[v1.4.0] Minimal changes in interpolate to support Keypointrcnn (pytorch#32010)

* Fix interpolate

* add keypointrcnn test

* update ort versio for test

* pin tv version

* Update test.sh

* Get rid of onnxruntime test changes.

* [v1.4.0] Added torchvision tests as part of ORT tests (pytorch#31835)

Summary:
Added torchvision tests as part of ORT tests
Pull Request resolved: pytorch#31835

Reviewed By: hl475

Differential Revision: D19278607

Pulled By: houseroad

fbshipit-source-id: 18a6a85ce3019bcc9aee9517af1378964b585afd

* Remove faster_rcnn and mask_rcnn tests.

Co-authored-by: Lara Haidar <[email protected]>
Co-authored-by: Negin Raoof <[email protected]>

Author: 3 people

.jenkins/caffe2/test.sh

@@ -133,15 +133,15 @@ pip install --user pytest-sugar
 # torchvision tests #
 #####################
 if [[ "$BUILD_ENVIRONMENT" == *onnx* ]]; then
-  pip install -q --user git+https://github.com/pytorch/vision.git
+  pip install -q --user git+https://github.com/pytorch/vision.git@v0.5.0
   pip install -q --user ninja
   # JIT C++ extensions require ninja, so put it into PATH.
   export PATH="/var/lib/jenkins/.local/bin:$PATH"
   if [[ "$BUILD_ENVIRONMENT" == *py3* ]]; then
     # default pip version is too old(9.0.2), unable to support tag `manylinux2010`.
     # Fix the pip error: Couldn't find a version that satisfies the requirement
     sudo pip install --upgrade pip
-    pip install -q --user -i https://test.pypi.org/simple/ ort-nightly==1.0.0.dev1104
+    pip install -q --user -i https://test.pypi.org/simple/ ort-nightly==1.1.0.dev1228
   fi
   "$ROOT_DIR/scripts/onnx/test.sh"
 fi

test/onnx/test_pytorch_onnx_onnxruntime.py

@@ -183,31 +183,6 @@ def test_deeplab(self):
         x = torch.randn(2, 3, 224, 224, requires_grad=True)
         self.run_test(model, (x,), rtol=1e-3, atol=1e-5)
 
-    def test_googlenet_quantization(self):
-        model = torchvision.models.quantization.googlenet(pretrained=True)
-        x = torch.randn(2, 3, 224, 224, requires_grad=True)
-        self.run_test(model, (x,), rtol=1e-3, atol=1e-5)
-
-    def test_inception_quantization(self):
-        model = torchvision.models.quantization.inception_v3(pretrained=True)
-        x = torch.randn(2, 3, 224, 224, requires_grad=True)
-        self.run_test(model, (x,), rtol=1e-3, atol=1e-5)
-
-    def test_mobilenet_quantization(self):
-        model = torchvision.models.quantization.mobilenet_v2(pretrained=True)
-        x = torch.randn(2, 3, 224, 224, requires_grad=True)
-        self.run_test(model, (x,), rtol=1e-3, atol=1e-5)
-
-    def test_resnet_quantization(self):
-        model = torchvision.models.quantization.resnet50(pretrained=True)
-        x = torch.randn(2, 3, 224, 224, requires_grad=True)
-        self.run_test(model, (x,))
-
-    def test_shufflenet_quantization(self):
-        model = torchvision.models.quantization.shufflenet_v2_x1_0(pretrained=True)
-        x = torch.randn(2, 3, 224, 224, requires_grad=True)
-        self.run_test(model, (x,), rtol=1e-3, atol=1e-5)
-
     def test_r3d_18_video(self):
         model = torchvision.models.video.r3d_18(pretrained=True)
         x = torch.randn(1, 3, 4, 112, 112, requires_grad=True)
@@ -238,6 +213,55 @@ def run_word_language_model(self, model_name):
         # Only support CPU version, since tracer is not working in GPU RNN.
         self.run_test(model, (x, model.hidden))
 
+    def get_image_from_url(self, url):
+        import sys
+        import os
+        if sys.version_info < (3,):
+            from urlparse import urlsplit
+            import urllib2
+            request = urllib2
+        else:
+            from urllib.parse import urlsplit
+            from urllib import request
+        from PIL import Image
+        from torchvision import transforms
+        from torch._utils_internal import get_writable_path
+
+        filename = os.path.basename(urlsplit(url)[2])
+        data_dir = get_writable_path(os.path.join(os.path.dirname(__file__)))
+        path = os.path.join(data_dir, filename)
+        data = request.urlopen(url, timeout=15).read()
+        with open(path, 'wb') as f:
+            f.write(data)
+        image = Image.open(path).convert("RGB")
+        image = image.resize((300, 200), Image.BILINEAR)
+        to_tensor = transforms.ToTensor()
+        return to_tensor(image)
+
+    def get_test_images(self):
+        image_url = "http://farm3.staticflickr.com/2469/3915380994_2e611b1779_z.jpg"
+        image = self.get_image_from_url(url=image_url)
+        images = [image]
+        return images
+
+    @skipIfUnsupportedMinOpsetVersion(11)
+    def test_keypoint_rcnn(self):
+        class KeyPointRCNN(torch.nn.Module):
+            def __init__(self):
+                super(KeyPointRCNN, self).__init__()
+                self.model = torchvision.models.detection.keypoint_rcnn.keypointrcnn_resnet50_fpn(pretrained=True,
+                                                                                                  min_size=200,
+                                                                                                  max_size=300)
+
+            def forward(self, images):
+                output = self.model(images)
+                # TODO: The keypoints_scores require the use of Argmax that is updated in ONNX.
+                #       For now we are testing all the output of KeypointRCNN except keypoints_scores.
+                #       Enable When Argmax is updated in ONNX Runtime.
+                return output[0]['boxes'], output[0]['labels'], output[0]['scores'], output[0]['keypoints']
+        images = self.get_test_images()
+        self.run_test(KeyPointRCNN(), (images,), rtol=1e-3, atol=1e-5)
+
     def test_word_language_model_RNN_TANH(self):
         self.run_word_language_model("RNN_TANH")
 

torch/onnx/symbolic_opset11.py

@@ -56,10 +56,10 @@ def symbolic_fn(g, input, output_size, align_corners=None):
         coordinate_transformation_mode = "asymmetric" if interpolate_mode == "nearest" \
             else "align_corners" if align_corners else "pytorch_half_pixel"
         empty_tensor = g.op("Constant", value_t=torch.tensor([], dtype=torch.float32))
-        input_size = input.type().sizes()
-        input_size = g.op("Constant", value_t=torch.tensor(input_size[0:2], dtype=torch.int64))
+        input_size = g.op("Shape", input)
+        input_size_beg = sym_help._slice_helper(g, input_size, axes=[0], ends=[2], starts=[0])
         output_size = g.op("Cast", output_size, to_i=sym_help.cast_pytorch_to_onnx["Long"])
-        output_size = g.op("Concat", input_size, output_size, axis_i=0)
+        output_size = g.op("Concat", input_size_beg, output_size, axis_i=0)
 
         return g.op("Resize",
                     input,
@@ -115,6 +115,7 @@ def __interpolate(g, input, size, scale_factor, mode, align_corners):
             size = unsqueeze(g, size, 0)
             size = [size for i in range(input.type().dim() - 2)]
             size = g.op("Concat", *size, axis_i=0)
+        size = g.op("Cast", size, to_i=sym_help.cast_pytorch_to_onnx['Long'])
         size = g.op("Concat", input_size, size, axis_i=0)
         scales = g.op("Constant", value_t=torch.tensor([], dtype=torch.float32))
         return g.op("Resize",