Setting up a UNet model (#46)

* Initial commit for setting up a UNet model

Co-authored-by: Maria Novitasari <mnovitasari@turing.ac.uk>
Co-authored-by: Louisa van Zeeland <louisa-ai2@protonmail.com>
Co-authored-by: James Robinson <james.em.robinson@gmail.com>

* Linting with pre-commit

* Remove redundant code lines

Co-authored-by: James Robinson <james.em.robinson@gmail.com>

* Remove redundant class

Co-authored-by: James Robinson <james.em.robinson@gmail.com>

* Remove unnecessary import

Co-authored-by: James Robinson <james.em.robinson@gmail.com>

* Remove print statement

Co-authored-by: James Robinson <james.em.robinson@gmail.com>

* Converting channels from dict to list

* Linting

* Removing redundant argument

Co-authored-by: James Robinson <james.em.robinson@gmail.com>

* 🏷️ Add type hints and shift defaults to yaml

* 🚚 Moving common code out of the unet file

* 🚨 Linting

* ♻️ Replace n_filters_factor / reduced channels with start_out_channels arg

* 💡 Adding comment on origin of the UNet model

* 👽 Use `nn.MaxPool2d` instead of `functional.max_pool2d`

---------

Co-authored-by: Maria Novitasari <mnovitasari@turing.ac.uk>
Co-authored-by: Louisa van Zeeland <louisa-ai2@protonmail.com>
Co-authored-by: James Robinson <james.em.robinson@gmail.com>

Author: 4 people

ice_station_zebra/config/model/encode_unet_decode.yaml

@@ -0,0 +1,19 @@
+_target_: ice_station_zebra.models.EncodeProcessDecode
+
+name: encode-unet-decode
+
+# Each dataset will be encoded into a latent space with these properties
+latent_space:
+  channels: 20
+  shape: [128, 128]
+
+encoder:
+  _target_: ice_station_zebra.models.encoders.NaiveLatentSpaceEncoder
+
+processor:
+  _target_: ice_station_zebra.models.processors.UNetProcessor
+  filter_size: 3 # Size of the kernel for convolutional layers
+  start_out_channels: 64 # Initial number of channels for the first convolutional layer
+
+decoder:
+  _target_: ice_station_zebra.models.decoders.NaiveLatentSpaceDecoder

ice_station_zebra/models/common/bottleneckblock.py

@@ -0,0 +1,28 @@
+import torch.nn as nn
+from torch import Tensor
+
+
+class BottleneckBlock(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        *,
+        filter_size: int,
+    ) -> None:
+        super().__init__()
+
+        self.model = nn.Sequential(
+            nn.Conv2d(
+                in_channels, out_channels, kernel_size=filter_size, padding="same"
+            ),
+            nn.ReLU(inplace=True),
+            nn.Conv2d(
+                out_channels, out_channels, kernel_size=filter_size, padding="same"
+            ),
+            nn.ReLU(inplace=True),
+            nn.BatchNorm2d(num_features=out_channels),
+        )
+
+    def forward(self, x: Tensor) -> Tensor:
+        return self.model(x)

ice_station_zebra/models/common/convblock.py

@@ -0,0 +1,45 @@
+import torch.nn as nn
+from torch import Tensor
+
+
+class ConvBlock(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        *,
+        filter_size: int,
+        final: bool = False,
+    ) -> None:
+        super().__init__()
+
+        layers = [
+            nn.Conv2d(
+                in_channels, out_channels, kernel_size=filter_size, padding="same"
+            ),
+            nn.ReLU(inplace=True),
+            nn.Conv2d(
+                out_channels, out_channels, kernel_size=filter_size, padding="same"
+            ),
+            nn.ReLU(inplace=True),
+        ]
+        if final:
+            layers += [
+                nn.Conv2d(
+                    out_channels,
+                    out_channels,
+                    kernel_size=filter_size,
+                    padding="same",
+                ),
+                nn.ReLU(inplace=True),
+            ]
+
+        else:
+            layers.append(
+                nn.BatchNorm2d(num_features=out_channels),
+            )
+
+        self.model = nn.Sequential(*layers)
+
+    def forward(self, x: Tensor) -> Tensor:
+        return self.model(x)

ice_station_zebra/models/common/upconvblock.py

@@ -0,0 +1,16 @@
+import torch.nn as nn
+from torch import Tensor
+
+
+class UpconvBlock(nn.Module):
+    def __init__(self, in_channels: int, out_channels: int) -> None:
+        super().__init__()
+
+        self.model = nn.Sequential(
+            nn.Upsample(scale_factor=2, mode="nearest"),
+            nn.Conv2d(in_channels, out_channels, kernel_size=2, padding="same"),
+            nn.ReLU(inplace=True),
+        )
+
+    def forward(self, x: Tensor) -> Tensor:
+        return self.model(x)

ice_station_zebra/models/processors/__init__.py

@@ -1,5 +1,7 @@
 from .null import NullProcessor
+from .unet import UNetProcessor
 
 __all__ = [
     "NullProcessor",
+    "UNetProcessor",
 ]

ice_station_zebra/models/processors/unet.py

@@ -0,0 +1,82 @@
+import torch.nn as nn
+from torch import Tensor
+import torch
+
+from ice_station_zebra.models.common.convblock import ConvBlock
+from ice_station_zebra.models.common.bottleneckblock import BottleneckBlock
+from ice_station_zebra.models.common.upconvblock import UpconvBlock
+
+
+class UNetProcessor(nn.Module):
+    """UNet model that processes input through a UNet architecture
+
+    Structure based on Andersson et al. (2021) Nature Communications
+    https://doi.org/10.1038/s41467-021-25257-4"""
+
+    def __init__(
+        self,
+        n_latent_channels: int,
+        filter_size: int,
+        start_out_channels: int,
+    ) -> None:
+        super().__init__()
+
+        channels = [start_out_channels * 2**pow for pow in range(4)]
+
+        # Encoder
+        self.conv1 = ConvBlock(n_latent_channels, channels[0], filter_size=filter_size)
+        self.maxpool1 = nn.MaxPool2d(kernel_size=2)
+        self.conv2 = ConvBlock(channels[0], channels[1], filter_size=filter_size)
+        self.maxpool2 = nn.MaxPool2d(kernel_size=2)
+        self.conv3 = ConvBlock(channels[1], channels[2], filter_size=filter_size)
+        self.maxpool3 = nn.MaxPool2d(kernel_size=2)
+        self.conv4 = ConvBlock(channels[2], channels[2], filter_size=filter_size)
+        self.maxpool4 = nn.MaxPool2d(kernel_size=2)
+
+        # Bottleneck
+        self.conv5 = BottleneckBlock(channels[2], channels[3], filter_size=filter_size)
+
+        # Decoder
+        self.up6 = UpconvBlock(channels[3], channels[2])
+        self.up7 = UpconvBlock(channels[2], channels[2])
+        self.up8 = UpconvBlock(channels[2], channels[1])
+        self.up9 = UpconvBlock(channels[1], channels[0])
+
+        self.up6b = ConvBlock(channels[3], channels[2], filter_size=filter_size)
+        self.up7b = ConvBlock(channels[3], channels[2], filter_size=filter_size)
+        self.up8b = ConvBlock(channels[2], channels[1], filter_size=filter_size)
+        self.up9b = ConvBlock(
+            channels[1], channels[0], filter_size=filter_size, final=True
+        )
+
+        # Final layer
+        self.final_layer = nn.Conv2d(
+            channels[0], n_latent_channels, kernel_size=1, padding="same"
+        )
+
+    def forward(self, x: Tensor) -> Tensor:
+        # Process in latent space: tensor with (batch_size, all_variables, latent_height, latent_width)
+
+        # Encoder
+        bn1 = self.conv1(x)
+        conv1 = self.maxpool1(bn1)
+        bn2 = self.conv2(conv1)
+        conv2 = self.maxpool1(bn2)
+        bn3 = self.conv3(conv2)
+        conv3 = self.maxpool3(bn3)
+        bn4 = self.conv4(conv3)
+        conv4 = self.maxpool4(bn4)
+
+        # Bottleneck
+        bn5 = self.conv5(conv4)
+
+        # Decoder
+        up6 = self.up6b(torch.cat([bn4, self.up6(bn5)], dim=1))
+        up7 = self.up7b(torch.cat([bn3, self.up7(up6)], dim=1))
+        up8 = self.up8b(torch.cat([bn2, self.up8(up7)], dim=1))
+        up9 = self.up9b(torch.cat([bn1, self.up9(up8)], dim=1))
+
+        # Final layer
+        output = self.final_layer(up9)
+
+        return output