Fix torch export issue

Summary:
X-link: fairinternal/detectron2#604

Pull Request resolved: #5334

Add a check to deal with dynamic shapes so that the model can be exported with `torch.export`. This check prevents the graph break caused by the SymInt by delaying the assertion to runtime.

Reviewed By: wat3rBro

Differential Revision: D60126415

fbshipit-source-id: a2a75530db523bfdde984b890595e02360d8e07f

detectron2/export/c10.py

@@ -84,6 +84,11 @@ def set(self, name, value):
         else:
             data_len = len(value)
         if len(self.batch_extra_fields):
+            # If we are tracing with Dynamo, the check here is needed since len(self)
+            # represents the number of bounding boxes detected in the image and thus is
+            # an unbounded SymInt.
+            if torch._utils.is_compiling():
+                torch._check(len(self) == data_len)
             assert (
                 len(self) == data_len
             ), "Adding a field of length {} to a Instances of length {}".format(data_len, len(self))