- Correct pytest complexity

- Make bounding box test utils use 256,256 image size

keras_cv/src/models/object_detection/__test_utils__.py

@@ -19,11 +19,13 @@
 
 
 def _create_bounding_box_dataset(
-    bounding_box_format, use_dictionary_box_format=False
+    bounding_box_format,
+    image_shape=(256, 256, 3),
+    use_dictionary_box_format=False,
 ):
     # Just about the easiest dataset you can have, all classes are 0, all boxes
     # are exactly the same. [1, 1, 2, 2] are the coordinates in xyxy.
-    xs = np.random.normal(size=(1, 512, 512, 3))
+    xs = np.random.normal(size=(1,) + image_shape)
     xs = np.tile(xs, [5, 1, 1, 1])
 
     y_classes = np.zeros((5, 3), "float32")

keras_cv/src/models/object_detection/faster_rcnn/faster_rcnn_test.py

@@ -39,7 +39,7 @@ def test_faster_rcnn_construction(self):
             num_classes=80,
             bounding_box_format="xyxy",
             backbone=keras_cv.models.ResNet18V2Backbone(
-                input_shape=(512, 512, 3)
+                input_shape=(256, 256, 3)
             ),
         )
         faster_rcnn.compile(
@@ -50,16 +50,17 @@ def test_faster_rcnn_construction(self):
             rpn_classification_loss="BinaryCrossentropy",
         )
 
+    @pytest.mark.large()
     @pytest.mark.skipif(not keras_3(), reason="disabling test for Keras 2")
     def test_faster_rcnn_call(self):
         faster_rcnn = FasterRCNN(
             num_classes=80,
             bounding_box_format="xywh",
             backbone=keras_cv.models.ResNet18V2Backbone(
-                input_shape=(512, 512, 3)
+                input_shape=(256, 256, 3)
             ),
         )
-        images = np.random.uniform(size=(2, 512, 512, 3))
+        images = np.random.uniform(size=(2, 256, 256, 3))
         _ = faster_rcnn(images)
         _ = faster_rcnn.predict(images)
 
@@ -69,7 +70,7 @@ def test_wrong_logits(self):
             num_classes=80,
             bounding_box_format="xywh",
             backbone=keras_cv.models.ResNet18V2Backbone(
-                input_shape=(512, 512, 3)
+                input_shape=(256, 256, 3)
             ),
         )
 
@@ -93,14 +94,15 @@ def test_wrong_logits(self):
                 ),
             )
 
+    @pytest.mark.large()
     @pytest.mark.skipif(not keras_3(), reason="disabling test for Keras 2")
     def test_weights_contained_in_trainable_variables(self):
         bounding_box_format = "xyxy"
         faster_rcnn = FasterRCNN(
             num_classes=80,
             bounding_box_format=bounding_box_format,
             backbone=keras_cv.models.ResNet18V2Backbone(
-                input_shape=(512, 512, 3)
+                input_shape=(256, 256, 3)
             ),
         )
         faster_rcnn.backbone.trainable = False
@@ -124,7 +126,7 @@ def test_no_nans(self):
             num_classes=80,
             bounding_box_format="xyxy",
             backbone=keras_cv.models.ResNet18V2Backbone(
-                input_shape=(512, 512, 3)
+                input_shape=(256, 256, 3)
             ),
         )
         faster_rcnn.compile(
@@ -136,7 +138,7 @@ def test_no_nans(self):
         )
 
         # only a -1 box
-        xs = np.ones((1, 512, 512, 3), "float32")
+        xs = np.ones((1, 256, 256, 3), "float32")
         ys = {
             "classes": np.array([[-1]], "float32"),
             "boxes": np.array([[[0, 0, 0, 0]]], "float32"),
@@ -157,7 +159,7 @@ def test_weights_change(self):
             num_classes=80,
             bounding_box_format="xyxy",
             backbone=keras_cv.models.ResNet18V2Backbone(
-                input_shape=(512, 512, 3)
+                input_shape=(256, 256, 3)
             ),
         )
         faster_rcnn.compile(
@@ -174,7 +176,7 @@ def test_weights_change(self):
         ).batch(5, drop_remainder=True)
 
         # call once
-        _ = faster_rcnn(ops.ones((1, 512, 512, 3)))
+        _ = faster_rcnn(ops.ones((1, 256, 256, 3)))
         original_fpn_weights = faster_rcnn.feature_pyramid.get_weights()
         original_rpn_head_weights = faster_rcnn.rpn_head.get_weights()
         original_rcnn_head_weights = faster_rcnn.rcnn_head.get_weights()
@@ -205,10 +207,10 @@ def test_saved_model(self):
             num_classes=80,
             bounding_box_format="xyxy",
             backbone=keras_cv.models.ResNet18V2Backbone(
-                input_shape=(512, 512, 3)
+                input_shape=(256, 256, 3)
             ),
         )
-        input_batch = ops.ones(shape=(1, 512, 512, 3))
+        input_batch = ops.ones(shape=(1, 256, 256, 3))
         model_output = model(input_batch)
         save_path = os.path.join(self.get_temp_dir(), "faster_rcnn.keras")
         model.save(save_path)
@@ -231,9 +233,10 @@ def test_saved_model(self):
     # https://github.com/keras-team/keras-cv/pull/1882
     @parameterized.parameters(
         ((2, 640, 384, 3),),
-        ((2, 512, 512, 3),),
+        ((2, 256, 256, 3),),
         ((2, 128, 128, 3),),
     )
+    @pytest.mark.large
     @pytest.mark.skipif(not keras_3(), reason="disabling test for Keras 2")
     def test_faster_rcnn_infer(self, batch_shape):
         batch_size = batch_shape[0]
@@ -254,9 +257,10 @@ def test_faster_rcnn_infer(self, batch_shape):
 
     @parameterized.parameters(
         ((2, 640, 384, 3),),
-        ((2, 512, 512, 3),),
+        ((2, 256, 256, 3),),
         ((2, 128, 128, 3),),
     )
+    @pytest.mark.large
     @pytest.mark.skipif(not keras_3(), reason="disabling test for Keras 2")
     def test_faster_rcnn_train(self, batch_shape):
         batch_size = batch_shape[0]
@@ -280,7 +284,7 @@ def test_invalid_compile(self):
             num_classes=80,
             bounding_box_format="yxyx",
             backbone=keras_cv.models.ResNet18V2Backbone(
-                input_shape=(512, 512, 3)
+                input_shape=(256, 256, 3)
             ),
         )
         with self.assertRaisesRegex(ValueError, "expects"):
@@ -302,7 +306,7 @@ def test_faster_rcnn_with_dictionary_input_format(self):
             num_classes=20,
             bounding_box_format="xywh",
             backbone=keras_cv.models.ResNet18V2Backbone(
-                input_shape=(512, 512, 3)
+                input_shape=(256, 256, 3)
             ),
         )
 
@@ -330,7 +334,7 @@ def test_fit_with_no_valid_gt_bbox(self):
             num_classes=20,
             bounding_box_format=bounding_box_format,
             backbone=keras_cv.models.ResNet18V2Backbone(
-                input_shape=(512, 512, 3)
+                input_shape=(256, 256, 3)
             ),
         )
 

keras_cv/src/models/object_detection/faster_rcnn/feature_pyamid_test.py

@@ -25,10 +25,10 @@ class FeaturePyramidTest(TestCase):
     @pytest.mark.skipif(not keras_3(), reason="disabling test for Keras 2")
     def test_return_type_dict(self):
         layer = FeaturePyramid(min_level=2, max_level=5)
-        c2 = np.ones([2, 64, 64, 3])
-        c3 = np.ones([2, 32, 32, 3])
-        c4 = np.ones([2, 16, 16, 3])
-        c5 = np.ones([2, 8, 8, 3])
+        c2 = np.ones([2, 32, 32, 3])
+        c3 = np.ones([2, 16, 16, 3])
+        c4 = np.ones([2, 8, 8, 3])
+        c5 = np.ones([2, 4, 4, 3])
 
         inputs = {"P2": c2, "P3": c3, "P4": c4, "P5": c5}
         output = layer(inputs)
@@ -38,10 +38,10 @@ def test_return_type_dict(self):
     @pytest.mark.skipif(not keras_3(), reason="disabling test for Keras 2")
     def test_result_shapes(self):
         layer = FeaturePyramid(min_level=2, max_level=5)
-        c2 = np.ones([2, 64, 64, 3])
-        c3 = np.ones([2, 32, 32, 3])
-        c4 = np.ones([2, 16, 16, 3])
-        c5 = np.ones([2, 8, 8, 3])
+        c2 = np.ones([2, 32, 32, 3])
+        c3 = np.ones([2, 16, 16, 3])
+        c4 = np.ones([2, 8, 8, 3])
+        c5 = np.ones([2, 4, 4, 3])
 
         inputs = {"P2": c2, "P3": c3, "P4": c4, "P5": c5}
         output = layer(inputs)
@@ -68,10 +68,10 @@ def test_result_shapes(self):
     def test_with_keras_input_tensor(self):
         # This mimic the model building with Backbone network
         layer = FeaturePyramid(min_level=2, max_level=5)
-        c2 = keras.layers.Input([64, 64, 3])
-        c3 = keras.layers.Input([32, 32, 3])
-        c4 = keras.layers.Input([16, 16, 3])
-        c5 = keras.layers.Input([8, 8, 3])
+        c2 = keras.layers.Input([32, 32, 3])
+        c3 = keras.layers.Input([16, 16, 3])
+        c4 = keras.layers.Input([8, 8, 3])
+        c5 = keras.layers.Input([4, 4, 3])
 
         inputs = {"P2": c2, "P3": c3, "P4": c4, "P5": c5}
         output = layer(inputs)
@@ -126,10 +126,10 @@ def test_invalid_output_layers(self):
     def test_invalid_input_features(self):
         layer = FeaturePyramid(min_level=2, max_level=5)
 
-        c2 = np.ones([2, 64, 64, 3])
-        c3 = np.ones([2, 32, 32, 3])
-        c4 = np.ones([2, 16, 16, 3])
-        c5 = np.ones([2, 8, 8, 3])
+        c2 = np.ones([2, 32, 32, 3])
+        c3 = np.ones([2, 16, 16, 3])
+        c4 = np.ones([2, 8, 8, 3])
+        c5 = np.ones([2, 4, 4, 3])
         inputs = {"P2": c2, "P3": c3, "P4": c4, "P5": c5}
         # Build required for Keas 3
         _ = layer(inputs)

keras_cv/src/models/object_detection/faster_rcnn/rcnn_head_test.py

@@ -23,8 +23,8 @@
 
 class RCNNHeadTest(TestCase):
     @parameterized.parameters(
-        (2, 512, 20, 7, 256),
-        (1, 1000, 80, 14, 512),
+        (2, 256, 20, 7, 256),
+        (1, 512, 80, 14, 512),
     )
     @pytest.mark.skipif(not keras_3(), reason="disabling test for Keras 2")
     def test_rcnn_head_output_shapes(

keras_cv/src/models/object_detection/faster_rcnn/rpn_head_test.py

@@ -28,11 +28,11 @@ def test_return_type_dict(
         self,
     ):
         layer = RPNHead()
-        c2 = ops.ones([2, 128, 128, 256])
-        c3 = ops.ones([2, 64, 64, 256])
-        c4 = ops.ones([2, 32, 32, 256])
-        c5 = ops.ones([2, 16, 16, 256])
-        c6 = ops.ones([2, 8, 8, 256])
+        c2 = ops.ones([2, 64, 64, 256])
+        c3 = ops.ones([2, 32, 32, 256])
+        c4 = ops.ones([2, 16, 16, 256])
+        c5 = ops.ones([2, 8, 8, 256])
+        c6 = ops.ones([2, 4, 4, 256])
 
         inputs = {"P2": c2, "P3": c3, "P4": c4, "P5": c5, "P6": c6}
         rpn_boxes, rpn_scores = layer(inputs)
@@ -48,11 +48,11 @@ def test_return_type_dict(
     @pytest.mark.skipif(not keras_3(), reason="disabling test for Keras 2")
     def test_return_type_list(self):
         layer = RPNHead()
-        c2 = ops.ones([2, 128, 128, 256])
-        c3 = ops.ones([2, 64, 64, 256])
-        c4 = ops.ones([2, 32, 32, 256])
-        c5 = ops.ones([2, 16, 16, 256])
-        c6 = ops.ones([2, 8, 8, 256])
+        c2 = ops.ones([2, 64, 64, 256])
+        c3 = ops.ones([2, 32, 32, 256])
+        c4 = ops.ones([2, 16, 16, 256])
+        c5 = ops.ones([2, 8, 8, 256])
+        c6 = ops.ones([2, 4, 4, 256])
 
         inputs = [c2, c3, c4, c5, c6]
         rpn_boxes, rpn_scores = layer(inputs)
@@ -66,11 +66,11 @@ def test_return_type_list(self):
     )
     def test_with_keras_input_tensor_and_num_anchors(self, num_anchors):
         layer = RPNHead(num_anchors_per_location=num_anchors)
-        c2 = keras.layers.Input([128, 128, 256])
-        c3 = keras.layers.Input([64, 64, 256])
-        c4 = keras.layers.Input([32, 32, 256])
-        c5 = keras.layers.Input([16, 16, 256])
-        c6 = keras.layers.Input([8, 8, 256])
+        c2 = keras.layers.Input([64, 64, 256])
+        c3 = keras.layers.Input([32, 32, 256])
+        c4 = keras.layers.Input([16, 16, 256])
+        c5 = keras.layers.Input([8, 8, 256])
+        c6 = keras.layers.Input([4, 4, 256])
 
         inputs = {"P2": c2, "P3": c3, "P4": c4, "P5": c5, "P6": c6}
         rpn_boxes, rpn_scores = layer(inputs)