Update tests

NielsRogge · Aug 25, 2024 · be0cf18 · be0cf18
1 parent edbd7eb
commit be0cf18
Show file tree

Hide file tree

Showing 17 changed files with 48 additions and 92 deletions.
diff --git a/src/transformers/models/deprecated/deta/convert_deta_swin_to_pytorch.py b/src/transformers/models/deprecated/deta/convert_deta_swin_to_pytorch.py
@@ -39,7 +39,7 @@ def get_deta_config(model_name):
         depths=(2, 2, 18, 2),
         num_heads=(6, 12, 24, 48),
         window_size=12,
-        out_features=["stage2", "stage3", "stage4"],
+        out_indices=[2, 3, 4],
     )
 
     config = DetaConfig(

diff --git a/src/transformers/models/dpt/convert_dpt_beit_to_hf.py b/src/transformers/models/dpt/convert_dpt_beit_to_hf.py
@@ -34,14 +34,14 @@ def get_dpt_config(model_name):
     num_hidden_layers = 12
     num_attention_heads = 12
     intermediate_size = 3072
-    out_features = ["stage3", "stage6", "stage9", "stage12"]  # beit-base-384 uses [2, 5, 8, 11]
+    out_indices = [3, 6, 9, 12]  # beit-base-384 uses [2, 5, 8, 11] in original implementation
 
     if "large" in model_name:
         hidden_size = 1024
         num_hidden_layers = 24
         num_attention_heads = 16
         intermediate_size = 4096
-        out_features = ["stage6", "stage12", "stage18", "stage24"]  # beit-large-512 uses [5, 11, 17, 23]
+        out_indices = [6, 12, 18, 24]  # beit-large-512 uses [5, 11, 17, 23] in original implementation
 
     if "512" in model_name:
         image_size = 512
@@ -58,7 +58,7 @@ def get_dpt_config(model_name):
         num_attention_heads=num_attention_heads,
         use_relative_position_bias=True,
         reshape_hidden_states=False,
-        out_features=out_features,
+        out_indices=out_indices,
     )
 
     neck_hidden_sizes = [256, 512, 1024, 1024] if "large" in model_name else [96, 192, 384, 768]

diff --git a/tests/models/depth_anything/test_modeling_depth_anything.py b/tests/models/depth_anything/test_modeling_depth_anything.py
@@ -53,7 +53,7 @@ def __init__(
         num_hidden_layers=2,
         num_attention_heads=2,
         intermediate_size=8,
-        out_features=["stage1", "stage2"],
+        out_indices=[1, 2],
         apply_layernorm=False,
         reshape_hidden_states=False,
         neck_hidden_sizes=[2, 2],
@@ -68,7 +68,7 @@ def __init__(
         self.num_hidden_layers = num_hidden_layers
         self.num_attention_heads = num_attention_heads
         self.intermediate_size = intermediate_size
-        self.out_features = out_features
+        self.out_indices = out_indices
         self.apply_layernorm = apply_layernorm
         self.reshape_hidden_states = reshape_hidden_states
         self.use_labels = use_labels
@@ -111,7 +111,7 @@ def get_backbone_config(self):
             num_attention_heads=self.num_attention_heads,
             intermediate_size=self.intermediate_size,
             is_training=self.is_training,
-            out_features=self.out_features,
+            out_indices=self.out_indices,
             reshape_hidden_states=self.reshape_hidden_states,
         )
 

diff --git a/tests/models/dinov2/test_modeling_dinov2.py b/tests/models/dinov2/test_modeling_dinov2.py
@@ -129,17 +129,17 @@ def create_and_check_backbone(self, config, pixel_values, labels):
         result = model(pixel_values)
 
         # verify hidden states
-        self.parent.assertEqual(len(result.feature_maps), len(config.out_features))
+        self.parent.assertEqual(len(result.feature_maps), len(config.out_indices))
         expected_size = self.image_size // config.patch_size
         self.parent.assertListEqual(
             list(result.feature_maps[0].shape), [self.batch_size, model.channels[0], expected_size, expected_size]
         )
 
         # verify channels
-        self.parent.assertEqual(len(model.channels), len(config.out_features))
+        self.parent.assertEqual(len(model.channels), len(config.out_indices))
 
-        # verify backbone works with out_features=None
-        config.out_features = None
+        # verify backbone works with out_indices=None
+        config.out_indices = None
         model = Dinov2Backbone(config=config)
         model.to(torch_device)
         model.eval()

diff --git a/tests/models/dpt/test_modeling_dpt_auto_backbone.py b/tests/models/dpt/test_modeling_dpt_auto_backbone.py
@@ -53,7 +53,7 @@ def __init__(
         num_hidden_layers=2,
         num_attention_heads=2,
         intermediate_size=8,
-        out_features=["stage1", "stage2"],
+        out_indices=[1, 2],
         apply_layernorm=False,
         reshape_hidden_states=False,
         neck_hidden_sizes=[2, 2],
@@ -68,7 +68,7 @@ def __init__(
         self.num_hidden_layers = num_hidden_layers
         self.num_attention_heads = num_attention_heads
         self.intermediate_size = intermediate_size
-        self.out_features = out_features
+        self.out_indices = out_indices
         self.apply_layernorm = apply_layernorm
         self.reshape_hidden_states = reshape_hidden_states
         self.use_labels = use_labels
@@ -108,7 +108,7 @@ def get_backbone_config(self):
             num_attention_heads=self.num_attention_heads,
             intermediate_size=self.intermediate_size,
             is_training=self.is_training,
-            out_features=self.out_features,
+            out_indices=self.out_indices,
             reshape_hidden_states=self.reshape_hidden_states,
         )
 

diff --git a/tests/models/focalnet/test_modeling_focalnet.py b/tests/models/focalnet/test_modeling_focalnet.py
@@ -72,7 +72,6 @@ def __init__(
         use_labels=True,
         type_sequence_label_size=10,
         encoder_stride=8,
-        out_features=["stage1", "stage2"],
         out_indices=[1, 2],
     ):
         self.parent = parent
@@ -100,7 +99,6 @@ def __init__(
         self.use_labels = use_labels
         self.type_sequence_label_size = type_sequence_label_size
         self.encoder_stride = encoder_stride
-        self.out_features = out_features
         self.out_indices = out_indices
 
     def prepare_config_and_inputs(self):
@@ -135,7 +133,6 @@ def get_config(self):
             layer_norm_eps=self.layer_norm_eps,
             initializer_range=self.initializer_range,
             encoder_stride=self.encoder_stride,
-            out_features=self.out_features,
             out_indices=self.out_indices,
         )
 
@@ -157,15 +154,15 @@ def create_and_check_backbone(self, config, pixel_values, labels):
         result = model(pixel_values)
 
         # verify feature maps
-        self.parent.assertEqual(len(result.feature_maps), len(config.out_features))
+        self.parent.assertEqual(len(result.feature_maps), len(config.out_indices))
         self.parent.assertListEqual(list(result.feature_maps[0].shape), [self.batch_size, self.image_size, 8, 8])
 
         # verify channels
-        self.parent.assertEqual(len(model.channels), len(config.out_features))
+        self.parent.assertEqual(len(model.channels), len(config.out_indices))
         self.parent.assertListEqual(model.channels, config.hidden_sizes[:-1])
 
-        # verify backbone works with out_features=None
-        config.out_features = None
+        # verify backbone works with out_indices=None
+        config.out_indices = None
         model = FocalNetBackbone(config=config)
         model.to(torch_device)
         model.eval()

diff --git a/tests/models/grounding_dino/test_modeling_grounding_dino.py b/tests/models/grounding_dino/test_modeling_grounding_dino.py
@@ -140,7 +140,6 @@ def get_config(self):
             depths=[1, 1, 1, 1],
             num_heads=[1, 1, 1, 1],
             image_size=self.image_size,
-            out_features=["stage2", "stage3", "stage4"],
             out_indices=[2, 3, 4],
         )
         text_backbone = {

diff --git a/tests/models/hiera/test_modeling_hiera.py b/tests/models/hiera/test_modeling_hiera.py
@@ -148,17 +148,17 @@ def create_and_check_backbone(self, config, pixel_values, labels):
         result = model(pixel_values)
 
         # verify hidden states
-        self.parent.assertEqual(len(result.feature_maps), len(config.out_features))
+        self.parent.assertEqual(len(result.feature_maps), len(config.out_indices))
         num_patches = config.image_size[0] // config.patch_stride[0] // config.masked_unit_size[0]
         self.parent.assertListEqual(
             list(result.feature_maps[0].shape), [self.batch_size, model.channels[0], num_patches, num_patches]
         )
 
         # verify channels
-        self.parent.assertEqual(len(model.channels), len(config.out_features))
+        self.parent.assertEqual(len(model.channels), len(config.out_indices))
 
-        # verify backbone works with out_features=None
-        config.out_features = None
+        # verify backbone works with out_indices=None
+        config.out_indices = None
         model = HieraBackbone(config=config)
         model.to(torch_device)
         model.eval()

diff --git a/tests/models/maskformer/test_modeling_maskformer_swin.py b/tests/models/maskformer/test_modeling_maskformer_swin.py
@@ -63,7 +63,6 @@ def __init__(
         use_labels=True,
         type_sequence_label_size=10,
         encoder_stride=8,
-        out_features=["stage1", "stage2", "stage3"],
         out_indices=[1, 2, 3],
     ):
         self.parent = parent
@@ -90,7 +89,6 @@ def __init__(
         self.use_labels = use_labels
         self.type_sequence_label_size = type_sequence_label_size
         self.encoder_stride = encoder_stride
-        self.out_features = out_features
         self.out_indices = out_indices
 
     def prepare_config_and_inputs(self):
@@ -124,7 +122,6 @@ def get_config(self):
             layer_norm_eps=self.layer_norm_eps,
             initializer_range=self.initializer_range,
             encoder_stride=self.encoder_stride,
-            out_features=self.out_features,
             out_indices=self.out_indices,
         )
 
@@ -146,18 +143,13 @@ def create_and_check_backbone(self, config, pixel_values, labels):
         result = model(pixel_values)
 
         # verify feature maps
-        self.parent.assertEqual(len(result.feature_maps), len(config.out_features))
+        self.parent.assertEqual(len(result.feature_maps), len(config.out_indices))
         self.parent.assertListEqual(list(result.feature_maps[0].shape), [13, 16, 16, 16])
 
         # verify channels
-        self.parent.assertEqual(len(model.channels), len(config.out_features))
+        self.parent.assertEqual(len(model.channels), len(config.out_indices))
         self.parent.assertListEqual(model.channels, [16, 32, 64])
 
-        # verify ValueError
-        with self.parent.assertRaises(ValueError):
-            config.out_features = ["stem"]
-            model = MaskFormerSwinBackbone(config=config)
-
     def prepare_config_and_inputs_for_common(self):
         config_and_inputs = self.prepare_config_and_inputs()
         config, pixel_values, labels = config_and_inputs

diff --git a/tests/models/pvt_v2/test_modeling_pvt_v2.py b/tests/models/pvt_v2/test_modeling_pvt_v2.py
@@ -142,8 +142,8 @@ def create_and_check_backbone(self, config, pixel_values, labels):
         self.parent.assertEqual(len(model.channels), len(config.out_features))
         self.parent.assertListEqual(model.channels, config.hidden_sizes[1:])
 
-        # verify backbone works with out_features=None
-        config.out_features = None
+        # verify backbone works with out_indices=None
+        config.out_indices = None
         model = PvtV2Backbone(config=config)
         model.to(torch_device)
         model.eval()
@@ -397,33 +397,12 @@ def test_config(self):
         num_stages = len(config.depths) if hasattr(config, "depths") else config.num_hidden_layers
         expected_stage_names = [f"stage{idx}" for idx in range(1, num_stages + 1)]
         self.assertEqual(config.stage_names, expected_stage_names)
-        self.assertTrue(set(config.out_features).issubset(set(config.stage_names)))
 
-        # Test out_features and out_indices are correctly set
-        # out_features and out_indices both None
-        config = config_class(out_features=None, out_indices=None)
-        self.assertEqual(config.out_features, [config.stage_names[-1]])
+        # Test out_indices are correctly set
+        # out_indices set to None
+        config = config_class(out_indices=None)
         self.assertEqual(config.out_indices, [len(config.stage_names) - 1])
 
-        # out_features and out_indices both set
-        config = config_class(out_features=["stage1", "stage2"], out_indices=[0, 1])
-        self.assertEqual(config.out_features, ["stage1", "stage2"])
-        self.assertEqual(config.out_indices, [0, 1])
-
-        # Only out_features set
-        config = config_class(out_features=["stage2", "stage4"])
-        self.assertEqual(config.out_features, ["stage2", "stage4"])
-        self.assertEqual(config.out_indices, [1, 3])
-
-        # Only out_indices set
-        config = config_class(out_indices=[0, 2])
-        self.assertEqual(config.out_features, [config.stage_names[0], config.stage_names[2]])
-        self.assertEqual(config.out_indices, [0, 2])
-
-        # Error raised when out_indices do not correspond to out_features
-        with self.assertRaises(ValueError):
-            config = config_class(out_features=["stage1", "stage2"], out_indices=[0, 2])
-
     def test_config_save_pretrained(self):
         config_class = self.config_class
         config_first = config_class(out_indices=[0, 1, 2, 3])

diff --git a/tests/models/resnet/test_modeling_resnet.py b/tests/models/resnet/test_modeling_resnet.py
@@ -54,7 +54,6 @@ def __init__(
         hidden_act="relu",
         num_labels=3,
         scope=None,
-        out_features=["stage2", "stage3", "stage4"],
         out_indices=[2, 3, 4],
     ):
         self.parent = parent
@@ -70,7 +69,6 @@ def __init__(
         self.num_labels = num_labels
         self.scope = scope
         self.num_stages = len(hidden_sizes)
-        self.out_features = out_features
         self.out_indices = out_indices
 
     def prepare_config_and_inputs(self):
@@ -92,7 +90,6 @@ def get_config(self):
             depths=self.depths,
             hidden_act=self.hidden_act,
             num_labels=self.num_labels,
-            out_features=self.out_features,
             out_indices=self.out_indices,
         )
 
@@ -122,15 +119,15 @@ def create_and_check_backbone(self, config, pixel_values, labels):
         result = model(pixel_values)
 
         # verify feature maps
-        self.parent.assertEqual(len(result.feature_maps), len(config.out_features))
+        self.parent.assertEqual(len(result.feature_maps), len(config.out_indices))
         self.parent.assertListEqual(list(result.feature_maps[0].shape), [self.batch_size, self.hidden_sizes[1], 4, 4])
 
         # verify channels
-        self.parent.assertEqual(len(model.channels), len(config.out_features))
+        self.parent.assertEqual(len(model.channels), len(config.out_indices))
         self.parent.assertListEqual(model.channels, config.hidden_sizes[1:])
 
-        # verify backbone works with out_features=None
-        config.out_features = None
+        # verify backbone works with out_indices=None
+        config.out_indices = None
         model = ResNetBackbone(config=config)
         model.to(torch_device)
         model.eval()

diff --git a/tests/models/rt_detr/test_modeling_rt_detr.py b/tests/models/rt_detr/test_modeling_rt_detr.py
@@ -170,7 +170,6 @@ def get_config(self):
             embeddings_size=10,
             hidden_sizes=hidden_sizes,
             depths=[1, 1, 2, 1],
-            out_features=["stage2", "stage3", "stage4"],
             out_indices=[2, 3, 4],
         )
         return RTDetrConfig.from_backbone_configs(

diff --git a/tests/models/rt_detr/test_modeling_rt_detr_resnet.py b/tests/models/rt_detr/test_modeling_rt_detr_resnet.py
@@ -42,7 +42,6 @@ def __init__(
         hidden_act="relu",
         num_labels=3,
         scope=None,
-        out_features=["stage2", "stage3", "stage4"],
         out_indices=[2, 3, 4],
     ):
         self.parent = parent
@@ -58,7 +57,6 @@ def __init__(
         self.num_labels = num_labels
         self.scope = scope
         self.num_stages = len(hidden_sizes)
-        self.out_features = out_features
         self.out_indices = out_indices
 
     def prepare_config_and_inputs(self):
@@ -80,7 +78,6 @@ def get_config(self):
             depths=self.depths,
             hidden_act=self.hidden_act,
             num_labels=self.num_labels,
-            out_features=self.out_features,
             out_indices=self.out_indices,
         )
 
@@ -91,15 +88,15 @@ def create_and_check_backbone(self, config, pixel_values, labels):
         result = model(pixel_values)
 
         # verify feature maps
-        self.parent.assertEqual(len(result.feature_maps), len(config.out_features))
+        self.parent.assertEqual(len(result.feature_maps), len(config.out_indices))
         self.parent.assertListEqual(list(result.feature_maps[0].shape), [self.batch_size, self.hidden_sizes[1], 4, 4])
 
         # verify channels
-        self.parent.assertEqual(len(model.channels), len(config.out_features))
+        self.parent.assertEqual(len(model.channels), len(config.out_indices))
         self.parent.assertListEqual(model.channels, config.hidden_sizes[1:])
 
-        # verify backbone works with out_features=None
-        config.out_features = None
+        # verify backbone works with out_indices=None
+        config.out_indices = None
         model = RTDetrResNetBackbone(config=config)
         model.to(torch_device)
         model.eval()