[LightGlue] Fixed attribute usage from descriptor_dim to keypoint_detector_descriptor_dim (huggingface#39021)

fix: fix descriptor dimension handling in LightGlue model

Author: sbucaille

src/transformers/models/lightglue/modeling_lightglue.py

@@ -516,16 +516,15 @@ def __init__(self, config: LightGlueConfig):
 
         self.keypoint_detector = AutoModelForKeypointDetection.from_config(config.keypoint_detector_config)
 
+        self.keypoint_detector_descriptor_dim = config.keypoint_detector_config.descriptor_decoder_dim
         self.descriptor_dim = config.descriptor_dim
         self.num_layers = config.num_hidden_layers
         self.filter_threshold = config.filter_threshold
         self.depth_confidence = config.depth_confidence
         self.width_confidence = config.width_confidence
 
-        if self.descriptor_dim != config.keypoint_detector_config.descriptor_decoder_dim:
-            self.input_projection = nn.Linear(
-                config.keypoint_detector_config.descriptor_decoder_dim, self.descriptor_dim, bias=True
-            )
+        if self.descriptor_dim != self.keypoint_detector_descriptor_dim:
+            self.input_projection = nn.Linear(self.keypoint_detector_descriptor_dim, self.descriptor_dim, bias=True)
         else:
             self.input_projection = nn.Identity()
 
@@ -721,7 +720,7 @@ def _match_image_pair(
         # (batch_size, 2, num_keypoints, 2) -> (batch_size * 2, num_keypoints, 2)
         keypoints = keypoints.reshape(batch_size * 2, initial_num_keypoints, 2)
         mask = mask.reshape(batch_size * 2, initial_num_keypoints) if mask is not None else None
-        descriptors = descriptors.reshape(batch_size * 2, initial_num_keypoints, self.descriptor_dim)
+        descriptors = descriptors.reshape(batch_size * 2, initial_num_keypoints, self.keypoint_detector_descriptor_dim)
         image_indices = torch.arange(batch_size * 2, device=device)
         # Keypoint normalization
         keypoints = normalize_keypoints(keypoints, height, width)
@@ -892,7 +891,7 @@ def forward(
 
         keypoints, _, descriptors, mask = keypoint_detections[:4]
         keypoints = keypoints.reshape(batch_size, 2, -1, 2).to(pixel_values)
-        descriptors = descriptors.reshape(batch_size, 2, -1, self.descriptor_dim).to(pixel_values)
+        descriptors = descriptors.reshape(batch_size, 2, -1, self.keypoint_detector_descriptor_dim).to(pixel_values)
         mask = mask.reshape(batch_size, 2, -1)
 
         absolute_keypoints = keypoints.clone()

src/transformers/models/lightglue/modular_lightglue.py

@@ -587,16 +587,15 @@ def __init__(self, config: LightGlueConfig):
 
         self.keypoint_detector = AutoModelForKeypointDetection.from_config(config.keypoint_detector_config)
 
+        self.keypoint_detector_descriptor_dim = config.keypoint_detector_config.descriptor_decoder_dim
         self.descriptor_dim = config.descriptor_dim
         self.num_layers = config.num_hidden_layers
         self.filter_threshold = config.filter_threshold
         self.depth_confidence = config.depth_confidence
         self.width_confidence = config.width_confidence
 
-        if self.descriptor_dim != config.keypoint_detector_config.descriptor_decoder_dim:
-            self.input_projection = nn.Linear(
-                config.keypoint_detector_config.descriptor_decoder_dim, self.descriptor_dim, bias=True
-            )
+        if self.descriptor_dim != self.keypoint_detector_descriptor_dim:
+            self.input_projection = nn.Linear(self.keypoint_detector_descriptor_dim, self.descriptor_dim, bias=True)
         else:
             self.input_projection = nn.Identity()
 
@@ -792,7 +791,7 @@ def _match_image_pair(
         # (batch_size, 2, num_keypoints, 2) -> (batch_size * 2, num_keypoints, 2)
         keypoints = keypoints.reshape(batch_size * 2, initial_num_keypoints, 2)
         mask = mask.reshape(batch_size * 2, initial_num_keypoints) if mask is not None else None
-        descriptors = descriptors.reshape(batch_size * 2, initial_num_keypoints, self.descriptor_dim)
+        descriptors = descriptors.reshape(batch_size * 2, initial_num_keypoints, self.keypoint_detector_descriptor_dim)
         image_indices = torch.arange(batch_size * 2, device=device)
         # Keypoint normalization
         keypoints = normalize_keypoints(keypoints, height, width)
@@ -963,7 +962,7 @@ def forward(
 
         keypoints, _, descriptors, mask = keypoint_detections[:4]
         keypoints = keypoints.reshape(batch_size, 2, -1, 2).to(pixel_values)
-        descriptors = descriptors.reshape(batch_size, 2, -1, self.descriptor_dim).to(pixel_values)
+        descriptors = descriptors.reshape(batch_size, 2, -1, self.keypoint_detector_descriptor_dim).to(pixel_values)
         mask = mask.reshape(batch_size, 2, -1)
 
         absolute_keypoints = keypoints.clone()