Modified Dtype of Labels to support cross-entropy on cpu

gcervantes8 · Dec 20, 2023 · 76ade85 · 76ade85
1 parent 9a51b12
commit 76ade85
Showing 1 changed file with 8 additions and 8 deletions.
diff --git a/src/models/gan_model.py b/src/models/gan_model.py
@@ -75,21 +75,21 @@ def create_real_labels(self, b_size, labels):
             # Real_label is of size [b_size, num_classes]
             real_label = torch.nn.functional.one_hot(labels, num_classes=self.num_classes)
             # Adds 2 columns, One with all values 1, the second with all values 0. Size [b_size, num_classes + 2]
-            real_label = torch.concat((real_label, torch.ones(b_size, 1, device=self.device, dtype=labels.dtype),
-                                    torch.zeros(b_size, 1, device=self.device, dtype=labels.dtype)), dim=1)
+            real_label = torch.concat((real_label, torch.ones(b_size, 1, device=self.device, dtype=self.torch_dtype),
+                                    torch.zeros(b_size, 1, device=self.device, dtype=self.torch_dtype)), dim=1)
             # Replace the 0s and 1s with value of the fake_label, or true_label
             real_label[real_label == 0] = self.fake_label
             real_label[real_label == 1] = self.real_label
         else:
-            real_label = torch.full((b_size,), self.real_label, dtype=labels.dtype, device=self.device)
+            real_label = torch.full((b_size,), self.real_label, dtype=self.torch_dtype, device=self.device)
         return real_label
 
-    def create_fake_labels(self, b_size, labels):
+    def create_fake_labels(self, b_size):
         if self.is_omni_loss:
-            fake_label = torch.full((b_size, self.num_classes + 1), self.fake_label, dtype=labels.dtype, device=self.device)
-            fake_label = torch.concat((fake_label, torch.ones(b_size, 1, device=self.device, dtype=labels.dtype)), dim=1)
+            fake_label = torch.full((b_size, self.num_classes + 1), self.fake_label, dtype=self.torch_dtype, device=self.device)
+            fake_label = torch.concat((fake_label, torch.ones(b_size, 1, device=self.device, dtype=self.torch_dtypee)), dim=1)
         else:
-            fake_label = torch.full((b_size,), self.fake_label, dtype=labels.dtype, device=self.device)
+            fake_label = torch.full((b_size,), self.fake_label, dtype=self.torch_dtype, device=self.device)
         return fake_label
 
     def train_step(self, batches_accumulated):
@@ -133,7 +133,7 @@ def discriminator_step(self, batches_accumulated):
                 fake_output = self.netD(fake.detach(), labels)
 
                 # Calculate D's loss on the all-fake batch
-                fake_label = self.create_fake_labels(b_size, labels)
+                fake_label = self.create_fake_labels(b_size)
                 discrim_on_fake_error = self.criterion(fake_output, fake_label.reshape_as(fake_output))
                 discrim_on_fake_error = discrim_on_fake_error / self.accumulation_iterations
                 self.accelerator.backward(discrim_on_fake_error)