hacked up

torchfix/visitors/deprecated_symbols/__init__.py

@@ -15,6 +15,7 @@
 from .chain_matmul import call_replacement_chain_matmul
 from .cholesky import call_replacement_cholesky
 from .qr import call_replacement_qr
+from .size_average import call_replacement_loss
 
 from .range import call_replacement_range
 
@@ -54,6 +55,7 @@ def _call_replacement(
             "torch.qr": call_replacement_qr,
             "torch.cuda.amp.autocast": call_replacement_cuda_amp_autocast,
             "torch.cpu.amp.autocast": call_replacement_cpu_amp_autocast,
+            "torch.nn.functional.soft_margin_loss": call_replacement_loss
         }
         replacement = None
 
@@ -103,7 +105,8 @@ def visit_Call(self, node) -> None:
         qualified_name = self.get_qualified_name_for_call(node)
         if qualified_name is None:
             return
-
+        self.deprecated_config["torch.nn.functional.soft_margin_loss"] = {}
+        self.deprecated_config["torch.nn.functional.soft_margin_loss"]["remove_pr"] = None
         if qualified_name in self.deprecated_config:
             if self.deprecated_config[qualified_name]["remove_pr"] is None:
                 error_code = self.ERRORS[1].error_code
@@ -112,7 +115,6 @@ def visit_Call(self, node) -> None:
                 error_code = self.ERRORS[0].error_code
                 message = self.ERRORS[0].message(old_name=qualified_name)
             replacement = self._call_replacement(node, qualified_name)
-
             reference = self.deprecated_config[qualified_name].get("reference")
             if reference is not None:
                 message = f"{message}: {reference}"

torchfix/visitors/deprecated_symbols/size_average.py

@@ -0,0 +1,60 @@
+"""size_average and reduce are deprecated, please use reduction='mean' instead."""
+
+import libcst as cst
+from ...common import TorchVisitor, get_module_name
+from torch.nn._reduction import legacy_get_string
+
+def call_replacement_loss(node: cst.Call) -> cst.CSTNode:
+    """
+    Replace loss function that contains size_average / reduce with a new loss function 
+    that uses reduction='mean' instead. Uses the logic from torch.nn._reduction to 
+    determine the correct reduction value.
+
+    Args:
+        node: The CST Call node representing the loss function call
+
+    Returns:
+        A new CST node with updated reduction parameter
+    """
+    # Extract existing arguments
+    input_arg = TorchVisitor.get_specific_arg(node, "input", 0)
+    target_arg = TorchVisitor.get_specific_arg(node, "target", 1)
+
+    size_average_arg = TorchVisitor.get_specific_arg(node, "size_average", 2)
+    reduce_arg = TorchVisitor.get_specific_arg(node, "reduce", 3)
+
+    # Ensure input and target args maintain their commas
+    input_arg = cst.ensure_type(input_arg, cst.Arg).with_changes(
+        comma=cst.MaybeSentinel.DEFAULT
+    )
+
+    target_arg = cst.ensure_type(target_arg, cst.Arg).with_changes(
+        comma=cst.MaybeSentinel.DEFAULT
+    )
+
+    # Extract size_average and reduce values
+    size_average_value = None
+    reduce_value = None
+
+    if size_average_arg:
+        size_average_value = getattr(size_average_arg.value, "value", True)
+    if reduce_arg:
+        reduce_value = getattr(reduce_arg.value, "value", True)
+
+    if size_average_value is None and reduce_value is None:
+        # We want to return the original call as is
+        return node
+    # Use legacy_get_string to determine the correct reduction value
+    reduction = legacy_get_string(size_average_value, reduce_value, emit_warning=False)
+
+    # Create new reduction argument
+    reduction_arg = cst.Arg(
+        value=cst.SimpleString(f"'{reduction}'"),
+        keyword=cst.Name("reduction"),
+        comma=cst.MaybeSentinel.DEFAULT,
+    )
+
+    # Build new arguments list
+    new_args = [input_arg, target_arg, reduction_arg]
+    replacement = node.with_changes(args=new_args)
+    return replacement