Add CI tests for detr-resnet50 inference

tests/test_object_detection.py

@@ -44,6 +44,21 @@ class GaudiDETRTester(TestCase):
     Tests for Object Detection - DETR
     """
 
+    def get_expected_loc(self, mode="default"):
+        expected_location_def = np.array([344.0622, 24.8543, 640.3398, 373.7401])
+        expected_location_ac = np.array([342, 25.25, 636, 376])
+        modeCheck = True if mode in ["default", "autocast"] else False
+        self.assertEqual(modeCheck, True)
+        if mode == "default":
+            return expected_location_def
+        else:
+            return expected_location_ac
+
+    def get_expected_num_boxes(self):
+        # For image http://images.cocodataset.org/val2017/000000039769.jpg
+        # the expected result is 5
+        return 5
+
     def prepare_model_and_processor(self):
         model = DetrForObjectDetection.from_pretrained("facebook/detr-resnet-101").to("hpu")
         model = model.eval()
@@ -62,9 +77,9 @@ def test_inference_default(self):
         target_sizes = torch.Tensor([image.size[::-1]])
         results = processor.post_process_object_detection(outputs, target_sizes=target_sizes, threshold=0.9)[0]
         boxes = results["boxes"]
-        self.assertEqual(len(boxes), 5)
-        expected_location = np.array([344.0622, 24.8543, 640.3398, 373.7401])
-        self.assertLess(np.abs(boxes[0].cpu().detach().numpy() - expected_location).max(), 1)
+        self.assertEqual(len(boxes), self.get_expected_num_boxes())
+        expected_loc = self.get_expected_loc()
+        self.assertLess(np.abs(boxes[0].cpu().detach().numpy() - expected_loc).max(), 1)
 
     def test_inference_autocast(self):
         model, processor = self.prepare_model_and_processor()
@@ -76,8 +91,9 @@ def test_inference_autocast(self):
             target_sizes = torch.Tensor([image.size[::-1]])
             results = processor.post_process_object_detection(outputs, target_sizes=target_sizes, threshold=0.9)[0]
             boxes = results["boxes"]
-            self.assertEqual(len(boxes), 5)
-            expected_location = np.array([342, 25.25, 636, 376])
+            self.assertEqual(len(boxes), self.get_expected_num_boxes())
+            print(f"boxes = {boxes}")
+            expected_location = self.get_expected_loc(mode="autocast")
             self.assertLess(np.abs(boxes[0].to(torch.float32).cpu().detach().numpy() - expected_location).max(), 5)
 
     def test_inference_hpu_graphs(self):
@@ -91,8 +107,9 @@ def test_inference_hpu_graphs(self):
         target_sizes = torch.Tensor([image.size[::-1]])
         results = processor.post_process_object_detection(outputs=outputs, target_sizes=target_sizes, threshold=0.1)
         boxes = results[0]["boxes"]
+        print(f"boxes = {boxes}")
         self.assertEqual(len(boxes), 5)
-        expected_location = np.array([344.0622, 24.8543, 640.3398, 373.7401])
+        expected_location = self.get_expected_loc()
         self.assertLess(np.abs(boxes[0].to(torch.float32).cpu().detach().numpy() - expected_location).max(), 1)
 
     def test_no_latency_regression_autocast(self):
@@ -121,3 +138,39 @@ def test_no_latency_regression_autocast(self):
 
         latency = total_model_time * 1000 / iterations  # in terms of ms
         self.assertLessEqual(latency, TIME_PERF_FACTOR * LATENCY_DETR_BF16_GRAPH_BASELINE)
+
+
+class GaudiDetrResnet50_Tester(GaudiDETRTester):
+    """
+    Tests for Custom Configuration of Detr-Resnet50 model
+    """
+
+    def get_num_labels(self):
+        # COCO-2017, the dataset the test is based on uses 91 labels
+        # TODO: automate this from dataset info. For now, just return this number
+        return 91
+
+    def prepare_model_and_processor(self):
+        model = DetrForObjectDetection.from_pretrained(
+            "facebook/detr-resnet-50",
+            revision="no_timm",
+            num_labels=self.get_num_labels(),
+            ignore_mismatched_sizes=True,
+        )
+        model = model.to("hpu")
+        model = model.eval()
+        processor = AutoProcessor.from_pretrained("facebook/detr-resnet-50")
+        return model, processor
+
+    def get_expected_loc(self, mode="default"):
+        # Reference: first box co-ordinates listed in model card:
+        # https://huggingface.co/facebook/detr-resnet-50#how-to-use
+        expected_location = np.array([40.16, 70.81, 175.55, 117.98])
+        modeCheck = True if mode in ["default", "autocast"] else False
+        self.assertEqual(modeCheck, True)
+        return expected_location
+
+    def get_expected_num_boxes(self):
+        # For image http://images.cocodataset.org/val2017/000000039769.jpg
+        # the expected result is 5
+        return 5