add unittest for onnx convert (#608)

* add unittest for onnx convert

* build onnx and onnxruntime in CI

* skip onnx op unit test while using CUDA

* fix offset==0 case in NMS

* remove tmp file used in test

* delete tmp file before assert so that we can remove the tmp file anyway

.github/workflows/build.yml

@@ -111,7 +111,7 @@ jobs:
       - name: Install PyTorch
         run: pip install torch==${{matrix.torch}}+cpu torchvision==${{matrix.torchvision}}+cpu -f https://download.pytorch.org/whl/torch_stable.html
       - name: Install unittest dependencies
-        run: pip install pytest coverage lmdb PyTurboJPEG
+        run: pip install pytest coverage lmdb PyTurboJPEG onnx==1.6.0 onnxruntime==1.2.0
       - name: Build and install
         run: rm -rf .eggs && pip install -e .
       - name: Run unittests and generate coverage report
@@ -181,7 +181,7 @@ jobs:
         run: rm -rf .eggs && pip install -e .
       - name: Run unittests and generate coverage report
         run: |
-          coverage run --branch --source=mmcv -m pytest tests/
+          coverage run --branch --source=mmcv -m pytest tests/ --ignore=tests/test_ops/test_onnx.py
           coverage xml
           coverage report -m
       # Only upload coverage report for python3.7 && pytorch1.5
@@ -220,6 +220,8 @@ jobs:
       - name: Install Pillow
         run: pip install Pillow==6.2.2
         if: ${{matrix.torchvision == '0.4.2'}}
+      - name: Install ONNX
+        run: pip install onnx==1.6.0 onnxruntime==1.2.0
       - name: Install PyTorch
         run: pip install torch==${{matrix.torch}} torchvision==${{matrix.torchvision}} --no-cache-dir
       - name: Build and install

mmcv/ops/nms.py

@@ -111,6 +111,9 @@ def nms(boxes, scores, iou_threshold, offset=0):
             # ONNX only support offset == 1
             boxes[:, -2:] -= 1
         inds = NMSop.apply(boxes, scores, iou_threshold, offset)
+        if torch.onnx.is_in_onnx_export() and offset == 0:
+            # ONNX only support offset == 1
+            boxes[:, -2:] += 1
     dets = torch.cat((boxes[inds], scores[inds].reshape(-1, 1)), dim=1)
     if is_numpy:
         dets = dets.cpu().numpy()

setup.cfg

@@ -14,6 +14,6 @@ line_length = 79
 multi_line_output = 0
 known_standard_library = pkg_resources,setuptools
 known_first_party = mmcv
-known_third_party = Cython,addict,cv2,m2r,numpy,pytest,recommonmark,resnet_cifar,torch,torchvision,yaml,yapf
+known_third_party = Cython,addict,cv2,m2r,numpy,onnx,onnxruntime,pytest,recommonmark,resnet_cifar,torch,torchvision,yaml,yapf
 no_lines_before = STDLIB,LOCALFOLDER
 default_section = THIRDPARTY

tests/test_ops/test_onnx.py

@@ -0,0 +1,62 @@
+import os
+from functools import partial
+
+import numpy as np
+import onnx
+import onnxruntime as rt
+import torch
+import torch.nn as nn
+
+onnx_file = 'tmp.onnx'
+
+
+class WrapFunction(nn.Module):
+
+    def __init__(self, wrapped_function):
+        super(WrapFunction, self).__init__()
+        self.wrapped_function = wrapped_function
+
+    def forward(self, *args, **kwargs):
+        return self.wrapped_function(*args, **kwargs)
+
+
+class Testonnx(object):
+
+    def test_nms(self):
+        from mmcv.ops import nms
+        np_boxes = np.array([[6.0, 3.0, 8.0, 7.0], [3.0, 6.0, 9.0, 11.0],
+                             [3.0, 7.0, 10.0, 12.0], [1.0, 4.0, 13.0, 7.0]],
+                            dtype=np.float32)
+        np_scores = np.array([0.6, 0.9, 0.7, 0.2], dtype=np.float32)
+        boxes = torch.from_numpy(np_boxes)
+        scores = torch.from_numpy(np_scores)
+        pytorch_dets, _ = nms(boxes, scores, iou_threshold=0.3, offset=0)
+        pytorch_score = pytorch_dets[:, 4]
+        nms = partial(nms, iou_threshold=0.3, offset=0)
+        wrapped_model = WrapFunction(nms)
+        wrapped_model.cpu().eval()
+        with torch.no_grad():
+            torch.onnx.export(
+                wrapped_model, (boxes, scores),
+                onnx_file,
+                export_params=True,
+                keep_initializers_as_inputs=True,
+                input_names=['boxes', 'scores'],
+                opset_version=11)
+        onnx_model = onnx.load(onnx_file)
+
+        # get onnx output
+        input_all = [node.name for node in onnx_model.graph.input]
+        input_initializer = [
+            node.name for node in onnx_model.graph.initializer
+        ]
+        net_feed_input = list(set(input_all) - set(input_initializer))
+        assert (len(net_feed_input) == 2)
+        sess = rt.InferenceSession(onnx_file)
+        onnx_dets, _ = sess.run(None, {
+            'scores': scores.detach().numpy(),
+            'boxes': boxes.detach().numpy()
+        })
+        onnx_score = onnx_dets[:, 4]
+        os.remove(onnx_file)
+        assert np.allclose(pytorch_score, onnx_score, atol=1e-3)