Introduce feature flag for Unet

face_swapper/README.md

@@ -47,6 +47,7 @@ motion_extractor_path = .models/motion_extractor.pt
 
 ```
 [training.model.generator]
+encoder_type = unet-pro
 num_blocks = 2
 id_channels = 512
 ```

face_swapper/src/models/generator.py

@@ -3,7 +3,7 @@ import configparser
 from torch import Tensor, nn
 
 from ..networks.attribute_modulator import AADGenerator
-from ..networks.unet import UNet
+from ..networks.unet import UNet, UNetPro
 from ..types import Attributes, Embedding
 
 CONFIG = configparser.ConfigParser()
@@ -13,21 +13,25 @@ CONFIG.read('config.ini')
 class Generator(nn.Module):
 	def __init__(self) -> None:
 		super(Generator, self).__init__()
+		encoder_type = CONFIG.getint('training.model.generator', 'encoder_type')
 		id_channels = CONFIG.getint('training.model.generator', 'id_channels')
 		num_blocks = CONFIG.getint('training.model.generator', 'num_blocks')
 
-		self.unet = UNet()
-		self.aad_generator = AADGenerator(id_channels, num_blocks)
-		self.unet.apply(init_weight)
-		self.aad_generator.apply(init_weight)
+		if encoder_type == 'unet':
+			self.encoder = UNet()
+		if encoder_type == 'unet-pro':
+			self.encoder = UNetPro()
+		self.generator = AADGenerator(id_channels, num_blocks)
+		self.encoder.apply(init_weight)
+		self.generator.apply(init_weight)
 
 	def forward(self, source_embedding : Embedding, target_tensor : Tensor) -> Tensor:
 		target_attributes = self.get_attributes(target_tensor)
-		output_tensor = self.aad_generator(target_attributes, source_embedding)
+		output_tensor = self.generator(target_attributes, source_embedding)
 		return output_tensor
 
 	def get_attributes(self, input_tensor : Tensor) -> Attributes:
-		return self.unet(input_tensor)
+		return self.encoder(input_tensor)
 
 
 def init_weight(module : nn.Module) -> None:

face_swapper/src/networks/unet.py

@@ -8,7 +8,6 @@ from torchvision import models
 class UNet(nn.Module):
 	def __init__(self) -> None:
 		super(UNet, self).__init__()
-		self.resnet = models.resnet34(pretrained = True)
 		self.down_samples = self.create_down_samples(self)
 		self.up_samples = self.create_up_samples()
 
@@ -16,18 +15,11 @@ class UNet(nn.Module):
 	def create_down_samples(self : nn.Module) -> nn.ModuleList:
 		return nn.ModuleList(
 		[
-			nn.Sequential(
-				self.resnet.conv1,
-				self.resnet.bn1,
-				self.resnet.relu,
-				nn.Conv2d(64, 32, kernel_size = 1, bias = False),
-				nn.BatchNorm2d(32),
-				nn.LeakyReLU(0.1, inplace = True)
-			),
+			DownSample(3, 32),
 			DownSample(32, 64),
-			self.resnet.layer2,
-			self.resnet.layer3,
-			self.resnet.layer4,
+			DownSample(64, 128),
+			DownSample(128, 256),
+			DownSample(256, 512),
 			DownSample(512, 1024),
 			DownSample(1024, 1024)
 		])
@@ -65,6 +57,34 @@ class UNet(nn.Module):
 		return bottleneck_tensor, *up_features, output_tensor
 
 
+class UNetPro(UNet):
+	def __init__(self) -> None:
+		super(UNetPro, self).__init__()
+		self.resnet = models.resnet34(pretrained = True)
+		self.down_samples = self.create_down_samples(self)
+		self.up_samples = self.create_up_samples()
+
+	@staticmethod
+	def create_down_samples(self : nn.Module) -> nn.ModuleList:
+		return nn.ModuleList(
+		[
+			nn.Sequential(
+				self.resnet.conv1,
+				self.resnet.bn1,
+				self.resnet.relu,
+				nn.Conv2d(64, 32, kernel_size = 1, bias = False),
+				nn.BatchNorm2d(32),
+				nn.LeakyReLU(0.1, inplace = True)
+			),
+			DownSample(32, 64),
+			self.resnet.layer2,
+			self.resnet.layer3,
+			self.resnet.layer4,
+			DownSample(512, 1024),
+			DownSample(1024, 1024)
+		])
+
+
 class UpSample(nn.Module):
 	def __init__(self, input_channels : int, output_channels : int) -> None:
 		super(UpSample, self).__init__()