Remove unnecessary constant outputs from ONNX exported graph

`TracingAdapter` creates extra outputs (through `flatten_to_tuple`) to hold
metadata information to rebuild the original data format during
deserialization.

When exporting a PyTorch model to ONNX, the support to de-serialize the
output to the original formatThis is unnecessary during ONNX export as the original data will never
be reconstructed to its original format using Schema.__call__ API.
This PR suppresses such extra output constants during
torch.onnx.export() execution. Outside this API, the behavior is not
changed, ensuring BC.

Although not stricly necessary to achieve the same numerical results as
PyTorch, when a ONNX model schema is compared to PyTorch's, the diffrent
number of outputs (ONNX model will have more outputs than PyTorch) may
not only confuse users, but also result in false negative when coding
model comparison helpers.

detectron2/export/c10.py

@@ -74,7 +74,13 @@ def has(self, name):
         return name in self.batch_extra_fields
 
     def set(self, name, value):
-        data_len = len(value)
+        # len(tensor) leads to constants during tracing mode
+        if isinstance(value, Boxes):
+            data_len = value.tensor.shape[0]
+        elif isinstance(value, torch.Tensor):
+            data_len = value.shape[0]
+        else:
+            data_len = len(value)
         if len(self.batch_extra_fields):
             assert (
                 len(self) == data_len

detectron2/export/caffe2_export.py

@@ -63,11 +63,6 @@ def _check_eval(module):
             )
             onnx_model = onnx.load_from_string(f.getvalue())
 
-    # Apply ONNX's Optimization
-    all_passes = onnx.optimizer.get_available_passes()
-    passes = ["fuse_bn_into_conv"]
-    assert all(p in all_passes for p in passes)
-    onnx_model = onnx.optimizer.optimize(onnx_model, passes)
     return onnx_model
 
 

docs/tutorials/deployment.md

@@ -90,7 +90,7 @@ and then export the model into Caffe2, TorchScript or ONNX format.
 The converted model is able to run in either Python or C++ without detectron2/torchvision dependency, on CPU or GPUs.
 It has a runtime optimized for CPU & mobile inference, but not optimized for GPU inference.
 
-This feature requires 1.9 > ONNX ≥ 1.6.
+This feature requires ONNX ≥ 1.6.
 
 ### Coverage
 

tests/export/test_c10.py

@@ -1,9 +1,16 @@
 # Copyright (c) Facebook, Inc. and its affiliates.
 import unittest
 
-from detectron2.config import get_cfg
-from detectron2.export.c10 import Caffe2RPN
-from detectron2.layers import ShapeSpec
+try:
+    # Caffe2 used to be included in PyTorch, but since PyTorch 1.10+,
+    # it is not included in pre-built packages. This is a safety BC check
+    from detectron2.config import get_cfg
+    from detectron2.export.c10 import Caffe2RPN
+    from detectron2.layers import ShapeSpec
+except ImportError:
+    raise unittest.SkipTest(
+        f"PyTorch does not have Caffe2 support. Skipping all tests in {__name__}"
+    )
 
 
 class TestCaffe2RPN(unittest.TestCase):

tests/test_export_caffe2.py

@@ -6,16 +6,26 @@
 import tempfile
 import unittest
 import torch
+from torch.hub import _check_module_exists
 
 from detectron2 import model_zoo
-from detectron2.export import Caffe2Model, Caffe2Tracer
 from detectron2.utils.logger import setup_logger
 from detectron2.utils.testing import get_sample_coco_image
 
+try:
+    # Caffe2 used to be included in PyTorch, but since PyTorch 1.10+,
+    # Caffe2 is not included in pre-built packages. This is a safety BC check
+    from detectron2.export import Caffe2Model, Caffe2Tracer
+except ImportError:
+    raise unittest.SkipTest(
+        f"PyTorch does not have Caffe2 support. Skipping all tests in {__name__}"
+    )
+
 
 # TODO: this test requires manifold access, see: T88318502
 # Running it on CircleCI causes crash, not sure why.
 @unittest.skipIf(os.environ.get("CIRCLECI"), "Caffe2 tests crash on CircleCI.")
+@unittest.skipIf(not _check_module_exists("onnx"), "ONNX not installed.")
 class TestCaffe2Export(unittest.TestCase):
     def setUp(self):
         setup_logger()