Drop LBFGS

descreen/training/proc.py

@@ -32,7 +32,6 @@ def train[
 
     optimizer = torch.optim.RAdam(model.parameters(), lr=0.001)
     # optimizer = torch.optim.SGD(model.parameters(), lr=0.002)
-    # optimizer = torch.optim.LBFGS(model.parameters(), lr=0.1)
 
     input_size = model.input_size(patch_size)
     padding = model.reduced_padding(input_size)
@@ -81,28 +80,13 @@ def interrupt(signum, frame):
                 test_step()
 
     def train_step(x, y):
-        if False:
-            loss = None
-
-            def clos():
-                global loss
-                pred = model(x)
-                loss = loss_fn(pred, y)
-                print(f"loss: {loss}")
-                # Backpropagation
-                optimizer.zero_grad()
-                loss.backward()
-                return loss
-
-            optimizer.step(clos)
-        else:
-            pred = model(x)
-            loss = descreen_loss(pred, y, tv=0.01)
-            optimizer.zero_grad()
-            loss.backward()
-            print(f"loss: {loss}")
-            optimizer.step()
-            ema_model.update_parameters(model)
+        pred = model(x)
+        loss = descreen_loss(pred, y, tv=0.01)
+        optimizer.zero_grad()
+        loss.backward()
+        print(f"loss: {loss}")
+        optimizer.step()
+        ema_model.update_parameters(model)
 
     def valid_step(dataloader):
         # Set the model to evaluation mode - important for batch normalization and dropout layers