chore: update files

python/demo.py

@@ -18,7 +18,7 @@ from rapidocr import RapidOCR, VisRes
 # engine = engine = RapidOCR(
 #     params={"Global.with_onnx": True, "EngineConfig.onnxruntime.use_cuda": True}
 # )
-engine = RapidOCR(params={"Global.with_openvino": True, "Global.lang": "ch"})
+engine = RapidOCR(params={"Global.with_torch": True, "Global.lang": "ch"})
 vis = VisRes()
 
 image_path = "tests/test_files/ch_en_num.jpg"

python/rapidocr/inference_engine/torch.py

@@ -2,44 +2,39 @@
 # @Author: SWHL
 # @Contact: liekkaskono@163.com
 from pathlib import Path
-from typing import Dict, Optional, Union
+from typing import Optional
 
 import numpy as np
 import torch
-import yaml
+from omegaconf import OmegaConf
+
+from ..networks.architectures.base_model import BaseModel
+from ..utils.logger import Logger
+from .base import InferSession
 
 root_dir = Path(__file__).resolve().parent.parent
-DEFAULT_CFG_PATH = root_dir / "arch_config.yaml"
+DEFAULT_CFG_PATH = root_dir / "networks" / "arch_config.yaml"
 
 
-def read_yaml(yaml_path: Union[str, Path]) -> Dict[str, Dict]:
-    with open(yaml_path, "rb") as f:
-        data = yaml.load(f, Loader=yaml.Loader)
-    return data
-
-
-from rapidocr_torch.modeling.architectures.base_model import BaseModel
-
-from ..utils.logger import Logger
-
-
-class TorchInferSession:
+class TorchInferSession(InferSession):
     def __init__(self, config, mode: Optional[str] = None) -> None:
         self.logger = Logger(logger_name=__name__).get_log()
 
-        all_arch_config = read_yaml(DEFAULT_CFG_PATH)
+        all_arch_config = OmegaConf.load(DEFAULT_CFG_PATH)
 
         self.mode = mode
         model_path = Path(config["model_path"])
         self._verify_model(model_path)
+
         file_name = model_path.stem
         if file_name not in all_arch_config:
-            raise ValueError(f"architecture {file_name} is not in config.yaml")
+            raise ValueError(f"architecture {file_name} is not in arch_config.yaml")
 
         arch_config = all_arch_config[file_name]
         self.predictor = BaseModel(arch_config)
         self.predictor.load_state_dict(torch.load(model_path, weights_only=True))
         self.predictor.eval()
+
         self.use_gpu = False
         if config["use_cuda"]:
             self.predictor.cuda()
@@ -50,17 +45,13 @@ class TorchInferSession:
             inp = torch.from_numpy(img)
             if self.use_gpu:
                 inp = inp.cuda()
+
             # 适配跟onnx对齐取值逻辑
             outputs = self.predictor(inp).unsqueeze(0)
             return outputs.cpu().numpy()
 
-    @staticmethod
-    def _verify_model(model_path):
-        model_path = Path(model_path)
-        if not model_path.exists():
-            raise FileNotFoundError(f"{model_path} does not exists.")
-        if not model_path.is_file():
-            raise FileExistsError(f"{model_path} is not a file.")
+    def have_key(self, key: str = "character") -> bool:
+        return False
 
 
 class TorchInferError(Exception):

python/rapidocr/networks/arch_config.yaml

@@ -0,0 +1,98 @@
+ch_ptocr_mobile_v2.0_cls_infer:
+  model_type: cls
+  algorithm: CLS
+  Transform:
+  Backbone:
+    name: MobileNetV3
+    scale: 0.35
+    model_name: small
+  Neck:
+  Head:
+    name: ClsHead
+    class_dim: 2
+
+ch_PP-OCRv4_det_infer:
+  model_type: det
+  algorithm: DB
+  Transform: null
+  Backbone:
+    name: PPLCNetV3
+    scale: 0.75
+    det: True
+  Neck:
+    name: RSEFPN
+    out_channels: 96
+    shortcut: True
+  Head:
+    name: DBHead
+    k: 50
+
+
+ch_PP-OCRv4_det_server_infer:
+  model_type: det
+  algorithm: DB
+  Transform: null
+  Backbone:
+    name: PPHGNet_small
+    det: True
+  Neck:
+    name: LKPAN
+    out_channels: 256
+    intracl: true
+  Head:
+    name: PFHeadLocal
+    k: 50
+    mode: "large"
+
+
+ch_PP-OCRv4_rec_infer:
+  model_type: rec
+  algorithm: SVTR_LCNet
+  Transform:
+  Backbone:
+    name: PPLCNetV3
+    scale: 0.95
+  Head:
+    name: MultiHead
+    out_channels_list:
+      CTCLabelDecode: 6625 #'blank' + ...(6623) + ' '
+    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_rec_server_infer:
+  model_type: rec
+  algorithm: SVTR_HGNet
+  Transform:
+  Backbone:
+    name: PPHGNet_small
+  Head:
+    name: MultiHead
+    out_channels_list:
+      CTCLabelDecode: 6625 #'blank' + ...(6623) + ' '
+    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

python/rapidocr/networks/architectures/__init__.py

@@ -0,0 +1,25 @@
+# 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 copy
+
+__all__ = ["build_model"]
+
+
+def build_model(config, **kwargs):
+    from .base_model import BaseModel
+
+    config = copy.deepcopy(config)
+    module_class = BaseModel(config, **kwargs)
+    return module_class

python/rapidocr/networks/architectures/base_model.py

@@ -0,0 +1,105 @@
+from torch import nn
+
+from ..backbones import build_backbone
+from ..heads import build_head
+from ..necks import build_neck
+
+
+class BaseModel(nn.Module):
+    def __init__(self, config, **kwargs):
+        """
+        the module for OCR.
+        args:
+            config (dict): the super parameters for module.
+        """
+        super(BaseModel, self).__init__()
+
+        in_channels = config.get("in_channels", 3)
+        model_type = config["model_type"]
+        # build backbone, backbone is need for del, rec and cls
+        if "Backbone" not in config or config["Backbone"] is None:
+            self.use_backbone = False
+        else:
+            self.use_backbone = True
+            config["Backbone"]["in_channels"] = in_channels
+            self.backbone = build_backbone(config["Backbone"], model_type)
+            in_channels = self.backbone.out_channels
+
+        # build neck
+        # for rec, neck can be cnn,rnn or reshape(None)
+        # for det, neck can be FPN, BIFPN and so on.
+        # for cls, neck should be none
+        if "Neck" not in config or config["Neck"] is None:
+            self.use_neck = False
+        else:
+            self.use_neck = True
+            config["Neck"]["in_channels"] = in_channels
+            self.neck = build_neck(config["Neck"])
+            in_channels = self.neck.out_channels
+
+        # # build head, head is need for det, rec and cls
+        if "Head" not in config or config["Head"] is None:
+            self.use_head = False
+        else:
+            self.use_head = True
+            config["Head"]["in_channels"] = in_channels
+            self.head = build_head(config["Head"], **kwargs)
+
+        self.return_all_feats = config.get("return_all_feats", False)
+
+        self._initialize_weights()
+
+    def _initialize_weights(self):
+        # weight initialization
+        for m in self.modules():
+            if isinstance(m, nn.Conv2d):
+                nn.init.kaiming_normal_(m.weight, mode="fan_out")
+                if m.bias is not None:
+                    nn.init.zeros_(m.bias)
+            elif isinstance(m, nn.BatchNorm2d):
+                nn.init.ones_(m.weight)
+                nn.init.zeros_(m.bias)
+            elif isinstance(m, nn.Linear):
+                nn.init.normal_(m.weight, 0, 0.01)
+                if m.bias is not None:
+                    nn.init.zeros_(m.bias)
+            elif isinstance(m, nn.ConvTranspose2d):
+                nn.init.kaiming_normal_(m.weight, mode="fan_out")
+                if m.bias is not None:
+                    nn.init.zeros_(m.bias)
+
+    def forward(self, x):
+        y = dict()
+        if self.use_backbone:
+            x = self.backbone(x)
+        if isinstance(x, dict):
+            y.update(x)
+        else:
+            y["backbone_out"] = x
+        final_name = "backbone_out"
+        if self.use_neck:
+            x = self.neck(x)
+            if isinstance(x, dict):
+                y.update(x)
+            else:
+                y["neck_out"] = x
+            final_name = "neck_out"
+        if self.use_head:
+            x = self.head(x)
+        # for multi head, save ctc neck out for udml
+        if isinstance(x, dict) and "ctc_nect" in x.keys():
+            y["neck_out"] = x["ctc_neck"]
+            y["head_out"] = x
+        elif isinstance(x, dict):
+            y.update(x)
+        else:
+            y["head_out"] = x
+        if self.return_all_feats:
+            if self.training:
+                return y
+            elif isinstance(x, dict):
+                return x
+            else:
+                return {final_name: x}
+        else:
+            return x

python/rapidocr/networks/backbones/__init__.py

@@ -0,0 +1,61 @@
+# 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.
+
+__all__ = ["build_backbone"]
+
+
+def build_backbone(config, model_type):
+    if model_type == "det":
+        from .det_mobilenet_v3 import MobileNetV3
+        from .rec_hgnet import PPHGNet_small
+        from .rec_lcnetv3 import PPLCNetV3
+
+        support_dict = [
+            "MobileNetV3",
+            "ResNet",
+            "ResNet_vd",
+            "ResNet_SAST",
+            "PPLCNetV3",
+            "PPHGNet_small",
+        ]
+    elif model_type == "rec" or model_type == "cls":
+        from .rec_hgnet import PPHGNet_small
+        from .rec_lcnetv3 import PPLCNetV3
+        from .rec_mobilenet_v3 import MobileNetV3
+        from .rec_svtrnet import SVTRNet
+
+        support_dict = [
+            "MobileNetV1Enhance",
+            "MobileNetV3",
+            "ResNet",
+            "ResNetFPN",
+            "MTB",
+            "ResNet31",
+            "SVTRNet",
+            "ViTSTR",
+            "DenseNet",
+            "PPLCNetV3",
+            "PPHGNet_small",
+        ]
+    else:
+        raise NotImplementedError
+
+    module_name = config.pop("name")
+    assert module_name in support_dict, Exception(
+        "when model typs is {}, backbone only support {}".format(
+            model_type, support_dict
+        )
+    )
+    module_class = eval(module_name)(**config)
+    return module_class

python/rapidocr/networks/backbones/det_mobilenet_v3.py

@@ -0,0 +1,269 @@
+from torch import nn
+
+from ..common import Activation
+
+
+def make_divisible(v, divisor=8, min_value=None):
+    if min_value is None:
+        min_value = divisor
+    new_v = max(min_value, int(v + divisor / 2) // divisor * divisor)
+    if new_v < 0.9 * v:
+        new_v += divisor
+    return new_v
+
+
+class ConvBNLayer(nn.Module):
+    def __init__(
+        self,
+        in_channels,
+        out_channels,
+        kernel_size,
+        stride,
+        padding,
+        groups=1,
+        if_act=True,
+        act=None,
+        name=None,
+    ):
+        super(ConvBNLayer, self).__init__()
+        self.if_act = if_act
+        self.conv = nn.Conv2d(
+            in_channels=in_channels,
+            out_channels=out_channels,
+            kernel_size=kernel_size,
+            stride=stride,
+            padding=padding,
+            groups=groups,
+            bias=False,
+        )
+
+        self.bn = nn.BatchNorm2d(
+            out_channels,
+        )
+        if self.if_act:
+            self.act = Activation(act_type=act, inplace=True)
+
+    def forward(self, x):
+        x = self.conv(x)
+        x = self.bn(x)
+        if self.if_act:
+            x = self.act(x)
+        return x
+
+
+class SEModule(nn.Module):
+    def __init__(self, in_channels, reduction=4, name=""):
+        super(SEModule, self).__init__()
+        self.avg_pool = nn.AdaptiveAvgPool2d(1)
+        self.conv1 = nn.Conv2d(
+            in_channels=in_channels,
+            out_channels=in_channels // reduction,
+            kernel_size=1,
+            stride=1,
+            padding=0,
+            bias=True,
+        )
+        self.relu1 = Activation(act_type="relu", inplace=True)
+        self.conv2 = nn.Conv2d(
+            in_channels=in_channels // reduction,
+            out_channels=in_channels,
+            kernel_size=1,
+            stride=1,
+            padding=0,
+            bias=True,
+        )
+        self.hard_sigmoid = Activation(act_type="hard_sigmoid", inplace=True)
+
+    def forward(self, inputs):
+        outputs = self.avg_pool(inputs)
+        outputs = self.conv1(outputs)
+        outputs = self.relu1(outputs)
+        outputs = self.conv2(outputs)
+        outputs = self.hard_sigmoid(outputs)
+        outputs = inputs * outputs
+        return outputs
+
+
+class ResidualUnit(nn.Module):
+    def __init__(
+        self,
+        in_channels,
+        mid_channels,
+        out_channels,
+        kernel_size,
+        stride,
+        use_se,
+        act=None,
+        name="",
+    ):
+        super(ResidualUnit, self).__init__()
+        self.if_shortcut = stride == 1 and in_channels == out_channels
+        self.if_se = use_se
+
+        self.expand_conv = ConvBNLayer(
+            in_channels=in_channels,
+            out_channels=mid_channels,
+            kernel_size=1,
+            stride=1,
+            padding=0,
+            if_act=True,
+            act=act,
+            name=name + "_expand",
+        )
+        self.bottleneck_conv = ConvBNLayer(
+            in_channels=mid_channels,
+            out_channels=mid_channels,
+            kernel_size=kernel_size,
+            stride=stride,
+            padding=int((kernel_size - 1) // 2),
+            groups=mid_channels,
+            if_act=True,
+            act=act,
+            name=name + "_depthwise",
+        )
+        if self.if_se:
+            self.mid_se = SEModule(mid_channels, name=name + "_se")
+        self.linear_conv = ConvBNLayer(
+            in_channels=mid_channels,
+            out_channels=out_channels,
+            kernel_size=1,
+            stride=1,
+            padding=0,
+            if_act=False,
+            act=None,
+            name=name + "_linear",
+        )
+
+    def forward(self, inputs):
+        x = self.expand_conv(inputs)
+        x = self.bottleneck_conv(x)
+        if self.if_se:
+            x = self.mid_se(x)
+        x = self.linear_conv(x)
+        if self.if_shortcut:
+            x = inputs + x
+        return x
+
+
+class MobileNetV3(nn.Module):
+    def __init__(
+        self, in_channels=3, model_name="large", scale=0.5, disable_se=False, **kwargs
+    ):
+        """
+        the MobilenetV3 backbone network for detection module.
+        Args:
+            params(dict): the super parameters for build network
+        """
+        super(MobileNetV3, self).__init__()
+
+        self.disable_se = disable_se
+
+        if model_name == "large":
+            cfg = [
+                # k, exp, c,  se,     nl,  s,
+                [3, 16, 16, False, "relu", 1],
+                [3, 64, 24, False, "relu", 2],
+                [3, 72, 24, False, "relu", 1],
+                [5, 72, 40, True, "relu", 2],
+                [5, 120, 40, True, "relu", 1],
+                [5, 120, 40, True, "relu", 1],
+                [3, 240, 80, False, "hard_swish", 2],
+                [3, 200, 80, False, "hard_swish", 1],
+                [3, 184, 80, False, "hard_swish", 1],
+                [3, 184, 80, False, "hard_swish", 1],
+                [3, 480, 112, True, "hard_swish", 1],
+                [3, 672, 112, True, "hard_swish", 1],
+                [5, 672, 160, True, "hard_swish", 2],
+                [5, 960, 160, True, "hard_swish", 1],
+                [5, 960, 160, True, "hard_swish", 1],
+            ]
+            cls_ch_squeeze = 960
+        elif model_name == "small":
+            cfg = [
+                # k, exp, c,  se,     nl,  s,
+                [3, 16, 16, True, "relu", 2],
+                [3, 72, 24, False, "relu", 2],
+                [3, 88, 24, False, "relu", 1],
+                [5, 96, 40, True, "hard_swish", 2],
+                [5, 240, 40, True, "hard_swish", 1],
+                [5, 240, 40, True, "hard_swish", 1],
+                [5, 120, 48, True, "hard_swish", 1],
+                [5, 144, 48, True, "hard_swish", 1],
+                [5, 288, 96, True, "hard_swish", 2],
+                [5, 576, 96, True, "hard_swish", 1],
+                [5, 576, 96, True, "hard_swish", 1],
+            ]
+            cls_ch_squeeze = 576
+        else:
+            raise NotImplementedError(
+                "mode[" + model_name + "_model] is not implemented!"
+            )
+
+        supported_scale = [0.35, 0.5, 0.75, 1.0, 1.25]
+        assert (
+            scale in supported_scale
+        ), "supported scale are {} but input scale is {}".format(supported_scale, scale)
+        inplanes = 16
+        # conv1
+        self.conv = ConvBNLayer(
+            in_channels=in_channels,
+            out_channels=make_divisible(inplanes * scale),
+            kernel_size=3,
+            stride=2,
+            padding=1,
+            groups=1,
+            if_act=True,
+            act="hard_swish",
+            name="conv1",
+        )
+
+        self.stages = nn.ModuleList()
+        self.out_channels = []
+        block_list = []
+        i = 0
+        inplanes = make_divisible(inplanes * scale)
+        for k, exp, c, se, nl, s in cfg:
+            se = se and not self.disable_se
+            if s == 2 and i > 2:
+                self.out_channels.append(inplanes)
+                self.stages.append(nn.Sequential(*block_list))
+                block_list = []
+            block_list.append(
+                ResidualUnit(
+                    in_channels=inplanes,
+                    mid_channels=make_divisible(scale * exp),
+                    out_channels=make_divisible(scale * c),
+                    kernel_size=k,
+                    stride=s,
+                    use_se=se,
+                    act=nl,
+                    name="conv" + str(i + 2),
+                )
+            )
+            inplanes = make_divisible(scale * c)
+            i += 1
+        block_list.append(
+            ConvBNLayer(
+                in_channels=inplanes,
+                out_channels=make_divisible(scale * cls_ch_squeeze),
+                kernel_size=1,
+                stride=1,
+                padding=0,
+                groups=1,
+                if_act=True,
+                act="hard_swish",
+                name="conv_last",
+            )
+        )
+        self.stages.append(nn.Sequential(*block_list))
+        self.out_channels.append(make_divisible(scale * cls_ch_squeeze))
+        # for i, stage in enumerate(self.stages):
+        #     self.add_sublayer(sublayer=stage, name="stage{}".format(i))
+
+    def forward(self, x):
+        x = self.conv(x)
+        out_list = []
+        for stage in self.stages:
+            x = stage(x)
+            out_list.append(x)
+        return out_list

python/rapidocr/networks/backbones/rec_hgnet.py

@@ -0,0 +1,290 @@
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+
+class ConvBNAct(nn.Module):
+    def __init__(
+        self, in_channels, out_channels, kernel_size, stride, groups=1, use_act=True
+    ):
+        super().__init__()
+        self.use_act = use_act
+        self.conv = nn.Conv2d(
+            in_channels,
+            out_channels,
+            kernel_size,
+            stride,
+            padding=(kernel_size - 1) // 2,
+            groups=groups,
+            bias=False,
+        )
+        self.bn = nn.BatchNorm2d(out_channels)
+        if self.use_act:
+            self.act = nn.ReLU()
+
+    def forward(self, x):
+        x = self.conv(x)
+        x = self.bn(x)
+        if self.use_act:
+            x = self.act(x)
+        return x
+
+
+class ESEModule(nn.Module):
+    def __init__(self, channels):
+        super().__init__()
+        self.avg_pool = nn.AdaptiveAvgPool2d(1)
+        self.conv = nn.Conv2d(
+            in_channels=channels,
+            out_channels=channels,
+            kernel_size=1,
+            stride=1,
+            padding=0,
+        )
+        self.sigmoid = nn.Sigmoid()
+
+    def forward(self, x):
+        identity = x
+        x = self.avg_pool(x)
+        x = self.conv(x)
+        x = self.sigmoid(x)
+        return x * identity
+
+
+class HG_Block(nn.Module):
+    def __init__(
+        self,
+        in_channels,
+        mid_channels,
+        out_channels,
+        layer_num,
+        identity=False,
+    ):
+        super().__init__()
+        self.identity = identity
+
+        self.layers = nn.ModuleList()
+        self.layers.append(
+            ConvBNAct(
+                in_channels=in_channels,
+                out_channels=mid_channels,
+                kernel_size=3,
+                stride=1,
+            )
+        )
+        for _ in range(layer_num - 1):
+            self.layers.append(
+                ConvBNAct(
+                    in_channels=mid_channels,
+                    out_channels=mid_channels,
+                    kernel_size=3,
+                    stride=1,
+                )
+            )
+
+        # feature aggregation
+        total_channels = in_channels + layer_num * mid_channels
+        self.aggregation_conv = ConvBNAct(
+            in_channels=total_channels,
+            out_channels=out_channels,
+            kernel_size=1,
+            stride=1,
+        )
+        self.att = ESEModule(out_channels)
+
+    def forward(self, x):
+        identity = x
+        output = []
+        output.append(x)
+        for layer in self.layers:
+            x = layer(x)
+            output.append(x)
+        x = torch.cat(output, dim=1)
+        x = self.aggregation_conv(x)
+        x = self.att(x)
+        if self.identity:
+            x += identity
+        return x
+
+
+class HG_Stage(nn.Module):
+    def __init__(
+        self,
+        in_channels,
+        mid_channels,
+        out_channels,
+        block_num,
+        layer_num,
+        downsample=True,
+        stride=[2, 1],
+    ):
+        super().__init__()
+        self.downsample = downsample
+        if downsample:
+            self.downsample = ConvBNAct(
+                in_channels=in_channels,
+                out_channels=in_channels,
+                kernel_size=3,
+                stride=stride,
+                groups=in_channels,
+                use_act=False,
+            )
+
+        blocks_list = []
+        blocks_list.append(
+            HG_Block(in_channels, mid_channels, out_channels, layer_num, identity=False)
+        )
+        for _ in range(block_num - 1):
+            blocks_list.append(
+                HG_Block(
+                    out_channels, mid_channels, out_channels, layer_num, identity=True
+                )
+            )
+        self.blocks = nn.Sequential(*blocks_list)
+
+    def forward(self, x):
+        if self.downsample:
+            x = self.downsample(x)
+        x = self.blocks(x)
+        return x
+
+
+class PPHGNet(nn.Module):
+    """
+    PPHGNet
+    Args:
+        stem_channels: list. Stem channel list of PPHGNet.
+        stage_config: dict. The configuration of each stage of PPHGNet. such as the number of channels, stride, etc.
+        layer_num: int. Number of layers of HG_Block.
+        use_last_conv: boolean. Whether to use a 1x1 convolutional layer before the classification layer.
+        class_expand: int=2048. Number of channels for the last 1x1 convolutional layer.
+        dropout_prob: float. Parameters of dropout, 0.0 means dropout is not used.
+        class_num: int=1000. The number of classes.
+    Returns:
+        model: nn.Layer. Specific PPHGNet model depends on args.
+    """
+
+    def __init__(
+        self,
+        stem_channels,
+        stage_config,
+        layer_num,
+        in_channels=3,
+        det=False,
+        out_indices=None,
+    ):
+        super().__init__()
+        self.det = det
+        self.out_indices = out_indices if out_indices is not None else [0, 1, 2, 3]
+
+        # stem
+        stem_channels.insert(0, in_channels)
+        self.stem = nn.Sequential(
+            *[
+                ConvBNAct(
+                    in_channels=stem_channels[i],
+                    out_channels=stem_channels[i + 1],
+                    kernel_size=3,
+                    stride=2 if i == 0 else 1,
+                )
+                for i in range(len(stem_channels) - 1)
+            ]
+        )
+
+        if self.det:
+            self.pool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
+        # stages
+        self.stages = nn.ModuleList()
+        self.out_channels = []
+        for block_id, k in enumerate(stage_config):
+            (
+                in_channels,
+                mid_channels,
+                out_channels,
+                block_num,
+                downsample,
+                stride,
+            ) = stage_config[k]
+            self.stages.append(
+                HG_Stage(
+                    in_channels,
+                    mid_channels,
+                    out_channels,
+                    block_num,
+                    layer_num,
+                    downsample,
+                    stride,
+                )
+            )
+            if block_id in self.out_indices:
+                self.out_channels.append(out_channels)
+
+        if not self.det:
+            self.out_channels = stage_config["stage4"][2]
+
+        self._init_weights()
+
+    def _init_weights(self):
+        for m in self.modules():
+            if isinstance(m, nn.Conv2d):
+                nn.init.kaiming_normal_(m.weight)
+            elif isinstance(m, nn.BatchNorm2d):
+                nn.init.ones_(m.weight)
+                nn.init.zeros_(m.bias)
+            elif isinstance(m, nn.Linear):
+                nn.init.zeros_(m.bias)
+
+    def forward(self, x):
+        x = self.stem(x)
+        if self.det:
+            x = self.pool(x)
+
+        out = []
+        for i, stage in enumerate(self.stages):
+            x = stage(x)
+            if self.det and i in self.out_indices:
+                out.append(x)
+        if self.det:
+            return out
+
+        if self.training:
+            x = F.adaptive_avg_pool2d(x, [1, 40])
+        else:
+            x = F.avg_pool2d(x, [3, 2])
+        return x
+
+
+def PPHGNet_small(pretrained=False, use_ssld=False, det=False, **kwargs):
+    """
+    PPHGNet_small
+    Args:
+        pretrained: bool=False or str. If `True` load pretrained parameters, `False` otherwise.
+                    If str, means the path of the pretrained model.
+        use_ssld: bool=False. Whether using distillation pretrained model when pretrained=True.
+    Returns:
+        model: nn.Layer. Specific `PPHGNet_small` model depends on args.
+    """
+    stage_config_det = {
+        # in_channels, mid_channels, out_channels, blocks, downsample
+        "stage1": [128, 128, 256, 1, False, 2],
+        "stage2": [256, 160, 512, 1, True, 2],
+        "stage3": [512, 192, 768, 2, True, 2],
+        "stage4": [768, 224, 1024, 1, True, 2],
+    }
+
+    stage_config_rec = {
+        # in_channels, mid_channels, out_channels, blocks, downsample
+        "stage1": [128, 128, 256, 1, True, [2, 1]],
+        "stage2": [256, 160, 512, 1, True, [1, 2]],
+        "stage3": [512, 192, 768, 2, True, [2, 1]],
+        "stage4": [768, 224, 1024, 1, True, [2, 1]],
+    }
+
+    model = PPHGNet(
+        stem_channels=[64, 64, 128],
+        stage_config=stage_config_det if det else stage_config_rec,
+        layer_num=6,
+        det=det,
+        **kwargs
+    )
+    return model

python/rapidocr/networks/backbones/rec_lcnetv3.py

@@ -0,0 +1,516 @@
+# 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, division, print_function
+
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+from ..common import Activation
+
+NET_CONFIG_det = {
+    "blocks2":
+    # k, in_c, out_c, s, use_se
+    [[3, 16, 32, 1, False]],
+    "blocks3": [[3, 32, 64, 2, False], [3, 64, 64, 1, False]],
+    "blocks4": [[3, 64, 128, 2, False], [3, 128, 128, 1, False]],
+    "blocks5": [
+        [3, 128, 256, 2, False],
+        [5, 256, 256, 1, False],
+        [5, 256, 256, 1, False],
+        [5, 256, 256, 1, False],
+        [5, 256, 256, 1, False],
+    ],
+    "blocks6": [
+        [5, 256, 512, 2, True],
+        [5, 512, 512, 1, True],
+        [5, 512, 512, 1, False],
+        [5, 512, 512, 1, False],
+    ],
+}
+
+NET_CONFIG_rec = {
+    "blocks2":
+    # k, in_c, out_c, s, use_se
+    [[3, 16, 32, 1, False]],
+    "blocks3": [[3, 32, 64, 1, False], [3, 64, 64, 1, False]],
+    "blocks4": [[3, 64, 128, (2, 1), False], [3, 128, 128, 1, False]],
+    "blocks5": [
+        [3, 128, 256, (1, 2), False],
+        [5, 256, 256, 1, False],
+        [5, 256, 256, 1, False],
+        [5, 256, 256, 1, False],
+        [5, 256, 256, 1, False],
+    ],
+    "blocks6": [
+        [5, 256, 512, (2, 1), True],
+        [5, 512, 512, 1, True],
+        [5, 512, 512, (2, 1), False],
+        [5, 512, 512, 1, False],
+    ],
+}
+
+
+def make_divisible(v, divisor=16, min_value=None):
+    if min_value is None:
+        min_value = divisor
+    new_v = max(min_value, int(v + divisor / 2) // divisor * divisor)
+    if new_v < 0.9 * v:
+        new_v += divisor
+    return new_v
+
+
+class LearnableAffineBlock(nn.Module):
+    def __init__(self, scale_value=1.0, bias_value=0.0, lr_mult=1.0, lab_lr=0.1):
+        super().__init__()
+        self.scale = nn.Parameter(torch.Tensor([scale_value]))
+        self.bias = nn.Parameter(torch.Tensor([bias_value]))
+
+    def forward(self, x):
+        return self.scale * x + self.bias
+
+
+class ConvBNLayer(nn.Module):
+    def __init__(
+        self, in_channels, out_channels, kernel_size, stride, groups=1, lr_mult=1.0
+    ):
+        super().__init__()
+        self.conv = nn.Conv2d(
+            in_channels=in_channels,
+            out_channels=out_channels,
+            kernel_size=kernel_size,
+            stride=stride,
+            padding=(kernel_size - 1) // 2,
+            groups=groups,
+            bias=False,
+        )
+
+        self.bn = nn.BatchNorm2d(
+            out_channels,
+        )
+
+    def forward(self, x):
+        x = self.conv(x)
+        x = self.bn(x)
+        return x
+
+
+class Act(nn.Module):
+    def __init__(self, act="hswish", lr_mult=1.0, lab_lr=0.1):
+        super().__init__()
+        if act == "hswish":
+            self.act = nn.Hardswish(inplace=True)
+        else:
+            assert act == "relu"
+            self.act = Activation(act)
+        self.lab = LearnableAffineBlock(lr_mult=lr_mult, lab_lr=lab_lr)
+
+    def forward(self, x):
+        return self.lab(self.act(x))
+
+
+class LearnableRepLayer(nn.Module):
+    def __init__(
+        self,
+        in_channels,
+        out_channels,
+        kernel_size,
+        stride=1,
+        groups=1,
+        num_conv_branches=1,
+        lr_mult=1.0,
+        lab_lr=0.1,
+    ):
+        super().__init__()
+        self.is_repped = False
+        self.groups = groups
+        self.stride = stride
+        self.kernel_size = kernel_size
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+        self.num_conv_branches = num_conv_branches
+        self.padding = (kernel_size - 1) // 2
+
+        self.identity = (
+            nn.BatchNorm2d(
+                num_features=in_channels,
+            )
+            if out_channels == in_channels and stride == 1
+            else None
+        )
+
+        self.conv_kxk = nn.ModuleList(
+            [
+                ConvBNLayer(
+                    in_channels,
+                    out_channels,
+                    kernel_size,
+                    stride,
+                    groups=groups,
+                    lr_mult=lr_mult,
+                )
+                for _ in range(self.num_conv_branches)
+            ]
+        )
+
+        self.conv_1x1 = (
+            ConvBNLayer(
+                in_channels, out_channels, 1, stride, groups=groups, lr_mult=lr_mult
+            )
+            if kernel_size > 1
+            else None
+        )
+
+        self.lab = LearnableAffineBlock(lr_mult=lr_mult, lab_lr=lab_lr)
+        self.act = Act(lr_mult=lr_mult, lab_lr=lab_lr)
+
+    def forward(self, x):
+        # for export
+        if self.is_repped:
+            out = self.lab(self.reparam_conv(x))
+            if self.stride != 2:
+                out = self.act(out)
+            return out
+
+        out = 0
+        if self.identity is not None:
+            out += self.identity(x)
+
+        if self.conv_1x1 is not None:
+            out += self.conv_1x1(x)
+
+        for conv in self.conv_kxk:
+            out += conv(x)
+
+        out = self.lab(out)
+        if self.stride != 2:
+            out = self.act(out)
+        return out
+
+    def rep(self):
+        if self.is_repped:
+            return
+        kernel, bias = self._get_kernel_bias()
+        self.reparam_conv = nn.Conv2d(
+            in_channels=self.in_channels,
+            out_channels=self.out_channels,
+            kernel_size=self.kernel_size,
+            stride=self.stride,
+            padding=self.padding,
+            groups=self.groups,
+        )
+        self.reparam_conv.weight.data = kernel
+        self.reparam_conv.bias.data = bias
+        self.is_repped = True
+
+    def _pad_kernel_1x1_to_kxk(self, kernel1x1, pad):
+        if not isinstance(kernel1x1, torch.Tensor):
+            return 0
+        else:
+            return nn.functional.pad(kernel1x1, [pad, pad, pad, pad])
+
+    def _get_kernel_bias(self):
+        kernel_conv_1x1, bias_conv_1x1 = self._fuse_bn_tensor(self.conv_1x1)
+        kernel_conv_1x1 = self._pad_kernel_1x1_to_kxk(
+            kernel_conv_1x1, self.kernel_size // 2
+        )
+
+        kernel_identity, bias_identity = self._fuse_bn_tensor(self.identity)
+
+        kernel_conv_kxk = 0
+        bias_conv_kxk = 0
+        for conv in self.conv_kxk:
+            kernel, bias = self._fuse_bn_tensor(conv)
+            kernel_conv_kxk += kernel
+            bias_conv_kxk += bias
+
+        kernel_reparam = kernel_conv_kxk + kernel_conv_1x1 + kernel_identity
+        bias_reparam = bias_conv_kxk + bias_conv_1x1 + bias_identity
+        return kernel_reparam, bias_reparam
+
+    def _fuse_bn_tensor(self, branch):
+        if not branch:
+            return 0, 0
+        elif isinstance(branch, ConvBNLayer):
+            kernel = branch.conv.weight
+            running_mean = branch.bn._mean
+            running_var = branch.bn._variance
+            gamma = branch.bn.weight
+            beta = branch.bn.bias
+            eps = branch.bn._epsilon
+        else:
+            assert isinstance(branch, nn.BatchNorm2d)
+            if not hasattr(self, "id_tensor"):
+                input_dim = self.in_channels // self.groups
+                kernel_value = torch.zeros(
+                    (self.in_channels, input_dim, self.kernel_size, self.kernel_size),
+                    dtype=branch.weight.dtype,
+                )
+                for i in range(self.in_channels):
+                    kernel_value[
+                        i, i % input_dim, self.kernel_size // 2, self.kernel_size // 2
+                    ] = 1
+                self.id_tensor = kernel_value
+            kernel = self.id_tensor
+            running_mean = branch._mean
+            running_var = branch._variance
+            gamma = branch.weight
+            beta = branch.bias
+            eps = branch._epsilon
+        std = (running_var + eps).sqrt()
+        t = (gamma / std).reshape((-1, 1, 1, 1))
+        return kernel * t, beta - running_mean * gamma / std
+
+
+class SELayer(nn.Module):
+    def __init__(self, channel, reduction=4, lr_mult=1.0):
+        super().__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,
+        )
+        self.relu = nn.ReLU()
+        self.conv2 = nn.Conv2d(
+            in_channels=channel // reduction,
+            out_channels=channel,
+            kernel_size=1,
+            stride=1,
+            padding=0,
+        )
+        self.hardsigmoid = nn.Hardsigmoid(inplace=True)
+
+    def forward(self, x):
+        identity = x
+        x = self.avg_pool(x)
+        x = self.conv1(x)
+        x = self.relu(x)
+        x = self.conv2(x)
+        x = self.hardsigmoid(x)
+        x = identity * x
+        return x
+
+
+class LCNetV3Block(nn.Module):
+    def __init__(
+        self,
+        in_channels,
+        out_channels,
+        stride,
+        dw_size,
+        use_se=False,
+        conv_kxk_num=4,
+        lr_mult=1.0,
+        lab_lr=0.1,
+    ):
+        super().__init__()
+        self.use_se = use_se
+        self.dw_conv = LearnableRepLayer(
+            in_channels=in_channels,
+            out_channels=in_channels,
+            kernel_size=dw_size,
+            stride=stride,
+            groups=in_channels,
+            num_conv_branches=conv_kxk_num,
+            lr_mult=lr_mult,
+            lab_lr=lab_lr,
+        )
+        if use_se:
+            self.se = SELayer(in_channels, lr_mult=lr_mult)
+        self.pw_conv = LearnableRepLayer(
+            in_channels=in_channels,
+            out_channels=out_channels,
+            kernel_size=1,
+            stride=1,
+            num_conv_branches=conv_kxk_num,
+            lr_mult=lr_mult,
+            lab_lr=lab_lr,
+        )
+
+    def forward(self, x):
+        x = self.dw_conv(x)
+        if self.use_se:
+            x = self.se(x)
+        x = self.pw_conv(x)
+        return x
+
+
+class PPLCNetV3(nn.Module):
+    def __init__(
+        self,
+        scale=1.0,
+        conv_kxk_num=4,
+        lr_mult_list=[1.0, 1.0, 1.0, 1.0, 1.0, 1.0],
+        lab_lr=0.1,
+        det=False,
+        **kwargs
+    ):
+        super().__init__()
+        self.scale = scale
+        self.lr_mult_list = lr_mult_list
+        self.det = det
+
+        self.net_config = NET_CONFIG_det if self.det else NET_CONFIG_rec
+
+        assert isinstance(
+            self.lr_mult_list, (list, tuple)
+        ), "lr_mult_list should be in (list, tuple) but got {}".format(
+            type(self.lr_mult_list)
+        )
+        assert (
+            len(self.lr_mult_list) == 6
+        ), "lr_mult_list length should be 6 but got {}".format(len(self.lr_mult_list))
+
+        self.conv1 = ConvBNLayer(
+            in_channels=3,
+            out_channels=make_divisible(16 * scale),
+            kernel_size=3,
+            stride=2,
+            lr_mult=self.lr_mult_list[0],
+        )
+
+        self.blocks2 = nn.Sequential(
+            *[
+                LCNetV3Block(
+                    in_channels=make_divisible(in_c * scale),
+                    out_channels=make_divisible(out_c * scale),
+                    dw_size=k,
+                    stride=s,
+                    use_se=se,
+                    conv_kxk_num=conv_kxk_num,
+                    lr_mult=self.lr_mult_list[1],
+                    lab_lr=lab_lr,
+                )
+                for i, (k, in_c, out_c, s, se) in enumerate(self.net_config["blocks2"])
+            ]
+        )
+
+        self.blocks3 = nn.Sequential(
+            *[
+                LCNetV3Block(
+                    in_channels=make_divisible(in_c * scale),
+                    out_channels=make_divisible(out_c * scale),
+                    dw_size=k,
+                    stride=s,
+                    use_se=se,
+                    conv_kxk_num=conv_kxk_num,
+                    lr_mult=self.lr_mult_list[2],
+                    lab_lr=lab_lr,
+                )
+                for i, (k, in_c, out_c, s, se) in enumerate(self.net_config["blocks3"])
+            ]
+        )
+
+        self.blocks4 = nn.Sequential(
+            *[
+                LCNetV3Block(
+                    in_channels=make_divisible(in_c * scale),
+                    out_channels=make_divisible(out_c * scale),
+                    dw_size=k,
+                    stride=s,
+                    use_se=se,
+                    conv_kxk_num=conv_kxk_num,
+                    lr_mult=self.lr_mult_list[3],
+                    lab_lr=lab_lr,
+                )
+                for i, (k, in_c, out_c, s, se) in enumerate(self.net_config["blocks4"])
+            ]
+        )
+
+        self.blocks5 = nn.Sequential(
+            *[
+                LCNetV3Block(
+                    in_channels=make_divisible(in_c * scale),
+                    out_channels=make_divisible(out_c * scale),
+                    dw_size=k,
+                    stride=s,
+                    use_se=se,
+                    conv_kxk_num=conv_kxk_num,
+                    lr_mult=self.lr_mult_list[4],
+                    lab_lr=lab_lr,
+                )
+                for i, (k, in_c, out_c, s, se) in enumerate(self.net_config["blocks5"])
+            ]
+        )
+
+        self.blocks6 = nn.Sequential(
+            *[
+                LCNetV3Block(
+                    in_channels=make_divisible(in_c * scale),
+                    out_channels=make_divisible(out_c * scale),
+                    dw_size=k,
+                    stride=s,
+                    use_se=se,
+                    conv_kxk_num=conv_kxk_num,
+                    lr_mult=self.lr_mult_list[5],
+                    lab_lr=lab_lr,
+                )
+                for i, (k, in_c, out_c, s, se) in enumerate(self.net_config["blocks6"])
+            ]
+        )
+        self.out_channels = make_divisible(512 * scale)
+
+        if self.det:
+            mv_c = [16, 24, 56, 480]
+            self.out_channels = [
+                make_divisible(self.net_config["blocks3"][-1][2] * scale),
+                make_divisible(self.net_config["blocks4"][-1][2] * scale),
+                make_divisible(self.net_config["blocks5"][-1][2] * scale),
+                make_divisible(self.net_config["blocks6"][-1][2] * scale),
+            ]
+
+            self.layer_list = nn.ModuleList(
+                [
+                    nn.Conv2d(self.out_channels[0], int(mv_c[0] * scale), 1, 1, 0),
+                    nn.Conv2d(self.out_channels[1], int(mv_c[1] * scale), 1, 1, 0),
+                    nn.Conv2d(self.out_channels[2], int(mv_c[2] * scale), 1, 1, 0),
+                    nn.Conv2d(self.out_channels[3], int(mv_c[3] * scale), 1, 1, 0),
+                ]
+            )
+            self.out_channels = [
+                int(mv_c[0] * scale),
+                int(mv_c[1] * scale),
+                int(mv_c[2] * scale),
+                int(mv_c[3] * scale),
+            ]
+
+    def forward(self, x):
+        out_list = []
+        x = self.conv1(x)
+        x = self.blocks2(x)
+        x = self.blocks3(x)
+        out_list.append(x)
+        x = self.blocks4(x)
+        out_list.append(x)
+        x = self.blocks5(x)
+        out_list.append(x)
+        x = self.blocks6(x)
+        out_list.append(x)
+
+        if self.det:
+            out_list[0] = self.layer_list[0](out_list[0])
+            out_list[1] = self.layer_list[1](out_list[1])
+            out_list[2] = self.layer_list[2](out_list[2])
+            out_list[3] = self.layer_list[3](out_list[3])
+            return out_list
+
+        if self.training:
+            x = F.adaptive_avg_pool2d(x, [1, 40])
+        else:
+            x = F.avg_pool2d(x, [3, 2])
+        return x

python/rapidocr/networks/backbones/rec_mobilenet_v3.py

@@ -0,0 +1,136 @@
+from torch import nn
+
+from .det_mobilenet_v3 import ConvBNLayer, ResidualUnit, make_divisible
+
+
+class MobileNetV3(nn.Module):
+    def __init__(
+        self,
+        in_channels=3,
+        model_name="small",
+        scale=0.5,
+        large_stride=None,
+        small_stride=None,
+        **kwargs
+    ):
+        super(MobileNetV3, self).__init__()
+        if small_stride is None:
+            small_stride = [2, 2, 2, 2]
+        if large_stride is None:
+            large_stride = [1, 2, 2, 2]
+
+        assert isinstance(
+            large_stride, list
+        ), "large_stride type must " "be list but got {}".format(type(large_stride))
+        assert isinstance(
+            small_stride, list
+        ), "small_stride type must " "be list but got {}".format(type(small_stride))
+        assert (
+            len(large_stride) == 4
+        ), "large_stride length must be " "4 but got {}".format(len(large_stride))
+        assert (
+            len(small_stride) == 4
+        ), "small_stride length must be " "4 but got {}".format(len(small_stride))
+
+        if model_name == "large":
+            cfg = [
+                # k, exp, c,  se,     nl,  s,
+                [3, 16, 16, False, "relu", large_stride[0]],
+                [3, 64, 24, False, "relu", (large_stride[1], 1)],
+                [3, 72, 24, False, "relu", 1],
+                [5, 72, 40, True, "relu", (large_stride[2], 1)],
+                [5, 120, 40, True, "relu", 1],
+                [5, 120, 40, True, "relu", 1],
+                [3, 240, 80, False, "hard_swish", 1],
+                [3, 200, 80, False, "hard_swish", 1],
+                [3, 184, 80, False, "hard_swish", 1],
+                [3, 184, 80, False, "hard_swish", 1],
+                [3, 480, 112, True, "hard_swish", 1],
+                [3, 672, 112, True, "hard_swish", 1],
+                [5, 672, 160, True, "hard_swish", (large_stride[3], 1)],
+                [5, 960, 160, True, "hard_swish", 1],
+                [5, 960, 160, True, "hard_swish", 1],
+            ]
+            cls_ch_squeeze = 960
+        elif model_name == "small":
+            cfg = [
+                # k, exp, c,  se,     nl,  s,
+                [3, 16, 16, True, "relu", (small_stride[0], 1)],
+                [3, 72, 24, False, "relu", (small_stride[1], 1)],
+                [3, 88, 24, False, "relu", 1],
+                [5, 96, 40, True, "hard_swish", (small_stride[2], 1)],
+                [5, 240, 40, True, "hard_swish", 1],
+                [5, 240, 40, True, "hard_swish", 1],
+                [5, 120, 48, True, "hard_swish", 1],
+                [5, 144, 48, True, "hard_swish", 1],
+                [5, 288, 96, True, "hard_swish", (small_stride[3], 1)],
+                [5, 576, 96, True, "hard_swish", 1],
+                [5, 576, 96, True, "hard_swish", 1],
+            ]
+            cls_ch_squeeze = 576
+        else:
+            raise NotImplementedError(
+                "mode[" + model_name + "_model] is not implemented!"
+            )
+
+        supported_scale = [0.35, 0.5, 0.75, 1.0, 1.25]
+        assert (
+            scale in supported_scale
+        ), "supported scales are {} but input scale is {}".format(
+            supported_scale, scale
+        )
+
+        inplanes = 16
+        # conv1
+        self.conv1 = ConvBNLayer(
+            in_channels=in_channels,
+            out_channels=make_divisible(inplanes * scale),
+            kernel_size=3,
+            stride=2,
+            padding=1,
+            groups=1,
+            if_act=True,
+            act="hard_swish",
+            name="conv1",
+        )
+        i = 0
+        block_list = []
+        inplanes = make_divisible(inplanes * scale)
+        for k, exp, c, se, nl, s in cfg:
+            block_list.append(
+                ResidualUnit(
+                    in_channels=inplanes,
+                    mid_channels=make_divisible(scale * exp),
+                    out_channels=make_divisible(scale * c),
+                    kernel_size=k,
+                    stride=s,
+                    use_se=se,
+                    act=nl,
+                    name="conv" + str(i + 2),
+                )
+            )
+            inplanes = make_divisible(scale * c)
+            i += 1
+        self.blocks = nn.Sequential(*block_list)
+
+        self.conv2 = ConvBNLayer(
+            in_channels=inplanes,
+            out_channels=make_divisible(scale * cls_ch_squeeze),
+            kernel_size=1,
+            stride=1,
+            padding=0,
+            groups=1,
+            if_act=True,
+            act="hard_swish",
+            name="conv_last",
+        )
+
+        self.pool = nn.MaxPool2d(kernel_size=2, stride=2, padding=0)
+        self.out_channels = make_divisible(scale * cls_ch_squeeze)
+
+    def forward(self, x):
+        x = self.conv1(x)
+        x = self.blocks(x)
+        x = self.conv2(x)
+        x = self.pool(x)
+        return x

python/rapidocr/networks/backbones/rec_svtrnet.py

@@ -0,0 +1,638 @@
+import numpy as np
+import torch
+from torch import nn
+
+from ..common import Activation
+
+
+def drop_path(x, drop_prob=0.0, training=False):
+    """Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks).
+    the original name is misleading as 'Drop Connect' is a different form of dropout in a separate paper...
+    See discussion: https://github.com/tensorflow/tpu/issues/494#issuecomment-532968956 ...
+    """
+    if drop_prob == 0.0 or not training:
+        return x
+    keep_prob = torch.as_tensor(1 - drop_prob)
+    shape = (x.shape[0],) + (1,) * (x.ndim - 1)
+    random_tensor = keep_prob + torch.rand(shape, dtype=x.dtype)
+    random_tensor = torch.floor(random_tensor)  # binarize
+    output = x.divide(keep_prob) * random_tensor
+    return output
+
+
+class ConvBNLayer(nn.Module):
+    def __init__(
+        self,
+        in_channels,
+        out_channels,
+        kernel_size=3,
+        stride=1,
+        padding=0,
+        bias_attr=False,
+        groups=1,
+        act="gelu",
+    ):
+        super().__init__()
+        self.conv = nn.Conv2d(
+            in_channels=in_channels,
+            out_channels=out_channels,
+            kernel_size=kernel_size,
+            stride=stride,
+            padding=padding,
+            groups=groups,
+            bias=bias_attr,
+        )
+        self.norm = nn.BatchNorm2d(out_channels)
+        self.act = Activation(act_type=act, inplace=True)
+
+    def forward(self, inputs):
+        out = self.conv(inputs)
+        out = self.norm(out)
+        out = self.act(out)
+        return out
+
+
+class DropPath(nn.Module):
+    """Drop paths (Stochastic Depth) per sample  (when applied in main path of residual blocks)."""
+
+    def __init__(self, drop_prob=None):
+        super(DropPath, self).__init__()
+        self.drop_prob = drop_prob
+
+    def forward(self, x):
+        return drop_path(x, self.drop_prob, self.training)
+
+
+class Identity(nn.Module):
+    def __init__(self):
+        super(Identity, self).__init__()
+
+    def forward(self, input):
+        return input
+
+
+class Mlp(nn.Module):
+    def __init__(
+        self,
+        in_features,
+        hidden_features=None,
+        out_features=None,
+        act_layer="gelu",
+        drop=0.0,
+    ):
+        super().__init__()
+        out_features = out_features or in_features
+        hidden_features = hidden_features or in_features
+        self.fc1 = nn.Linear(in_features, hidden_features)
+        self.act = Activation(act_type=act_layer, inplace=True)
+        self.fc2 = nn.Linear(hidden_features, out_features)
+        self.drop = nn.Dropout(drop)
+
+    def forward(self, x):
+        x = self.fc1(x)
+        x = self.act(x)
+        x = self.drop(x)
+        x = self.fc2(x)
+        x = self.drop(x)
+        return x
+
+
+class ConvMixer(nn.Module):
+    def __init__(
+        self,
+        dim,
+        num_heads=8,
+        HW=[8, 25],
+        local_k=[3, 3],
+    ):
+        super().__init__()
+        self.HW = HW
+        self.dim = dim
+        self.local_mixer = nn.Conv2d(
+            dim,
+            dim,
+            local_k,
+            1,
+            [local_k[0] // 2, local_k[1] // 2],
+            groups=num_heads,
+        )
+
+    def forward(self, x):
+        h = self.HW[0]
+        w = self.HW[1]
+        x = x.transpose([0, 2, 1]).reshape([0, self.dim, h, w])
+        x = self.local_mixer(x)
+        x = x.flatten(2).permute(0, 2, 1)
+        return x
+
+
+class Attention(nn.Module):
+    def __init__(
+        self,
+        dim,
+        num_heads=8,
+        mixer="Global",
+        HW=[8, 25],
+        local_k=[7, 11],
+        qkv_bias=False,
+        qk_scale=None,
+        attn_drop=0.0,
+        proj_drop=0.0,
+    ):
+        super().__init__()
+        self.num_heads = num_heads
+        head_dim = dim // num_heads
+        self.scale = qk_scale or head_dim**-0.5
+        self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
+        self.attn_drop = nn.Dropout(attn_drop)
+        self.proj = nn.Linear(dim, dim)
+        self.proj_drop = nn.Dropout(proj_drop)
+        self.HW = HW
+        if HW is not None:
+            H = HW[0]
+            W = HW[1]
+            self.N = H * W
+            self.C = dim
+        if mixer == "Local" and HW is not None:
+            hk = local_k[0]
+            wk = local_k[1]
+            mask = torch.ones(H * W, H + hk - 1, W + wk - 1, dtype=torch.float32)
+            for h in range(0, H):
+                for w in range(0, W):
+                    mask[h * W + w, h : h + hk, w : w + wk] = 0.0
+            mask_paddle = mask[:, hk // 2 : H + hk // 2, wk // 2 : W + wk // 2].flatten(
+                1
+            )
+            mask_inf = torch.full(
+                [H * W, H * W], fill_value=float("-Inf"), dtype=torch.float32
+            )
+            mask = torch.where(mask_paddle < 1, mask_paddle, mask_inf)
+            self.mask = mask.unsqueeze(0).unsqueeze(1)
+            # self.mask = mask[None, None, :]
+        self.mixer = mixer
+
+    def forward(self, x):
+        if self.HW is not None:
+            N = self.N
+            C = self.C
+        else:
+            _, N, C = x.shape
+        qkv = self.qkv(x)
+        qkv = qkv.reshape((-1, N, 3, self.num_heads, C // self.num_heads)).permute(
+            2, 0, 3, 1, 4
+        )
+        q, k, v = qkv[0] * self.scale, qkv[1], qkv[2]
+
+        attn = q.matmul(k.permute(0, 1, 3, 2))
+        if self.mixer == "Local":
+            attn += self.mask
+        attn = nn.functional.softmax(attn, dim=-1)
+        attn = self.attn_drop(attn)
+
+        x = (attn.matmul(v)).permute(0, 2, 1, 3).reshape((-1, N, C))
+        x = self.proj(x)
+        x = self.proj_drop(x)
+        return x
+
+
+class Block(nn.Module):
+    def __init__(
+        self,
+        dim,
+        num_heads,
+        mixer="Global",
+        local_mixer=[7, 11],
+        HW=None,
+        mlp_ratio=4.0,
+        qkv_bias=False,
+        qk_scale=None,
+        drop=0.0,
+        attn_drop=0.0,
+        drop_path=0.0,
+        act_layer="gelu",
+        norm_layer="nn.LayerNorm",
+        epsilon=1e-6,
+        prenorm=True,
+    ):
+        super().__init__()
+        if isinstance(norm_layer, str):
+            self.norm1 = eval(norm_layer)(dim, eps=epsilon)
+        else:
+            self.norm1 = norm_layer(dim)
+        if mixer == "Global" or mixer == "Local":
+            self.mixer = Attention(
+                dim,
+                num_heads=num_heads,
+                mixer=mixer,
+                HW=HW,
+                local_k=local_mixer,
+                qkv_bias=qkv_bias,
+                qk_scale=qk_scale,
+                attn_drop=attn_drop,
+                proj_drop=drop,
+            )
+        elif mixer == "Conv":
+            self.mixer = ConvMixer(dim, num_heads=num_heads, HW=HW, local_k=local_mixer)
+        else:
+            raise TypeError("The mixer must be one of [Global, Local, Conv]")
+
+        self.drop_path = DropPath(drop_path) if drop_path > 0.0 else Identity()
+        if isinstance(norm_layer, str):
+            self.norm2 = eval(norm_layer)(dim, eps=epsilon)
+        else:
+            self.norm2 = norm_layer(dim)
+        mlp_hidden_dim = int(dim * mlp_ratio)
+        self.mlp_ratio = mlp_ratio
+        self.mlp = Mlp(
+            in_features=dim,
+            hidden_features=mlp_hidden_dim,
+            act_layer=act_layer,
+            drop=drop,
+        )
+        self.prenorm = prenorm
+
+    def forward(self, x):
+        if self.prenorm:
+            x = self.norm1(x + self.drop_path(self.mixer(x)))
+            x = self.norm2(x + self.drop_path(self.mlp(x)))
+        else:
+            x = x + self.drop_path(self.mixer(self.norm1(x)))
+            x = x + self.drop_path(self.mlp(self.norm2(x)))
+        return x
+
+
+class PatchEmbed(nn.Module):
+    """Image to Patch Embedding"""
+
+    def __init__(
+        self,
+        img_size=[32, 100],
+        in_channels=3,
+        embed_dim=768,
+        sub_num=2,
+        patch_size=[4, 4],
+        mode="pope",
+    ):
+        super().__init__()
+        num_patches = (img_size[1] // (2**sub_num)) * (img_size[0] // (2**sub_num))
+        self.img_size = img_size
+        self.num_patches = num_patches
+        self.embed_dim = embed_dim
+        self.norm = None
+        if mode == "pope":
+            if sub_num == 2:
+                self.proj = nn.Sequential(
+                    ConvBNLayer(
+                        in_channels=in_channels,
+                        out_channels=embed_dim // 2,
+                        kernel_size=3,
+                        stride=2,
+                        padding=1,
+                        act="gelu",
+                        bias_attr=True,
+                    ),
+                    ConvBNLayer(
+                        in_channels=embed_dim // 2,
+                        out_channels=embed_dim,
+                        kernel_size=3,
+                        stride=2,
+                        padding=1,
+                        act="gelu",
+                        bias_attr=True,
+                    ),
+                )
+            if sub_num == 3:
+                self.proj = nn.Sequential(
+                    ConvBNLayer(
+                        in_channels=in_channels,
+                        out_channels=embed_dim // 4,
+                        kernel_size=3,
+                        stride=2,
+                        padding=1,
+                        act="gelu",
+                        bias_attr=True,
+                    ),
+                    ConvBNLayer(
+                        in_channels=embed_dim // 4,
+                        out_channels=embed_dim // 2,
+                        kernel_size=3,
+                        stride=2,
+                        padding=1,
+                        act="gelu",
+                        bias_attr=True,
+                    ),
+                    ConvBNLayer(
+                        in_channels=embed_dim // 2,
+                        out_channels=embed_dim,
+                        kernel_size=3,
+                        stride=2,
+                        padding=1,
+                        act="gelu",
+                        bias_attr=True,
+                    ),
+                )
+        elif mode == "linear":
+            self.proj = nn.Conv2d(
+                1, embed_dim, kernel_size=patch_size, stride=patch_size
+            )
+            self.num_patches = (
+                img_size[0] // patch_size[0] * img_size[1] // patch_size[1]
+            )
+
+    def forward(self, x):
+        B, C, H, W = x.shape
+        assert (
+            H == self.img_size[0] and W == self.img_size[1]
+        ), "Input image size ({}*{}) doesn't match model ({}*{}).".format(
+            H, W, self.img_size[0], self.img_size[1]
+        )
+        x = self.proj(x).flatten(2).permute(0, 2, 1)
+        return x
+
+
+class SubSample(nn.Module):
+    def __init__(
+        self,
+        in_channels,
+        out_channels,
+        types="Pool",
+        stride=[2, 1],
+        sub_norm="nn.LayerNorm",
+        act=None,
+    ):
+        super().__init__()
+        self.types = types
+        if types == "Pool":
+            self.avgpool = nn.AvgPool2d(
+                kernel_size=[3, 5], stride=stride, padding=[1, 2]
+            )
+            self.maxpool = nn.MaxPool2d(
+                kernel_size=[3, 5], stride=stride, padding=[1, 2]
+            )
+            self.proj = nn.Linear(in_channels, out_channels)
+        else:
+            self.conv = nn.Conv2d(
+                in_channels,
+                out_channels,
+                kernel_size=3,
+                stride=stride,
+                padding=1,
+            )
+        self.norm = eval(sub_norm)(out_channels)
+        if act is not None:
+            self.act = act()
+        else:
+            self.act = None
+
+    def forward(self, x):
+        if self.types == "Pool":
+            x1 = self.avgpool(x)
+            x2 = self.maxpool(x)
+            x = (x1 + x2) * 0.5
+            out = self.proj(x.flatten(2).permute(0, 2, 1))
+        else:
+            x = self.conv(x)
+            out = x.flatten(2).permute(0, 2, 1)
+        out = self.norm(out)
+        if self.act is not None:
+            out = self.act(out)
+
+        return out
+
+
+class SVTRNet(nn.Module):
+    def __init__(
+        self,
+        img_size=[32, 100],
+        in_channels=3,
+        embed_dim=[64, 128, 256],
+        depth=[3, 6, 3],
+        num_heads=[2, 4, 8],
+        mixer=["Local"] * 6 + ["Global"] * 6,  # Local atten, Global atten, Conv
+        local_mixer=[[7, 11], [7, 11], [7, 11]],
+        patch_merging="Conv",  # Conv, Pool, None
+        mlp_ratio=4,
+        qkv_bias=True,
+        qk_scale=None,
+        drop_rate=0.0,
+        last_drop=0.0,
+        attn_drop_rate=0.0,
+        drop_path_rate=0.1,
+        norm_layer="nn.LayerNorm",
+        sub_norm="nn.LayerNorm",
+        epsilon=1e-6,
+        out_channels=192,
+        out_char_num=25,
+        block_unit="Block",
+        act="gelu",
+        last_stage=True,
+        sub_num=2,
+        prenorm=True,
+        use_lenhead=False,
+        **kwargs
+    ):
+        super().__init__()
+        self.img_size = img_size
+        self.embed_dim = embed_dim
+        self.out_channels = out_channels
+        self.prenorm = prenorm
+        patch_merging = (
+            None
+            if patch_merging != "Conv" and patch_merging != "Pool"
+            else patch_merging
+        )
+        self.patch_embed = PatchEmbed(
+            img_size=img_size,
+            in_channels=in_channels,
+            embed_dim=embed_dim[0],
+            sub_num=sub_num,
+        )
+        num_patches = self.patch_embed.num_patches
+        self.HW = [img_size[0] // (2**sub_num), img_size[1] // (2**sub_num)]
+        self.pos_embed = nn.Parameter(torch.zeros(1, num_patches, embed_dim[0]))
+        self.pos_drop = nn.Dropout(p=drop_rate)
+        Block_unit = eval(block_unit)
+
+        dpr = np.linspace(0, drop_path_rate, sum(depth))
+        self.blocks1 = nn.ModuleList(
+            [
+                Block_unit(
+                    dim=embed_dim[0],
+                    num_heads=num_heads[0],
+                    mixer=mixer[0 : depth[0]][i],
+                    HW=self.HW,
+                    local_mixer=local_mixer[0],
+                    mlp_ratio=mlp_ratio,
+                    qkv_bias=qkv_bias,
+                    qk_scale=qk_scale,
+                    drop=drop_rate,
+                    act_layer=act,
+                    attn_drop=attn_drop_rate,
+                    drop_path=dpr[0 : depth[0]][i],
+                    norm_layer=norm_layer,
+                    epsilon=epsilon,
+                    prenorm=prenorm,
+                )
+                for i in range(depth[0])
+            ]
+        )
+        if patch_merging is not None:
+            self.sub_sample1 = SubSample(
+                embed_dim[0],
+                embed_dim[1],
+                sub_norm=sub_norm,
+                stride=[2, 1],
+                types=patch_merging,
+            )
+            HW = [self.HW[0] // 2, self.HW[1]]
+        else:
+            HW = self.HW
+        self.patch_merging = patch_merging
+        self.blocks2 = nn.ModuleList(
+            [
+                Block_unit(
+                    dim=embed_dim[1],
+                    num_heads=num_heads[1],
+                    mixer=mixer[depth[0] : depth[0] + depth[1]][i],
+                    HW=HW,
+                    local_mixer=local_mixer[1],
+                    mlp_ratio=mlp_ratio,
+                    qkv_bias=qkv_bias,
+                    qk_scale=qk_scale,
+                    drop=drop_rate,
+                    act_layer=act,
+                    attn_drop=attn_drop_rate,
+                    drop_path=dpr[depth[0] : depth[0] + depth[1]][i],
+                    norm_layer=norm_layer,
+                    epsilon=epsilon,
+                    prenorm=prenorm,
+                )
+                for i in range(depth[1])
+            ]
+        )
+        if patch_merging is not None:
+            self.sub_sample2 = SubSample(
+                embed_dim[1],
+                embed_dim[2],
+                sub_norm=sub_norm,
+                stride=[2, 1],
+                types=patch_merging,
+            )
+            HW = [self.HW[0] // 4, self.HW[1]]
+        else:
+            HW = self.HW
+        self.blocks3 = nn.ModuleList(
+            [
+                Block_unit(
+                    dim=embed_dim[2],
+                    num_heads=num_heads[2],
+                    mixer=mixer[depth[0] + depth[1] :][i],
+                    HW=HW,
+                    local_mixer=local_mixer[2],
+                    mlp_ratio=mlp_ratio,
+                    qkv_bias=qkv_bias,
+                    qk_scale=qk_scale,
+                    drop=drop_rate,
+                    act_layer=act,
+                    attn_drop=attn_drop_rate,
+                    drop_path=dpr[depth[0] + depth[1] :][i],
+                    norm_layer=norm_layer,
+                    epsilon=epsilon,
+                    prenorm=prenorm,
+                )
+                for i in range(depth[2])
+            ]
+        )
+        self.last_stage = last_stage
+        if last_stage:
+            self.avg_pool = nn.AdaptiveAvgPool2d([1, out_char_num])
+            self.last_conv = nn.Conv2d(
+                in_channels=embed_dim[2],
+                out_channels=self.out_channels,
+                kernel_size=1,
+                stride=1,
+                padding=0,
+                bias=False,
+            )
+            self.hardswish = Activation("hard_swish", inplace=True)  # nn.Hardswish()
+            # self.dropout = nn.Dropout(p=last_drop, mode="downscale_in_infer")
+            self.dropout = nn.Dropout(p=last_drop)
+        if not prenorm:
+            self.norm = eval(norm_layer)(embed_dim[-1], eps=epsilon)
+        self.use_lenhead = use_lenhead
+        if use_lenhead:
+            self.len_conv = nn.Linear(embed_dim[2], self.out_channels)
+            self.hardswish_len = Activation(
+                "hard_swish", inplace=True
+            )  # nn.Hardswish()
+            self.dropout_len = nn.Dropout(p=last_drop)
+
+        torch.nn.init.xavier_normal_(self.pos_embed)
+        self.apply(self._init_weights)
+
+    def _init_weights(self, m):
+        # weight initialization
+        if isinstance(m, nn.Conv2d):
+            nn.init.kaiming_normal_(m.weight, mode="fan_out")
+            if m.bias is not None:
+                nn.init.zeros_(m.bias)
+        elif isinstance(m, nn.BatchNorm2d):
+            nn.init.ones_(m.weight)
+            nn.init.zeros_(m.bias)
+        elif isinstance(m, nn.Linear):
+            nn.init.normal_(m.weight, 0, 0.01)
+            if m.bias is not None:
+                nn.init.zeros_(m.bias)
+        elif isinstance(m, nn.ConvTranspose2d):
+            nn.init.kaiming_normal_(m.weight, mode="fan_out")
+            if m.bias is not None:
+                nn.init.zeros_(m.bias)
+        elif isinstance(m, nn.LayerNorm):
+            nn.init.ones_(m.weight)
+            nn.init.zeros_(m.bias)
+
+    def forward_features(self, x):
+        x = self.patch_embed(x)
+        x = x + self.pos_embed
+        x = self.pos_drop(x)
+        for blk in self.blocks1:
+            x = blk(x)
+        if self.patch_merging is not None:
+            x = self.sub_sample1(
+                x.permute(0, 2, 1).reshape(
+                    [-1, self.embed_dim[0], self.HW[0], self.HW[1]]
+                )
+            )
+        for blk in self.blocks2:
+            x = blk(x)
+        if self.patch_merging is not None:
+            x = self.sub_sample2(
+                x.permute(0, 2, 1).reshape(
+                    [-1, self.embed_dim[1], self.HW[0] // 2, self.HW[1]]
+                )
+            )
+        for blk in self.blocks3:
+            x = blk(x)
+        if not self.prenorm:
+            x = self.norm(x)
+        return x
+
+    def forward(self, x):
+        x = self.forward_features(x)
+        if self.use_lenhead:
+            len_x = self.len_conv(x.mean(1))
+            len_x = self.dropout_len(self.hardswish_len(len_x))
+        if self.last_stage:
+            if self.patch_merging is not None:
+                h = self.HW[0] // 4
+            else:
+                h = self.HW[0]
+            x = self.avg_pool(
+                x.permute(0, 2, 1).reshape([-1, self.embed_dim[2], h, self.HW[1]])
+            )
+            x = self.last_conv(x)
+            x = self.hardswish(x)
+            x = self.dropout(x)
+        if self.use_lenhead:
+            return x, len_x
+        return x

python/rapidocr/networks/common.py

@@ -0,0 +1,76 @@
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+
+class Hswish(nn.Module):
+    def __init__(self, inplace=True):
+        super(Hswish, self).__init__()
+        self.inplace = inplace
+
+    def forward(self, x):
+        return x * F.relu6(x + 3.0, inplace=self.inplace) / 6.0
+
+
+# out = max(0, min(1, slop*x+offset))
+# paddle.fluid.layers.hard_sigmoid(x, slope=0.2, offset=0.5, name=None)
+class Hsigmoid(nn.Module):
+    def __init__(self, inplace=True):
+        super(Hsigmoid, self).__init__()
+        self.inplace = inplace
+
+    def forward(self, x):
+        # torch: F.relu6(x + 3., inplace=self.inplace) / 6.
+        # paddle: F.relu6(1.2 * x + 3., inplace=self.inplace) / 6.
+        return F.relu6(1.2 * x + 3.0, inplace=self.inplace) / 6.0
+
+
+class GELU(nn.Module):
+    def __init__(self, inplace=True):
+        super(GELU, self).__init__()
+        self.inplace = inplace
+
+    def forward(self, x):
+        return torch.nn.functional.gelu(x)
+
+
+class Swish(nn.Module):
+    def __init__(self, inplace=True):
+        super(Swish, self).__init__()
+        self.inplace = inplace
+
+    def forward(self, x):
+        if self.inplace:
+            x.mul_(torch.sigmoid(x))
+            return x
+        else:
+            return x * torch.sigmoid(x)
+
+
+class Activation(nn.Module):
+    def __init__(self, act_type, inplace=True):
+        super(Activation, self).__init__()
+        act_type = act_type.lower()
+        if act_type == "relu":
+            self.act = nn.ReLU(inplace=inplace)
+        elif act_type == "relu6":
+            self.act = nn.ReLU6(inplace=inplace)
+        elif act_type == "sigmoid":
+            raise NotImplementedError
+        elif act_type == "hard_sigmoid":
+            self.act = Hsigmoid(
+                inplace
+            )  # nn.Hardsigmoid(inplace=inplace)#Hsigmoid(inplace)#
+        elif act_type == "hard_swish" or act_type == "hswish":
+            self.act = Hswish(inplace=inplace)
+        elif act_type == "leakyrelu":
+            self.act = nn.LeakyReLU(inplace=inplace)
+        elif act_type == "gelu":
+            self.act = GELU(inplace=inplace)
+        elif act_type == "swish":
+            self.act = Swish(inplace=inplace)
+        else:
+            raise NotImplementedError
+
+    def forward(self, inputs):
+        return self.act(inputs)

python/rapidocr/networks/heads/__init__.py

@@ -0,0 +1,43 @@
+# 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.
+
+__all__ = ["build_head"]
+
+
+def build_head(config, **kwargs):
+    # det head
+    from .det_db_head import DBHead, PFHeadLocal
+
+    # rec head
+    from .rec_ctc_head import CTCHead
+    from .rec_multi_head import MultiHead
+
+    # cls head
+    from .cls_head import ClsHead
+
+    support_dict = [
+        "DBHead",
+        "CTCHead",
+        "ClsHead",
+        "MultiHead",
+        "PFHeadLocal",
+    ]
+
+    module_name = config.pop("name")
+    char_num = config.pop("char_num", 6625)
+    assert module_name in support_dict, Exception(
+        "head only support {}".format(support_dict)
+    )
+    module_class = eval(module_name)(**config, **kwargs)
+    return module_class

python/rapidocr/networks/heads/cls_head.py

@@ -0,0 +1,23 @@
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+
+class ClsHead(nn.Module):
+    """
+    Class orientation
+    Args:
+        params(dict): super parameters for build Class network
+    """
+
+    def __init__(self, in_channels, class_dim, **kwargs):
+        super(ClsHead, self).__init__()
+        self.pool = nn.AdaptiveAvgPool2d(1)
+        self.fc = nn.Linear(in_channels, class_dim, bias=True)
+
+    def forward(self, x):
+        x = self.pool(x)
+        x = torch.reshape(x, shape=[x.shape[0], x.shape[1]])
+        x = self.fc(x)
+        x = F.softmax(x, dim=1)
+        return x

python/rapidocr/networks/heads/det_db_head.py

@@ -0,0 +1,131 @@
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+from ..backbones.det_mobilenet_v3 import ConvBNLayer
+from ..common import Activation
+
+
+class Head(nn.Module):
+    def __init__(self, in_channels, **kwargs):
+        super(Head, self).__init__()
+        self.conv1 = nn.Conv2d(
+            in_channels=in_channels,
+            out_channels=in_channels // 4,
+            kernel_size=3,
+            padding=1,
+            bias=False,
+        )
+        self.conv_bn1 = nn.BatchNorm2d(in_channels // 4)
+        self.relu1 = Activation(act_type="relu")
+
+        self.conv2 = nn.ConvTranspose2d(
+            in_channels=in_channels // 4,
+            out_channels=in_channels // 4,
+            kernel_size=2,
+            stride=2,
+        )
+        self.conv_bn2 = nn.BatchNorm2d(in_channels // 4)
+        self.relu2 = Activation(act_type="relu")
+
+        self.conv3 = nn.ConvTranspose2d(
+            in_channels=in_channels // 4, out_channels=1, kernel_size=2, stride=2
+        )
+
+    def forward(self, x, return_f=False):
+        x = self.conv1(x)
+        x = self.conv_bn1(x)
+        x = self.relu1(x)
+        x = self.conv2(x)
+        x = self.conv_bn2(x)
+        x = self.relu2(x)
+        if return_f is True:
+            f = x
+        x = self.conv3(x)
+        x = torch.sigmoid(x)
+        if return_f is True:
+            return x, f
+        return x
+
+
+class DBHead(nn.Module):
+    """
+    Differentiable Binarization (DB) for text detection:
+        see https://arxiv.org/abs/1911.08947
+    args:
+        params(dict): super parameters for build DB network
+    """
+
+    def __init__(self, in_channels, k=50, **kwargs):
+        super(DBHead, self).__init__()
+        self.k = k
+        binarize_name_list = [
+            "conv2d_56",
+            "batch_norm_47",
+            "conv2d_transpose_0",
+            "batch_norm_48",
+            "conv2d_transpose_1",
+            "binarize",
+        ]
+        thresh_name_list = [
+            "conv2d_57",
+            "batch_norm_49",
+            "conv2d_transpose_2",
+            "batch_norm_50",
+            "conv2d_transpose_3",
+            "thresh",
+        ]
+        self.binarize = Head(in_channels, **kwargs)  # binarize_name_list)
+        self.thresh = Head(in_channels, **kwargs)  # thresh_name_list)
+
+    def step_function(self, x, y):
+        return torch.reciprocal(1 + torch.exp(-self.k * (x - y)))
+
+    def forward(self, x):
+        # shrink_maps = self.binarize(x)
+        return self.binarize(x)
+        # if not self.training:
+        #     return {'maps': shrink_maps}
+
+        # threshold_maps = self.thresh(x)
+        # binary_maps = self.step_function(shrink_maps, threshold_maps)
+        # y = torch.cat([shrink_maps, threshold_maps, binary_maps], dim=1)
+        # return y
+
+
+class LocalModule(nn.Module):
+    def __init__(self, in_c, mid_c, use_distance=True):
+        super(self.__class__, self).__init__()
+        self.last_3 = ConvBNLayer(in_c + 1, mid_c, 3, 1, 1, act="relu")
+        self.last_1 = nn.Conv2d(mid_c, 1, 1, 1, 0)
+
+    def forward(self, x, init_map, distance_map):
+        outf = torch.cat([init_map, x], dim=1)
+        # last Conv
+        out = self.last_1(self.last_3(outf))
+        return out
+
+
+class PFHeadLocal(DBHead):
+    def __init__(self, in_channels, k=50, mode="small", **kwargs):
+        super(PFHeadLocal, self).__init__(in_channels, k, **kwargs)
+        self.mode = mode
+
+        self.up_conv = nn.Upsample(scale_factor=2, mode="nearest")
+        if self.mode == "large":
+            self.cbn_layer = LocalModule(in_channels // 4, in_channels // 4)
+        elif self.mode == "small":
+            self.cbn_layer = LocalModule(in_channels // 4, in_channels // 8)
+
+    def forward(self, x, targets=None):
+        shrink_maps, f = self.binarize(x, return_f=True)
+        base_maps = shrink_maps
+        cbn_maps = self.cbn_layer(self.up_conv(f), shrink_maps, None)
+        cbn_maps = F.sigmoid(cbn_maps)
+        return 0.5 * (base_maps + cbn_maps)
+        # return {'maps': 0.5 * (base_maps + cbn_maps), 'cbn_maps': cbn_maps}
+
+        # threshold_maps = self.thresh(x)
+        # binary_maps = self.step_function(shrink_maps, threshold_maps)
+        # y = torch.cat([cbn_maps, threshold_maps, binary_maps], dim=1)
+        # return y

python/rapidocr/networks/heads/rec_ctc_head.py

@@ -0,0 +1,54 @@
+import torch.nn.functional as F
+from torch import nn
+
+
+class CTCHead(nn.Module):
+    def __init__(
+        self,
+        in_channels,
+        out_channels=6625,
+        fc_decay=0.0004,
+        mid_channels=None,
+        return_feats=False,
+        **kwargs
+    ):
+        super(CTCHead, self).__init__()
+        if mid_channels is None:
+            self.fc = nn.Linear(
+                in_channels,
+                out_channels,
+                bias=True,
+            )
+        else:
+            self.fc1 = nn.Linear(
+                in_channels,
+                mid_channels,
+                bias=True,
+            )
+            self.fc2 = nn.Linear(
+                mid_channels,
+                out_channels,
+                bias=True,
+            )
+
+        self.out_channels = out_channels
+        self.mid_channels = mid_channels
+        self.return_feats = return_feats
+
+    def forward(self, x, labels=None):
+        if self.mid_channels is None:
+            predicts = self.fc(x)
+        else:
+            x = self.fc1(x)
+            predicts = self.fc2(x)
+
+        if self.return_feats:
+            result = (x, predicts)
+        else:
+            result = predicts
+
+        if not self.training:
+            predicts = F.softmax(predicts, dim=2)
+            result = predicts
+
+        return result

python/rapidocr/networks/heads/rec_multi_head.py

@@ -0,0 +1,58 @@
+from torch import nn
+
+from ..necks.rnn import Im2Seq, SequenceEncoder
+
+
+class FCTranspose(nn.Module):
+    def __init__(self, in_channels, out_channels, only_transpose=False):
+        super().__init__()
+        self.only_transpose = only_transpose
+        if not self.only_transpose:
+            self.fc = nn.Linear(in_channels, out_channels, bias=False)
+
+    def forward(self, x):
+        if self.only_transpose:
+            return x.permute([0, 2, 1])
+        else:
+            return self.fc(x.permute([0, 2, 1]))
+
+
+class MultiHead(nn.Module):
+    def __init__(self, in_channels, out_channels_list, **kwargs):
+        super().__init__()
+        self.head_list = kwargs.pop("head_list")
+
+        self.gtc_head = "sar"
+        assert len(self.head_list) >= 2
+        for idx, head_name in enumerate(self.head_list):
+            name = list(head_name)[0]
+            if name == "SARHead":
+                pass
+
+            elif name == "NRTRHead":
+                pass
+            elif name == "CTCHead":
+                # ctc neck
+                self.encoder_reshape = Im2Seq(in_channels)
+                neck_args = self.head_list[idx][name]["Neck"]
+                encoder_type = neck_args.pop("name")
+                self.ctc_encoder = SequenceEncoder(
+                    in_channels=in_channels, encoder_type=encoder_type, **neck_args
+                )
+                # ctc head
+                head_args = self.head_list[idx][name].get("Head", {})
+                if head_args is None:
+                    head_args = {}
+                self.ctc_head = eval(name)(
+                    in_channels=self.ctc_encoder.out_channels,
+                    out_channels=out_channels_list["CTCLabelDecode"],
+                    **head_args
+                )
+            else:
+                raise NotImplementedError(
+                    "{} is not supported in MultiHead yet".format(name)
+                )
+
+    def forward(self, x, data=None):
+        ctc_encoder = self.ctc_encoder(x)
+        return self.ctc_head(ctc_encoder)

python/rapidocr/networks/necks/__init__.py

@@ -0,0 +1,28 @@
+# 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.
+
+__all__ = ["build_neck"]
+
+
+def build_neck(config):
+    from .db_fpn import DBFPN, LKPAN, RSEFPN
+
+    support_dict = ["DBFPN", "SequenceEncoder", "RSEFPN", "LKPAN"]
+
+    module_name = config.pop("name")
+    assert module_name in support_dict, Exception(
+        "neck only support {}".format(support_dict)
+    )
+    module_class = eval(module_name)(**config)
+    return module_class

python/rapidocr/networks/necks/db_fpn.py

@@ -0,0 +1,456 @@
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+from ..backbones.det_mobilenet_v3 import SEModule
+from ..necks.intracl import IntraCLBlock
+
+
+def hard_swish(x, inplace=True):
+    return x * F.relu6(x + 3.0, inplace=inplace) / 6.0
+
+
+class DSConv(nn.Module):
+    def __init__(
+        self,
+        in_channels,
+        out_channels,
+        kernel_size,
+        padding,
+        stride=1,
+        groups=None,
+        if_act=True,
+        act="relu",
+        **kwargs
+    ):
+        super(DSConv, self).__init__()
+        if groups == None:
+            groups = in_channels
+        self.if_act = if_act
+        self.act = act
+        self.conv1 = nn.Conv2d(
+            in_channels=in_channels,
+            out_channels=in_channels,
+            kernel_size=kernel_size,
+            stride=stride,
+            padding=padding,
+            groups=groups,
+            bias=False,
+        )
+
+        self.bn1 = nn.BatchNorm2d(in_channels)
+
+        self.conv2 = nn.Conv2d(
+            in_channels=in_channels,
+            out_channels=int(in_channels * 4),
+            kernel_size=1,
+            stride=1,
+            bias=False,
+        )
+
+        self.bn2 = nn.BatchNorm2d(int(in_channels * 4))
+
+        self.conv3 = nn.Conv2d(
+            in_channels=int(in_channels * 4),
+            out_channels=out_channels,
+            kernel_size=1,
+            stride=1,
+            bias=False,
+        )
+        self._c = [in_channels, out_channels]
+        if in_channels != out_channels:
+            self.conv_end = nn.Conv2d(
+                in_channels=in_channels,
+                out_channels=out_channels,
+                kernel_size=1,
+                stride=1,
+                bias=False,
+            )
+
+    def forward(self, inputs):
+        x = self.conv1(inputs)
+        x = self.bn1(x)
+
+        x = self.conv2(x)
+        x = self.bn2(x)
+        if self.if_act:
+            if self.act == "relu":
+                x = F.relu(x)
+            elif self.act == "hardswish":
+                x = hard_swish(x)
+            else:
+                print(
+                    "The activation function({}) is selected incorrectly.".format(
+                        self.act
+                    )
+                )
+                exit()
+
+        x = self.conv3(x)
+        if self._c[0] != self._c[1]:
+            x = x + self.conv_end(inputs)
+        return x
+
+
+class DBFPN(nn.Module):
+    def __init__(self, in_channels, out_channels, use_asf=False, **kwargs):
+        super(DBFPN, self).__init__()
+        self.out_channels = out_channels
+        self.use_asf = use_asf
+
+        self.in2_conv = nn.Conv2d(
+            in_channels=in_channels[0],
+            out_channels=self.out_channels,
+            kernel_size=1,
+            bias=False,
+        )
+        self.in3_conv = nn.Conv2d(
+            in_channels=in_channels[1],
+            out_channels=self.out_channels,
+            kernel_size=1,
+            bias=False,
+        )
+        self.in4_conv = nn.Conv2d(
+            in_channels=in_channels[2],
+            out_channels=self.out_channels,
+            kernel_size=1,
+            bias=False,
+        )
+        self.in5_conv = nn.Conv2d(
+            in_channels=in_channels[3],
+            out_channels=self.out_channels,
+            kernel_size=1,
+            bias=False,
+        )
+        self.p5_conv = nn.Conv2d(
+            in_channels=self.out_channels,
+            out_channels=self.out_channels // 4,
+            kernel_size=3,
+            padding=1,
+            bias=False,
+        )
+        self.p4_conv = nn.Conv2d(
+            in_channels=self.out_channels,
+            out_channels=self.out_channels // 4,
+            kernel_size=3,
+            padding=1,
+            bias=False,
+        )
+        self.p3_conv = nn.Conv2d(
+            in_channels=self.out_channels,
+            out_channels=self.out_channels // 4,
+            kernel_size=3,
+            padding=1,
+            bias=False,
+        )
+        self.p2_conv = nn.Conv2d(
+            in_channels=self.out_channels,
+            out_channels=self.out_channels // 4,
+            kernel_size=3,
+            padding=1,
+            bias=False,
+        )
+
+        if self.use_asf is True:
+            self.asf = ASFBlock(self.out_channels, self.out_channels // 4)
+
+    def forward(self, x):
+        c2, c3, c4, c5 = x
+
+        in5 = self.in5_conv(c5)
+        in4 = self.in4_conv(c4)
+        in3 = self.in3_conv(c3)
+        in2 = self.in2_conv(c2)
+
+        out4 = in4 + F.interpolate(
+            in5,
+            scale_factor=2,
+            mode="nearest",
+        )  # align_mode=1)  # 1/16
+        out3 = in3 + F.interpolate(
+            out4,
+            scale_factor=2,
+            mode="nearest",
+        )  # align_mode=1)  # 1/8
+        out2 = in2 + F.interpolate(
+            out3,
+            scale_factor=2,
+            mode="nearest",
+        )  # align_mode=1)  # 1/4
+
+        p5 = self.p5_conv(in5)
+        p4 = self.p4_conv(out4)
+        p3 = self.p3_conv(out3)
+        p2 = self.p2_conv(out2)
+        p5 = F.interpolate(
+            p5,
+            scale_factor=8,
+            mode="nearest",
+        )  # align_mode=1)
+        p4 = F.interpolate(
+            p4,
+            scale_factor=4,
+            mode="nearest",
+        )  # align_mode=1)
+        p3 = F.interpolate(
+            p3,
+            scale_factor=2,
+            mode="nearest",
+        )  # align_mode=1)
+
+        fuse = torch.cat([p5, p4, p3, p2], dim=1)
+
+        if self.use_asf is True:
+            fuse = self.asf(fuse, [p5, p4, p3, p2])
+
+        return fuse
+
+
+class RSELayer(nn.Module):
+    def __init__(self, in_channels, out_channels, kernel_size, shortcut=True):
+        super(RSELayer, self).__init__()
+        self.out_channels = out_channels
+        self.in_conv = nn.Conv2d(
+            in_channels=in_channels,
+            out_channels=self.out_channels,
+            kernel_size=kernel_size,
+            padding=int(kernel_size // 2),
+            bias=False,
+        )
+        self.se_block = SEModule(self.out_channels)
+        self.shortcut = shortcut
+
+    def forward(self, ins):
+        x = self.in_conv(ins)
+        if self.shortcut:
+            out = x + self.se_block(x)
+        else:
+            out = self.se_block(x)
+        return out
+
+
+class RSEFPN(nn.Module):
+    def __init__(self, in_channels, out_channels, shortcut=True, **kwargs):
+        super(RSEFPN, self).__init__()
+        self.out_channels = out_channels
+        self.ins_conv = nn.ModuleList()
+        self.inp_conv = nn.ModuleList()
+        self.intracl = False
+        if "intracl" in kwargs.keys() and kwargs["intracl"] is True:
+            self.intracl = kwargs["intracl"]
+            self.incl1 = IntraCLBlock(self.out_channels // 4, reduce_factor=2)
+            self.incl2 = IntraCLBlock(self.out_channels // 4, reduce_factor=2)
+            self.incl3 = IntraCLBlock(self.out_channels // 4, reduce_factor=2)
+            self.incl4 = IntraCLBlock(self.out_channels // 4, reduce_factor=2)
+
+        for i in range(len(in_channels)):
+            self.ins_conv.append(
+                RSELayer(in_channels[i], out_channels, kernel_size=1, shortcut=shortcut)
+            )
+            self.inp_conv.append(
+                RSELayer(
+                    out_channels, out_channels // 4, kernel_size=3, shortcut=shortcut
+                )
+            )
+
+    def forward(self, x):
+        c2, c3, c4, c5 = x
+
+        in5 = self.ins_conv[3](c5)
+        in4 = self.ins_conv[2](c4)
+        in3 = self.ins_conv[1](c3)
+        in2 = self.ins_conv[0](c2)
+
+        out4 = in4 + F.interpolate(in5, scale_factor=2, mode="nearest")  # 1/16
+        out3 = in3 + F.interpolate(out4, scale_factor=2, mode="nearest")  # 1/8
+        out2 = in2 + F.interpolate(out3, scale_factor=2, mode="nearest")  # 1/4
+
+        p5 = self.inp_conv[3](in5)
+        p4 = self.inp_conv[2](out4)
+        p3 = self.inp_conv[1](out3)
+        p2 = self.inp_conv[0](out2)
+
+        if self.intracl is True:
+            p5 = self.incl4(p5)
+            p4 = self.incl3(p4)
+            p3 = self.incl2(p3)
+            p2 = self.incl1(p2)
+
+        p5 = F.interpolate(p5, scale_factor=8, mode="nearest")
+        p4 = F.interpolate(p4, scale_factor=4, mode="nearest")
+        p3 = F.interpolate(p3, scale_factor=2, mode="nearest")
+
+        fuse = torch.cat([p5, p4, p3, p2], dim=1)
+        return fuse
+
+
+class LKPAN(nn.Module):
+    def __init__(self, in_channels, out_channels, mode="large", **kwargs):
+        super(LKPAN, self).__init__()
+        self.out_channels = out_channels
+
+        self.ins_conv = nn.ModuleList()
+        self.inp_conv = nn.ModuleList()
+        # pan head
+        self.pan_head_conv = nn.ModuleList()
+        self.pan_lat_conv = nn.ModuleList()
+
+        if mode.lower() == "lite":
+            p_layer = DSConv
+        elif mode.lower() == "large":
+            p_layer = nn.Conv2d
+        else:
+            raise ValueError(
+                "mode can only be one of ['lite', 'large'], but received {}".format(
+                    mode
+                )
+            )
+
+        for i in range(len(in_channels)):
+            self.ins_conv.append(
+                nn.Conv2d(
+                    in_channels=in_channels[i],
+                    out_channels=self.out_channels,
+                    kernel_size=1,
+                    bias=False,
+                )
+            )
+
+            self.inp_conv.append(
+                p_layer(
+                    in_channels=self.out_channels,
+                    out_channels=self.out_channels // 4,
+                    kernel_size=9,
+                    padding=4,
+                    bias=False,
+                )
+            )
+
+            if i > 0:
+                self.pan_head_conv.append(
+                    nn.Conv2d(
+                        in_channels=self.out_channels // 4,
+                        out_channels=self.out_channels // 4,
+                        kernel_size=3,
+                        padding=1,
+                        stride=2,
+                        bias=False,
+                    )
+                )
+            self.pan_lat_conv.append(
+                p_layer(
+                    in_channels=self.out_channels // 4,
+                    out_channels=self.out_channels // 4,
+                    kernel_size=9,
+                    padding=4,
+                    bias=False,
+                )
+            )
+            self.intracl = False
+            if "intracl" in kwargs.keys() and kwargs["intracl"] is True:
+                self.intracl = kwargs["intracl"]
+                self.incl1 = IntraCLBlock(self.out_channels // 4, reduce_factor=2)
+                self.incl2 = IntraCLBlock(self.out_channels // 4, reduce_factor=2)
+                self.incl3 = IntraCLBlock(self.out_channels // 4, reduce_factor=2)
+                self.incl4 = IntraCLBlock(self.out_channels // 4, reduce_factor=2)
+
+    def forward(self, x):
+        c2, c3, c4, c5 = x
+
+        in5 = self.ins_conv[3](c5)
+        in4 = self.ins_conv[2](c4)
+        in3 = self.ins_conv[1](c3)
+        in2 = self.ins_conv[0](c2)
+
+        out4 = in4 + F.interpolate(in5, scale_factor=2, mode="nearest")  # 1/16
+        out3 = in3 + F.interpolate(out4, scale_factor=2, mode="nearest")  # 1/8
+        out2 = in2 + F.interpolate(out3, scale_factor=2, mode="nearest")  # 1/4
+
+        f5 = self.inp_conv[3](in5)
+        f4 = self.inp_conv[2](out4)
+        f3 = self.inp_conv[1](out3)
+        f2 = self.inp_conv[0](out2)
+
+        pan3 = f3 + self.pan_head_conv[0](f2)
+        pan4 = f4 + self.pan_head_conv[1](pan3)
+        pan5 = f5 + self.pan_head_conv[2](pan4)
+
+        p2 = self.pan_lat_conv[0](f2)
+        p3 = self.pan_lat_conv[1](pan3)
+        p4 = self.pan_lat_conv[2](pan4)
+        p5 = self.pan_lat_conv[3](pan5)
+
+        if self.intracl is True:
+            p5 = self.incl4(p5)
+            p4 = self.incl3(p4)
+            p3 = self.incl2(p3)
+            p2 = self.incl1(p2)
+
+        p5 = F.interpolate(p5, scale_factor=8, mode="nearest")
+        p4 = F.interpolate(p4, scale_factor=4, mode="nearest")
+        p3 = F.interpolate(p3, scale_factor=2, mode="nearest")
+
+        fuse = torch.cat([p5, p4, p3, p2], dim=1)
+        return fuse
+
+
+class ASFBlock(nn.Module):
+    """
+    This code is refered from:
+        https://github.com/MhLiao/DB/blob/master/decoders/feature_attention.py
+    """
+
+    def __init__(self, in_channels, inter_channels, out_features_num=4):
+        """
+        Adaptive Scale Fusion (ASF) block of DBNet++
+        Args:
+            in_channels: the number of channels in the input data
+            inter_channels: the number of middle channels
+            out_features_num: the number of fused stages
+        """
+        super(ASFBlock, self).__init__()
+        self.in_channels = in_channels
+        self.inter_channels = inter_channels
+        self.out_features_num = out_features_num
+        self.conv = nn.Conv2d(in_channels, inter_channels, 3, padding=1)
+
+        self.spatial_scale = nn.Sequential(
+            # Nx1xHxW
+            nn.Conv2d(
+                in_channels=1,
+                out_channels=1,
+                kernel_size=3,
+                bias=False,
+                padding=1,
+            ),
+            nn.ReLU(),
+            nn.Conv2d(
+                in_channels=1,
+                out_channels=1,
+                kernel_size=1,
+                bias=False,
+            ),
+            nn.Sigmoid(),
+        )
+
+        self.channel_scale = nn.Sequential(
+            nn.Conv2d(
+                in_channels=inter_channels,
+                out_channels=out_features_num,
+                kernel_size=1,
+                bias=False,
+            ),
+            nn.Sigmoid(),
+        )
+
+    def forward(self, fuse_features, features_list):
+        fuse_features = self.conv(fuse_features)
+        spatial_x = torch.mean(fuse_features, dim=1, keepdim=True)
+        attention_scores = self.spatial_scale(spatial_x) + fuse_features
+        attention_scores = self.channel_scale(attention_scores)
+        assert len(features_list) == self.out_features_num
+
+        out_list = []
+        for i in range(self.out_features_num):
+            out_list.append(attention_scores[:, i : i + 1] * features_list[i])
+        return torch.cat(out_list, dim=1)

python/rapidocr/networks/necks/intracl.py

@@ -0,0 +1,117 @@
+from torch import nn
+
+
+class IntraCLBlock(nn.Module):
+    def __init__(self, in_channels=96, reduce_factor=4):
+        super(IntraCLBlock, self).__init__()
+        self.channels = in_channels
+        self.rf = reduce_factor
+        self.conv1x1_reduce_channel = nn.Conv2d(
+            self.channels, self.channels // self.rf, kernel_size=1, stride=1, padding=0
+        )
+        self.conv1x1_return_channel = nn.Conv2d(
+            self.channels // self.rf, self.channels, kernel_size=1, stride=1, padding=0
+        )
+
+        self.v_layer_7x1 = nn.Conv2d(
+            self.channels // self.rf,
+            self.channels // self.rf,
+            kernel_size=(7, 1),
+            stride=(1, 1),
+            padding=(3, 0),
+        )
+        self.v_layer_5x1 = nn.Conv2d(
+            self.channels // self.rf,
+            self.channels // self.rf,
+            kernel_size=(5, 1),
+            stride=(1, 1),
+            padding=(2, 0),
+        )
+        self.v_layer_3x1 = nn.Conv2d(
+            self.channels // self.rf,
+            self.channels // self.rf,
+            kernel_size=(3, 1),
+            stride=(1, 1),
+            padding=(1, 0),
+        )
+
+        self.q_layer_1x7 = nn.Conv2d(
+            self.channels // self.rf,
+            self.channels // self.rf,
+            kernel_size=(1, 7),
+            stride=(1, 1),
+            padding=(0, 3),
+        )
+        self.q_layer_1x5 = nn.Conv2d(
+            self.channels // self.rf,
+            self.channels // self.rf,
+            kernel_size=(1, 5),
+            stride=(1, 1),
+            padding=(0, 2),
+        )
+        self.q_layer_1x3 = nn.Conv2d(
+            self.channels // self.rf,
+            self.channels // self.rf,
+            kernel_size=(1, 3),
+            stride=(1, 1),
+            padding=(0, 1),
+        )
+
+        # base
+        self.c_layer_7x7 = nn.Conv2d(
+            self.channels // self.rf,
+            self.channels // self.rf,
+            kernel_size=(7, 7),
+            stride=(1, 1),
+            padding=(3, 3),
+        )
+        self.c_layer_5x5 = nn.Conv2d(
+            self.channels // self.rf,
+            self.channels // self.rf,
+            kernel_size=(5, 5),
+            stride=(1, 1),
+            padding=(2, 2),
+        )
+        self.c_layer_3x3 = nn.Conv2d(
+            self.channels // self.rf,
+            self.channels // self.rf,
+            kernel_size=(3, 3),
+            stride=(1, 1),
+            padding=(1, 1),
+        )
+
+        self.bn = nn.BatchNorm2d(self.channels)
+        self.relu = nn.ReLU()
+
+    def forward(self, x):
+        x_new = self.conv1x1_reduce_channel(x)
+
+        x_7_c = self.c_layer_7x7(x_new)
+        x_7_v = self.v_layer_7x1(x_new)
+        x_7_q = self.q_layer_1x7(x_new)
+        x_7 = x_7_c + x_7_v + x_7_q
+
+        x_5_c = self.c_layer_5x5(x_7)
+        x_5_v = self.v_layer_5x1(x_7)
+        x_5_q = self.q_layer_1x5(x_7)
+        x_5 = x_5_c + x_5_v + x_5_q
+
+        x_3_c = self.c_layer_3x3(x_5)
+        x_3_v = self.v_layer_3x1(x_5)
+        x_3_q = self.q_layer_1x3(x_5)
+        x_3 = x_3_c + x_3_v + x_3_q
+
+        x_relation = self.conv1x1_return_channel(x_3)
+
+        x_relation = self.bn(x_relation)
+        x_relation = self.relu(x_relation)
+
+        return x + x_relation
+
+
+def build_intraclblock_list(num_block):
+    IntraCLBlock_list = nn.ModuleList()
+    for i in range(num_block):
+        IntraCLBlock_list.append(IntraCLBlock())
+
+    return IntraCLBlock_list

python/rapidocr/networks/necks/rnn.py

@@ -0,0 +1,228 @@
+import torch
+from torch import nn
+
+from ..backbones.rec_svtrnet import Block, ConvBNLayer
+
+
+class Im2Seq(nn.Module):
+    def __init__(self, in_channels, **kwargs):
+        super().__init__()
+        self.out_channels = in_channels
+
+    def forward(self, x):
+        B, C, H, W = x.shape
+        # assert H == 1
+        x = x.squeeze(dim=2)
+        # x = x.transpose([0, 2, 1])  # paddle (NTC)(batch, width, channels)
+        x = x.permute(0, 2, 1)
+        return x
+
+
+class EncoderWithRNN_(nn.Module):
+    def __init__(self, in_channels, hidden_size):
+        super(EncoderWithRNN_, self).__init__()
+        self.out_channels = hidden_size * 2
+        self.rnn1 = nn.LSTM(
+            in_channels,
+            hidden_size,
+            bidirectional=False,
+            batch_first=True,
+            num_layers=2,
+        )
+        self.rnn2 = nn.LSTM(
+            in_channels,
+            hidden_size,
+            bidirectional=False,
+            batch_first=True,
+            num_layers=2,
+        )
+
+    def forward(self, x):
+        self.rnn1.flatten_parameters()
+        self.rnn2.flatten_parameters()
+        out1, h1 = self.rnn1(x)
+        out2, h2 = self.rnn2(torch.flip(x, [1]))
+        return torch.cat([out1, torch.flip(out2, [1])], 2)
+
+
+class EncoderWithRNN(nn.Module):
+    def __init__(self, in_channels, hidden_size):
+        super(EncoderWithRNN, self).__init__()
+        self.out_channels = hidden_size * 2
+        self.lstm = nn.LSTM(
+            in_channels, hidden_size, num_layers=2, batch_first=True, bidirectional=True
+        )  # batch_first:=True
+
+    def forward(self, x):
+        x, _ = self.lstm(x)
+        return x
+
+
+class EncoderWithFC(nn.Module):
+    def __init__(self, in_channels, hidden_size):
+        super(EncoderWithFC, self).__init__()
+        self.out_channels = hidden_size
+        self.fc = nn.Linear(
+            in_channels,
+            hidden_size,
+            bias=True,
+        )
+
+    def forward(self, x):
+        x = self.fc(x)
+        return x
+
+
+class EncoderWithSVTR(nn.Module):
+    def __init__(
+        self,
+        in_channels,
+        dims=64,  # XS
+        depth=2,
+        hidden_dims=120,
+        use_guide=False,
+        num_heads=8,
+        qkv_bias=True,
+        mlp_ratio=2.0,
+        drop_rate=0.1,
+        kernel_size=[3, 3],
+        attn_drop_rate=0.1,
+        drop_path=0.0,
+        qk_scale=None,
+    ):
+        super(EncoderWithSVTR, self).__init__()
+        self.depth = depth
+        self.use_guide = use_guide
+        self.conv1 = ConvBNLayer(
+            in_channels,
+            in_channels // 8,
+            kernel_size=kernel_size,
+            padding=[kernel_size[0] // 2, kernel_size[1] // 2],
+            act="swish",
+        )
+        self.conv2 = ConvBNLayer(
+            in_channels // 8, hidden_dims, kernel_size=1, act="swish"
+        )
+
+        self.svtr_block = nn.ModuleList(
+            [
+                Block(
+                    dim=hidden_dims,
+                    num_heads=num_heads,
+                    mixer="Global",
+                    HW=None,
+                    mlp_ratio=mlp_ratio,
+                    qkv_bias=qkv_bias,
+                    qk_scale=qk_scale,
+                    drop=drop_rate,
+                    act_layer="swish",
+                    attn_drop=attn_drop_rate,
+                    drop_path=drop_path,
+                    norm_layer="nn.LayerNorm",
+                    epsilon=1e-05,
+                    prenorm=False,
+                )
+                for i in range(depth)
+            ]
+        )
+        self.norm = nn.LayerNorm(hidden_dims, eps=1e-6)
+        self.conv3 = ConvBNLayer(hidden_dims, in_channels, kernel_size=1, act="swish")
+        # last conv-nxn, the input is concat of input tensor and conv3 output tensor
+        self.conv4 = ConvBNLayer(
+            2 * in_channels, in_channels // 8, padding=1, act="swish"
+        )
+
+        self.conv1x1 = ConvBNLayer(in_channels // 8, dims, kernel_size=1, act="swish")
+        self.out_channels = dims
+        self.apply(self._init_weights)
+
+    def _init_weights(self, m):
+        # weight initialization
+        if isinstance(m, nn.Conv2d):
+            nn.init.kaiming_normal_(m.weight, mode="fan_out")
+            if m.bias is not None:
+                nn.init.zeros_(m.bias)
+        elif isinstance(m, nn.BatchNorm2d):
+            nn.init.ones_(m.weight)
+            nn.init.zeros_(m.bias)
+        elif isinstance(m, nn.Linear):
+            nn.init.normal_(m.weight, 0, 0.01)
+            if m.bias is not None:
+                nn.init.zeros_(m.bias)
+        elif isinstance(m, nn.ConvTranspose2d):
+            nn.init.kaiming_normal_(m.weight, mode="fan_out")
+            if m.bias is not None:
+                nn.init.zeros_(m.bias)
+        elif isinstance(m, nn.LayerNorm):
+            nn.init.ones_(m.weight)
+            nn.init.zeros_(m.bias)
+
+    def forward(self, x):
+        # for use guide
+        if self.use_guide:
+            z = x.clone()
+            z.stop_gradient = True
+        else:
+            z = x
+        # for short cut
+        h = z
+        # reduce dim
+        z = self.conv1(z)
+        z = self.conv2(z)
+        # SVTR global block
+        B, C, H, W = z.shape
+        z = z.flatten(2).permute(0, 2, 1)
+
+        for blk in self.svtr_block:
+            z = blk(z)
+
+        z = self.norm(z)
+        # last stage
+        z = z.reshape([-1, H, W, C]).permute(0, 3, 1, 2)
+        z = self.conv3(z)
+        z = torch.cat((h, z), dim=1)
+        z = self.conv1x1(self.conv4(z))
+
+        return z
+
+
+class SequenceEncoder(nn.Module):
+    def __init__(self, in_channels, encoder_type, hidden_size=48, **kwargs):
+        super(SequenceEncoder, self).__init__()
+        self.encoder_reshape = Im2Seq(in_channels)
+        self.out_channels = self.encoder_reshape.out_channels
+        self.encoder_type = encoder_type
+        if encoder_type == "reshape":
+            self.only_reshape = True
+        else:
+            support_encoder_dict = {
+                "reshape": Im2Seq,
+                "fc": EncoderWithFC,
+                "rnn": EncoderWithRNN,
+                "svtr": EncoderWithSVTR,
+            }
+            assert encoder_type in support_encoder_dict, "{} must in {}".format(
+                encoder_type, support_encoder_dict.keys()
+            )
+
+            if encoder_type == "svtr":
+                self.encoder = support_encoder_dict[encoder_type](
+                    self.encoder_reshape.out_channels, **kwargs
+                )
+            else:
+                self.encoder = support_encoder_dict[encoder_type](
+                    self.encoder_reshape.out_channels, hidden_size
+                )
+            self.out_channels = self.encoder.out_channels
+            self.only_reshape = False
+
+    def forward(self, x):
+        if self.encoder_type != "svtr":
+            x = self.encoder_reshape(x)
+            if not self.only_reshape:
+                x = self.encoder(x)
+            return x
+        else:
+            x = self.encoder(x)
+            x = self.encoder_reshape(x)
+            return x

python/rapidocr/utils/parse_parameters.py

@@ -100,6 +100,10 @@ def parse_lang(lang):
         lang = "cyrillic"
     elif lang in devanagari_lang:
         lang = "devanagari"
+    elif lang in ["ch", "en"]:
+        pass
+    else:
+        raise ValueError(f"lang: {lang} is not in the supported list.")
 
     if lang == "ch":
         det_lang = "ch"

python/rapidocr_torch/modeling/common.py

@@ -1,16 +1,16 @@
-
-
 import torch
 import torch.nn as nn
 import torch.nn.functional as F
 
+
 class Hswish(nn.Module):
     def __init__(self, inplace=True):
         super(Hswish, self).__init__()
         self.inplace = inplace
 
     def forward(self, x):
-        return x * F.relu6(x + 3., inplace=self.inplace) / 6.
+        return x * F.relu6(x + 3.0, inplace=self.inplace) / 6.0
+
 
 # out = max(0, min(1, slop*x+offset))
 # paddle.fluid.layers.hard_sigmoid(x, slope=0.2, offset=0.5, name=None)
@@ -22,7 +22,8 @@ class Hsigmoid(nn.Module):
     def forward(self, x):
         # torch: F.relu6(x + 3., inplace=self.inplace) / 6.
         # paddle: F.relu6(1.2 * x + 3., inplace=self.inplace) / 6.
-        return F.relu6(1.2 * x + 3., inplace=self.inplace) / 6.
+        return F.relu6(1.2 * x + 3.0, inplace=self.inplace) / 6.0
+
 
 class GELU(nn.Module):
     def __init__(self, inplace=True):
@@ -43,31 +44,33 @@ class Swish(nn.Module):
             x.mul_(torch.sigmoid(x))
             return x
         else:
-            return x*torch.sigmoid(x)
+            return x * torch.sigmoid(x)
 
 
 class Activation(nn.Module):
     def __init__(self, act_type, inplace=True):
         super(Activation, self).__init__()
         act_type = act_type.lower()
-        if act_type == 'relu':
+        if act_type == "relu":
             self.act = nn.ReLU(inplace=inplace)
-        elif act_type == 'relu6':
+        elif act_type == "relu6":
             self.act = nn.ReLU6(inplace=inplace)
-        elif act_type == 'sigmoid':
+        elif act_type == "sigmoid":
             raise NotImplementedError
-        elif act_type == 'hard_sigmoid':
-            self.act = Hsigmoid(inplace)#nn.Hardsigmoid(inplace=inplace)#Hsigmoid(inplace)#
-        elif act_type == 'hard_swish' or act_type == 'hswish':
+        elif act_type == "hard_sigmoid":
+            self.act = Hsigmoid(
+                inplace
+            )  # nn.Hardsigmoid(inplace=inplace)#Hsigmoid(inplace)#
+        elif act_type == "hard_swish" or act_type == "hswish":
             self.act = Hswish(inplace=inplace)
-        elif act_type == 'leakyrelu':
+        elif act_type == "leakyrelu":
             self.act = nn.LeakyReLU(inplace=inplace)
-        elif act_type == 'gelu':
+        elif act_type == "gelu":
             self.act = GELU(inplace=inplace)
-        elif act_type == 'swish':
+        elif act_type == "swish":
             self.act = Swish(inplace=inplace)
         else:
             raise NotImplementedError
 
     def forward(self, inputs):
-        return self.act(inputs)
+        return self.act(inputs)

python/rapidocr_torch/modeling/necks/intracl.py

@@ -7,55 +7,55 @@ class IntraCLBlock(nn.Module):
         self.channels = in_channels
         self.rf = reduce_factor
         self.conv1x1_reduce_channel = nn.Conv2d(
-            self.channels,
-            self.channels // self.rf,
-            kernel_size=1,
-            stride=1,
-            padding=0)
+            self.channels, self.channels // self.rf, kernel_size=1, stride=1, padding=0
+        )
         self.conv1x1_return_channel = nn.Conv2d(
-            self.channels // self.rf,
-            self.channels,
-            kernel_size=1,
-            stride=1,
-            padding=0)
+            self.channels // self.rf, self.channels, kernel_size=1, stride=1, padding=0
+        )
 
         self.v_layer_7x1 = nn.Conv2d(
             self.channels // self.rf,
             self.channels // self.rf,
             kernel_size=(7, 1),
             stride=(1, 1),
-            padding=(3, 0))
+            padding=(3, 0),
+        )
         self.v_layer_5x1 = nn.Conv2d(
             self.channels // self.rf,
             self.channels // self.rf,
             kernel_size=(5, 1),
             stride=(1, 1),
-            padding=(2, 0))
+            padding=(2, 0),
+        )
         self.v_layer_3x1 = nn.Conv2d(
             self.channels // self.rf,
             self.channels // self.rf,
             kernel_size=(3, 1),
             stride=(1, 1),
-            padding=(1, 0))
+            padding=(1, 0),
+        )
 
         self.q_layer_1x7 = nn.Conv2d(
             self.channels // self.rf,
             self.channels // self.rf,
             kernel_size=(1, 7),
             stride=(1, 1),
-            padding=(0, 3))
+            padding=(0, 3),
+        )
         self.q_layer_1x5 = nn.Conv2d(
             self.channels // self.rf,
             self.channels // self.rf,
             kernel_size=(1, 5),
             stride=(1, 1),
-            padding=(0, 2))
+            padding=(0, 2),
+        )
         self.q_layer_1x3 = nn.Conv2d(
             self.channels // self.rf,
             self.channels // self.rf,
             kernel_size=(1, 3),
             stride=(1, 1),
-            padding=(0, 1))
+            padding=(0, 1),
+        )
 
         # base
         self.c_layer_7x7 = nn.Conv2d(
@@ -63,19 +63,22 @@ class IntraCLBlock(nn.Module):
             self.channels // self.rf,
             kernel_size=(7, 7),
             stride=(1, 1),
-            padding=(3, 3))
+            padding=(3, 3),
+        )
         self.c_layer_5x5 = nn.Conv2d(
             self.channels // self.rf,
             self.channels // self.rf,
             kernel_size=(5, 5),
             stride=(1, 1),
-            padding=(2, 2))
+            padding=(2, 2),
+        )
         self.c_layer_3x3 = nn.Conv2d(
             self.channels // self.rf,
             self.channels // self.rf,
             kernel_size=(3, 3),
             stride=(1, 1),
-            padding=(1, 1))
+            padding=(1, 1),
+        )
 
         self.bn = nn.BatchNorm2d(self.channels)
         self.relu = nn.ReLU()
@@ -111,4 +114,4 @@ def build_intraclblock_list(num_block):
     for i in range(num_block):
         IntraCLBlock_list.append(IntraCLBlock())
 
-    return IntraCLBlock_list
+    return IntraCLBlock_list

python/tests/test_files/config.yaml

@@ -1,62 +0,0 @@
-Global:
-    lang: "ch"
-    text_score: 0.5
-    use_det: true
-    use_cls: true
-    use_rec: true
-
-    print_verbose: false
-    min_height: 30
-    width_height_ratio: 8
-    max_side_len: 2000
-    min_side_len: 30
-    return_word_box: false
-
-    with_onnx: false
-    with_openvino: false
-    with_paddle: false
-    with_torch: false
-
-EngineConfig:
-    onnxruntime:
-        intra_op_num_threads: -1
-        inter_op_num_threads: -1
-        use_cuda: false
-        use_dml: false
-
-    paddlepaddle:
-        cpu_math_library_num_threads: -1
-        gpu_id: 0
-        gpu_mem: 500
-
-Det:
-    model_path: models/ch_PP-OCRv4_det_infer.onnx
-    paddle_model_dir: models/paddle/ch_PP-OCRv4_det_infer
-
-    limit_side_len: 736
-    limit_type: min
-    std: [ 0.5, 0.5, 0.5 ]
-    mean: [ 0.5, 0.5, 0.5 ]
-
-    thresh: 0.3
-    box_thresh: 0.5
-    max_candidates: 1000
-    unclip_ratio: 1.6
-    use_dilation: true
-    score_mode: fast
-
-Cls:
-    model_path: models/ch_ppocr_mobile_v2.0_cls_infer.onnx
-    paddle_model_dir: models/paddle/ch_ppocr_mobile_v2_cls_infer
-
-    cls_image_shape: [3, 48, 192]
-    cls_batch_num: 6
-    cls_thresh: 0.9
-    label_list: ['0', '180']
-
-Rec:
-    model_path: models/ch_PP-OCRv4_rec_infer.onnx
-    paddle_model_dir: models/paddle/ch_PP-OCRv4_rec_infer
-
-    rec_img_shape: [3, 48, 320]
-    rec_batch_num: 6

python/tests/test_main.py

@@ -34,11 +34,18 @@ def get_engine(params: Optional[Dict[str, Any]] = None):
     return engine
 
 
+def test_error_lang():
+    with pytest.raises(ValueError) as exc_info:
+        engine = get_engine(params={"Global.lang": "eh"})
+    assert exc_info.type is ValueError
+
+
 def test_lang():
-    engine = get_engine(params={"Global.lang": "ch", "Global.with_openvino": True})
+    engine = get_engine(params={"Global.lang": "en"})
+    img_path = tests_dir / "en.jpg"
     result = engine(img_path)
     assert result.txts is not None
-    assert result.txts[0] == "正品促销"
+    assert result.txts[0] == "MovieShots Dataset"
 
 
 def test_engine_openvino():