write to_instances as separate functions instead

Summary: jit.unused methods cannot be used on scripted classes after pytorch/pytorch#44324

Reviewed By: theschnitz

Differential Revision: D26273064

fbshipit-source-id: 6858a07cf93c373bd62c86869f7ef9c7a88869df

Author: ppwwyyxx

detectron2/export/torchscript_patch.py

@@ -27,22 +27,10 @@ def _clear_jit_cache():
 
 def _add_instances_conversion_methods(newInstances):
     """
-    Add to_instances/from_instances methods to the scripted Instances class.
+    Add from_instances methods to the scripted Instances class.
     """
     cls_name = newInstances.__name__
 
-    @torch.jit.unused
-    def to_instances(self):
-        """
-        Convert scripted Instances to original Instances
-        """
-        ret = Instances(self.image_size)
-        for name in self._field_names:
-            val = getattr(self, "_" + name, None)
-            if val is not None:
-                ret.set(name, val)
-        return ret
-
     @torch.jit.unused
     def from_instances(instances: Instances):
         """
@@ -56,7 +44,6 @@ def from_instances(instances: Instances):
             setattr(ret, name, deepcopy(val))
         return ret
 
-    newInstances.to_instances = to_instances
     newInstances.from_instances = from_instances
 
 

detectron2/utils/testing.py

@@ -7,7 +7,7 @@
 from detectron2.data import DatasetCatalog
 from detectron2.data.detection_utils import read_image
 from detectron2.modeling import build_model
-from detectron2.structures import Boxes
+from detectron2.structures import Boxes, Instances
 from detectron2.utils.file_io import PathManager
 
 
@@ -50,7 +50,7 @@ def get_sample_coco_image(tensor=True):
         an image, in BGR color.
     """
     try:
-        file_name = DatasetCatalog.get("coco_2017_train")[0]["file_name"]
+        file_name = DatasetCatalog.get("coco_2017_val_100")[0]["file_name"]
         if not PathManager.exists(file_name):
             raise FileNotFoundError()
     except IOError:
@@ -62,13 +62,30 @@ def get_sample_coco_image(tensor=True):
     return ret
 
 
+def convert_scripted_instances(instances):
+    """
+    Convert a scripted Instances object to a regular :class:`Instances` object
+    """
+    ret = Instances(instances.image_size)
+    for name in instances._field_names:
+        val = getattr(instances, "_" + name, None)
+        if val is not None:
+            ret.set(name, val)
+    return ret
+
+
 def assert_instances_allclose(input, other, *, rtol=1e-5, msg="", size_as_tensor=False):
     """
     Args:
         input, other (Instances):
         size_as_tensor: compare image_size of the Instances as tensors (instead of tuples).
              Useful for comparing outputs of tracing.
     """
+    if not isinstance(input, Instances):
+        input = convert_scripted_instances(input)
+    if not isinstance(other, Instances):
+        other = convert_scripted_instances(other)
+
     if not msg:
         msg = "Two Instances are different! "
     else:

tests/modeling/test_roi_heads.py

@@ -190,7 +190,7 @@ def test_mask_head_scriptability(self):
             script_outputs = sciript_mask_head(mask_features, [pred_instance0, pred_instance1])
 
         for origin_ins, script_ins in zip(origin_outputs, script_outputs):
-            assert_instances_allclose(origin_ins, script_ins.to_instances(), rtol=0)
+            assert_instances_allclose(origin_ins, script_ins, rtol=0)
 
     @unittest.skipIf(TORCH_VERSION < (1, 8), "Insufficient pytorch version")
     def test_keypoint_head_scriptability(self):
@@ -226,7 +226,7 @@ def test_keypoint_head_scriptability(self):
             )
 
         for origin_ins, script_ins in zip(origin_outputs, script_outputs):
-            assert_instances_allclose(origin_ins, script_ins.to_instances(), rtol=0)
+            assert_instances_allclose(origin_ins, script_ins, rtol=0)
 
     @unittest.skipIf(TORCH_VERSION < (1, 8), "Insufficient pytorch version")
     def test_StandardROIHeads_scriptability(self):
@@ -278,7 +278,7 @@ def test_StandardROIHeads_scriptability(self):
             scripted_pred_instances, _ = scripted_rot_heads(images, features, proposals)
 
         for instance, scripted_instance in zip(pred_instances, scripted_pred_instances):
-            assert_instances_allclose(instance, scripted_instance.to_instances(), rtol=0)
+            assert_instances_allclose(instance, scripted_instance, rtol=0)
 
     @unittest.skipIf(TORCH_VERSION < (1, 8), "Insufficient pytorch version")
     def test_PointRend_mask_head_tracing(self):

tests/structures/test_instances.py

@@ -6,6 +6,7 @@
 from detectron2.export.torchscript import patch_instances
 from detectron2.structures import Boxes, Instances
 from detectron2.utils.env import TORCH_VERSION
+from detectron2.utils.testing import convert_scripted_instances
 
 
 class TestInstances(unittest.TestCase):
@@ -182,7 +183,7 @@ def test_from_to_instances(self):
         with patch_instances(fields) as NewInstances:
             # convert to NewInstances and back
             new1 = NewInstances.from_instances(orig)
-            new2 = new1.to_instances()
+            new2 = convert_scripted_instances(new1)
         self.assertTrue(torch.equal(orig.proposal_boxes.tensor, new1.proposal_boxes.tensor))
         self.assertTrue(torch.equal(orig.proposal_boxes.tensor, new2.proposal_boxes.tensor))
 

tests/test_export_torchscript.py

@@ -17,7 +17,12 @@
 from detectron2.modeling.roi_heads import KRCNNConvDeconvUpsampleHead
 from detectron2.structures import Boxes, Instances
 from detectron2.utils.env import TORCH_VERSION
-from detectron2.utils.testing import assert_instances_allclose, get_sample_coco_image, random_boxes
+from detectron2.utils.testing import (
+    assert_instances_allclose,
+    convert_scripted_instances,
+    get_sample_coco_image,
+    random_boxes,
+)
 
 
 """
@@ -53,9 +58,7 @@ def _test_rcnn_model(self, config_path):
         inputs = [{"image": get_sample_coco_image()}]
         with torch.no_grad():
             instance = model.inference(inputs, do_postprocess=False)[0]
-            scripted_instance = script_model.inference(inputs, do_postprocess=False)[
-                0
-            ].to_instances()
+            scripted_instance = script_model.inference(inputs, do_postprocess=False)[0]
         assert_instances_allclose(instance, scripted_instance)
 
     def _test_retinanet_model(self, config_path):
@@ -73,7 +76,7 @@ def _test_retinanet_model(self, config_path):
         inputs = [{"image": img}]
         with torch.no_grad():
             instance = model(inputs)[0]["instances"]
-            scripted_instance = script_model(inputs)[0].to_instances()
+            scripted_instance = convert_scripted_instances(script_model(inputs)[0])
             scripted_instance = detector_postprocess(scripted_instance, img.shape[1], img.shape[2])
         assert_instances_allclose(instance, scripted_instance)
         # Note that the model currently cannot be saved and loaded into a new process: