UNet's complete architecture

ivy_models/UNET/UNET.py

@@ -1,7 +1,5 @@
 import ivy
 
-ivy.set_backend("tensorflow")
-
 
 def double_conv(in_c, out_c):
     conv = ivy.Sequential(
@@ -12,21 +10,60 @@ def double_conv(in_c, out_c):
     )
     return conv
 
+def crop_img(tensor, target_tensor):
+    target_size = target_tensor.shape[2]
+    tensor_size = tensor.shape[2]
+    delta = tensor_size - target_size
+    delta  = delta // 2
+    return tensor[:, delta:tensor_size-delta, delta:tensor_size-delta, :]
 
 class UNet(ivy.Module):
     def __init__(self):
         super(UNet, self).__init__()
-
         self.pool = ivy.MaxPool2D(2, 2, 0)
         self.down_conv_1 = double_conv(1, 64)
         self.down_conv_2 = double_conv(64, 128)
         self.down_conv_3 = double_conv(128, 256)
         self.down_conv_4 = double_conv(256, 512)
         self.down_conv_5 = double_conv(512, 1024)
+        self.up_trans_1 = ivy.Conv2DTranspose(
+            1024,
+            512,
+            [2, 2],
+            2,
+            "VALID"
+        )
+        self.up_conv_1 = double_conv(1024, 512)
+        self.up_trans_2 = ivy.Conv2DTranspose(
+            512,
+            256,
+            [2, 2],
+            2,
+            "VALID"
+        )
+        self.up_conv_2 = double_conv(512, 256)
+        self.up_trans_3 = ivy.Conv2DTranspose(
+            256,
+            128,
+            [2, 2],
+            2,
+            "VALID"
+        )
+        self.up_conv_3 = double_conv(256, 128)
+        self.up_trans_4 = ivy.Conv2DTranspose(
+            128,
+            64,
+            [2, 2],
+            2,
+            "VALID"
+        )
+        self.up_conv_4 = double_conv(128, 64)
+        self.out = ivy.Conv2D(64, 2, [1, 1], 1, 0)
 
     def _forward(self, image):
+        # B, H, W, C
+        #encoder
         x1 = self.down_conv_1(image)
-        print(x1.shape)
         x2 = self.pool(x1)
         x3 = self.down_conv_2(x2)
         x4 = self.pool(x3)
@@ -35,9 +72,23 @@ def _forward(self, image):
         x7 = self.down_conv_4(x6)
         x8 = self.pool(x7)
         x9 = self.down_conv_5(x8)
-        print(x9.shape)
 
+        #decoder
+        x = self.up_trans_1(x9)
+        y = crop_img(x7, x)
+        x = self.up_conv_1(ivy.concat([x, y], axis=-1))
+        x = self.up_trans_2(x)
+        y = crop_img(x5, x)
+        x = self.up_conv_2(ivy.concat([x, y], axis=-1))
+        x = self.up_trans_3(x)
+        y = crop_img(x3, x)
+        x = self.up_conv_3(ivy.concat([x, y], axis=-1))
+        x = self.up_trans_4(x)
+        y = crop_img(x1, x)
+        x = self.up_conv_4(ivy.concat([x, y], axis=-1))
+        x = self.out(x)
+        return x
 
 image = ivy.random_normal(shape=(1, 572, 572, 1))
 model = UNet()
-print(model(image))
+model(image)