fix sem_seg inference without "height" and "width"

Summary: fix #3491

Reviewed By: sstsai-adl

Differential Revision: D31061615

fbshipit-source-id: 1080687340721a31d3689ea5b2450eef891e6e33

Author: ppwwyyxx

detectron2/modeling/meta_arch/semantic_seg.py

@@ -117,8 +117,8 @@ def forward(self, batched_inputs):
 
         processed_results = []
         for result, input_per_image, image_size in zip(results, batched_inputs, images.image_sizes):
-            height = input_per_image.get("height")
-            width = input_per_image.get("width")
+            height = input_per_image.get("height", image_size[0])
+            width = input_per_image.get("width", image_size[1])
             r = sem_seg_postprocess(result, image_size, height, width)
             processed_results.append({"sem_seg": r})
         return processed_results

tests/modeling/test_model_e2e.py

@@ -73,7 +73,7 @@ def get_regular_bitmask_instances(h, w):
     return inst
 
 
-class ModelE2ETest:
+class InstanceModelE2ETest:
     def setUp(self):
         torch.manual_seed(43)
         self.model = get_model_no_weights(self.CONFIG_PATH)
@@ -115,7 +115,7 @@ def test_eval_tocpu(self):
         model(inputs)
 
 
-class MaskRCNNE2ETest(ModelE2ETest, unittest.TestCase):
+class MaskRCNNE2ETest(InstanceModelE2ETest, unittest.TestCase):
     CONFIG_PATH = "COCO-InstanceSegmentation/mask_rcnn_R_50_FPN_1x.yaml"
 
     def test_half_empty_data(self):
@@ -171,7 +171,7 @@ def test_autocast(self):
             self.assertEqual(out.scores.dtype, torch.float32)  # scores comes from softmax
 
 
-class RetinaNetE2ETest(ModelE2ETest, unittest.TestCase):
+class RetinaNetE2ETest(InstanceModelE2ETest, unittest.TestCase):
     CONFIG_PATH = "COCO-Detection/retinanet_R_50_FPN_1x.yaml"
 
     def test_inf_nan_data(self):
@@ -209,3 +209,19 @@ def test_autocast(self):
             out = self.model(inputs)[0]["instances"]
             self.assertEqual(out.pred_boxes.tensor.dtype, torch.float32)
             self.assertEqual(out.scores.dtype, torch.float16)
+
+
+class SemSegE2ETest(unittest.TestCase):
+    CONFIG_PATH = "Misc/semantic_R_50_FPN_1x.yaml"
+
+    def setUp(self):
+        torch.manual_seed(43)
+        self.model = get_model_no_weights(self.CONFIG_PATH)
+
+    def _test_eval(self, input_sizes):
+        inputs = [create_model_input(torch.rand(3, s[0], s[1])) for s in input_sizes]
+        self.model.eval()
+        self.model(inputs)
+
+    def test_forward(self):
+        self._test_eval([(200, 250), (200, 249)])