Upload DiskForKeypointDetection
Browse files- config.json +2 -1
- modeling_disk.py +64 -0
config.json
CHANGED
@@ -3,7 +3,8 @@
|
|
3 |
"DiskForKeypointDetection"
|
4 |
],
|
5 |
"auto_map": {
|
6 |
-
"AutoConfig": "configuration_disk.DiskConfig"
|
|
|
7 |
},
|
8 |
"descriptor_decoder_dim": 128,
|
9 |
"detection_threshold": 0.0,
|
|
|
3 |
"DiskForKeypointDetection"
|
4 |
],
|
5 |
"auto_map": {
|
6 |
+
"AutoConfig": "configuration_disk.DiskConfig",
|
7 |
+
"AutoModelForKeypointDetection": "modeling_disk.DiskForKeypointDetection"
|
8 |
},
|
9 |
"descriptor_decoder_dim": 128,
|
10 |
"detection_threshold": 0.0,
|
modeling_disk.py
ADDED
@@ -0,0 +1,64 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import kornia
|
2 |
+
import torch
|
3 |
+
|
4 |
+
from .configuration_disk import DiskConfig
|
5 |
+
from transformers import AutoConfig, AutoModelForKeypointDetection, PreTrainedModel
|
6 |
+
from transformers.models.superpoint.modeling_superpoint import (
|
7 |
+
SuperPointKeypointDescriptionOutput,
|
8 |
+
)
|
9 |
+
|
10 |
+
|
11 |
+
class DiskForKeypointDetection(PreTrainedModel):
|
12 |
+
config_class = DiskConfig
|
13 |
+
|
14 |
+
def __init__(self, config: DiskConfig):
|
15 |
+
super().__init__(config)
|
16 |
+
|
17 |
+
self.config = config
|
18 |
+
self.model = kornia.feature.DISK.from_pretrained(self.config.weights)
|
19 |
+
|
20 |
+
def forward(
|
21 |
+
self, pixel_values: torch.Tensor
|
22 |
+
) -> SuperPointKeypointDescriptionOutput:
|
23 |
+
detections = self.model(
|
24 |
+
pixel_values,
|
25 |
+
n=self.config.max_num_keypoints,
|
26 |
+
window_size=self.config.nms_window_size,
|
27 |
+
score_threshold=self.config.detection_threshold,
|
28 |
+
pad_if_not_divisible=self.config.pad_if_not_divisible,
|
29 |
+
)
|
30 |
+
max_num_keypoints = max(
|
31 |
+
detection.keypoints.shape[0] for detection in detections
|
32 |
+
)
|
33 |
+
keypoints = torch.zeros(
|
34 |
+
len(detections), max_num_keypoints, 2, device=pixel_values.device
|
35 |
+
)
|
36 |
+
descriptors = torch.zeros(
|
37 |
+
len(detections),
|
38 |
+
max_num_keypoints,
|
39 |
+
self.config.descriptor_decoder_dim,
|
40 |
+
device=pixel_values.device,
|
41 |
+
)
|
42 |
+
scores = torch.zeros(
|
43 |
+
len(detections), max_num_keypoints, device=pixel_values.device
|
44 |
+
)
|
45 |
+
mask = torch.zeros(
|
46 |
+
len(detections), max_num_keypoints, device=pixel_values.device
|
47 |
+
)
|
48 |
+
for i, detection in enumerate(detections):
|
49 |
+
keypoints[i, : detection.keypoints.shape[0]] = detection.keypoints
|
50 |
+
descriptors[i, : detection.descriptors.shape[0]] = detection.descriptors
|
51 |
+
scores[i, : detection.detection_scores.shape[0]] = (
|
52 |
+
detection.detection_scores
|
53 |
+
)
|
54 |
+
mask[i, : detection.detection_scores.shape[0]] = 1
|
55 |
+
width, height = pixel_values.shape[-1], pixel_values.shape[-2]
|
56 |
+
keypoints[:, :, 0] = keypoints[:, :, 0] / width
|
57 |
+
keypoints[:, :, 1] = keypoints[:, :, 1] / height
|
58 |
+
|
59 |
+
return SuperPointKeypointDescriptionOutput(
|
60 |
+
keypoints=keypoints,
|
61 |
+
scores=scores,
|
62 |
+
descriptors=descriptors,
|
63 |
+
mask=mask,
|
64 |
+
)
|