fix tensor_to_int in avg_pool2d

frontend/guard_tracker.py

@@ -267,6 +267,18 @@ def record_function(self,
             func = torch._C._set_grad_enabled
             kwargs = {}
         pargs, pkwargs = self.as_node_args_kwargs(args, kwargs)
+        if func == torch.nn.functional.avg_pool2d:
+            # avg_pool2d only supports integer or tuple(with two int values) as inputs
+            if isinstance(pargs[1], tuple):
+                for i in pargs[1]:
+                    if isinstance(i, torch.fx.Node):
+                        raise ValueError("cannot convert tensor in avg_pool2d")
+            if isinstance(args[1], tuple):
+                for i in args[1]:
+                    if torch.is_tensor(i):
+                        raise ValueError("cannot convert tensor in avg_pool2d")
+            elif torch.is_tensor(args[1]):
+                raise ValueError("cannot convert tensor in avg_pool2d")
         if func in fx_graph_inplace_functions:
             scalar = None
             node = None