Initial commit: Add DISK model implementation with fixed imports

.gitignore

@@ -0,0 +1,21 @@
+__pycache__/
+*.py[cod]
+*$py.class
+*.so
+.Python
+build/
+develop-eggs/
+dist/
+downloads/
+eggs/
+.eggs/
+lib/
+lib64/
+parts/
+sdist/
+var/
+wheels/
+*.egg-info/
+.installed.cfg
+*.egg
+MANIFEST

disk_model/configuration_disk.py

@@ -0,0 +1,30 @@
+from typing import Optional
+
+from transformers import PretrainedConfig
+
+
+class DiskConfig(PretrainedConfig):
+    model_type = "disk"
+
+    def __init__(
+        self,
+        weights: str = "depth",
+        max_num_keypoints: Optional[int] = None,
+        descriptor_decoder_dim: int = 128,
+        nms_window_size: int = 5,
+        detection_threshold: float = 0.0,
+        pad_if_not_divisible: bool = True,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+        self.weights = weights
+        self.max_num_keypoints = max_num_keypoints
+        self.descriptor_decoder_dim = descriptor_decoder_dim
+        self.nms_window_size = nms_window_size
+        self.detection_threshold = detection_threshold
+        self.pad_if_not_divisible = pad_if_not_divisible
+
+
+if __name__ == "__main__":
+    config = DiskConfig()
+    config.save_pretrained("stevenbucaille/disk", push_to_hub=True)

disk_model/modeling_disk.py

@@ -0,0 +1,76 @@
+import kornia
+import torch
+
+from configuration_disk import DiskConfig
+from transformers import AutoConfig, AutoModelForKeypointDetection, PreTrainedModel
+from transformers.models.superpoint.modeling_superpoint import (
+    SuperPointKeypointDescriptionOutput,
+)
+
+
+class DiskForKeypointDetection(PreTrainedModel):
+    config_class = DiskConfig
+
+    def __init__(self, config: DiskConfig):
+        super().__init__(config)
+
+        self.config = config
+        self.model = kornia.feature.DISK.from_pretrained(self.config.weights)
+
+    def forward(
+        self, pixel_values: torch.Tensor
+    ) -> SuperPointKeypointDescriptionOutput:
+        detections = self.model(
+            pixel_values,
+            n=self.config.max_num_keypoints,
+            window_size=self.config.nms_window_size,
+            score_threshold=self.config.detection_threshold,
+            pad_if_not_divisible=self.config.pad_if_not_divisible,
+        )
+        max_num_keypoints = max(
+            detection.keypoints.shape[0] for detection in detections
+        )
+        keypoints = torch.zeros(
+            len(detections), max_num_keypoints, 2, device=pixel_values.device
+        )
+        descriptors = torch.zeros(
+            len(detections),
+            max_num_keypoints,
+            self.config.descriptor_decoder_dim,
+            device=pixel_values.device,
+        )
+        scores = torch.zeros(
+            len(detections), max_num_keypoints, device=pixel_values.device
+        )
+        mask = torch.zeros(
+            len(detections), max_num_keypoints, device=pixel_values.device
+        )
+        for i, detection in enumerate(detections):
+            keypoints[i, : detection.keypoints.shape[0]] = detection.keypoints
+            descriptors[i, : detection.descriptors.shape[0]] = detection.descriptors
+            scores[i, : detection.detection_scores.shape[0]] = (
+                detection.detection_scores
+            )
+            mask[i, : detection.detection_scores.shape[0]] = 1
+        width, height = pixel_values.shape[-1], pixel_values.shape[-2]
+        keypoints[:, :, 0] = keypoints[:, :, 0] / width
+        keypoints[:, :, 1] = keypoints[:, :, 1] / height
+
+        return SuperPointKeypointDescriptionOutput(
+            keypoints=keypoints,
+            scores=scores,
+            descriptors=descriptors,
+            mask=mask,
+        )
+
+
+if __name__ == "__main__":
+    config = DiskConfig()
+    model = DiskForKeypointDetection(config)
+    model.save_pretrained("stevenbucaille/disk", push_to_hub=True)
+
+    AutoConfig.register("disk", DiskConfig)
+    AutoModelForKeypointDetection.register(DiskConfig, DiskForKeypointDetection)
+
+    DiskConfig.register_for_auto_class()
+    DiskForKeypointDetection.register_for_auto_class("AutoModelForKeypointDetection")