[Feature] Add UNetEarlyFusion

paddlers/models/cd/models/__init__.py

@@ -12,4 +12,5 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from .cdnet import CDNet
+from .cdnet import CDNet
+from .unet_ef import UNetEarlyFusion

paddlers/models/cd/models/layers/__init__.py

@@ -0,0 +1,15 @@
+# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#    http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from .blocks import *

paddlers/models/cd/models/layers/blocks.py

@@ -0,0 +1,140 @@
+# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#    http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import paddle.nn as nn
+
+
+__all__ = [
+    'BasicConv', 'Conv1x1', 'Conv3x3', 'Conv7x7', 
+    'MaxPool2x2', 'MaxUnPool2x2', 
+    'ConvTransposed3x3',
+    'Identity'
+]
+
+
+def get_norm_layer():
+    # TODO: select appropriate norm layer.
+    return nn.BatchNorm2D
+
+
+def get_act_layer():
+    # TODO: select appropriate activation layer.
+    return nn.ReLU
+
+
+def make_norm(*args, **kwargs):
+    norm_layer = get_norm_layer()
+    return norm_layer(*args, **kwargs)
+
+
+def make_act(*args, **kwargs):
+    act_layer = get_act_layer()
+    return act_layer(*args, **kwargs)
+
+
+class BasicConv(nn.Layer):
+    def __init__(
+        self, in_ch, out_ch, 
+        kernel_size, pad_mode='constant', 
+        bias='auto', norm=False, act=False, 
+        **kwargs
+    ):
+        super().__init__()
+        seq = []
+        if kernel_size >= 2:
+            seq.append(nn.Pad2D(kernel_size//2, mode=pad_mode))
+        seq.append(
+            nn.Conv2D(
+                in_ch, out_ch, kernel_size,
+                stride=1, padding=0,
+                bias_attr=(False if norm else None) if bias=='auto' else bias,
+                **kwargs
+            )
+        )
+        if norm:
+            if norm is True:
+                norm = make_norm(out_ch)
+            seq.append(norm)
+        if act:
+            if act is True:
+                act = make_act()
+            seq.append(act)
+        self.seq = nn.Sequential(*seq)
+
+    def forward(self, x):
+        return self.seq(x)
+
+
+class Conv1x1(BasicConv):
+    def __init__(self, in_ch, out_ch, pad_mode='constant', bias='auto', norm=False, act=False, **kwargs):
+        super().__init__(in_ch, out_ch, 1, pad_mode=pad_mode, bias=bias, norm=norm, act=act, **kwargs)
+
+
+class Conv3x3(BasicConv):
+    def __init__(self, in_ch, out_ch, pad_mode='constant', bias='auto', norm=False, act=False, **kwargs):
+        super().__init__(in_ch, out_ch, 3, pad_mode=pad_mode, bias=bias, norm=norm, act=act, **kwargs)
+
+
+class Conv7x7(BasicConv):
+    def __init__(self, in_ch, out_ch, pad_mode='constant', bias='auto', norm=False, act=False, **kwargs):
+        super().__init__(in_ch, out_ch, 7, pad_mode=pad_mode, bias=bias, norm=norm, act=act, **kwargs)
+
+
+class MaxPool2x2(nn.MaxPool2D):
+    def __init__(self, **kwargs):
+        super().__init__(kernel_size=2, stride=(2,2), padding=(0,0), **kwargs)
+
+
+class MaxUnPool2x2(nn.MaxUnPool2D):
+    def __init__(self, **kwargs):
+        super().__init__(kernel_size=2, stride=(2,2), padding=(0,0), **kwargs)
+
+
+class ConvTransposed3x3(nn.Layer):
+    def __init__(
+        self, in_ch, out_ch,
+        bias='auto', norm=False, act=False, 
+        **kwargs
+    ):
+        super().__init__()
+        seq = []
+        seq.append(
+            nn.Conv2DTranspose(
+                in_ch, out_ch, 3,
+                stride=2, padding=1,
+                bias_attr=(False if norm else None) if bias=='auto' else bias,
+                **kwargs
+            )
+        )
+        if norm:
+            if norm is True:
+                norm = make_norm(out_ch)
+            seq.append(norm)
+        if act:
+            if act is True:
+                act = make_act()
+            seq.append(act)
+        self.seq = nn.Sequential(*seq)
+
+    def forward(self, x):
+        return self.seq(x)
+
+
+class Identity(nn.Layer):
+    """A placeholder identity operator that accepts exactly one argument."""
+    def __init__(self, *args, **kwargs):
+        super().__init__()
+    
+    def forward(self, x):
+        return x

paddlers/models/cd/models/param_init.py

@@ -0,0 +1,82 @@
+# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#    http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import paddle.nn as nn
+import paddle.nn.functional as F
+
+
+def normal_init(param, *args, **kwargs):
+    """
+    Initialize parameters with a normal distribution.
+
+    Args:
+        param (Tensor): The tensor that needs to be initialized.
+
+    Returns:
+        The initialized parameters.
+    """
+    return nn.initializer.Normal(*args, **kwargs)(param)
+
+
+def kaiming_normal_init(param, *args, **kwargs):
+    """
+    Initialize parameters with the Kaiming normal distribution.
+
+    For more information about the Kaiming initialization method, please refer to
+    https://arxiv.org/abs/1502.01852
+
+    Args:
+        param (Tensor): The tensor that needs to be initialized.
+
+    Returns:
+        The initialized parameters.
+    """
+    return nn.initializer.KaimingNormal(*args, **kwargs)(param)
+
+
+def constant_init(param, *args, **kwargs):
+    """
+    Initialize parameters with constants.
+
+    Args:
+        param (Tensor): The tensor that needs to be initialized.
+
+    Returns:
+        The initialized parameters.
+    """
+    return nn.initializer.Constant(*args, **kwargs)(param)
+
+
+class KaimingInitMixin:
+    """
+    A mix-in that provides the Kaiming initialization functionality.
+
+    Examples:
+
+        from paddlers.models.cd.models.param_init import KaimingInitMixin
+
+        class CustomNet(nn.Layer, KaimingInitMixin):
+            def __init__(self, num_channels, num_classes):
+                super().__init__()
+                self.conv = nn.Conv2D(num_channels, num_classes, 3, 1, 0, bias_attr=False)
+                self.bn = nn.BatchNorm2D(num_classes)
+                self.init_weight()
+    """
+    def init_weight(self):
+        for layer in self.sublayers():
+            if isinstance(layer, nn.Conv2D):
+                kaiming_normal_init(layer.weight)
+            elif isinstance(layer, (nn.BatchNorm, nn.SyncBatchNorm)):
+                constant_init(layer.weight, value=1)
+                constant_init(layer.bias, value=0)

paddlers/models/cd/models/unet_ef.py

@@ -0,0 +1,201 @@
+# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#    http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import paddle
+import paddle.nn as nn
+import paddle.nn.functional as F
+
+
+from .layers import Conv3x3, MaxPool2x2, ConvTransposed3x3, Identity
+from .param_init import normal_init, constant_init
+
+
+class UNetEarlyFusion(nn.Layer):
+    """
+    The FC-EF implementation based on PaddlePaddle.
+
+    The original article refers to
+    Caye Daudt, R., et al. "Fully convolutional siamese networks for change detection"
+    (https://arxiv.org/abs/1810.08462)
+
+    Args:
+        in_channels (int): The number of bands of the input images.
+        num_classes (int): The number of target classes.
+        use_dropout (bool, optional): A bool value that indicates whether to use dropout layers. 
+            When the model is trained on a relatively small dataset, the dropout layers help prevent 
+            overfitting. Default: False.
+    """
+    def __init__(
+        self, 
+        in_channels, 
+        num_classes, 
+        use_dropout=False
+    ):
+        super().__init__()
+
+        C1, C2, C3, C4, C5 = 16, 32, 64, 128, 256
+
+        self.use_dropout = use_dropout
+
+        self.conv11 = Conv3x3(in_channels, C1, norm=True, act=True)
+        self.do11 = self._make_dropout()
+        self.conv12 = Conv3x3(C1, C1, norm=True, act=True)
+        self.do12 = self._make_dropout()
+        self.pool1 = MaxPool2x2()
+
+        self.conv21 = Conv3x3(C1, C2, norm=True, act=True)
+        self.do21 = self._make_dropout()
+        self.conv22 = Conv3x3(C2, C2, norm=True, act=True)
+        self.do22 = self._make_dropout()
+        self.pool2 = MaxPool2x2()
+
+        self.conv31 = Conv3x3(C2, C3, norm=True, act=True)
+        self.do31 = self._make_dropout()
+        self.conv32 = Conv3x3(C3, C3, norm=True, act=True)
+        self.do32 = self._make_dropout()
+        self.conv33 = Conv3x3(C3, C3, norm=True, act=True)
+        self.do33 = self._make_dropout()
+        self.pool3 = MaxPool2x2()
+
+        self.conv41 = Conv3x3(C3, C4, norm=True, act=True)
+        self.do41 = self._make_dropout()
+        self.conv42 = Conv3x3(C4, C4, norm=True, act=True)
+        self.do42 = self._make_dropout()
+        self.conv43 = Conv3x3(C4, C4, norm=True, act=True)
+        self.do43 = self._make_dropout()
+        self.pool4 = MaxPool2x2()
+
+        self.upconv4 = ConvTransposed3x3(C4, C4, output_padding=1)
+
+        self.conv43d = Conv3x3(C5, C4, norm=True, act=True)
+        self.do43d = self._make_dropout()
+        self.conv42d = Conv3x3(C4, C4, norm=True, act=True)
+        self.do42d = self._make_dropout()
+        self.conv41d = Conv3x3(C4, C3, norm=True, act=True)
+        self.do41d = self._make_dropout()
+
+        self.upconv3 = ConvTransposed3x3(C3, C3, output_padding=1)
+
+        self.conv33d = Conv3x3(C4, C3, norm=True, act=True)
+        self.do33d = self._make_dropout()
+        self.conv32d = Conv3x3(C3, C3, norm=True, act=True)
+        self.do32d = self._make_dropout()
+        self.conv31d = Conv3x3(C3, C2, norm=True, act=True)
+        self.do31d = self._make_dropout()
+
+        self.upconv2 = ConvTransposed3x3(C2, C2, output_padding=1)
+
+        self.conv22d = Conv3x3(C3, C2, norm=True, act=True)
+        self.do22d = self._make_dropout()
+        self.conv21d = Conv3x3(C2, C1, norm=True, act=True)
+        self.do21d = self._make_dropout()
+
+        self.upconv1 = ConvTransposed3x3(C1, C1, output_padding=1)
+
+        self.conv12d = Conv3x3(C2, C1, norm=True, act=True)
+        self.do12d = self._make_dropout()
+        self.conv11d = Conv3x3(C1, num_classes)
+
+        self.init_weight()
+
+    def forward(self, t1, t2):
+        x = paddle.concat([t1, t2], axis=1)
+
+        # Stage 1
+        x11 = self.do11(self.conv11(x))
+        x12 = self.do12(self.conv12(x11))
+        x1p = self.pool1(x12)
+
+        # Stage 2
+        x21 = self.do21(self.conv21(x1p))
+        x22 = self.do22(self.conv22(x21))
+        x2p = self.pool2(x22)
+
+        # Stage 3
+        x31 = self.do31(self.conv31(x2p))
+        x32 = self.do32(self.conv32(x31))
+        x33 = self.do33(self.conv33(x32))
+        x3p = self.pool3(x33)
+
+        # Stage 4
+        x41 = self.do41(self.conv41(x3p))
+        x42 = self.do42(self.conv42(x41))
+        x43 = self.do43(self.conv43(x42))
+        x4p = self.pool4(x43)
+
+        # Stage 4d
+        x4d = self.upconv4(x4p)
+        pad4 = (
+            0, 
+            paddle.shape(x43)[3]-paddle.shape(x4d)[3], 
+            0, 
+            paddle.shape(x43)[2]-paddle.shape(x4d)[2]
+        )
+        x4d = paddle.concat([F.pad(x4d, pad=pad4, mode='replicate'), x43], 1)
+        x43d = self.do43d(self.conv43d(x4d))
+        x42d = self.do42d(self.conv42d(x43d))
+        x41d = self.do41d(self.conv41d(x42d))
+
+        # Stage 3d
+        x3d = self.upconv3(x41d)
+        pad3 = (
+            0, 
+            paddle.shape(x33)[3]-paddle.shape(x3d)[3], 
+            0, 
+            paddle.shape(x33)[2]-paddle.shape(x3d)[2]
+        )
+        x3d = paddle.concat([F.pad(x3d, pad=pad3, mode='replicate'), x33], 1)
+        x33d = self.do33d(self.conv33d(x3d))
+        x32d = self.do32d(self.conv32d(x33d))
+        x31d = self.do31d(self.conv31d(x32d))
+
+        # Stage 2d
+        x2d = self.upconv2(x31d)
+        pad2 = (
+            0, 
+            paddle.shape(x22)[3]-paddle.shape(x2d)[3], 
+            0, 
+            paddle.shape(x22)[2]-paddle.shape(x2d)[2]
+        )
+        x2d = paddle.concat([F.pad(x2d, pad=pad2, mode='replicate'), x22], 1)
+        x22d = self.do22d(self.conv22d(x2d))
+        x21d = self.do21d(self.conv21d(x22d))
+
+        # Stage 1d
+        x1d = self.upconv1(x21d)
+        pad1 = (
+            0, 
+            paddle.shape(x12)[3]-paddle.shape(x1d)[3], 
+            0, 
+            paddle.shape(x12)[2]-paddle.shape(x1d)[2]
+        )
+        x1d = paddle.concat([F.pad(x1d, pad=pad1, mode='replicate'), x12], 1)
+        x12d = self.do12d(self.conv12d(x1d))
+        x11d = self.conv11d(x12d)
+
+        return x11d,
+
+    def init_weight(self):
+        for sublayer in self.sublayers():
+            if isinstance(sublayer, nn.Conv2D):
+                normal_init(sublayer.weight, std=0.001)
+            elif isinstance(sublayer, (nn.BatchNorm, nn.SyncBatchNorm)):
+                constant_init(sublayer.weight, value=1.0)
+                constant_init(sublayer.bias, value=0.0)
+
+    def _make_dropout(self):
+        if self.use_dropout:
+            return nn.Dropout2D(p=0.2)
+        else:
+            return Identity()

paddlers/tasks/changedetector.py

@@ -31,7 +31,7 @@ from paddlers.utils.checkpoint import seg_pretrain_weights_dict
 from paddlers.transforms import ImgDecoder, Resize
 import paddlers.models.cd as cd
 
-__all__ = ["CDNet"]
+__all__ = ["CDNet", "UNetEarlyFusion"]
 
 
 class BaseChangeDetector(BaseModel):
@@ -663,3 +663,21 @@ class CDNet(BaseChangeDetector):
             num_classes=num_classes,
             use_mixed_loss=use_mixed_loss,
             **params)
+
+
+class UNetEarlyFusion(BaseChangeDetector):
+    def __init__(self,
+                 num_classes=2,
+                 use_mixed_loss=False,
+                 in_channels=6,
+                 use_dropout=False,
+                 **params):
+        params.update({
+            'in_channels': in_channels,
+            'use_dropout': use_dropout
+        })
+        super(UNetEarlyFusion, self).__init__(
+            model_name='UNetEarlyFusion',
+            num_classes=num_classes,
+            use_mixed_loss=use_mixed_loss,
+            **params)