Skip to content

πŸ“„ Awesome OCR multiple programing languages toolkits based on ONNX Runtime, OpenVINO, MNN, PaddlePaddle, TensorRT and PyTorch.

feat(rapidocr): sup more lang model for torch inference (#374)

Joker121254570e63be5a
diff --git a/python/rapidocr/ch_ppocr_rec/main.py b/python/rapidocr/ch_ppocr_rec/main.py
index 7d7f988..933247b 100644
--- a/python/rapidocr/ch_ppocr_rec/main.py
+++ b/python/rapidocr/ch_ppocr_rec/main.py
@@ -26,21 +26,14 @@ from .utils import CTCLabelDecode, TextRecInput, TextRecOutput
DEFAULT_DICT_PATH = Path(__file__).parent.parent / "models" / "ppocr_keys_v1.txt"
DEFAULT_DICT_URL = "https://www.modelscope.cn/models/RapidAI/RapidOCR/resolve/master/paddle/PP-OCRv4/rec/ch_PP-OCRv4_rec_infer/ppocr_keys_v1.txt"
-
+DEFAULT_MODE_PATH = Path(__file__).parent.parent / "models"
class TextRecognizer:
def __init__(self, config: Dict[str, Any]):
self.session = get_engine(config.engine_name)(config, mode="rec")
self.logger = Logger(logger_name=__name__).get_log()
-
- character = None
- if self.session.have_key():
- character = self.session.get_character_list()
-
- dict_path = config.get("rec_keys_path", None)
- character_dict_path = dict_path if dict_path else DEFAULT_DICT_PATH
- if not Path(character_dict_path).exists():
- download_file(DEFAULT_DICT_URL, character_dict_path, self.logger)
+ # onnx has inner character, other engine get or download character_dict_path
+ character, character_dict_path = self.get_character_dict(config)
self.postprocess_op = CTCLabelDecode(
character=character, character_path=character_dict_path
@@ -49,6 +42,25 @@ class TextRecognizer:
self.rec_batch_num = config["rec_batch_num"]
self.rec_image_shape = config["rec_img_shape"]
+ def get_character_dict(self, config):
+ character = None
+ dict_path = config.get("rec_keys_path", None)
+ if self.session.have_key():
+ character = self.session.get_character_list()
+ return character, dict_path
+
+ # onnx has character, other engine need dict_path
+ if (not dict_path and not character) or (not Path(dict_path).exists()):
+ dict_download_url = self.session.get_dict_key_url(
+ self.session.__class__, config.engine_name, config.task_type, config.lang
+ )
+ dict_download_url = dict_download_url if dict_download_url is not None else DEFAULT_DICT_URL
+ dict_path = DEFAULT_MODE_PATH / Path(dict_download_url).name
+ if not Path(dict_path).exists():
+ download_file(dict_download_url, dict_path, self.logger)
+
+ return character, dict_path
+
def __call__(self, args: TextRecInput) -> TextRecOutput:
img_list = [args.img] if isinstance(args.img, np.ndarray) else args.img
return_word_box = args.return_word_box
diff --git a/python/rapidocr/default_models.yaml b/python/rapidocr/default_models.yaml
index c9c0988..fd315db 100644
--- a/python/rapidocr/default_models.yaml
+++ b/python/rapidocr/default_models.yaml
@@ -127,13 +127,48 @@ torch:
det:
ch_PP-OCRv4_det_infer.pth: https://www.modelscope.cn/models/RapidAI/RapidOCR/resolve/master/torch/ch_PP-OCRv4_det_infer.pth
ch_PP-OCRv4_det_server_infer.pth: https://www.modelscope.cn/models/RapidAI/RapidOCR/resolve/master/torch/ch_PP-OCRv4_det_server_infer.pth
+ en_PP-OCRv3_det_infer.pth: https://www.modelscope.cn/models/RapidAI/RapidOCR/resolve/master/torch/en_PP-OCRv3_det_infer.pth
+ Multilingual_PP-OCRv3_det_infer.pth: https://www.modelscope.cn/models/RapidAI/RapidOCR/resolve/master/torch/Multilingual_PP-OCRv3_det_infer.pth
rec:
+ arabic_PP-OCRv4_rec_infer:
+ model_dir: https://www.modelscope.cn/models/RapidAI/RapidOCR/resolve/master/torch/arabic_PP-OCRv4_rec_infer.pth
+ dict_url: https://www.modelscope.cn/models/RapidAI/RapidOCR/resolve/master/paddle/PP-OCRv4/rec/arabic_PP-OCRv4_rec_infer/arabic_dict.txt
ch_PP-OCRv4_rec_infer.pth:
model_dir: https://www.modelscope.cn/models/RapidAI/RapidOCR/resolve/master/torch/ch_PP-OCRv4_rec_infer.pth
dict_url: https://www.modelscope.cn/models/RapidAI/RapidOCR/resolve/master/paddle/PP-OCRv4/rec/ch_PP-OCRv4_rec_infer/ppocr_keys_v1.txt
ch_PP-OCRv4_rec_server_infer.pth:
model_dir: https://www.modelscope.cn/models/RapidAI/RapidOCR/resolve/master/torch/ch_PP-OCRv4_rec_server_infer.pth
dict_url: https://www.modelscope.cn/models/RapidAI/RapidOCR/resolve/master/paddle/PP-OCRv4/rec/ch_PP-OCRv4_rec_infer/ppocr_keys_v1.txt
+ chinese_cht_PP-OCRv3_rec_infer.pth:
+ model_dir: https://www.modelscope.cn/models/RapidAI/RapidOCR/resolve/master/torch/chinese_cht_PP-OCRv3_rec_infer.pth
+ dict_url: https://www.modelscope.cn/models/RapidAI/RapidOCR/resolve/master/paddle/PP-OCRv4/rec/chinese_cht_PP-OCRv3_rec_infer/chinese_cht_dict.txt
+ cyrillic_PP-OCRv3_rec_infer.pth:
+ model_dir: https://www.modelscope.cn/models/RapidAI/RapidOCR/resolve/master/torch/cyrillic_PP-OCRv3_rec_infer.pth
+ dict_url: https://www.modelscope.cn/models/RapidAI/RapidOCR/resolve/master/paddle/PP-OCRv4/rec/cyrillic_PP-OCRv3_rec_infer/cyrillic_dict.txt
+ devanagari_PP-OCRv4_rec_infer:
+ model_dir: https://www.modelscope.cn/models/RapidAI/RapidOCR/resolve/master/torch/devanagari_PP-OCRv4_rec_infer.pth
+ dict_url: https://www.modelscope.cn/models/RapidAI/RapidOCR/resolve/master/paddle/PP-OCRv4/rec/devanagari_PP-OCRv4_rec_infer/devanagari_dict.txt
+ en_PP-OCRv4_rec_infer.pth:
+ model_dir: https://www.modelscope.cn/models/RapidAI/RapidOCR/resolve/master/torch/en_PP-OCRv4_rec_infer.pth
+ dict_url: https://www.modelscope.cn/models/RapidAI/RapidOCR/resolve/master/paddle/PP-OCRv4/rec/en_PP-OCRv4_rec_infer/en_dict.txt
+ japan_PP-OCRv4_rec_infer:
+ model_dir: https://www.modelscope.cn/models/RapidAI/RapidOCR/resolve/master/torch/japan_PP-OCRv4_rec_infer.pth
+ dict_url: https://www.modelscope.cn/models/RapidAI/RapidOCR/resolve/master/paddle/PP-OCRv4/rec/japan_PP-OCRv4_rec_infer/japan_dict.txt
+ ka_PP-OCRv4_rec_infer:
+ model_dir: https://www.modelscope.cn/models/RapidAI/RapidOCR/resolve/master/torch/ka_PP-OCRv4_rec_infer.pth
+ dict_url: https://www.modelscope.cn/models/RapidAI/RapidOCR/resolve/master/paddle/PP-OCRv4/rec/kannada_PP-OCRv4_rec_infer/ka_dict.txt
+ korean_PP-OCRv4_rec_infer:
+ model_dir: https://www.modelscope.cn/models/RapidAI/RapidOCR/resolve/master/torch/korean_PP-OCRv4_rec_infer.pth
+ dict_url: https://www.modelscope.cn/models/RapidAI/RapidOCR/resolve/master/paddle/PP-OCRv4/rec/korean_PP-OCRv4_rec_infer/korean_dict.txt
+ latin_PP-OCRv3_rec_infer.pth:
+ model_dir: https://www.modelscope.cn/models/RapidAI/RapidOCR/resolve/master/torch/latin_PP-OCRv3_rec_infer.pth
+ dict_url: https://www.modelscope.cn/models/RapidAI/RapidOCR/resolve/master/paddle/PP-OCRv4/rec/latin_PP-OCRv3_rec_infer/latin_dict.txt
+ ta_PP-OCRv4_rec_infer:
+ model_dir: https://www.modelscope.cn/models/RapidAI/RapidOCR/resolve/master/paddle/PP-OCRv4/rec/ta_PP-OCRv4_rec_infer.pth
+ dict_url: https://www.modelscope.cn/models/RapidAI/RapidOCR/resolve/master/paddle/PP-OCRv4/rec/ta_PP-OCRv4_rec_infer/ta_dict.txt
+ te_PP-OCRv4_rec_infer:
+ model_dir: https://www.modelscope.cn/models/RapidAI/RapidOCR/resolve/master/torch/te_PP-OCRv4_rec_infer.pth
+ dict_url: https://www.modelscope.cn/models/RapidAI/RapidOCR/resolve/master/paddle/PP-OCRv4/rec/te_PP-OCRv4_rec_infer/te_dict.txt
cls:
ch_ptocr_mobile_v2.0_cls_infer.pth: https://www.modelscope.cn/models/RapidAI/RapidOCR/resolve/master/torch/ch_ptocr_mobile_v2.0_cls_infer.pth
diff --git a/python/rapidocr/inference_engine/base.py b/python/rapidocr/inference_engine/base.py
index 5378513..e305efd 100644
--- a/python/rapidocr/inference_engine/base.py
+++ b/python/rapidocr/inference_engine/base.py
@@ -132,3 +132,20 @@ class InferSession(abc.ABC):
return model_dict[k]
raise KeyError("Model not found")
+
+ @staticmethod
+ def get_dict_key_url(cls, engine_name: str, task_type: str, lang: str) -> str:
+ lang, model_type = lang.split("_")
+ model_dict = cls.model_info[engine_name]["PP-OCRv4"][task_type]
+
+ # δΌ˜ε…ˆζŸ₯ζ‰Ύ server ζ¨‘εž‹
+ if model_type == "server":
+ for k in model_dict:
+ prefix = k.split("_")[0]
+ if lang == prefix and "server" in k:
+ return model_dict[k]['dict_url']
+
+ for k in model_dict:
+ prefix = k.split("_")[0]
+ if lang == prefix:
+ return model_dict[k]['dict_url']
diff --git a/python/rapidocr/networks/arch_config.yaml b/python/rapidocr/networks/arch_config.yaml
index 0879157..bdb423a 100644
--- a/python/rapidocr/networks/arch_config.yaml
+++ b/python/rapidocr/networks/arch_config.yaml
@@ -11,6 +11,66 @@ ch_ptocr_mobile_v2.0_cls_infer:
name: ClsHead
class_dim: 2
+Multilingual_PP-OCRv3_det_infer:
+ model_type: det
+ algorithm: DB
+ Transform:
+ Backbone:
+ name: MobileNetV3
+ scale: 0.5
+ model_name: large
+ disable_se: True
+ Neck:
+ name: RSEFPN
+ out_channels: 96
+ shortcut: True
+ Head:
+ name: DBHead
+ k: 50
+
+en_PP-OCRv3_det_infer:
+ model_type: det
+ algorithm: DB
+ Transform:
+ Backbone:
+ name: MobileNetV3
+ scale: 0.5
+ model_name: large
+ disable_se: True
+ Neck:
+ name: RSEFPN
+ out_channels: 96
+ shortcut: True
+ Head:
+ name: DBHead
+ k: 50
+
+en_PP-OCRv4_rec_infer:
+ model_type: rec
+ algorithm: SVTR_LCNet
+ Transform:
+ Backbone:
+ name: PPLCNetV3
+ scale: 0.95
+ Head:
+ name: MultiHead
+ out_channels_list:
+ CTCLabelDecode: 97 #'blank' + ...(62) + ' '
+ head_list:
+ - CTCHead:
+ Neck:
+ name: svtr
+ dims: 120
+ depth: 2
+ hidden_dims: 120
+ kernel_size: [ 1, 3 ]
+ use_guide: True
+ Head:
+ fc_decay: 0.00001
+ - NRTRHead:
+ nrtr_dim: 384
+ max_text_length: 25
+
ch_PP-OCRv4_det_infer:
model_type: det
algorithm: DB
@@ -27,7 +87,6 @@ ch_PP-OCRv4_det_infer:
name: DBHead
k: 50
-
ch_PP-OCRv4_det_server_infer:
model_type: det
algorithm: DB
@@ -44,7 +103,6 @@ ch_PP-OCRv4_det_server_infer:
k: 50
mode: "large"
-
ch_PP-OCRv4_rec_infer:
model_type: rec
algorithm: SVTR_LCNet
@@ -71,7 +129,6 @@ ch_PP-OCRv4_rec_infer:
nrtr_dim: 384
max_text_length: 25
-
ch_PP-OCRv4_rec_server_infer:
model_type: rec
algorithm: SVTR_HGNet
@@ -95,4 +152,215 @@ ch_PP-OCRv4_rec_server_infer:
fc_decay: 0.00001
- NRTRHead:
nrtr_dim: 384
- max_text_length: 25
\ No newline at end of file
+ max_text_length: 25
+
+chinese_cht_PP-OCRv3_rec_infer:
+ model_type: rec
+ algorithm: SVTR
+ Transform:
+ Backbone:
+ name: MobileNetV1Enhance
+ scale: 0.5
+ last_conv_stride: [1, 2]
+ last_pool_type: avg
+ Neck:
+ name: SequenceEncoder
+ encoder_type: svtr
+ dims: 64
+ depth: 2
+ hidden_dims: 120
+ use_guide: True
+ Head:
+ name: CTCHead
+ out_channels: 8423
+ fc_decay: 0.00001
+
+latin_PP-OCRv3_rec_infer:
+ model_type: rec
+ algorithm: SVTR
+ Transform:
+ Backbone:
+ name: MobileNetV1Enhance
+ scale: 0.5
+ last_conv_stride: [ 1, 2 ]
+ last_pool_type: avg
+ Neck:
+ name: SequenceEncoder
+ encoder_type: svtr
+ dims: 64
+ depth: 2
+ hidden_dims: 120
+ use_guide: True
+ Head:
+ name: CTCHead
+ out_channels: 187
+ fc_decay: 0.00001
+
+cyrillic_PP-OCRv3_rec_infer:
+ model_type: rec
+ algorithm: SVTR
+ Transform:
+ Backbone:
+ name: MobileNetV1Enhance
+ scale: 0.5
+ last_conv_stride: [ 1, 2 ]
+ last_pool_type: avg
+ Neck:
+ name: SequenceEncoder
+ encoder_type: svtr
+ dims: 64
+ depth: 2
+ hidden_dims: 120
+ use_guide: True
+ Head:
+ name: CTCHead
+ out_channels: 165
+ fc_decay: 0.00001
+
+arabic_PP-OCRv4_rec_infer:
+ model_type: rec
+ algorithm: SVTR
+ Transform:
+ Backbone:
+ name: MobileNetV1Enhance
+ scale: 0.5
+ last_conv_stride: [ 1, 2 ]
+ last_pool_type: avg
+ Neck:
+ name: SequenceEncoder
+ encoder_type: svtr
+ dims: 64
+ depth: 2
+ hidden_dims: 120
+ use_guide: True
+ Head:
+ name: CTCHead
+ out_channels: 164
+ fc_decay: 0.00001
+
+korean_PP-OCRv4_rec_infer:
+ model_type: rec
+ algorithm: SVTR
+ Transform:
+ Backbone:
+ name: MobileNetV1Enhance
+ scale: 0.5
+ last_conv_stride: [ 1, 2 ]
+ last_pool_type: avg
+ Neck:
+ name: SequenceEncoder
+ encoder_type: svtr
+ dims: 64
+ depth: 2
+ hidden_dims: 120
+ use_guide: True
+ Head:
+ name: CTCHead
+ out_channels: 3690
+ fc_decay: 0.00001
+
+japan_PP-OCRv4_rec_infer:
+ model_type: rec
+ algorithm: SVTR
+ Transform:
+ Backbone:
+ name: MobileNetV1Enhance
+ scale: 0.5
+ last_conv_stride: [ 1, 2 ]
+ last_pool_type: avg
+ Neck:
+ name: SequenceEncoder
+ encoder_type: svtr
+ dims: 64
+ depth: 2
+ hidden_dims: 120
+ use_guide: True
+ Head:
+ name: CTCHead
+ out_channels: 4401
+ fc_decay: 0.00001
+
+ta_PP-OCRv4_rec_infer:
+ model_type: rec
+ algorithm: SVTR
+ Transform:
+ Backbone:
+ name: MobileNetV1Enhance
+ scale: 0.5
+ last_conv_stride: [ 1, 2 ]
+ last_pool_type: avg
+ Neck:
+ name: SequenceEncoder
+ encoder_type: svtr
+ dims: 64
+ depth: 2
+ hidden_dims: 120
+ use_guide: True
+ Head:
+ name: CTCHead
+ out_channels: 130
+ fc_decay: 0.00001
+
+te_PP-OCRv4_rec_infer:
+ model_type: rec
+ algorithm: SVTR
+ Transform:
+ Backbone:
+ name: MobileNetV1Enhance
+ scale: 0.5
+ last_conv_stride: [ 1, 2 ]
+ last_pool_type: avg
+ Neck:
+ name: SequenceEncoder
+ encoder_type: svtr
+ dims: 64
+ depth: 2
+ hidden_dims: 120
+ use_guide: True
+ Head:
+ name: CTCHead
+ out_channels: 153
+ fc_decay: 0.00001
+
+ka_PP-OCRv4_rec_infer:
+ model_type: rec
+ algorithm: SVTR
+ Transform:
+ Backbone:
+ name: MobileNetV1Enhance
+ scale: 0.5
+ last_conv_stride: [ 1, 2 ]
+ last_pool_type: avg
+ Neck:
+ name: SequenceEncoder
+ encoder_type: svtr
+ dims: 64
+ depth: 2
+ hidden_dims: 120
+ use_guide: True
+ Head:
+ name: CTCHead
+ out_channels: 155
+ fc_decay: 0.00001
+
+devanagari_PP-OCRv4_rec_infer:
+ model_type: rec
+ algorithm: SVTR
+ Transform:
+ Backbone:
+ name: MobileNetV1Enhance
+ scale: 0.5
+ last_conv_stride: [ 1, 2 ]
+ last_pool_type: avg
+ Neck:
+ name: SequenceEncoder
+ encoder_type: svtr
+ dims: 64
+ depth: 2
+ hidden_dims: 120
+ use_guide: True
+ Head:
+ name: CTCHead
+ out_channels: 169
+ fc_decay: 0.00001
+
diff --git a/python/rapidocr/networks/backbones/__init__.py b/python/rapidocr/networks/backbones/__init__.py
index 8e7c5b3..79f2e51 100644
--- a/python/rapidocr/networks/backbones/__init__.py
+++ b/python/rapidocr/networks/backbones/__init__.py
@@ -34,6 +34,7 @@ def build_backbone(config, model_type):
from .rec_lcnetv3 import PPLCNetV3
from .rec_mobilenet_v3 import MobileNetV3
from .rec_svtrnet import SVTRNet
+ from .rec_mv1_enhance import MobileNetV1Enhance
support_dict = [
"MobileNetV1Enhance",
diff --git a/python/rapidocr/networks/backbones/rec_mv1_enhance.py b/python/rapidocr/networks/backbones/rec_mv1_enhance.py
new file mode 100644
index 0000000..447c48f
--- /dev/null
+++ b/python/rapidocr/networks/backbones/rec_mv1_enhance.py
@@ -0,0 +1,234 @@
+import os, sys
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from ..common import Activation
+
+
+class ConvBNLayer(nn.Module):
+ def __init__(self,
+ num_channels,
+ filter_size,
+ num_filters,
+ stride,
+ padding,
+ channels=None,
+ num_groups=1,
+ act='hard_swish'):
+ super(ConvBNLayer, self).__init__()
+ self.act = act
+ self._conv = nn.Conv2d(
+ in_channels=num_channels,
+ out_channels=num_filters,
+ kernel_size=filter_size,
+ stride=stride,
+ padding=padding,
+ groups=num_groups,
+ bias=False)
+
+ self._batch_norm = nn.BatchNorm2d(
+ num_filters,
+ )
+ if self.act is not None:
+ self._act = Activation(act_type=act, inplace=True)
+
+ def forward(self, inputs):
+ y = self._conv(inputs)
+ y = self._batch_norm(y)
+ if self.act is not None:
+ y = self._act(y)
+ return y
+
+
+class DepthwiseSeparable(nn.Module):
+ def __init__(self,
+ num_channels,
+ num_filters1,
+ num_filters2,
+ num_groups,
+ stride,
+ scale,
+ dw_size=3,
+ padding=1,
+ use_se=False):
+ super(DepthwiseSeparable, self).__init__()
+ self.use_se = use_se
+ self._depthwise_conv = ConvBNLayer(
+ num_channels=num_channels,
+ num_filters=int(num_filters1 * scale),
+ filter_size=dw_size,
+ stride=stride,
+ padding=padding,
+ num_groups=int(num_groups * scale))
+ if use_se:
+ self._se = SEModule(int(num_filters1 * scale))
+ self._pointwise_conv = ConvBNLayer(
+ num_channels=int(num_filters1 * scale),
+ filter_size=1,
+ num_filters=int(num_filters2 * scale),
+ stride=1,
+ padding=0)
+
+ def forward(self, inputs):
+ y = self._depthwise_conv(inputs)
+ if self.use_se:
+ y = self._se(y)
+ y = self._pointwise_conv(y)
+ return y
+
+
+class MobileNetV1Enhance(nn.Module):
+ def __init__(self,
+ in_channels=3,
+ scale=0.5,
+ last_conv_stride=1,
+ last_pool_type='max',
+ **kwargs):
+ super().__init__()
+ self.scale = scale
+ self.block_list = []
+
+ self.conv1 = ConvBNLayer(
+ num_channels=in_channels,
+ filter_size=3,
+ channels=3,
+ num_filters=int(32 * scale),
+ stride=2,
+ padding=1)
+
+ conv2_1 = DepthwiseSeparable(
+ num_channels=int(32 * scale),
+ num_filters1=32,
+ num_filters2=64,
+ num_groups=32,
+ stride=1,
+ scale=scale)
+ self.block_list.append(conv2_1)
+
+ conv2_2 = DepthwiseSeparable(
+ num_channels=int(64 * scale),
+ num_filters1=64,
+ num_filters2=128,
+ num_groups=64,
+ stride=1,
+ scale=scale)
+ self.block_list.append(conv2_2)
+
+ conv3_1 = DepthwiseSeparable(
+ num_channels=int(128 * scale),
+ num_filters1=128,
+ num_filters2=128,
+ num_groups=128,
+ stride=1,
+ scale=scale)
+ self.block_list.append(conv3_1)
+
+ conv3_2 = DepthwiseSeparable(
+ num_channels=int(128 * scale),
+ num_filters1=128,
+ num_filters2=256,
+ num_groups=128,
+ stride=(2, 1),
+ scale=scale)
+ self.block_list.append(conv3_2)
+
+ conv4_1 = DepthwiseSeparable(
+ num_channels=int(256 * scale),
+ num_filters1=256,
+ num_filters2=256,
+ num_groups=256,
+ stride=1,
+ scale=scale)
+ self.block_list.append(conv4_1)
+
+ conv4_2 = DepthwiseSeparable(
+ num_channels=int(256 * scale),
+ num_filters1=256,
+ num_filters2=512,
+ num_groups=256,
+ stride=(2, 1),
+ scale=scale)
+ self.block_list.append(conv4_2)
+
+ for _ in range(5):
+ conv5 = DepthwiseSeparable(
+ num_channels=int(512 * scale),
+ num_filters1=512,
+ num_filters2=512,
+ num_groups=512,
+ stride=1,
+ dw_size=5,
+ padding=2,
+ scale=scale,
+ use_se=False)
+ self.block_list.append(conv5)
+
+ conv5_6 = DepthwiseSeparable(
+ num_channels=int(512 * scale),
+ num_filters1=512,
+ num_filters2=1024,
+ num_groups=512,
+ stride=(2, 1),
+ dw_size=5,
+ padding=2,
+ scale=scale,
+ use_se=True)
+ self.block_list.append(conv5_6)
+
+ conv6 = DepthwiseSeparable(
+ num_channels=int(1024 * scale),
+ num_filters1=1024,
+ num_filters2=1024,
+ num_groups=1024,
+ stride=last_conv_stride,
+ dw_size=5,
+ padding=2,
+ use_se=True,
+ scale=scale)
+ self.block_list.append(conv6)
+
+ self.block_list = nn.Sequential(*self.block_list)
+ if last_pool_type == 'avg':
+ self.pool = nn.AvgPool2d(kernel_size=2, stride=2, padding=0)
+ else:
+ self.pool = nn.MaxPool2d(kernel_size=2, stride=2, padding=0)
+ self.out_channels = int(1024 * scale)
+
+ def forward(self, inputs):
+ y = self.conv1(inputs)
+ y = self.block_list(y)
+ y = self.pool(y)
+ return y
+
+def hardsigmoid(x):
+ return F.relu6(x + 3., inplace=True) / 6.
+
+class SEModule(nn.Module):
+ def __init__(self, channel, reduction=4):
+ super(SEModule, self).__init__()
+ self.avg_pool = nn.AdaptiveAvgPool2d(1)
+ self.conv1 = nn.Conv2d(
+ in_channels=channel,
+ out_channels=channel // reduction,
+ kernel_size=1,
+ stride=1,
+ padding=0,
+ bias=True)
+ self.conv2 = nn.Conv2d(
+ in_channels=channel // reduction,
+ out_channels=channel,
+ kernel_size=1,
+ stride=1,
+ padding=0,
+ bias=True)
+
+ def forward(self, inputs):
+ outputs = self.avg_pool(inputs)
+ outputs = self.conv1(outputs)
+ outputs = F.relu(outputs)
+ outputs = self.conv2(outputs)
+ outputs = hardsigmoid(outputs)
+ x = torch.mul(inputs, outputs)
+
+ return x
diff --git a/python/rapidocr/networks/necks/__init__.py b/python/rapidocr/networks/necks/__init__.py
index 178ba76..bbe85bc 100644
--- a/python/rapidocr/networks/necks/__init__.py
+++ b/python/rapidocr/networks/necks/__init__.py
@@ -17,6 +17,7 @@ __all__ = ["build_neck"]
def build_neck(config):
from .db_fpn import DBFPN, LKPAN, RSEFPN
+ from .rnn import SequenceEncoder
support_dict = ["DBFPN", "SequenceEncoder", "RSEFPN", "LKPAN"]