enable (fake) inference for bolt exported model

Summary:
Enable the inference for boltnn (via running torchscript).
- merge rcnn's boltnn test with other export types.
- misc fixes.

Differential Revision: D30610386

fbshipit-source-id: 7b78136f8ca640b5fc179cb47e3218e709418d71

d2go/export/torchscript.py

@@ -23,6 +23,8 @@ from torch.utils.mobile_optimizer import optimize_for_mobile
 
 logger = logging.getLogger(__name__)
 
+TORCHSCRIPT_FILENAME_KEY: str = "torchscript_filename"
+
 
 class MobileOptimizationConfig(NamedTuple):
     # optimize_for_mobile
@@ -191,7 +193,7 @@ class DefaultTorchscriptExport(ModelExportMethod):
         torchscript_filename = trace_and_save_torchscript(
             model, input_args, save_path, **export_kwargs
         )
-        return {"torchscript_filename": torchscript_filename}
+        return {TORCHSCRIPT_FILENAME_KEY: torchscript_filename}
 
     @classmethod
     def load(cls, save_path, *, torchscript_filename="model.jit"):

d2go/utils/testing/rcnn_helper.py

@@ -310,21 +310,17 @@ class RCNNBaseTestCases:
                 )
 
                 predictor = create_predictor(predictor_path)
-                # This check is needed for models with unsupported backends.
-                # For these, predictor.model_or_models will be None -- as the
-                # model can't be loaded or run. (e.g. BoltNN).
-                if predictor.model_or_models is not None:
-                    predictor_outputs = predictor(inputs)
-                    _validate_outputs(inputs, predictor_outputs)
-
-                    if compare_match:
-                        with torch.no_grad():
-                            pytorch_outputs = self.test_model(inputs)
-
-                        assert_instances_allclose(
-                            predictor_outputs[0]["instances"],
-                            pytorch_outputs[0]["instances"],
-                        )
+                predictor_outputs = predictor(inputs)
+                _validate_outputs(inputs, predictor_outputs)
+
+                if compare_match:
+                    with torch.no_grad():
+                        pytorch_outputs = self.test_model(inputs)
+
+                    assert_instances_allclose(
+                        predictor_outputs[0]["instances"],
+                        pytorch_outputs[0]["instances"],
+                    )
 
             return predictor_path
 

tests/modeling/test_meta_arch_rcnn.py

@@ -19,7 +19,7 @@ from mobile_cv.common.misc.file_utils import make_temp_directory
 patch_d2_meta_arch()
 
 
-class TestFBNetV3MaskRCNNNormal(RCNNBaseTestCases.TemplateTestCase):
+class TestFBNetV3MaskRCNNFP32(RCNNBaseTestCases.TemplateTestCase):
     def setup_custom_test(self):
         super().setup_custom_test()
         self.cfg.merge_from_file("detectron2go://mask_rcnn_fbnetv3a_dsmask_C4.yaml")
@@ -66,7 +66,7 @@ class TestFBNetV3MaskRCNNQATEager(RCNNBaseTestCases.TemplateTestCase):
         self._test_export(predictor_type, compare_match=compare_match)
 
 
-class TestFBNetV3KeypointRCNNNormal(RCNNBaseTestCases.TemplateTestCase):
+class TestFBNetV3KeypointRCNNFP32(RCNNBaseTestCases.TemplateTestCase):
     def setup_custom_test(self):
         super().setup_custom_test()
         self.cfg.merge_from_file("detectron2go://keypoint_rcnn_fbnetv3a_dsmask_C4.yaml")