Attention unet (#1)

dacapo/experiments/architectures/cnnectome_unet.py

@@ -25,6 +25,7 @@ def __init__(self, architecture_config):
         self.upsample_factors = (
             self.upsample_factors if self.upsample_factors is not None else []
         )
+        self.use_attention = architecture_config.use_attention
 
         self.unet = self.module()
 
@@ -64,6 +65,7 @@ def module(self):
             activation_on_upsample=True,
             upsample_channel_contraction=[False]
             + [True] * (len(downsample_factors) - 1),
+            use_attention=self.use_attention,
         )
         if len(self.upsample_factors) > 0:
             layers = [unet]
@@ -125,6 +127,7 @@ def __init__(
         padding="valid",
         upsample_channel_contraction=False,
         activation_on_upsample=False,
+        use_attention=False,
     ):
         """Create a U-Net::
 
@@ -244,6 +247,7 @@ def __init__(
         )
 
         self.dims = len(downsample_factors[0])
+        self.use_attention = use_attention
 
         # default arguments
 
@@ -316,6 +320,29 @@ def __init__(
                 for _ in range(num_heads)
             ]
         )
+        #  if num_fmaps_out is None or level != self.num_levels-1 else num_fmaps_out
+        if self.use_attention:
+            self.attention = nn.ModuleList(
+                [
+                    nn.ModuleList(
+                        [
+                            AttentionBlockModule(
+                                F_g=num_fmaps * fmap_inc_factor ** (level + 1),
+                                F_l=num_fmaps * fmap_inc_factor**level,
+                                F_int=num_fmaps
+                                * fmap_inc_factor
+                                ** (level + (1 - upsample_channel_contraction[level]))
+                                if num_fmaps_out is None or level != 0
+                                else num_fmaps_out,
+                                dims=self.dims,
+                                upsample_factor=downsample_factors[level],
+                            )
+                            for level in range(self.num_levels - 1)
+                        ]
+                    )
+                    for _ in range(num_heads)
+                ]
+            )
 
         # right convolutional passes
         self.r_conv = nn.ModuleList(
@@ -359,10 +386,19 @@ def rec_forward(self, level, f_in):
             # nested levels
             gs_out = self.rec_forward(level - 1, g_in)
 
-            # up, concat, and crop
-            fs_right = [
-                self.r_up[h][i](gs_out[h], f_left) for h in range(self.num_heads)
-            ]
+            if self.use_attention:
+                f_left_attented = [
+                    self.attention[h][i](gs_out[h], f_left)
+                    for h in range(self.num_heads)
+                ]
+                fs_right = [
+                    self.r_up[h][i](gs_out[h], f_left_attented[h])
+                    for h in range(self.num_heads)
+                ]
+            else:  # up, concat, and crop
+                fs_right = [
+                    self.r_up[h][i](gs_out[h], f_left) for h in range(self.num_heads)
+                ]
 
             # convolve
             fs_out = [self.r_conv[h][i](fs_right[h]) for h in range(self.num_heads)]
@@ -580,3 +616,112 @@ def forward(self, g_out, f_left=None):
             return torch.cat([f_cropped, g_cropped], dim=1)
         else:
             return g_cropped
+
+
+class AttentionBlockModule(nn.Module):
+    def __init__(self, F_g, F_l, F_int, dims, upsample_factor=None):
+        """Attention Block Module::
+
+            The attention block takes two inputs: 'g' (gating signal) and 'x' (input features).
+
+                [g] --> W_g --\                 /--> psi --> * --> [output]
+                                \               /
+                [x] --> W_x --> [+] --> relu --
+
+        Where:
+        - W_g and W_x are 1x1 Convolution followed by Batch Normalization
+        - [+] indicates element-wise addition
+        - relu is the Rectified Linear Unit activation function
+        - psi is a sequence of 1x1 Convolution, Batch Normalization, and Sigmoid activation
+        - * indicates element-wise multiplication between the output of psi and input feature 'x'
+        - [output] has the same dimensions as input 'x', selectively emphasized by attention weights
+
+        Args:
+        F_g (int): The number of feature channels in the gating signal (g).
+                   This is the input channel dimension for the W_g convolutional layer.
+
+        F_l (int): The number of feature channels in the input features (x).
+                   This is the input channel dimension for the W_x convolutional layer.
+
+        F_int (int): The number of intermediate feature channels.
+                     This represents the output channel dimension of the W_g and W_x convolutional layers
+                     and the input channel dimension for the psi layer. Typically, F_int is smaller
+                     than F_g and F_l, as it serves to compress the feature representations before
+                     applying the attention mechanism.
+
+        The AttentionBlock uses two separate pathways to process 'g' and 'x', combines them,
+        and applies a sigmoid activation to generate an attention map. This map is then used
+        to scale the input features 'x', resulting in an output that focuses on important
+        features as dictated by the gating signal 'g'.
+
+        """
+
+        super(AttentionBlockModule, self).__init__()
+        self.dims = dims
+        self.kernel_sizes = [(1,) * self.dims, (1,) * self.dims]
+        if upsample_factor is not None:
+            self.upsample_factor = upsample_factor
+        else:
+            self.upsample_factor = (2,) * self.dims
+
+        self.W_g = ConvPass(
+            F_g, F_int, kernel_sizes=self.kernel_sizes, activation=None, padding="same"
+        )
+
+        self.W_x = nn.Sequential(
+            ConvPass(
+                F_l,
+                F_int,
+                kernel_sizes=self.kernel_sizes,
+                activation=None,
+                padding="same",
+            ),
+            Downsample(upsample_factor),
+        )
+
+        self.psi = ConvPass(
+            F_int,
+            1,
+            kernel_sizes=self.kernel_sizes,
+            activation="Sigmoid",
+            padding="same",
+        )
+
+        up_mode = {2: "bilinear", 3: "trilinear"}[self.dims]
+
+        self.up = nn.Upsample(
+            scale_factor=upsample_factor, mode=up_mode, align_corners=True
+        )
+
+        self.relu = nn.ReLU(inplace=True)
+
+    def calculate_and_apply_padding(self, smaller_tensor, larger_tensor):
+        """
+        Calculate and apply symmetric padding to the smaller tensor to match the dimensions of the larger tensor.
+
+        Args:
+        smaller_tensor (Tensor): The tensor to be padded.
+        larger_tensor (Tensor): The tensor whose dimensions the smaller tensor needs to match.
+
+        Returns:
+        Tensor: The padded smaller tensor with the same dimensions as the larger tensor.
+        """
+        padding = []
+        for i in range(2, 2 + self.dims):
+            diff = larger_tensor.size(i) - smaller_tensor.size(i)
+            padding.extend([diff // 2, diff - diff // 2])
+
+        # Reverse padding to match the 'pad' function's expectation
+        padding = padding[::-1]
+
+        # Apply symmetric padding
+        return nn.functional.pad(smaller_tensor, padding, mode="constant", value=0)
+
+    def forward(self, g, x):
+        g1 = self.W_g(g)
+        x1 = self.W_x(x)
+        g1 = self.calculate_and_apply_padding(g1, x1)
+        psi = self.relu(g1 + x1)
+        psi = self.psi(psi)
+        psi = self.up(psi)
+        return x * psi

dacapo/experiments/architectures/cnnectome_unet_config.py

@@ -82,3 +82,9 @@ class CNNectomeUNetConfig(ArchitectureConfig):
         default="valid",
         metadata={"help_text": "The padding to use in convolution operations."},
     )
+    use_attention: bool = attr.ib(
+        default=False,
+        metadata={
+            "help_text": "Whether to use attention blocks in the UNet. This is supported for 2D and  3D."
+        },
+    )