PaddleOCR 文字检测部分源码学习(3)-模型(1)

2021SC@SDUSC
代码位置：ppocr->modeling->architectures->base_model.py

class BaseModel(nn.Layer):
    def __init__(self, config):
        """
        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 transfrom,
        # for rec, transfrom can be TPS,None
        # for det and cls, transfrom shoule to be None,
        # if you make model differently, you can use transfrom in det and cls
        if 'Transform' not in config or config['Transform'] is None:
            self.use_transform = False
        else:
            self.use_transform = True
            config['Transform']['in_channels'] = in_channels
            self.transform = build_transform(config['Transform'])
            in_channels = self.transform.out_channels

        # build backbone, backbone is need for del, rec and cls
        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
        config["Head"]['in_channels'] = in_channels
        self.head = build_head(config["Head"])

        self.return_all_feats = config.get("return_all_feats", False)

    def forward(self, x, data=None):
        y = dict()
        if self.use_transform:
            x = self.transform(x)
        x = self.backbone(x)
        y["backbone_out"] = x
        if self.use_neck:
            x = self.neck(x)
        y["neck_out"] = x
        x = self.head(x, targets=data)
        if isinstance(x, dict):
            y.update(x)
        else:
            y["head_out"] = x
        if self.return_all_feats:
            return y
        else:
            return x

_init_函数的作用是选择合适的模型框架，分别选择
transform 图像变换算法 TPS或STN
backbone 特征提取 Vgg RestNet MobileNet
neck 特征增强 FRN RNN
head 输出 Attention DBHead

对图像变换来说，文本识别可以是TPS算法或者没有，对检测和分类则一定是空参。
对backbone来说，三者都是必须的
对neck来说
det 检测：可以是FPN，BIFPN等等
cls 分类：应该是空参
rec 识别：应该是cnn rnn或者空参
对于head来说 三者都是必须的