SAM_DataAnnotation/src/predictor.py at main · BioIntelligence-Lab/SAM_DataAnnotation · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
from segment_anything import sam_model_registry, SamPredictor
from .preprocessing import *
from .bbox import *
import torch

def sample_slices(
  mask,
  num_slices = 10,
  seed = 1337
):
  np.random.seed(seed)
  # Filter out slices w/o segmentation
  slice_sum = mask.reshape(-1, mask.shape[-1]).sum(axis=0)
  idx = np.where(slice_sum > 0)[0]
  # Weigh slices by occurence of segmentation
  proba = slice_sum[idx]/slice_sum[idx].sum()
  # Randomly select slices
  if idx.shape[0] < num_slices:
    try:
      slices = np.random.choice(idx, idx.shape[0], p=proba, replace=False)
    except:
      slices = np.random.choice(idx, idx.shape[0], replace=False)
  else:
    try:
      slices = np.random.choice(idx, num_slices, p=proba, replace=False)
    except:
      slices = np.random.choice(idx, num_slices, replace=False)
  return slices

# TODO: Too many points affect performance, use Kmeans to cluster groups of annotations into fewer (for curatation at scale!)
def sample_points(
  mask,
  num_points = 10,
  seed = 1337
):
  np.random.seed(seed)
  # Determine points corresponding to mask
  points = np.array(np.where(mask > 0)).transpose()[:,[1,0]]
  # Randomly select points
  if points.shape[0] < num_points:
    points = points[np.random.randint(0, points.shape[0], points.shape[0])]
  else:
    points = points[np.random.randint(0, points.shape[0], num_points)]
  return points

class Predictor:
  def __init__(self, model='MedSAM', device='cuda'):
    if model == 'SAM-vit_h' or model == 'SAM':
      model = sam_model_registry['vit_h'](checkpoint='checkpoints/sam_vit_h_4b8939.pth')
    elif model == 'SAM-vit_b':
      model = sam_model_registry['vit_b'](checkpoint='checkpoints/sam_vit_b_01ec64.pth')
    elif model == 'SAM-vit_l':
      model = sam_model_registry['vit_l'](checkpoint='checkpoints/sam_vit_l_0b3195.pth')
    elif model == 'MedSAM-vit_b' or model == 'MedSAM':
      model = sam_model_registry['vit_b'](checkpoint='checkpoints/medsam_vit_b.pth')
    else:
      raise ValueError('Whoops!')
    model.to(device=device)
    self.predictor = SamPredictor(model)

  def set_image(self, img):
    self.predictor.set_image(img)

  def set_torch_image(self, img):
    self.predictor.set_torch_image(img)

  def predict(self, *args, **kwargs):
    return self.predictor.predict(*args, **kwargs)

  def predict_torch(self, *args, **kwargs):
    return self.predictor.predict_torch(*args, **kwargs)

  def get_image_embdding(self):
    return self.predictor.get_image_embedding()

  @property
  def device(self):
    return self.predictor.device

  def reset_image(self):
    self.predictor.reset_image()

  def generate_mask_from_points(
    self,
    points,
    targets
  ):
    masks, scores, _ = self.predictor.predict(
      points,
      targets,
      multimask_output = True
    )
    return masks[np.argmax(scores)]

  def generate_mask_from_bbox(
    self,
    bbox
  ):
    # Prepare bounding boxes
    batched_bbox = preprocess_bbox(bbox)
    # Determine number of batches by number of bboxes
    n_steps = batched_bbox.shape[0]
    masks = []
    for i in range(n_steps):
      box = batched_bbox[i]
      mask_i, scores_i, _ = self.predictor.predict(
        box = box,
        multimask_output = True
      )
      masks += [mask_i[np.argmax(scores_i)]]
    masks = np.sum(masks, axis=0)
    masks[masks > 1] = 1
    return masks.astype(bool)

  def generate_mask_from_points_bbox(
    self,
    points,
    targets,
    bbox
  ):
    # Prepare bounding boxes
    batched_bbox = preprocess_bbox(bbox)
    # Determine number of batches by number of bboxes
    n_steps = batched_bbox.shape[0]
    masks = []
    for i in range(n_steps):
      box = batched_bbox[i]
      points_in_box, targets_in_box = points_inside_bbox(points, targets, box)
      mask_i, scores_i, _ = self.predictor.predict(
        points_in_box,
        targets_in_box,
        box = box,
        multimask_output = True
      )
      masks += [mask_i[np.argmax(scores_i)]]
    masks = np.sum(masks, axis=0)
    masks[masks > 1] = 1
    return masks.astype(bool)