0]
# point pair
point_pair_list = []
for idx in range(point_num // 2):
point_pair = [poly[idx], poly[point_num - 1 - idx]]
point_pair_list.append(point_pair)
quad_num = point_num // 2 - 1
for idx in range(quad_num):
# reshape and adjust to clock-wise
quad_list.append((np.array(point_pair_list)[[idx, idx + 1]]
).reshape(4, 2)[[0, 2, 3, 1]])
return np.array(quad_list)
def rotate_im_poly(self, im, text_polys):
"""
rotate image with 90 / 180 / 270 degre
"""
im_w, im_h = im.shape[1], im.shape[0]
dst_im = im.copy()
dst_polys = []
rand_degree_ratio = np.random.rand()
rand_degree_cnt = 1
if rand_degree_ratio > 0.5:
rand_degree_cnt = 3
for i in range(rand_degree_cnt):
dst_im = np.rot90(dst_im)
rot_degree = -90 * rand_degree_cnt
rot_angle = rot_degree * math.pi / 180.0
n_poly = text_polys.shape[0]
cx, cy = 0.5 * im_w, 0.5 * im_h
ncx, ncy = 0.5 * dst_im.shape[1], 0.5 * dst_im.shape[0]
for i in range(n_poly):
wordBB = text_polys[i]
poly = []
for j in range(4): # 16->4
sx, sy = wordBB[j][0], wordBB[j][1]
dx = math.cos(rot_angle) * (sx - cx) - math.sin(rot_angle) * (
sy - cy) + ncx
dy = math.sin(rot_angle) * (sx - cx) + math.cos(rot_angle) * (
sy - cy) + ncy
poly.append([dx, dy])
dst_polys.append(poly)
return dst_im, np.array(dst_polys, dtype=np.float32)
def __call__(self, data):
input_size = 512
im = data['image']
text_polys = data['polys']
text_tags = data['ignore_tags']
text_strs = data['texts']
h, w, _ = im.shape
text_polys, text_tags, hv_tags = self.check_and_validate_polys(
text_polys, text_tags, (h, w))
if text_polys.shape[0] <= 0:
return None
# set aspect ratio and keep area fix
asp_scales = np.arange(1.0, 1.55, 0.1)
asp_scale = np.random.choice(asp_scales)
if np.random.rand() < 0.5:
asp_scale = 1.0 / asp_scale
asp_scale = math.sqrt(asp_scale)
asp_wx = asp_scale
asp_hy = 1.0 / asp_scale
im = cv2.resize(im, dsize=None, fx=asp_wx, fy=asp_hy)
text_polys[:, :, 0] *= asp_wx
text_polys[:, :, 1] *= asp_hy
h, w, _ = im.shape
if max(h, w) > 2048:
rd_scale = 2048.0 / max(h, w)
im = cv2.resize(im, dsize=None, fx=rd_scale, fy=rd_scale)
text_polys *= rd_scale
h, w, _ = im.shape
if min(h, w) < 16:
return None
# no background
im, text_polys, text_tags, hv_tags, text_strs = self.crop_area(
im,
text_polys,
text_tags,
hv_tags,
text_strs,
crop_background=False)
if text_polys.shape[0] == 0:
return None
# # continue for all ignore case
if np.sum((text_tags * 1.0)) >= text_tags.size:
return None
new_h, new_w, _ = im.shape
if (new_h is None) or (new_w is None):
return None
# resize image
std_ratio = float(input_size) / max(new_w, new_h)
rand_scales = np.array(
[0.25, 0.375, 0.5, 0.625, 0.75, 0.875, 1.0, 1.0, 1.0, 1.0, 1.0])
rz_scale = std_ratio * np.random.choice(rand_scales)
im = cv2.resize(im, dsize=None, fx=rz_scale, fy=rz_scale)
text_polys[:, :, 0] *= rz_scale
text_polys[:, :, 1] *= rz_scale
# add gaussian blur
if np.random.rand() < 0.1 * 0.5:
ks = np.random.permutation(5)[0] + 1
ks = int(ks / 2) * 2 + 1
im = cv2.GaussianBlur(im, ksize=(ks, ks), sigmaX=0, sigmaY=0)
# add brighter
if np.random.rand() < 0.1 * 0.5:
im = im * (1.0 + np.random.rand() * 0.5)
im = np.clip(im, 0.0, 255.0)
# add darker
if np.random.rand() < 0.1 * 0.5:
im = im * (1.0 - np.random.rand() * 0.5)
im = np.clip(im, 0.0, 255.0)
# Padding the im to [input_size, input_size]
new_h, new_w, _ = im.shape
if min(new_w, new_h) < input_size * 0.5:
return None
im_padded = np.ones((input_size, input_size, 3), dtype=np.float32)
im_padded[:, :, 2] = 0.485 * 255
im_padded[:, :, 1] = 0.456 * 255
im_padded[:, :, 0] = 0.406 * 255
# Random the start position
del_h = input_size - new_h
del_w = input_size - new_w
sh, sw = 0, 0
if del_h > 1:
sh = int(np.random.rand() * del_h)
if del_w > 1:
sw = int(np.random.rand() * del_w)
# Padding
im_padded[sh:sh + new_h, sw:sw + new_w, :] = im.copy()
text_polys[:, :, 0] += sw
text_polys[:, :, 1] += sh
score_map, score_label_map, border_map, direction_map, training_mask, \
pos_list, pos_mask, label_list, score_label_map_text_label = self.generate_tcl_ctc_label(input_size,
input_size,
text_polys,
text_tags,
text_strs, 0.25)
if len(label_list) <= 0: # eliminate negative samples
return None
pos_list_temp = np.zeros([64, 3])
pos_mask_temp = np.zeros([64, 1])
label_list_temp = np.zeros([self.max_text_length, 1]) + self.pad_num
for i, label in enumerate(label_list):
n = len(label)
if n > self.max_text_length:
label_list[i] = label[:self.max_text_length]
continue
while n < self.max_text_length:
label.append([self.pad_num])
n += 1
for i in range(len(label_list)):
label_list[i] = np.array(label_list[i])
if len(pos_list) <= 0 or len(pos_list) > self.max_text_nums:
return None
for __ in range(self.max_text_nums - len(pos_list), 0, -1):
pos_list.append(pos_list_temp)
pos_mask.append(pos_mask_temp)
label_list.append(label_list_temp)
if self.img_id == self.batch_size - 1:
self.img_id = 0
else:
self.img_id += 1
im_padded[:, :, 2] -= 0.485 * 255
im_padded[:, :, 1] -= 0.456 * 255
im_padded[:, :, 0] -= 0.406 * 255
im_padded[:, :, 2] /= (255.0 * 0.229)
im_padded[:, :, 1] /= (255.0 * 0.224)
im_padded[:, :, 0] /= (255.0 * 0.225)
im_padded = im_padded.transpose((2, 0, 1))
images = im_padded[::-1, :, :]
tcl_maps = score_map[np.newaxis, :, :]
tcl_label_maps = score_label_map[np.newaxis, :, :]
border_maps = border_map.transpose((2, 0, 1))
direction_maps = direction_map.transpose((2, 0, 1))
training_masks = training_mask[np.newaxis, :, :]
pos_list = np.array(pos_list)
pos_mask = np.array(pos_mask)
label_list = np.array(label_list)
data['images'] = images
data['tcl_maps'] = tcl_maps
data['tcl_label_maps'] = tcl_label_maps
data['border_maps'] = border_maps
data['direction_maps'] = direction_maps
data['training_masks'] = training_masks
data['label_list'] = label_list
data['pos_list'] = pos_list
data['pos_mask'] = pos_mask
return data
ppocr/data/imaug/rec_img_aug.py
View file @ 19eb7eb8
......@@ -16,7 +16,7 @@ import math
import cv2
import numpy as np
import random
from PIL import Image
from .text_image_aug import tia_perspective, tia_stretch, tia_distort
......@@ -43,6 +43,25 @@ class ClsResizeImg(object):
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):
def __init__(self,
image_shape,
......
ppocr/data/pgnet_dataset.py 0 → 100644
View file @ 19eb7eb8
# 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 numpy as np
import os
from paddle.io import Dataset
from .imaug import transform, create_operators
import random
class PGDataSet(Dataset):
def __init__(self, config, mode, logger, seed=None):
super(PGDataSet, self).__init__()
self.logger = logger
self.seed = seed
self.mode = mode
global_config = config['Global']
dataset_config = config[mode]['dataset']
loader_config = config[mode]['loader']
self.delimiter = dataset_config.get('delimiter', '\t')
label_file_list = dataset_config.pop('label_file_list')
data_source_num = len(label_file_list)
ratio_list = dataset_config.get("ratio_list", [1.0])
if isinstance(ratio_list, (float, int)):
ratio_list = [float(ratio_list)] * int(data_source_num)
assert len(
ratio_list
) == data_source_num, "The length of ratio_list should be the same as the file_list."
self.data_dir = dataset_config['data_dir']
self.do_shuffle = loader_config['shuffle']
logger.info("Initialize indexs of datasets:%s" % label_file_list)
self.data_lines = self.get_image_info_list(label_file_list, ratio_list)
self.data_idx_order_list = list(range(len(self.data_lines)))
if mode.lower() == "train":
self.shuffle_data_random()
self.ops = create_operators(dataset_config['transforms'], global_config)
def shuffle_data_random(self):
if self.do_shuffle:
random.seed(self.seed)
random.shuffle(self.data_lines)
return
def get_image_info_list(self, file_list, ratio_list):
if isinstance(file_list, str):
file_list = [file_list]
data_lines = []
for idx, file in enumerate(file_list):
with open(file, "rb") as f:
lines = f.readlines()
if self.mode == "train" or ratio_list[idx] < 1.0:
random.seed(self.seed)
lines = random.sample(lines,
round(len(lines) * ratio_list[idx]))
data_lines.extend(lines)
return data_lines
def __getitem__(self, idx):
file_idx = self.data_idx_order_list[idx]
data_line = self.data_lines[file_idx]
img_id = 0
try:
data_line = data_line.decode('utf-8')
substr = data_line.strip("\n").split(self.delimiter)
file_name = substr[0]
label = substr[1]
img_path = os.path.join(self.data_dir, file_name)
if self.mode.lower() == 'eval':
try:
img_id = int(data_line.split(".")[0][7:])
except:
img_id = 0
data = {'img_path': img_path, 'label': label, 'img_id': img_id}
if not os.path.exists(img_path):
raise Exception("{} does not exist!".format(img_path))
with open(data['img_path'], 'rb') as f:
img = f.read()
data['image'] = img
outs = transform(data, self.ops)
except Exception as e:
self.logger.error(
"When parsing line {}, error happened with msg: {}".format(
self.data_idx_order_list[idx], e))
outs = None
if outs is None:
return self.__getitem__(np.random.randint(self.__len__()))
return outs
def __len__(self):
return len(self.data_idx_order_list)
ppocr/data/pubtab_dataset.py 0 → 100644
View file @ 19eb7eb8
# 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 numpy as np
import os
import random
from paddle.io import Dataset
import json
from .imaug import transform, create_operators
class PubTabDataSet(Dataset):
def __init__(self, config, mode, logger, seed=None):
super(PubTabDataSet, self).__init__()
self.logger = logger
global_config = config['Global']
dataset_config = config[mode]['dataset']
loader_config = config[mode]['loader']
label_file_path = dataset_config.pop('label_file_path')
self.data_dir = dataset_config['data_dir']
self.do_shuffle = loader_config['shuffle']
self.do_hard_select = False
if 'hard_select' in loader_config:
self.do_hard_select = loader_config['hard_select']
self.hard_prob = loader_config['hard_prob']
if self.do_hard_select:
self.img_select_prob = self.load_hard_select_prob()
self.table_select_type = None
if 'table_select_type' in loader_config:
self.table_select_type = loader_config['table_select_type']
self.table_select_prob = loader_config['table_select_prob']
self.seed = seed
logger.info("Initialize indexs of datasets:%s" % label_file_path)
with open(label_file_path, "rb") as f:
self.data_lines = f.readlines()
self.data_idx_order_list = list(range(len(self.data_lines)))
if mode.lower() == "train":
self.shuffle_data_random()
self.ops = create_operators(dataset_config['transforms'], global_config)
def shuffle_data_random(self):
if self.do_shuffle:
random.seed(self.seed)
random.shuffle(self.data_lines)
return
def __getitem__(self, idx):
try:
data_line = self.data_lines[idx]
data_line = data_line.decode('utf-8').strip("\n")
info = json.loads(data_line)
file_name = info['filename']
select_flag = True
if self.do_hard_select:
prob = self.img_select_prob[file_name]
if prob < random.uniform(0, 1):
select_flag = False
if self.table_select_type:
structure = info['html']['structure']['tokens'].copy()
structure_str = ''.join(structure)
table_type = "simple"
if 'colspan' in structure_str or 'rowspan' in structure_str:
table_type = "complex"
if table_type == "complex":
if self.table_select_prob < random.uniform(0, 1):
select_flag = False
if select_flag:
cells = info['html']['cells'].copy()
structure = info['html']['structure'].copy()
img_path = os.path.join(self.data_dir, file_name)
data = {'img_path': img_path, 'cells': cells, 'structure':structure}
if not os.path.exists(img_path):
raise Exception("{} does not exist!".format(img_path))
with open(data['img_path'], 'rb') as f:
img = f.read()
data['image'] = img
outs = transform(data, self.ops)
else:
outs = None
except Exception as e:
self.logger.error(
"When parsing line {}, error happened with msg: {}".format(
data_line, e))
outs = None
if outs is None:
return self.__getitem__(np.random.randint(self.__len__()))
return outs
def __len__(self):
return len(self.data_idx_order_list)
ppocr/data/simple_dataset.py
View file @ 19eb7eb8
......@@ -69,6 +69,36 @@ class SimpleDataSet(Dataset):
random.shuffle(self.data_lines)
return
def get_ext_data(self):
ext_data_num = 0
for op in self.ops:
if hasattr(op, 'ext_data_num'):
ext_data_num = getattr(op, 'ext_data_num')
break
load_data_ops = self.ops[:2]
ext_data = []
while len(ext_data) < ext_data_num:
file_idx = self.data_idx_order_list[np.random.randint(self.__len__(
))]
data_line = self.data_lines[file_idx]
data_line = data_line.decode('utf-8')
substr = data_line.strip("\n").split(self.delimiter)
file_name = substr[0]
label = substr[1]
img_path = os.path.join(self.data_dir, file_name)
data = {'img_path': img_path, 'label': label}
if not os.path.exists(img_path):
continue
with open(data['img_path'], 'rb') as f:
img = f.read()
data['image'] = img
data = transform(data, load_data_ops)
if data is None:
continue
ext_data.append(data)
return ext_data
def __getitem__(self, idx):
file_idx = self.data_idx_order_list[idx]
data_line = self.data_lines[file_idx]
......@@ -84,6 +114,7 @@ class SimpleDataSet(Dataset):
with open(data['img_path'], 'rb') as f:
img = f.read()
data['image'] = img
data['ext_data'] = self.get_ext_data()
outs = transform(data, self.ops)
except Exception as e:
self.logger.error(
......
ppocr/losses/__init__.py
View file @ 19eb7eb8
......@@ -13,25 +13,38 @@
# limitations under the License.
import copy
import paddle
import paddle.nn as nn
# det loss
from .det_db_loss import DBLoss
from .det_east_loss import EASTLoss
from .det_sast_loss import SASTLoss
def build_loss(config):
# det loss
from .det_db_loss import DBLoss
from .det_east_loss import EASTLoss
from .det_sast_loss import SASTLoss
# rec loss
from .rec_ctc_loss import CTCLoss
from .rec_att_loss import AttentionLoss
from .rec_srn_loss import SRNLoss
from .rec_nrtr_loss import NRTRLoss
# cls loss
from .cls_loss import ClsLoss
# e2e loss
from .e2e_pg_loss import PGLoss
# rec loss
from .rec_ctc_loss import CTCLoss
from .rec_att_loss import AttentionLoss
from .rec_srn_loss import SRNLoss
# basic loss function
from .basic_loss import DistanceLoss
# cls loss
from .cls_loss import ClsLoss
# combined loss function
from .combined_loss import CombinedLoss
# table loss
from .table_att_loss import TableAttentionLoss
def build_loss(config):
support_dict = [
'DBLoss', 'EASTLoss', 'SASTLoss', 'CTCLoss', 'ClsLoss', 'AttentionLoss',
'SRNLoss'
'SRNLoss', 'PGLoss', 'CombinedLoss', 'NRTRLoss', 'TableAttentionLoss'
]
config = copy.deepcopy(config)
......
ppocr/losses/basic_loss.py 0 → 100644
View file @ 19eb7eb8
#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
import paddle.nn as nn
import paddle.nn.functional as F
from paddle.nn import L1Loss
from paddle.nn import MSELoss as L2Loss
from paddle.nn import SmoothL1Loss
class CELoss(nn.Layer):
def __init__(self, epsilon=None):
super().__init__()
if epsilon is not None and (epsilon <= 0 or epsilon >= 1):
epsilon = None
self.epsilon = epsilon
def _labelsmoothing(self, target, class_num):
if target.shape[-1] != class_num:
one_hot_target = F.one_hot(target, class_num)
else:
one_hot_target = target
soft_target = F.label_smooth(one_hot_target, epsilon=self.epsilon)
soft_target = paddle.reshape(soft_target, shape=[-1, class_num])
return soft_target
def forward(self, x, label):
loss_dict = {}
if self.epsilon is not None:
class_num = x.shape[-1]
label = self._labelsmoothing(label, class_num)
x = -F.log_softmax(x, axis=-1)
loss = paddle.sum(x * label, axis=-1)
else:
if label.shape[-1] == x.shape[-1]:
label = F.softmax(label, axis=-1)
soft_label = True
else:
soft_label = False
loss = F.cross_entropy(x, label=label, soft_label=soft_label)
return loss
class KLJSLoss(object):
def __init__(self, mode='kl'):
assert mode in ['kl', 'js', 'KL', 'JS'], "mode can only be one of ['kl', 'js', 'KL', 'JS']"
self.mode = mode
def __call__(self, p1, p2, reduction="mean"):
loss = paddle.multiply(p2, paddle.log( (p2+1e-5)/(p1+1e-5) + 1e-5))
if self.mode.lower() == "js":
loss += paddle.multiply(p1, paddle.log((p1+1e-5)/(p2+1e-5) + 1e-5))
loss *= 0.5
if reduction == "mean":
loss = paddle.mean(loss, axis=[1,2])
elif reduction=="none" or reduction is None:
return loss
else:
loss = paddle.sum(loss, axis=[1,2])
return loss
class DMLLoss(nn.Layer):
"""
DMLLoss
"""
def __init__(self, act=None):
super().__init__()
if act is not None:
assert act in ["softmax", "sigmoid"]
if act == "softmax":
self.act = nn.Softmax(axis=-1)
elif act == "sigmoid":
self.act = nn.Sigmoid()
else:
self.act = None
self.jskl_loss = KLJSLoss(mode="js")
def forward(self, out1, out2):
if self.act is not None:
out1 = self.act(out1)
out2 = self.act(out2)
if len(out1.shape) < 2:
log_out1 = paddle.log(out1)
log_out2 = paddle.log(out2)
loss = (F.kl_div(
log_out1, out2, reduction='batchmean') + F.kl_div(
log_out2, out1, reduction='batchmean')) / 2.0
else:
loss = self.jskl_loss(out1, out2)
return loss
class DistanceLoss(nn.Layer):
"""
DistanceLoss:
mode: loss mode
"""
def __init__(self, mode="l2", **kargs):
super().__init__()
assert mode in ["l1", "l2", "smooth_l1"]
if mode == "l1":
self.loss_func = nn.L1Loss(**kargs)
elif mode == "l2":
self.loss_func = nn.MSELoss(**kargs)
elif mode == "smooth_l1":
self.loss_func = nn.SmoothL1Loss(**kargs)
def forward(self, x, y):
return self.loss_func(x, y)
ppocr/losses/cls_loss.py
View file @ 19eb7eb8
......@@ -24,7 +24,7 @@ class ClsLoss(nn.Layer):
super(ClsLoss, self).__init__()
self.loss_func = nn.CrossEntropyLoss(reduction='mean')
def __call__(self, predicts, batch):
label = batch[1]
def forward(self, predicts, batch):
label = batch[1].astype("int64")
loss = self.loss_func(input=predicts, label=label)
return {'loss': loss}
ppocr/losses/combined_loss.py 0 → 100644
View file @ 19eb7eb8
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
#
# 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
import paddle.nn as nn
from .distillation_loss import DistillationCTCLoss
from .distillation_loss import DistillationDMLLoss
from .distillation_loss import DistillationDistanceLoss, DistillationDBLoss, DistillationDilaDBLoss
class CombinedLoss(nn.Layer):
"""
CombinedLoss:
a combionation of loss function
"""
def __init__(self, loss_config_list=None):
super().__init__()
self.loss_func = []
self.loss_weight = []
assert isinstance(loss_config_list, list), (
'operator config should be a list')
for config in loss_config_list:
assert isinstance(config,
dict) and len(config) == 1, "yaml format error"
name = list(config)[0]
param = config[name]
assert "weight" in param, "weight must be in param, but param just contains {}".format(
param.keys())
self.loss_weight.append(param.pop("weight"))
self.loss_func.append(eval(name)(**param))
def forward(self, input, batch, **kargs):
loss_dict = {}
loss_all = 0.
for idx, loss_func in enumerate(self.loss_func):
loss = loss_func(input, batch, **kargs)
if isinstance(loss, paddle.Tensor):
loss = {"loss_{}_{}".format(str(loss), idx): loss}
weight = self.loss_weight[idx]
for key in loss.keys():
if key == "loss":
loss_all += loss[key] * weight
else:
loss_dict["{}_{}".format(key, idx)] = loss[key]
loss_dict["loss"] = loss_all
return loss_dict
ppocr/losses/distillation_loss.py 0 → 100644
View file @ 19eb7eb8
#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
import paddle.nn as nn
import numpy as np
import cv2
from .rec_ctc_loss import CTCLoss
from .basic_loss import DMLLoss
from .basic_loss import DistanceLoss
from .det_db_loss import DBLoss
from .det_basic_loss import BalanceLoss, MaskL1Loss, DiceLoss
def _sum_loss(loss_dict):
if "loss" in loss_dict.keys():
return loss_dict
else:
loss_dict["loss"] = 0.
for k, value in loss_dict.items():
if k == "loss":
continue
else:
loss_dict["loss"] += value
return loss_dict
class DistillationDMLLoss(DMLLoss):
"""
"""
def __init__(self,
model_name_pairs=[],
act=None,
key=None,
maps_name=None,
name="dml"):
super().__init__(act=act)
assert isinstance(model_name_pairs, list)
self.key = key
self.model_name_pairs = self._check_model_name_pairs(model_name_pairs)
self.name = name
self.maps_name = self._check_maps_name(maps_name)
def _check_model_name_pairs(self, model_name_pairs):
if not isinstance(model_name_pairs, list):
return []
elif isinstance(model_name_pairs[0], list) and isinstance(model_name_pairs[0][0], str):
return model_name_pairs
else:
return [model_name_pairs]
def _check_maps_name(self, maps_name):
if maps_name is None:
return None
elif type(maps_name) == str:
return [maps_name]
elif type(maps_name) == list:
return [maps_name]
else:
return None
def _slice_out(self, outs):
new_outs = {}
for k in self.maps_name:
if k == "thrink_maps":
new_outs[k] = outs[:, 0, :, :]
elif k == "threshold_maps":
new_outs[k] = outs[:, 1, :, :]
elif k == "binary_maps":
new_outs[k] = outs[:, 2, :, :]
else:
continue
return new_outs
def forward(self, predicts, batch):
loss_dict = dict()
for idx, pair in enumerate(self.model_name_pairs):
out1 = predicts[pair[0]]
out2 = predicts[pair[1]]
if self.key is not None:
out1 = out1[self.key]
out2 = out2[self.key]
if self.maps_name is None:
loss = super().forward(out1, out2)
if isinstance(loss, dict):
for key in loss:
loss_dict["{}_{}_{}_{}".format(key, pair[0], pair[1],
idx)] = loss[key]
else:
loss_dict["{}_{}".format(self.name, idx)] = loss
else:
outs1 = self._slice_out(out1)
outs2 = self._slice_out(out2)
for _c, k in enumerate(outs1.keys()):
loss = super().forward(outs1[k], outs2[k])
if isinstance(loss, dict):
for key in loss:
loss_dict["{}_{}_{}_{}_{}".format(key, pair[
0], pair[1], map_name, idx)] = loss[key]
else:
loss_dict["{}_{}_{}".format(self.name, self.maps_name[_c],
idx)] = loss
loss_dict = _sum_loss(loss_dict)
return loss_dict
class DistillationCTCLoss(CTCLoss):
def __init__(self, model_name_list=[], key=None, name="loss_ctc"):
super().__init__()
self.model_name_list = model_name_list
self.key = key
self.name = name
def forward(self, predicts, batch):
loss_dict = dict()
for idx, model_name in enumerate(self.model_name_list):
out = predicts[model_name]
if self.key is not None:
out = out[self.key]
loss = super().forward(out, batch)
if isinstance(loss, dict):
for key in loss:
loss_dict["{}_{}_{}".format(self.name, model_name,
idx)] = loss[key]
else:
loss_dict["{}_{}".format(self.name, model_name)] = loss
return loss_dict
class DistillationDBLoss(DBLoss):
def __init__(self,
model_name_list=[],
balance_loss=True,
main_loss_type='DiceLoss',
alpha=5,
beta=10,
ohem_ratio=3,
eps=1e-6,
name="db",
**kwargs):
super().__init__()
self.model_name_list = model_name_list
self.name = name
self.key = None
def forward(self, predicts, batch):
loss_dict = {}
for idx, model_name in enumerate(self.model_name_list):
out = predicts[model_name]
if self.key is not None:
out = out[self.key]
loss = super().forward(out, batch)
if isinstance(loss, dict):
for key in loss.keys():
if key == "loss":
continue
name = "{}_{}_{}".format(self.name, model_name, key)
loss_dict[name] = loss[key]
else:
loss_dict["{}_{}".format(self.name, model_name)] = loss
loss_dict = _sum_loss(loss_dict)
return loss_dict
class DistillationDilaDBLoss(DBLoss):
def __init__(self,
model_name_pairs=[],
key=None,
balance_loss=True,
main_loss_type='DiceLoss',
alpha=5,
beta=10,
ohem_ratio=3,
eps=1e-6,
name="dila_dbloss"):
super().__init__()
self.model_name_pairs = model_name_pairs
self.name = name
self.key = key
def forward(self, predicts, batch):
loss_dict = dict()
for idx, pair in enumerate(self.model_name_pairs):
stu_outs = predicts[pair[0]]
tch_outs = predicts[pair[1]]
if self.key is not None:
stu_preds = stu_outs[self.key]
tch_preds = tch_outs[self.key]
stu_shrink_maps = stu_preds[:, 0, :, :]
stu_binary_maps = stu_preds[:, 2, :, :]
# dilation to teacher prediction
dilation_w = np.array([[1, 1], [1, 1]])
th_shrink_maps = tch_preds[:, 0, :, :]
th_shrink_maps = th_shrink_maps.numpy() > 0.3 # thresh = 0.3
dilate_maps = np.zeros_like(th_shrink_maps).astype(np.float32)
for i in range(th_shrink_maps.shape[0]):
dilate_maps[i] = cv2.dilate(
th_shrink_maps[i, :, :].astype(np.uint8), dilation_w)
th_shrink_maps = paddle.to_tensor(dilate_maps)
label_threshold_map, label_threshold_mask, label_shrink_map, label_shrink_mask = batch[
1:]
# calculate the shrink map loss
bce_loss = self.alpha * self.bce_loss(
stu_shrink_maps, th_shrink_maps, label_shrink_mask)
loss_binary_maps = self.dice_loss(stu_binary_maps, th_shrink_maps,
label_shrink_mask)
# k = f"{self.name}_{pair[0]}_{pair[1]}"
k = "{}_{}_{}".format(self.name, pair[0], pair[1])
loss_dict[k] = bce_loss + loss_binary_maps
loss_dict = _sum_loss(loss_dict)
return loss_dict
class DistillationDistanceLoss(DistanceLoss):
"""
"""
def __init__(self,
mode="l2",
model_name_pairs=[],
key=None,
name="loss_distance",
**kargs):
super().__init__(mode=mode, **kargs)
assert isinstance(model_name_pairs, list)
self.key = key
self.model_name_pairs = model_name_pairs
self.name = name + "_l2"
def forward(self, predicts, batch):
loss_dict = dict()
for idx, pair in enumerate(self.model_name_pairs):
out1 = predicts[pair[0]]
out2 = predicts[pair[1]]
if self.key is not None:
out1 = out1[self.key]
out2 = out2[self.key]
loss = super().forward(out1, out2)
if isinstance(loss, dict):
for key in loss:
loss_dict["{}_{}_{}".format(self.name, key, idx)] = loss[
key]
else:
loss_dict["{}_{}_{}_{}".format(self.name, pair[0], pair[1],
idx)] = loss
return loss_dict
ppocr/losses/e2e_pg_loss.py 0 → 100644
View file @ 19eb7eb8
# 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.
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
from paddle import nn
import paddle
from .det_basic_loss import DiceLoss
from ppocr.utils.e2e_utils.extract_batchsize import pre_process
class PGLoss(nn.Layer):
def __init__(self,
tcl_bs,
max_text_length,
max_text_nums,
pad_num,
eps=1e-6,
**kwargs):
super(PGLoss, self).__init__()
self.tcl_bs = tcl_bs
self.max_text_nums = max_text_nums
self.max_text_length = max_text_length
self.pad_num = pad_num
self.dice_loss = DiceLoss(eps=eps)
def border_loss(self, f_border, l_border, l_score, l_mask):
l_border_split, l_border_norm = paddle.tensor.split(
l_border, num_or_sections=[4, 1], axis=1)
f_border_split = f_border
b, c, h, w = l_border_norm.shape
l_border_norm_split = paddle.expand(
x=l_border_norm, shape=[b, 4 * c, h, w])
b, c, h, w = l_score.shape
l_border_score = paddle.expand(x=l_score, shape=[b, 4 * c, h, w])
b, c, h, w = l_mask.shape
l_border_mask = paddle.expand(x=l_mask, shape=[b, 4 * c, h, w])
border_diff = l_border_split - f_border_split
abs_border_diff = paddle.abs(border_diff)
border_sign = abs_border_diff < 1.0
border_sign = paddle.cast(border_sign, dtype='float32')
border_sign.stop_gradient = True
border_in_loss = 0.5 * abs_border_diff * abs_border_diff * border_sign + \
(abs_border_diff - 0.5) * (1.0 - border_sign)
border_out_loss = l_border_norm_split * border_in_loss
border_loss = paddle.sum(border_out_loss * l_border_score * l_border_mask) / \
(paddle.sum(l_border_score * l_border_mask) + 1e-5)
return border_loss
def direction_loss(self, f_direction, l_direction, l_score, l_mask):
l_direction_split, l_direction_norm = paddle.tensor.split(
l_direction, num_or_sections=[2, 1], axis=1)
f_direction_split = f_direction
b, c, h, w = l_direction_norm.shape
l_direction_norm_split = paddle.expand(
x=l_direction_norm, shape=[b, 2 * c, h, w])
b, c, h, w = l_score.shape
l_direction_score = paddle.expand(x=l_score, shape=[b, 2 * c, h, w])
b, c, h, w = l_mask.shape
l_direction_mask = paddle.expand(x=l_mask, shape=[b, 2 * c, h, w])
direction_diff = l_direction_split - f_direction_split
abs_direction_diff = paddle.abs(direction_diff)
direction_sign = abs_direction_diff < 1.0
direction_sign = paddle.cast(direction_sign, dtype='float32')
direction_sign.stop_gradient = True
direction_in_loss = 0.5 * abs_direction_diff * abs_direction_diff * direction_sign + \
(abs_direction_diff - 0.5) * (1.0 - direction_sign)
direction_out_loss = l_direction_norm_split * direction_in_loss
direction_loss = paddle.sum(direction_out_loss * l_direction_score * l_direction_mask) / \
(paddle.sum(l_direction_score * l_direction_mask) + 1e-5)
return direction_loss
def ctcloss(self, f_char, tcl_pos, tcl_mask, tcl_label, label_t):
f_char = paddle.transpose(f_char, [0, 2, 3, 1])
tcl_pos = paddle.reshape(tcl_pos, [-1, 3])
tcl_pos = paddle.cast(tcl_pos, dtype=int)
f_tcl_char = paddle.gather_nd(f_char, tcl_pos)
f_tcl_char = paddle.reshape(f_tcl_char,
[-1, 64, 37]) # len(Lexicon_Table)+1
f_tcl_char_fg, f_tcl_char_bg = paddle.split(f_tcl_char, [36, 1], axis=2)
f_tcl_char_bg = f_tcl_char_bg * tcl_mask + (1.0 - tcl_mask) * 20.0
b, c, l = tcl_mask.shape
tcl_mask_fg = paddle.expand(x=tcl_mask, shape=[b, c, 36 * l])
tcl_mask_fg.stop_gradient = True
f_tcl_char_fg = f_tcl_char_fg * tcl_mask_fg + (1.0 - tcl_mask_fg) * (
-20.0)
f_tcl_char_mask = paddle.concat([f_tcl_char_fg, f_tcl_char_bg], axis=2)
f_tcl_char_ld = paddle.transpose(f_tcl_char_mask, (1, 0, 2))
N, B, _ = f_tcl_char_ld.shape
input_lengths = paddle.to_tensor([N] * B, dtype='int64')
cost = paddle.nn.functional.ctc_loss(
log_probs=f_tcl_char_ld,
labels=tcl_label,
input_lengths=input_lengths,
label_lengths=label_t,
blank=self.pad_num,
reduction='none')
cost = cost.mean()
return cost
def forward(self, predicts, labels):
images, tcl_maps, tcl_label_maps, border_maps \
, direction_maps, training_masks, label_list, pos_list, pos_mask = labels
# for all the batch_size
pos_list, pos_mask, label_list, label_t = pre_process(
label_list, pos_list, pos_mask, self.max_text_length,
self.max_text_nums, self.pad_num, self.tcl_bs)
f_score, f_border, f_direction, f_char = predicts['f_score'], predicts['f_border'], predicts['f_direction'], \
predicts['f_char']
score_loss = self.dice_loss(f_score, tcl_maps, training_masks)
border_loss = self.border_loss(f_border, border_maps, tcl_maps,
training_masks)
direction_loss = self.direction_loss(f_direction, direction_maps,
tcl_maps, training_masks)
ctc_loss = self.ctcloss(f_char, pos_list, pos_mask, label_list, label_t)
loss_all = score_loss + border_loss + direction_loss + 5 * ctc_loss
losses = {
'loss': loss_all,
"score_loss": score_loss,
"border_loss": border_loss,
"direction_loss": direction_loss,
"ctc_loss": ctc_loss
}
return losses
ppocr/losses/rec_ctc_loss.py
View file @ 19eb7eb8
......@@ -25,7 +25,7 @@ class CTCLoss(nn.Layer):
super(CTCLoss, self).__init__()
self.loss_func = nn.CTCLoss(blank=0, reduction='none')
def __call__(self, predicts, batch):
def forward(self, predicts, batch):
predicts = predicts.transpose((1, 0, 2))
N, B, _ = predicts.shape
preds_lengths = paddle.to_tensor([N] * B, dtype='int64')
......
ppocr/losses/rec_nrtr_loss.py 0 → 100644
View file @ 19eb7eb8
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/losses/table_att_loss.py 0 → 100644
View file @ 19eb7eb8
# 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.
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import paddle
from paddle import nn
from paddle.nn import functional as F
from paddle import fluid
class TableAttentionLoss(nn.Layer):
def __init__(self, structure_weight, loc_weight, use_giou=False, giou_weight=1.0, **kwargs):
super(TableAttentionLoss, self).__init__()
self.loss_func = nn.CrossEntropyLoss(weight=None, reduction='none')
self.structure_weight = structure_weight
self.loc_weight = loc_weight
self.use_giou = use_giou
self.giou_weight = giou_weight
def giou_loss(self, preds, bbox, eps=1e-7, reduction='mean'):
'''
:param preds:[[x1,y1,x2,y2], [x1,y1,x2,y2],,,]
:param bbox:[[x1,y1,x2,y2], [x1,y1,x2,y2],,,]
:return: loss
'''
ix1 = fluid.layers.elementwise_max(preds[:, 0], bbox[:, 0])
iy1 = fluid.layers.elementwise_max(preds[:, 1], bbox[:, 1])
ix2 = fluid.layers.elementwise_min(preds[:, 2], bbox[:, 2])
iy2 = fluid.layers.elementwise_min(preds[:, 3], bbox[:, 3])
iw = fluid.layers.clip(ix2 - ix1 + 1e-3, 0., 1e10)
ih = fluid.layers.clip(iy2 - iy1 + 1e-3, 0., 1e10)
# overlap
inters = iw * ih
# union
uni = (preds[:, 2] - preds[:, 0] + 1e-3) * (preds[:, 3] - preds[:, 1] + 1e-3
) + (bbox[:, 2] - bbox[:, 0] + 1e-3) * (
bbox[:, 3] - bbox[:, 1] + 1e-3) - inters + eps
# ious
ious = inters / uni
ex1 = fluid.layers.elementwise_min(preds[:, 0], bbox[:, 0])
ey1 = fluid.layers.elementwise_min(preds[:, 1], bbox[:, 1])
ex2 = fluid.layers.elementwise_max(preds[:, 2], bbox[:, 2])
ey2 = fluid.layers.elementwise_max(preds[:, 3], bbox[:, 3])
ew = fluid.layers.clip(ex2 - ex1 + 1e-3, 0., 1e10)
eh = fluid.layers.clip(ey2 - ey1 + 1e-3, 0., 1e10)
# enclose erea
enclose = ew * eh + eps
giou = ious - (enclose - uni) / enclose
loss = 1 - giou
if reduction == 'mean':
loss = paddle.mean(loss)
elif reduction == 'sum':
loss = paddle.sum(loss)
else:
raise NotImplementedError
return loss
def forward(self, predicts, batch):
structure_probs = predicts['structure_probs']
structure_targets = batch[1].astype("int64")
structure_targets = structure_targets[:, 1:]
if len(batch) == 6:
structure_mask = batch[5].astype("int64")
structure_mask = structure_mask[:, 1:]
structure_mask = paddle.reshape(structure_mask, [-1])
structure_probs = paddle.reshape(structure_probs, [-1, structure_probs.shape[-1]])
structure_targets = paddle.reshape(structure_targets, [-1])
structure_loss = self.loss_func(structure_probs, structure_targets)
if len(batch) == 6:
structure_loss = structure_loss * structure_mask
# structure_loss = paddle.sum(structure_loss) * self.structure_weight
structure_loss = paddle.mean(structure_loss) * self.structure_weight
loc_preds = predicts['loc_preds']
loc_targets = batch[2].astype("float32")
loc_targets_mask = batch[4].astype("float32")
loc_targets = loc_targets[:, 1:, :]
loc_targets_mask = loc_targets_mask[:, 1:, :]
loc_loss = F.mse_loss(loc_preds * loc_targets_mask, loc_targets) * self.loc_weight
if self.use_giou:
loc_loss_giou = self.giou_loss(loc_preds * loc_targets_mask, loc_targets) * self.giou_weight
total_loss = structure_loss + loc_loss + loc_loss_giou
return {'loss':total_loss, "structure_loss":structure_loss, "loc_loss":loc_loss, "loc_loss_giou":loc_loss_giou}
else:
total_loss = structure_loss + loc_loss
return {'loss':total_loss, "structure_loss":structure_loss, "loc_loss":loc_loss}
\ No newline at end of file
ppocr/metrics/__init__.py
View file @ 19eb7eb8
......@@ -19,19 +19,23 @@ from __future__ import unicode_literals
import copy
__all__ = ['build_metric']
__all__ = ["build_metric"]
from .det_metric import DetMetric
from .rec_metric import RecMetric
from .cls_metric import ClsMetric
from .e2e_metric import E2EMetric
from .distillation_metric import DistillationMetric
from .table_metric import TableMetric
def build_metric(config):
from .det_metric import DetMetric
from .rec_metric import RecMetric
from .cls_metric import ClsMetric
support_dict = ['DetMetric', 'RecMetric', 'ClsMetric']
support_dict = [
"DetMetric", "RecMetric", "ClsMetric", "E2EMetric", "DistillationMetric", "TableMetric"
]
config = copy.deepcopy(config)
module_name = config.pop('name')
module_name = config.pop("name")
assert module_name in support_dict, Exception(
'metric only support {}'.format(support_dict))
"metric only support {}".format(support_dict))
module_class = eval(module_name)(**config)
return module_class
ppocr/metrics/det_metric.py
View file @ 19eb7eb8
......@@ -55,6 +55,7 @@ class DetMetric(object):
result = self.evaluator.evaluate_image(gt_info_list, det_info_list)
self.results.append(result)
def get_metric(self):
"""
return metrics {
......
ppocr/metrics/distillation_metric.py 0 → 100644
View file @ 19eb7eb8
# 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.
import importlib
import copy
from .rec_metric import RecMetric
from .det_metric import DetMetric
from .e2e_metric import E2EMetric
from .cls_metric import ClsMetric
class DistillationMetric(object):
def __init__(self,
key=None,
base_metric_name=None,
main_indicator=None,
**kwargs):
self.main_indicator = main_indicator
self.key = key
self.main_indicator = main_indicator
self.base_metric_name = base_metric_name
self.kwargs = kwargs
self.metrics = None
def _init_metrcis(self, preds):
self.metrics = dict()
mod = importlib.import_module(__name__)
for key in preds:
self.metrics[key] = getattr(mod, self.base_metric_name)(
main_indicator=self.main_indicator, **self.kwargs)
self.metrics[key].reset()
def __call__(self, preds, batch, **kwargs):
assert isinstance(preds, dict)
if self.metrics is None:
self._init_metrcis(preds)
output = dict()
for key in preds:
self.metrics[key].__call__(preds[key], batch, **kwargs)
def get_metric(self):
"""
return metrics {
'acc': 0,
'norm_edit_dis': 0,
}
"""
output = dict()
for key in self.metrics:
metric = self.metrics[key].get_metric()
# main indicator
if key == self.key:
output.update(metric)
else:
for sub_key in metric:
output["{}_{}".format(key, sub_key)] = metric[sub_key]
return output
def reset(self):
for key in self.metrics:
self.metrics[key].reset()
ppocr/metrics/e2e_metric.py 0 → 100644
View file @ 19eb7eb8
# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
#
# 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 __future__ import absolute_import
from __future__ import division
from __future__ import print_function
__all__ = ['E2EMetric']
from ppocr.utils.e2e_metric.Deteval import get_socre_A, get_socre_B, combine_results
from ppocr.utils.e2e_utils.extract_textpoint_slow import get_dict
class E2EMetric(object):
def __init__(self,
mode,
gt_mat_dir,
character_dict_path,
main_indicator='f_score_e2e',
**kwargs):
self.mode = mode
self.gt_mat_dir = gt_mat_dir
self.label_list = get_dict(character_dict_path)
self.max_index = len(self.label_list)
self.main_indicator = main_indicator
self.reset()
def __call__(self, preds, batch, **kwargs):
if self.mode == 'A':
gt_polyons_batch = batch[2]
temp_gt_strs_batch = batch[3][0]
ignore_tags_batch = batch[4]
gt_strs_batch = []
for temp_list in temp_gt_strs_batch:
t = ""
for index in temp_list:
if index < self.max_index:
t += self.label_list[index]
gt_strs_batch.append(t)
for pred, gt_polyons, gt_strs, ignore_tags in zip(
[preds], gt_polyons_batch, [gt_strs_batch], ignore_tags_batch):
# prepare gt
gt_info_list = [{
'points': gt_polyon,
'text': gt_str,
'ignore': ignore_tag
} for gt_polyon, gt_str, ignore_tag in
zip(gt_polyons, gt_strs, ignore_tags)]
# prepare det
e2e_info_list = [{
'points': det_polyon,
'texts': pred_str
} for det_polyon, pred_str in
zip(pred['points'], pred['texts'])]
result = get_socre_A(gt_info_list, e2e_info_list)
self.results.append(result)
else:
img_id = batch[5][0]
e2e_info_list = [{
'points': det_polyon,
'texts': pred_str
} for det_polyon, pred_str in zip(preds['points'], preds['texts'])]
result = get_socre_B(self.gt_mat_dir, img_id, e2e_info_list)
self.results.append(result)
def get_metric(self):
metircs = combine_results(self.results)
self.reset()
return metircs
def reset(self):
self.results = [] # clear results
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