first commit

benchmark/PaddleOCR_DBNet/models/__init__.py

@@ -0,0 +1,22 @@
+# -*- coding: utf-8 -*-
+# @Time    : 2019/8/23 21:55
+# @Author  : zhoujun
+import copy
+from .model import Model
+from .losses import build_loss
+
+__all__ = ["build_loss", "build_model"]
+support_model = ["Model"]
+
+
+def build_model(config):
+    """
+    get architecture model class
+    """
+    copy_config = copy.deepcopy(config)
+    arch_type = copy_config.pop("type")
+    assert (
+        arch_type in support_model
+    ), f"{arch_type} is not developed yet!, only {support_model} are support now"
+    arch_model = eval(arch_type)(copy_config)
+    return arch_model

benchmark/PaddleOCR_DBNet/models/backbone/__init__.py

@@ -0,0 +1,25 @@
+# -*- coding: utf-8 -*-
+# @Time    : 2019/8/23 21:54
+# @Author  : zhoujun
+
+from .resnet import *
+
+__all__ = ["build_backbone"]
+
+support_backbone = [
+    "resnet18",
+    "deformable_resnet18",
+    "deformable_resnet50",
+    "resnet50",
+    "resnet34",
+    "resnet101",
+    "resnet152",
+]
+
+
+def build_backbone(backbone_name, **kwargs):
+    assert (
+        backbone_name in support_backbone
+    ), f"all support backbone is {support_backbone}"
+    backbone = eval(backbone_name)(**kwargs)
+    return backbone

benchmark/PaddleOCR_DBNet/models/backbone/resnet.py

@@ -0,0 +1,366 @@
+import math
+import paddle
+from paddle import nn
+
+BatchNorm2d = nn.BatchNorm2D
+
+__all__ = [
+    "ResNet",
+    "resnet18",
+    "resnet34",
+    "resnet50",
+    "resnet101",
+    "deformable_resnet18",
+    "deformable_resnet50",
+    "resnet152",
+]
+
+model_urls = {
+    "resnet18": "https://download.pytorch.org/models/resnet18-5c106cde.pth",
+    "resnet34": "https://download.pytorch.org/models/resnet34-333f7ec4.pth",
+    "resnet50": "https://download.pytorch.org/models/resnet50-19c8e357.pth",
+    "resnet101": "https://download.pytorch.org/models/resnet101-5d3b4d8f.pth",
+    "resnet152": "https://download.pytorch.org/models/resnet152-b121ed2d.pth",
+}
+
+
+def constant_init(module, constant, bias=0):
+    module.weight = paddle.create_parameter(
+        shape=module.weight.shape,
+        dtype="float32",
+        default_initializer=paddle.nn.initializer.Constant(constant),
+    )
+    if hasattr(module, "bias"):
+        module.bias = paddle.create_parameter(
+            shape=module.bias.shape,
+            dtype="float32",
+            default_initializer=paddle.nn.initializer.Constant(bias),
+        )
+
+
+def conv3x3(in_planes, out_planes, stride=1):
+    """3x3 convolution with padding"""
+    return nn.Conv2D(
+        in_planes, out_planes, kernel_size=3, stride=stride, padding=1, bias_attr=False
+    )
+
+
+class BasicBlock(nn.Layer):
+    expansion = 1
+
+    def __init__(self, inplanes, planes, stride=1, downsample=None, dcn=None):
+        super(BasicBlock, self).__init__()
+        self.with_dcn = dcn is not None
+        self.conv1 = conv3x3(inplanes, planes, stride)
+        self.bn1 = BatchNorm2d(planes, momentum=0.1)
+        self.relu = nn.ReLU()
+        self.with_modulated_dcn = False
+        if not self.with_dcn:
+            self.conv2 = nn.Conv2D(
+                planes, planes, kernel_size=3, padding=1, bias_attr=False
+            )
+        else:
+            from paddle.vision.ops import DeformConv2D
+
+            deformable_groups = dcn.get("deformable_groups", 1)
+            offset_channels = 18
+            self.conv2_offset = nn.Conv2D(
+                planes, deformable_groups * offset_channels, kernel_size=3, padding=1
+            )
+            self.conv2 = DeformConv2D(
+                planes, planes, kernel_size=3, padding=1, bias_attr=False
+            )
+        self.bn2 = BatchNorm2d(planes, momentum=0.1)
+        self.downsample = downsample
+        self.stride = stride
+
+    def forward(self, x):
+        residual = x
+
+        out = self.conv1(x)
+        out = self.bn1(out)
+        out = self.relu(out)
+
+        # out = self.conv2(out)
+        if not self.with_dcn:
+            out = self.conv2(out)
+        else:
+            offset = self.conv2_offset(out)
+            out = self.conv2(out, offset)
+        out = self.bn2(out)
+
+        if self.downsample is not None:
+            residual = self.downsample(x)
+
+        out += residual
+        out = self.relu(out)
+
+        return out
+
+
+class Bottleneck(nn.Layer):
+    expansion = 4
+
+    def __init__(self, inplanes, planes, stride=1, downsample=None, dcn=None):
+        super(Bottleneck, self).__init__()
+        self.with_dcn = dcn is not None
+        self.conv1 = nn.Conv2D(inplanes, planes, kernel_size=1, bias_attr=False)
+        self.bn1 = BatchNorm2d(planes, momentum=0.1)
+        self.with_modulated_dcn = False
+        if not self.with_dcn:
+            self.conv2 = nn.Conv2D(
+                planes, planes, kernel_size=3, stride=stride, padding=1, bias_attr=False
+            )
+        else:
+            deformable_groups = dcn.get("deformable_groups", 1)
+            from paddle.vision.ops import DeformConv2D
+
+            offset_channels = 18
+            self.conv2_offset = nn.Conv2D(
+                planes,
+                deformable_groups * offset_channels,
+                stride=stride,
+                kernel_size=3,
+                padding=1,
+            )
+            self.conv2 = DeformConv2D(
+                planes, planes, kernel_size=3, padding=1, stride=stride, bias_attr=False
+            )
+        self.bn2 = BatchNorm2d(planes, momentum=0.1)
+        self.conv3 = nn.Conv2D(planes, planes * 4, kernel_size=1, bias_attr=False)
+        self.bn3 = BatchNorm2d(planes * 4, momentum=0.1)
+        self.relu = nn.ReLU()
+        self.downsample = downsample
+        self.stride = stride
+        self.dcn = dcn
+        self.with_dcn = dcn is not None
+
+    def forward(self, x):
+        residual = x
+
+        out = self.conv1(x)
+        out = self.bn1(out)
+        out = self.relu(out)
+
+        # out = self.conv2(out)
+        if not self.with_dcn:
+            out = self.conv2(out)
+        else:
+            offset = self.conv2_offset(out)
+            out = self.conv2(out, offset)
+        out = self.bn2(out)
+        out = self.relu(out)
+
+        out = self.conv3(out)
+        out = self.bn3(out)
+
+        if self.downsample is not None:
+            residual = self.downsample(x)
+
+        out += residual
+        out = self.relu(out)
+
+        return out
+
+
+class ResNet(nn.Layer):
+    def __init__(self, block, layers, in_channels=3, dcn=None):
+        self.dcn = dcn
+        self.inplanes = 64
+        super(ResNet, self).__init__()
+        self.out_channels = []
+        self.conv1 = nn.Conv2D(
+            in_channels, 64, kernel_size=7, stride=2, padding=3, bias_attr=False
+        )
+        self.bn1 = BatchNorm2d(64, momentum=0.1)
+        self.relu = nn.ReLU()
+        self.maxpool = nn.MaxPool2D(kernel_size=3, stride=2, padding=1)
+        self.layer1 = self._make_layer(block, 64, layers[0])
+        self.layer2 = self._make_layer(block, 128, layers[1], stride=2, dcn=dcn)
+        self.layer3 = self._make_layer(block, 256, layers[2], stride=2, dcn=dcn)
+        self.layer4 = self._make_layer(block, 512, layers[3], stride=2, dcn=dcn)
+
+        if self.dcn is not None:
+            for m in self.modules():
+                if isinstance(m, Bottleneck) or isinstance(m, BasicBlock):
+                    if hasattr(m, "conv2_offset"):
+                        constant_init(m.conv2_offset, 0)
+
+    def _make_layer(self, block, planes, blocks, stride=1, dcn=None):
+        downsample = None
+        if stride != 1 or self.inplanes != planes * block.expansion:
+            downsample = nn.Sequential(
+                nn.Conv2D(
+                    self.inplanes,
+                    planes * block.expansion,
+                    kernel_size=1,
+                    stride=stride,
+                    bias_attr=False,
+                ),
+                BatchNorm2d(planes * block.expansion, momentum=0.1),
+            )
+
+        layers = []
+        layers.append(block(self.inplanes, planes, stride, downsample, dcn=dcn))
+        self.inplanes = planes * block.expansion
+        for i in range(1, blocks):
+            layers.append(block(self.inplanes, planes, dcn=dcn))
+        self.out_channels.append(planes * block.expansion)
+        return nn.Sequential(*layers)
+
+    def forward(self, x):
+        x = self.conv1(x)
+        x = self.bn1(x)
+        x = self.relu(x)
+        x = self.maxpool(x)
+
+        x2 = self.layer1(x)
+        x3 = self.layer2(x2)
+        x4 = self.layer3(x3)
+        x5 = self.layer4(x4)
+
+        return x2, x3, x4, x5
+
+
+def load_torch_params(paddle_model, torch_patams):
+    paddle_params = paddle_model.state_dict()
+
+    fc_names = ["classifier"]
+    for key, torch_value in torch_patams.items():
+        if "num_batches_tracked" in key:
+            continue
+        key = (
+            key.replace("running_var", "_variance")
+            .replace("running_mean", "_mean")
+            .replace("module.", "")
+        )
+        torch_value = torch_value.detach().cpu().numpy()
+        if key in paddle_params:
+            flag = [i in key for i in fc_names]
+            if any(flag) and "weight" in key:  # ignore bias
+                new_shape = [1, 0] + list(range(2, torch_value.ndim))
+                print(
+                    f"name: {key}, ori shape: {torch_value.shape}, new shape: {torch_value.transpose(new_shape).shape}"
+                )
+                torch_value = torch_value.transpose(new_shape)
+            paddle_params[key] = torch_value
+        else:
+            print(f"{key} not in paddle")
+    paddle_model.set_state_dict(paddle_params)
+
+
+def load_models(model, model_name):
+    import torch.utils.model_zoo as model_zoo
+
+    torch_patams = model_zoo.load_url(model_urls[model_name])
+    load_torch_params(model, torch_patams)
+
+
+def resnet18(pretrained=True, **kwargs):
+    """Constructs a ResNet-18 model.
+    Args:
+        pretrained (bool): If True, returns a model pre-trained on ImageNet
+    """
+    model = ResNet(BasicBlock, [2, 2, 2, 2], **kwargs)
+    if pretrained:
+        assert (
+            kwargs.get("in_channels", 3) == 3
+        ), "in_channels must be 3 when pretrained is True"
+        print("load from imagenet")
+        load_models(model, "resnet18")
+    return model
+
+
+def deformable_resnet18(pretrained=True, **kwargs):
+    """Constructs a ResNet-18 model.
+    Args:
+        pretrained (bool): If True, returns a model pre-trained on ImageNet
+    """
+    model = ResNet(BasicBlock, [2, 2, 2, 2], dcn=dict(deformable_groups=1), **kwargs)
+    if pretrained:
+        assert (
+            kwargs.get("in_channels", 3) == 3
+        ), "in_channels must be 3 when pretrained is True"
+        print("load from imagenet")
+        model.load_state_dict(model_zoo.load_url(model_urls["resnet18"]), strict=False)
+    return model
+
+
+def resnet34(pretrained=True, **kwargs):
+    """Constructs a ResNet-34 model.
+    Args:
+        pretrained (bool): If True, returns a model pre-trained on ImageNet
+    """
+    model = ResNet(BasicBlock, [3, 4, 6, 3], **kwargs)
+    if pretrained:
+        assert (
+            kwargs.get("in_channels", 3) == 3
+        ), "in_channels must be 3 when pretrained is True"
+        model.load_state_dict(model_zoo.load_url(model_urls["resnet34"]), strict=False)
+    return model
+
+
+def resnet50(pretrained=True, **kwargs):
+    """Constructs a ResNet-50 model.
+    Args:
+        pretrained (bool): If True, returns a model pre-trained on ImageNet
+    """
+    model = ResNet(Bottleneck, [3, 4, 6, 3], **kwargs)
+    if pretrained:
+        assert (
+            kwargs.get("in_channels", 3) == 3
+        ), "in_channels must be 3 when pretrained is True"
+        load_models(model, "resnet50")
+    return model
+
+
+def deformable_resnet50(pretrained=True, **kwargs):
+    """Constructs a ResNet-50 model with deformable conv.
+    Args:
+        pretrained (bool): If True, returns a model pre-trained on ImageNet
+    """
+    model = ResNet(Bottleneck, [3, 4, 6, 3], dcn=dict(deformable_groups=1), **kwargs)
+    if pretrained:
+        assert (
+            kwargs.get("in_channels", 3) == 3
+        ), "in_channels must be 3 when pretrained is True"
+        model.load_state_dict(model_zoo.load_url(model_urls["resnet50"]), strict=False)
+    return model
+
+
+def resnet101(pretrained=True, **kwargs):
+    """Constructs a ResNet-101 model.
+    Args:
+        pretrained (bool): If True, returns a model pre-trained on ImageNet
+    """
+    model = ResNet(Bottleneck, [3, 4, 23, 3], **kwargs)
+    if pretrained:
+        assert (
+            kwargs.get("in_channels", 3) == 3
+        ), "in_channels must be 3 when pretrained is True"
+        model.load_state_dict(model_zoo.load_url(model_urls["resnet101"]), strict=False)
+    return model
+
+
+def resnet152(pretrained=True, **kwargs):
+    """Constructs a ResNet-152 model.
+    Args:
+        pretrained (bool): If True, returns a model pre-trained on ImageNet
+    """
+    model = ResNet(Bottleneck, [3, 8, 36, 3], **kwargs)
+    if pretrained:
+        assert (
+            kwargs.get("in_channels", 3) == 3
+        ), "in_channels must be 3 when pretrained is True"
+        model.load_state_dict(model_zoo.load_url(model_urls["resnet152"]), strict=False)
+    return model
+
+
+if __name__ == "__main__":
+    x = paddle.zeros([2, 3, 640, 640])
+    net = resnet50(pretrained=True)
+    y = net(x)
+    for u in y:
+        print(u.shape)
+
+    print(net.out_channels)

benchmark/PaddleOCR_DBNet/models/basic.py

@@ -0,0 +1,40 @@
+# -*- coding: utf-8 -*-
+# @Time    : 2019/12/6 11:19
+# @Author  : zhoujun
+from paddle import nn
+
+
+class ConvBnRelu(nn.Layer):
+    def __init__(
+        self,
+        in_channels,
+        out_channels,
+        kernel_size,
+        stride=1,
+        padding=0,
+        dilation=1,
+        groups=1,
+        bias=True,
+        padding_mode="zeros",
+        inplace=True,
+    ):
+        super().__init__()
+        self.conv = nn.Conv2D(
+            in_channels=in_channels,
+            out_channels=out_channels,
+            kernel_size=kernel_size,
+            stride=stride,
+            padding=padding,
+            dilation=dilation,
+            groups=groups,
+            bias_attr=bias,
+            padding_mode=padding_mode,
+        )
+        self.bn = nn.BatchNorm2D(out_channels)
+        self.relu = nn.ReLU()
+
+    def forward(self, x):
+        x = self.conv(x)
+        x = self.bn(x)
+        x = self.relu(x)
+        return x

benchmark/PaddleOCR_DBNet/models/head/DBHead.py

@@ -0,0 +1,132 @@
+# -*- coding: utf-8 -*-
+# @Time    : 2019/12/4 14:54
+# @Author  : zhoujun
+import paddle
+from paddle import nn, ParamAttr
+
+
+class DBHead(nn.Layer):
+    def __init__(self, in_channels, out_channels, k=50):
+        super().__init__()
+        self.k = k
+        self.binarize = nn.Sequential(
+            nn.Conv2D(
+                in_channels,
+                in_channels // 4,
+                3,
+                padding=1,
+                weight_attr=ParamAttr(initializer=nn.initializer.KaimingNormal()),
+            ),
+            nn.BatchNorm2D(
+                in_channels // 4,
+                weight_attr=ParamAttr(initializer=nn.initializer.Constant(1)),
+                bias_attr=ParamAttr(initializer=nn.initializer.Constant(1e-4)),
+            ),
+            nn.ReLU(),
+            nn.Conv2DTranspose(
+                in_channels // 4,
+                in_channels // 4,
+                2,
+                2,
+                weight_attr=ParamAttr(initializer=nn.initializer.KaimingNormal()),
+            ),
+            nn.BatchNorm2D(
+                in_channels // 4,
+                weight_attr=ParamAttr(initializer=nn.initializer.Constant(1)),
+                bias_attr=ParamAttr(initializer=nn.initializer.Constant(1e-4)),
+            ),
+            nn.ReLU(),
+            nn.Conv2DTranspose(
+                in_channels // 4, 1, 2, 2, weight_attr=nn.initializer.KaimingNormal()
+            ),
+            nn.Sigmoid(),
+        )
+
+        self.thresh = self._init_thresh(in_channels)
+
+    def forward(self, x):
+        shrink_maps = self.binarize(x)
+        threshold_maps = self.thresh(x)
+        if self.training:
+            binary_maps = self.step_function(shrink_maps, threshold_maps)
+            y = paddle.concat((shrink_maps, threshold_maps, binary_maps), axis=1)
+        else:
+            y = paddle.concat((shrink_maps, threshold_maps), axis=1)
+        return y
+
+    def _init_thresh(self, inner_channels, serial=False, smooth=False, bias=False):
+        in_channels = inner_channels
+        if serial:
+            in_channels += 1
+        self.thresh = nn.Sequential(
+            nn.Conv2D(
+                in_channels,
+                inner_channels // 4,
+                3,
+                padding=1,
+                bias_attr=bias,
+                weight_attr=ParamAttr(initializer=nn.initializer.KaimingNormal()),
+            ),
+            nn.BatchNorm2D(
+                inner_channels // 4,
+                weight_attr=ParamAttr(initializer=nn.initializer.Constant(1)),
+                bias_attr=ParamAttr(initializer=nn.initializer.Constant(1e-4)),
+            ),
+            nn.ReLU(),
+            self._init_upsample(
+                inner_channels // 4, inner_channels // 4, smooth=smooth, bias=bias
+            ),
+            nn.BatchNorm2D(
+                inner_channels // 4,
+                weight_attr=ParamAttr(initializer=nn.initializer.Constant(1)),
+                bias_attr=ParamAttr(initializer=nn.initializer.Constant(1e-4)),
+            ),
+            nn.ReLU(),
+            self._init_upsample(inner_channels // 4, 1, smooth=smooth, bias=bias),
+            nn.Sigmoid(),
+        )
+        return self.thresh
+
+    def _init_upsample(self, in_channels, out_channels, smooth=False, bias=False):
+        if smooth:
+            inter_out_channels = out_channels
+            if out_channels == 1:
+                inter_out_channels = in_channels
+            module_list = [
+                nn.Upsample(scale_factor=2, mode="nearest"),
+                nn.Conv2D(
+                    in_channels,
+                    inter_out_channels,
+                    3,
+                    1,
+                    1,
+                    bias_attr=bias,
+                    weight_attr=ParamAttr(initializer=nn.initializer.KaimingNormal()),
+                ),
+            ]
+            if out_channels == 1:
+                module_list.append(
+                    nn.Conv2D(
+                        in_channels,
+                        out_channels,
+                        kernel_size=1,
+                        stride=1,
+                        padding=1,
+                        bias_attr=True,
+                        weight_attr=ParamAttr(
+                            initializer=nn.initializer.KaimingNormal()
+                        ),
+                    )
+                )
+            return nn.Sequential(module_list)
+        else:
+            return nn.Conv2DTranspose(
+                in_channels,
+                out_channels,
+                2,
+                2,
+                weight_attr=ParamAttr(initializer=nn.initializer.KaimingNormal()),
+            )
+
+    def step_function(self, x, y):
+        return paddle.reciprocal(1 + paddle.exp(-self.k * (x - y)))

benchmark/PaddleOCR_DBNet/models/head/__init__.py

@@ -0,0 +1,13 @@
+# -*- coding: utf-8 -*-
+# @Time    : 2020/6/5 11:35
+# @Author  : zhoujun
+from .DBHead import DBHead
+
+__all__ = ["build_head"]
+support_head = ["DBHead"]
+
+
+def build_head(head_name, **kwargs):
+    assert head_name in support_head, f"all support head is {support_head}"
+    head = eval(head_name)(**kwargs)
+    return head

benchmark/PaddleOCR_DBNet/models/losses/DB_loss.py

@@ -0,0 +1,50 @@
+import paddle
+from models.losses.basic_loss import BalanceCrossEntropyLoss, MaskL1Loss, DiceLoss
+
+
+class DBLoss(paddle.nn.Layer):
+    def __init__(self, alpha=1.0, beta=10, ohem_ratio=3, reduction="mean", eps=1e-06):
+        """
+        Implement PSE Loss.
+        :param alpha: binary_map loss 前面的系数
+        :param beta: threshold_map loss 前面的系数
+        :param ohem_ratio: OHEM的比例
+        :param reduction: 'mean' or 'sum'对 batch里的loss 算均值或求和
+        """
+        super().__init__()
+        assert reduction in ["mean", "sum"], " reduction must in ['mean','sum']"
+        self.alpha = alpha
+        self.beta = beta
+        self.bce_loss = BalanceCrossEntropyLoss(negative_ratio=ohem_ratio)
+        self.dice_loss = DiceLoss(eps=eps)
+        self.l1_loss = MaskL1Loss(eps=eps)
+        self.ohem_ratio = ohem_ratio
+        self.reduction = reduction
+
+    def forward(self, pred, batch):
+        shrink_maps = pred[:, 0, :, :]
+        threshold_maps = pred[:, 1, :, :]
+        binary_maps = pred[:, 2, :, :]
+        loss_shrink_maps = self.bce_loss(
+            shrink_maps, batch["shrink_map"], batch["shrink_mask"]
+        )
+        loss_threshold_maps = self.l1_loss(
+            threshold_maps, batch["threshold_map"], batch["threshold_mask"]
+        )
+        metrics = dict(
+            loss_shrink_maps=loss_shrink_maps, loss_threshold_maps=loss_threshold_maps
+        )
+        if pred.shape[1] > 2:
+            loss_binary_maps = self.dice_loss(
+                binary_maps, batch["shrink_map"], batch["shrink_mask"]
+            )
+            metrics["loss_binary_maps"] = loss_binary_maps
+            loss_all = (
+                self.alpha * loss_shrink_maps
+                + self.beta * loss_threshold_maps
+                + loss_binary_maps
+            )
+            metrics["loss"] = loss_all
+        else:
+            metrics["loss"] = loss_shrink_maps
+        return metrics

benchmark/PaddleOCR_DBNet/models/losses/__init__.py

@@ -0,0 +1,16 @@
+# -*- coding: utf-8 -*-
+# @Time    : 2020/6/5 11:36
+# @Author  : zhoujun
+import copy
+from .DB_loss import DBLoss
+
+__all__ = ["build_loss"]
+support_loss = ["DBLoss"]
+
+
+def build_loss(config):
+    copy_config = copy.deepcopy(config)
+    loss_type = copy_config.pop("type")
+    assert loss_type in support_loss, f"all support loss is {support_loss}"
+    criterion = eval(loss_type)(**copy_config)
+    return criterion

benchmark/PaddleOCR_DBNet/models/losses/basic_loss.py

@@ -0,0 +1,101 @@
+# -*- coding: utf-8 -*-
+# @Time    : 2019/12/4 14:39
+# @Author  : zhoujun
+import paddle
+import paddle.nn as nn
+
+
+class BalanceCrossEntropyLoss(nn.Layer):
+    """
+    Balanced cross entropy loss.
+    Shape:
+        - Input: :math:`(N, 1, H, W)`
+        - GT: :math:`(N, 1, H, W)`, same shape as the input
+        - Mask: :math:`(N, H, W)`, same spatial shape as the input
+        - Output: scalar.
+
+    """
+
+    def __init__(self, negative_ratio=3.0, eps=1e-6):
+        super(BalanceCrossEntropyLoss, self).__init__()
+        self.negative_ratio = negative_ratio
+        self.eps = eps
+
+    def forward(
+        self,
+        pred: paddle.Tensor,
+        gt: paddle.Tensor,
+        mask: paddle.Tensor,
+        return_origin=False,
+    ):
+        """
+        Args:
+            pred: shape :math:`(N, 1, H, W)`, the prediction of network
+            gt: shape :math:`(N, 1, H, W)`, the target
+            mask: shape :math:`(N, H, W)`, the mask indicates positive regions
+        """
+        positive = gt * mask
+        negative = (1 - gt) * mask
+        positive_count = int(positive.sum())
+        negative_count = min(
+            int(negative.sum()), int(positive_count * self.negative_ratio)
+        )
+        loss = nn.functional.binary_cross_entropy(pred, gt, reduction="none")
+        positive_loss = loss * positive
+        negative_loss = loss * negative
+        negative_loss, _ = negative_loss.reshape([-1]).topk(negative_count)
+
+        balance_loss = (positive_loss.sum() + negative_loss.sum()) / (
+            positive_count + negative_count + self.eps
+        )
+
+        if return_origin:
+            return balance_loss, loss
+        return balance_loss
+
+
+class DiceLoss(nn.Layer):
+    """
+    Loss function from https://arxiv.org/abs/1707.03237,
+    where iou computation is introduced heatmap manner to measure the
+    diversity between tow heatmaps.
+    """
+
+    def __init__(self, eps=1e-6):
+        super(DiceLoss, self).__init__()
+        self.eps = eps
+
+    def forward(self, pred: paddle.Tensor, gt, mask, weights=None):
+        """
+        pred: one or two heatmaps of shape (N, 1, H, W),
+            the losses of tow heatmaps are added together.
+        gt: (N, 1, H, W)
+        mask: (N, H, W)
+        """
+        return self._compute(pred, gt, mask, weights)
+
+    def _compute(self, pred, gt, mask, weights):
+        if len(pred.shape) == 4:
+            pred = pred[:, 0, :, :]
+            gt = gt[:, 0, :, :]
+        assert pred.shape == gt.shape
+        assert pred.shape == mask.shape
+        if weights is not None:
+            assert weights.shape == mask.shape
+            mask = weights * mask
+        intersection = (pred * gt * mask).sum()
+
+        union = (pred * mask).sum() + (gt * mask).sum() + self.eps
+        loss = 1 - 2.0 * intersection / union
+        assert loss <= 1
+        return loss
+
+
+class MaskL1Loss(nn.Layer):
+    def __init__(self, eps=1e-6):
+        super(MaskL1Loss, self).__init__()
+        self.eps = eps
+
+    def forward(self, pred: paddle.Tensor, gt, mask):
+        loss = (paddle.abs(pred - gt) * mask).sum() / (mask.sum() + self.eps)
+        return loss

benchmark/PaddleOCR_DBNet/models/model.py

@@ -0,0 +1,39 @@
+# -*- coding: utf-8 -*-
+# @Time    : 2019/8/23 21:57
+# @Author  : zhoujun
+from addict import Dict
+from paddle import nn
+import paddle.nn.functional as F
+
+from models.backbone import build_backbone
+from models.neck import build_neck
+from models.head import build_head
+
+
+class Model(nn.Layer):
+    def __init__(self, model_config: dict):
+        """
+        PANnet
+        :param model_config: 模型配置
+        """
+        super().__init__()
+        model_config = Dict(model_config)
+        backbone_type = model_config.backbone.pop("type")
+        neck_type = model_config.neck.pop("type")
+        head_type = model_config.head.pop("type")
+        self.backbone = build_backbone(backbone_type, **model_config.backbone)
+        self.neck = build_neck(
+            neck_type, in_channels=self.backbone.out_channels, **model_config.neck
+        )
+        self.head = build_head(
+            head_type, in_channels=self.neck.out_channels, **model_config.head
+        )
+        self.name = f"{backbone_type}_{neck_type}_{head_type}"
+
+    def forward(self, x):
+        _, _, H, W = x.shape
+        backbone_out = self.backbone(x)
+        neck_out = self.neck(backbone_out)
+        y = self.head(neck_out)
+        y = F.interpolate(y, size=(H, W), mode="bilinear", align_corners=True)
+        return y

benchmark/PaddleOCR_DBNet/models/neck/FPN.py

@@ -0,0 +1,75 @@
+# -*- coding: utf-8 -*-
+# @Time    : 2019/9/13 10:29
+# @Author  : zhoujun
+import paddle
+import paddle.nn.functional as F
+from paddle import nn
+
+from models.basic import ConvBnRelu
+
+
+class FPN(nn.Layer):
+    def __init__(self, in_channels, inner_channels=256, **kwargs):
+        """
+        :param in_channels: 基础网络输出的维度
+        :param kwargs:
+        """
+        super().__init__()
+        inplace = True
+        self.conv_out = inner_channels
+        inner_channels = inner_channels // 4
+        # reduce layers
+        self.reduce_conv_c2 = ConvBnRelu(
+            in_channels[0], inner_channels, kernel_size=1, inplace=inplace
+        )
+        self.reduce_conv_c3 = ConvBnRelu(
+            in_channels[1], inner_channels, kernel_size=1, inplace=inplace
+        )
+        self.reduce_conv_c4 = ConvBnRelu(
+            in_channels[2], inner_channels, kernel_size=1, inplace=inplace
+        )
+        self.reduce_conv_c5 = ConvBnRelu(
+            in_channels[3], inner_channels, kernel_size=1, inplace=inplace
+        )
+        # Smooth layers
+        self.smooth_p4 = ConvBnRelu(
+            inner_channels, inner_channels, kernel_size=3, padding=1, inplace=inplace
+        )
+        self.smooth_p3 = ConvBnRelu(
+            inner_channels, inner_channels, kernel_size=3, padding=1, inplace=inplace
+        )
+        self.smooth_p2 = ConvBnRelu(
+            inner_channels, inner_channels, kernel_size=3, padding=1, inplace=inplace
+        )
+
+        self.conv = nn.Sequential(
+            nn.Conv2D(self.conv_out, self.conv_out, kernel_size=3, padding=1, stride=1),
+            nn.BatchNorm2D(self.conv_out),
+            nn.ReLU(),
+        )
+        self.out_channels = self.conv_out
+
+    def forward(self, x):
+        c2, c3, c4, c5 = x
+        # Top-down
+        p5 = self.reduce_conv_c5(c5)
+        p4 = self._upsample_add(p5, self.reduce_conv_c4(c4))
+        p4 = self.smooth_p4(p4)
+        p3 = self._upsample_add(p4, self.reduce_conv_c3(c3))
+        p3 = self.smooth_p3(p3)
+        p2 = self._upsample_add(p3, self.reduce_conv_c2(c2))
+        p2 = self.smooth_p2(p2)
+
+        x = self._upsample_cat(p2, p3, p4, p5)
+        x = self.conv(x)
+        return x
+
+    def _upsample_add(self, x, y):
+        return F.interpolate(x, size=y.shape[2:]) + y
+
+    def _upsample_cat(self, p2, p3, p4, p5):
+        h, w = p2.shape[2:]
+        p3 = F.interpolate(p3, size=(h, w))
+        p4 = F.interpolate(p4, size=(h, w))
+        p5 = F.interpolate(p5, size=(h, w))
+        return paddle.concat([p2, p3, p4, p5], axis=1)

benchmark/PaddleOCR_DBNet/models/neck/__init__.py

@@ -0,0 +1,13 @@
+# -*- coding: utf-8 -*-
+# @Time    : 2020/6/5 11:34
+# @Author  : zhoujun
+from .FPN import FPN
+
+__all__ = ["build_neck"]
+support_neck = ["FPN"]
+
+
+def build_neck(neck_name, **kwargs):
+    assert neck_name in support_neck, f"all support neck is {support_neck}"
+    neck = eval(neck_name)(**kwargs)
+    return neck