dillonlaird commited on
Commit
6723494
1 Parent(s): 35f0dee

initial commit

Browse files
Dockerfile ADDED
@@ -0,0 +1,17 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ FROM node:slim
2
+ WORKDIR /app
3
+
4
+ COPY . .
5
+
6
+ # python setup
7
+ RUN apt-get update && apt-get install -y python3 python3-pip python3-venv ffmpeg libsm6 libxext6 pkg-config libpixman-1-dev libcairo2-dev libpango1.0-dev libgif-dev
8
+ RUN python3 -m venv .venv
9
+ RUN /app/.venv/bin/pip3 install --no-cache --upgrade -r /app/requirements.txt
10
+
11
+ # nextjs setup
12
+ WORKDIR /app/instance-labeler
13
+ RUN yarn install && yarn build
14
+
15
+ WORKDIR /app
16
+
17
+ CMD ["/app/.venv/bin/python3", "main.py"]
app/__init__.py ADDED
File without changes
app/configs.py ADDED
@@ -0,0 +1,7 @@
 
 
 
 
 
 
 
 
1
+ import torch
2
+
3
+ from pathlib import Path
4
+
5
+ DATA_ROOT = Path("data")
6
+ DEVICE = torch.device("cpu")
7
+ MODEL = "mobile_sam"
app/data.py ADDED
@@ -0,0 +1,132 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import pickle as pkl
2
+ import numpy as np
3
+ import numpy.typing as npt
4
+
5
+ from PIL import Image
6
+ from PIL.Image import Image as ImageType
7
+ from pathlib import Path
8
+
9
+
10
+ def build_data(data_path: Path) -> dict:
11
+ data = {}
12
+ image_paths = (
13
+ list(data_path.glob("*.png"))
14
+ + list(data_path.glob("*.jpg"))
15
+ + list(data_path.glob("*.jpeg"))
16
+ )
17
+ for image_path in image_paths:
18
+ image_name = image_path.stem
19
+ data[image_name] = {
20
+ "image": image_path,
21
+ "labels": [],
22
+ "emb": None,
23
+ "meta_data": None,
24
+ }
25
+ return data
26
+
27
+
28
+ class Data:
29
+ def __init__(self, data_path: Path):
30
+ self.data_path = data_path
31
+ if Path(data_path).exists():
32
+ with open(data_path, "rb") as f:
33
+ self.data = pkl.load(f)
34
+ else:
35
+ data_path.parent.mkdir(parents=True, exist_ok=True)
36
+ with open(data_path, "wb") as f:
37
+ pkl.dump({}, f)
38
+ self.data = {}
39
+
40
+ def _save_data(self) -> None:
41
+ with open(self.data_path, "wb") as f:
42
+ pkl.dump(self.data, f)
43
+
44
+ def __contains__(self, image: str) -> bool:
45
+ return image in self.data
46
+
47
+ def emb_exists(self, image: str) -> bool:
48
+ return "emb" in self.data[image] and self.data[image]["emb"] is not None
49
+
50
+ def save_labels(
51
+ self, image: str, masks: list[ImageType], bboxes: list[tuple[int, ...]], labels: list[str]
52
+ ) -> None:
53
+ self.clear_labels(image)
54
+ label_paths = []
55
+ for i, (mask, label) in enumerate(zip(masks, labels)):
56
+ label_path = self.data_path.parent / f"{image}.{label}.{i}.png"
57
+ mask.save(label_path)
58
+ label_paths.append(str(label_path))
59
+ self.data[image]["masks"] = label_paths
60
+ self.data[image]["labels"] = labels
61
+ self.data[image]["bboxes"] = bboxes
62
+ self._save_data()
63
+
64
+ def save_meta_data(self, image: str, meta_data: dict) -> None:
65
+ self.data[image]["meta_data"] = meta_data
66
+ self._save_data()
67
+
68
+ def save_emb(self, image: str, emb: npt.NDArray) -> None:
69
+ emb_path = self.data_path.parent / f"{image}.emb.npy"
70
+ np.save(emb_path, emb)
71
+ self.data[image]["emb"] = emb_path
72
+ self._save_data()
73
+
74
+ def save_hq_emb(self, image: str, embs: list[npt.NDArray]) -> None:
75
+ for i, emb in enumerate(embs):
76
+ emb_path = self.data_path.parent / f"{image}.emb.{i}.npy"
77
+ np.save(emb_path, emb)
78
+ self.data[image][f"emb.{i}"] = emb_path
79
+ self._save_data()
80
+
81
+ def save_image(self, image: str, image_pil: ImageType) -> None:
82
+ image_path = self.data_path.parent / f"{image}.png"
83
+ image_pil.save(image_path)
84
+ self.data[image] = {}
85
+ self.data[image]["image"] = image_path
86
+ self._save_data()
87
+
88
+ def clear_labels(self, image: str) -> None:
89
+ if "masks" in self.data[image]:
90
+ for label_path in self.data[image]["masks"]:
91
+ Path(label_path).unlink(missing_ok=True)
92
+ if "labels" in self.data[image]:
93
+ self.data[image]["labels"] = []
94
+ self._save_data()
95
+
96
+ def get_all_images(self) -> list:
97
+ return list(self.data.keys())
98
+
99
+ def get_image(self, image: str) -> ImageType:
100
+ return Image.open(self.data[image]["image"])
101
+
102
+ def get_emb(self, image: str) -> npt.NDArray:
103
+ return np.load(self.data[image]["emb"])
104
+
105
+ def get_hq_emb(self, image: str) -> list[npt.NDArray]:
106
+ embs = []
107
+ i = 0
108
+ while True:
109
+ if f"emb.{i}" in self.data[image]:
110
+ embs.append(np.load(self.data[image][f"emb.{i}"]))
111
+ i += 1
112
+ else:
113
+ break
114
+ return embs
115
+
116
+ def get_labels(
117
+ self, image: str
118
+ ) -> tuple[list[ImageType], list[tuple[int, ...]], list[str]]:
119
+ if (
120
+ "masks" not in self.data[image]
121
+ or "labels" not in self.data[image]
122
+ or "bboxes" not in self.data[image]
123
+ ):
124
+ return [], [], []
125
+ return (
126
+ [Image.open(mask) for mask in self.data[image]["masks"]],
127
+ [tuple(e) for e in self.data[image]["bboxes"]],
128
+ self.data[image]["labels"],
129
+ )
130
+
131
+ def get_meta_data(self, image: str) -> dict:
132
+ return self.data[image]["meta_data"]
app/main.py ADDED
@@ -0,0 +1,276 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import io
2
+ import base64
3
+ import numpy as np
4
+ import torch
5
+ import time
6
+
7
+ from PIL import Image
8
+ from pydantic import BaseModel
9
+ from fastapi import FastAPI
10
+ from fastapi.responses import Response, JSONResponse
11
+ from fastapi.exceptions import HTTPException
12
+ from fastapi.middleware.cors import CORSMiddleware
13
+ from fastapi.staticfiles import StaticFiles
14
+ from fastapi.responses import FileResponse
15
+ from torchvision.transforms.functional import resize
16
+ from .model import (
17
+ build_sam_predictor,
18
+ build_sam_hq_predictor,
19
+ build_mobile_sam_predictor,
20
+ get_multi_label_predictor,
21
+ )
22
+ from .data import Data
23
+ from .configs import DATA_ROOT, DEVICE, MODEL
24
+ from .transforms import ResizeLongestSide
25
+ from .mobile_sam.utils import batched_mask_to_box
26
+
27
+
28
+ app = FastAPI()
29
+ app.add_middleware(
30
+ CORSMiddleware,
31
+ allow_origins=["*"],
32
+ allow_methods=["*"],
33
+ allow_headers=["*"],
34
+ )
35
+ if MODEL == "sam":
36
+ SAM = build_sam_predictor(checkpoint="sam_vit_h_4b8939.pth")
37
+ elif MODEL == "sam_hq":
38
+ SAM = build_sam_hq_predictor(checkpoint="sam_hq_vit_h.pth")
39
+ elif MODEL == "mobile_sam":
40
+ SAM = build_mobile_sam_predictor(checkpoint="mobile_sam.pth")
41
+ else:
42
+ raise ValueError(f"MODEL must be one of sam, sam_hq, got {MODEL}")
43
+
44
+
45
+ DATA = Data(DATA_ROOT / "data.pkl")
46
+ T = ResizeLongestSide(1024)
47
+
48
+
49
+ class SamQuery(BaseModel):
50
+ points: list[list[int]]
51
+ labels: list[int]
52
+
53
+
54
+ class MaskLabel(BaseModel):
55
+ mask: str
56
+ label: str
57
+
58
+
59
+ class Masks(BaseModel):
60
+ masks: list[str]
61
+
62
+
63
+ class MaskLabels(BaseModel):
64
+ masks: list[str]
65
+ labels: list[str]
66
+
67
+
68
+ class Box(BaseModel):
69
+ x: int
70
+ y: int
71
+ width: int
72
+ height: int
73
+
74
+
75
+ class Boxes(BaseModel):
76
+ bboxes: list[Box]
77
+
78
+
79
+ class MaskBoxes(BaseModel):
80
+ masks: list[str]
81
+ bboxes: list[Box]
82
+
83
+
84
+ class MaskBoxLabels(BaseModel):
85
+ masks: list[str]
86
+ bboxes: list[Box]
87
+ labels: list[str]
88
+
89
+
90
+ class ImageData(BaseModel):
91
+ image: str
92
+
93
+
94
+ @app.get("/")
95
+ async def index():
96
+ return FileResponse(path="/app/instance-labeler/out/index.html", media_type="text/html")
97
+
98
+
99
+ @app.post("/v1/get_label_preds/{image}")
100
+ async def get_label_preds(image: str, q: SamQuery) -> MaskBoxes:
101
+ if image not in DATA:
102
+ raise HTTPException(status_code=404, detail="Image not found")
103
+
104
+ if MODEL == "sam" or MODEL == "mobile_sam":
105
+ SAM.features = torch.from_numpy(DATA.get_emb(image)).to(DEVICE)
106
+ elif MODEL == "sam_hq":
107
+ features = DATA.get_hq_emb(image)
108
+ SAM.features = torch.from_numpy(features[0]).to(DEVICE)
109
+ SAM.interm_features = [torch.from_numpy(f).to(DEVICE) for f in features[1:]]
110
+ meta_data = DATA.get_meta_data(image)
111
+ SAM.original_size = meta_data["original_size"]
112
+ SAM.input_size = meta_data["input_size"]
113
+ SAM.is_image_set = True # type: ignore
114
+
115
+ masks, _, _ = SAM.predict( # type: ignore
116
+ point_coords=np.array(q.points),
117
+ point_labels=np.array(q.labels),
118
+ multimask_output=False,
119
+ )
120
+ bboxes = batched_mask_to_box(torch.as_tensor(masks).to(DEVICE)).cpu().numpy()
121
+ bboxes = [
122
+ Box(x=x1, y=y1, width=y2 - y1, height=x2 - x1)
123
+ for x1, y1, x2, y2 in bboxes.tolist()
124
+ ]
125
+ masks_out = []
126
+ for i in range(masks.shape[0]):
127
+ mask_i = masks[i, :, :]
128
+ mask_i = Image.fromarray(mask_i)
129
+ with io.BytesIO() as buf:
130
+ mask_i.save(buf, format="PNG")
131
+ mask_i = buf.getvalue()
132
+ masks_i_b64 = base64.b64encode(mask_i).decode("utf-8")
133
+ masks_out.append(masks_i_b64)
134
+ return MaskBoxes(masks=masks_out, bboxes=bboxes)
135
+
136
+
137
+ @app.get("/v1/get_labels/{image}")
138
+ async def get_labels(image: str) -> MaskBoxLabels:
139
+ if image not in DATA:
140
+ raise HTTPException(status_code=404, detail="Image not found")
141
+
142
+ masks, bboxes, labels = DATA.get_labels(image)
143
+ if not masks:
144
+ raise HTTPException(status_code=404, detail="Label not found")
145
+
146
+ if len(masks) != len(labels):
147
+ raise HTTPException(
148
+ status_code=400, detail="Currupted data, masks not equal to labels"
149
+ )
150
+
151
+ out_masks = []
152
+ for mask in masks:
153
+ with io.BytesIO() as buf:
154
+ mask.save(buf, format="PNG")
155
+ mask = buf.getvalue()
156
+ mask_b64 = base64.b64encode(mask).decode("utf-8")
157
+ out_masks.append(mask_b64)
158
+ bboxes = [Box(x=x1, y=y1, width=w, height=h) for x1, y1, h, w in bboxes]
159
+ return MaskBoxLabels(masks=out_masks, bboxes=bboxes, labels=labels)
160
+
161
+
162
+ @app.post("/v1/get_multi_label_preds/{image}")
163
+ async def get_multi_label_preds(image: str, q: MaskLabel) -> MaskBoxLabels:
164
+ if image not in DATA:
165
+ raise HTTPException(status_code=404, detail="Image not found")
166
+
167
+ image_pil = DATA.get_image(image)
168
+ image_np = np.array(image_pil.convert("RGB"))
169
+ mask_data = q.mask.replace("data:image/png;base64,", "")
170
+ mask = np.array(Image.open(io.BytesIO(base64.b64decode(mask_data))).convert("L"))
171
+
172
+ if mask.sum() == 0:
173
+ raise HTTPException(status_code=422, detail="Mask is empty")
174
+
175
+ per_sam_model = get_multi_label_predictor(SAM, image_np, mask)
176
+ start = time.perf_counter()
177
+ masks, bboxes, _ = per_sam_model(image_np)
178
+ print(f"inference time {time.perf_counter() - start}")
179
+ masks_out = []
180
+ for i in range(len(masks)):
181
+ mask_i = Image.fromarray(masks[i])
182
+ with io.BytesIO() as buf:
183
+ mask_i.save(buf, format="PNG")
184
+ mask_i = buf.getvalue()
185
+ masks_i_b64 = base64.b64encode(mask_i).decode("utf-8")
186
+ masks_out.append(masks_i_b64)
187
+ bboxes = [
188
+ Box(x=x1, y=y1, width=y2 - y1, height=x2 - x1)
189
+ for x1, y1, x2, y2 in bboxes.tolist()
190
+ ]
191
+ return MaskBoxLabels(
192
+ masks=masks_out, bboxes=bboxes, labels=[q.label for _ in range(len(masks))]
193
+ )
194
+
195
+
196
+ @app.put("/v1/label_image/{image}")
197
+ async def label_image(image: str, mask_labels: MaskLabels) -> Response:
198
+ if image not in DATA:
199
+ raise HTTPException(status_code=404, detail="Image not found")
200
+
201
+ if len(mask_labels.masks) != len(mask_labels.labels):
202
+ raise HTTPException(status_code=400, detail="Invalid input")
203
+
204
+ save_masks = []
205
+ for i in range(len(mask_labels.masks)):
206
+ mask_i = mask_labels.masks[i]
207
+ mask_i = mask_i.replace("data:image/png;base64,", "")
208
+ save_masks.append(Image.open(io.BytesIO(base64.b64decode(mask_i))).convert("L"))
209
+ bboxes = (
210
+ batched_mask_to_box(
211
+ torch.as_tensor(np.array([np.array(m) for m in save_masks]))
212
+ .to(DEVICE)
213
+ .bool()
214
+ )
215
+ .cpu()
216
+ .numpy()
217
+ )
218
+ bboxes = [(x1, y1, (y2 - y1), (x2 - x1)) for x1, y1, x2, y2 in bboxes.tolist()]
219
+ DATA.save_labels(image, save_masks, bboxes, mask_labels.labels)
220
+ return Response(content="saved", media_type="text/plain")
221
+
222
+
223
+ @app.get("/v1/get_image/{image}")
224
+ async def get_image(image: str) -> Response:
225
+ if image not in DATA:
226
+ raise HTTPException(status_code=404, detail="Image not found")
227
+
228
+ image_ = DATA.get_image(image)
229
+
230
+ if not DATA.emb_exists(image):
231
+ SAM.set_image(np.asarray(image_.convert("RGB"))) # type: ignore
232
+ if MODEL == "sam" or MODEL == "mobile_sam":
233
+ features = SAM.get_image_embedding().detach().cpu().numpy() # type: ignore
234
+ DATA.save_emb(image, features)
235
+ elif MODEL == "sam_hq":
236
+ features = [SAM.features] + SAM.interm_features # type: ignore
237
+ DATA.save_hq_emb(image, [f.detach().cpu().numpy() for f in features])
238
+ DATA.save_meta_data(
239
+ image,
240
+ {"original_size": SAM.original_size, "input_size": SAM.input_size},
241
+ )
242
+
243
+ with io.BytesIO() as buf:
244
+ image_.save(buf, format="PNG")
245
+ image_ = buf.getvalue()
246
+ image_b64 = base64.b64encode(image_).decode("utf-8")
247
+ return Response(content=image_b64, media_type="image/png")
248
+
249
+
250
+ @app.put("/v1/upload_image/{image}")
251
+ async def upload_image(image: str, image_data: ImageData) -> Response:
252
+ image_b64 = image_data.image
253
+ image_b64 = image_b64.replace("data:image/png;base64,", "")
254
+ image_b64 = image_b64.replace("data:image/jpeg;base64,", "")
255
+ if "data:image/" in image_b64:
256
+ raise HTTPException(
257
+ status_code=400, detail="Invalid image format, only accepts png and jpeg"
258
+ )
259
+
260
+ image_pil = Image.open(io.BytesIO(base64.b64decode(image_b64))).convert("RGB")
261
+ # image_bytes = io.BytesIO(base64.b64decode(image_b64))
262
+ # image_pil = Image.open(image_bytes)
263
+ target_size = T.get_preprocess_shape(
264
+ image_pil.size[1], image_pil.size[0], T.target_length
265
+ )
266
+ image_pil = resize(image_pil, target_size)
267
+ image_id = DATA.save_image(image, image_pil)
268
+ return Response(content=image_id, media_type="text/plain")
269
+
270
+
271
+ @app.get("/v1/get_all_images")
272
+ async def get_all_images() -> Response:
273
+ return JSONResponse(content={"images": DATA.get_all_images()})
274
+
275
+
276
+ app.mount("/", StaticFiles(directory="/app/instance-labeler/out", html=True), name="static")
app/mobile_sam/__init__.py ADDED
@@ -0,0 +1,16 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Copyright (c) Meta Platforms, Inc. and affiliates.
2
+ # All rights reserved.
3
+
4
+ # This source code is licensed under the license found in the
5
+ # LICENSE file in the root directory of this source tree.
6
+
7
+ from .build_sam import (
8
+ build_sam,
9
+ build_sam_vit_h,
10
+ build_sam_vit_l,
11
+ build_sam_vit_b,
12
+ build_sam_vit_t,
13
+ sam_model_registry,
14
+ )
15
+ from .predictor import SamPredictor
16
+ from .automatic_mask_generator import SamAutomaticMaskGenerator
app/mobile_sam/automatic_mask_generator.py ADDED
@@ -0,0 +1,372 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Copyright (c) Meta Platforms, Inc. and affiliates.
2
+ # All rights reserved.
3
+
4
+ # This source code is licensed under the license found in the
5
+ # LICENSE file in the root directory of this source tree.
6
+
7
+ import numpy as np
8
+ import torch
9
+ from torchvision.ops.boxes import batched_nms, box_area # type: ignore
10
+
11
+ from typing import Any, Dict, List, Optional, Tuple
12
+
13
+ from .modeling import Sam
14
+ from .predictor import SamPredictor
15
+ from .utils.amg import (
16
+ MaskData,
17
+ area_from_rle,
18
+ batch_iterator,
19
+ batched_mask_to_box,
20
+ box_xyxy_to_xywh,
21
+ build_all_layer_point_grids,
22
+ calculate_stability_score,
23
+ coco_encode_rle,
24
+ generate_crop_boxes,
25
+ is_box_near_crop_edge,
26
+ mask_to_rle_pytorch,
27
+ remove_small_regions,
28
+ rle_to_mask,
29
+ uncrop_boxes_xyxy,
30
+ uncrop_masks,
31
+ uncrop_points,
32
+ )
33
+
34
+
35
+ class SamAutomaticMaskGenerator:
36
+ def __init__(
37
+ self,
38
+ model: Sam,
39
+ points_per_side: Optional[int] = 32,
40
+ points_per_batch: int = 64,
41
+ pred_iou_thresh: float = 0.88,
42
+ stability_score_thresh: float = 0.95,
43
+ stability_score_offset: float = 1.0,
44
+ box_nms_thresh: float = 0.7,
45
+ crop_n_layers: int = 0,
46
+ crop_nms_thresh: float = 0.7,
47
+ crop_overlap_ratio: float = 512 / 1500,
48
+ crop_n_points_downscale_factor: int = 1,
49
+ point_grids: Optional[List[np.ndarray]] = None,
50
+ min_mask_region_area: int = 0,
51
+ output_mode: str = "binary_mask",
52
+ ) -> None:
53
+ """
54
+ Using a SAM model, generates masks for the entire image.
55
+ Generates a grid of point prompts over the image, then filters
56
+ low quality and duplicate masks. The default settings are chosen
57
+ for SAM with a ViT-H backbone.
58
+
59
+ Arguments:
60
+ model (Sam): The SAM model to use for mask prediction.
61
+ points_per_side (int or None): The number of points to be sampled
62
+ along one side of the image. The total number of points is
63
+ points_per_side**2. If None, 'point_grids' must provide explicit
64
+ point sampling.
65
+ points_per_batch (int): Sets the number of points run simultaneously
66
+ by the model. Higher numbers may be faster but use more GPU memory.
67
+ pred_iou_thresh (float): A filtering threshold in [0,1], using the
68
+ model's predicted mask quality.
69
+ stability_score_thresh (float): A filtering threshold in [0,1], using
70
+ the stability of the mask under changes to the cutoff used to binarize
71
+ the model's mask predictions.
72
+ stability_score_offset (float): The amount to shift the cutoff when
73
+ calculated the stability score.
74
+ box_nms_thresh (float): The box IoU cutoff used by non-maximal
75
+ suppression to filter duplicate masks.
76
+ crop_n_layers (int): If >0, mask prediction will be run again on
77
+ crops of the image. Sets the number of layers to run, where each
78
+ layer has 2**i_layer number of image crops.
79
+ crop_nms_thresh (float): The box IoU cutoff used by non-maximal
80
+ suppression to filter duplicate masks between different crops.
81
+ crop_overlap_ratio (float): Sets the degree to which crops overlap.
82
+ In the first crop layer, crops will overlap by this fraction of
83
+ the image length. Later layers with more crops scale down this overlap.
84
+ crop_n_points_downscale_factor (int): The number of points-per-side
85
+ sampled in layer n is scaled down by crop_n_points_downscale_factor**n.
86
+ point_grids (list(np.ndarray) or None): A list over explicit grids
87
+ of points used for sampling, normalized to [0,1]. The nth grid in the
88
+ list is used in the nth crop layer. Exclusive with points_per_side.
89
+ min_mask_region_area (int): If >0, postprocessing will be applied
90
+ to remove disconnected regions and holes in masks with area smaller
91
+ than min_mask_region_area. Requires opencv.
92
+ output_mode (str): The form masks are returned in. Can be 'binary_mask',
93
+ 'uncompressed_rle', or 'coco_rle'. 'coco_rle' requires pycocotools.
94
+ For large resolutions, 'binary_mask' may consume large amounts of
95
+ memory.
96
+ """
97
+
98
+ assert (points_per_side is None) != (
99
+ point_grids is None
100
+ ), "Exactly one of points_per_side or point_grid must be provided."
101
+ if points_per_side is not None:
102
+ self.point_grids = build_all_layer_point_grids(
103
+ points_per_side,
104
+ crop_n_layers,
105
+ crop_n_points_downscale_factor,
106
+ )
107
+ elif point_grids is not None:
108
+ self.point_grids = point_grids
109
+ else:
110
+ raise ValueError("Can't have both points_per_side and point_grid be None.")
111
+
112
+ assert output_mode in [
113
+ "binary_mask",
114
+ "uncompressed_rle",
115
+ "coco_rle",
116
+ ], f"Unknown output_mode {output_mode}."
117
+ if output_mode == "coco_rle":
118
+ from pycocotools import mask as mask_utils # type: ignore # noqa: F401
119
+
120
+ if min_mask_region_area > 0:
121
+ import cv2 # type: ignore # noqa: F401
122
+
123
+ self.predictor = SamPredictor(model)
124
+ self.points_per_batch = points_per_batch
125
+ self.pred_iou_thresh = pred_iou_thresh
126
+ self.stability_score_thresh = stability_score_thresh
127
+ self.stability_score_offset = stability_score_offset
128
+ self.box_nms_thresh = box_nms_thresh
129
+ self.crop_n_layers = crop_n_layers
130
+ self.crop_nms_thresh = crop_nms_thresh
131
+ self.crop_overlap_ratio = crop_overlap_ratio
132
+ self.crop_n_points_downscale_factor = crop_n_points_downscale_factor
133
+ self.min_mask_region_area = min_mask_region_area
134
+ self.output_mode = output_mode
135
+
136
+ @torch.no_grad()
137
+ def generate(self, image: np.ndarray) -> List[Dict[str, Any]]:
138
+ """
139
+ Generates masks for the given image.
140
+
141
+ Arguments:
142
+ image (np.ndarray): The image to generate masks for, in HWC uint8 format.
143
+
144
+ Returns:
145
+ list(dict(str, any)): A list over records for masks. Each record is
146
+ a dict containing the following keys:
147
+ segmentation (dict(str, any) or np.ndarray): The mask. If
148
+ output_mode='binary_mask', is an array of shape HW. Otherwise,
149
+ is a dictionary containing the RLE.
150
+ bbox (list(float)): The box around the mask, in XYWH format.
151
+ area (int): The area in pixels of the mask.
152
+ predicted_iou (float): The model's own prediction of the mask's
153
+ quality. This is filtered by the pred_iou_thresh parameter.
154
+ point_coords (list(list(float))): The point coordinates input
155
+ to the model to generate this mask.
156
+ stability_score (float): A measure of the mask's quality. This
157
+ is filtered on using the stability_score_thresh parameter.
158
+ crop_box (list(float)): The crop of the image used to generate
159
+ the mask, given in XYWH format.
160
+ """
161
+
162
+ # Generate masks
163
+ mask_data = self._generate_masks(image)
164
+
165
+ # Filter small disconnected regions and holes in masks
166
+ if self.min_mask_region_area > 0:
167
+ mask_data = self.postprocess_small_regions(
168
+ mask_data,
169
+ self.min_mask_region_area,
170
+ max(self.box_nms_thresh, self.crop_nms_thresh),
171
+ )
172
+
173
+ # Encode masks
174
+ if self.output_mode == "coco_rle":
175
+ mask_data["segmentations"] = [coco_encode_rle(rle) for rle in mask_data["rles"]]
176
+ elif self.output_mode == "binary_mask":
177
+ mask_data["segmentations"] = [rle_to_mask(rle) for rle in mask_data["rles"]]
178
+ else:
179
+ mask_data["segmentations"] = mask_data["rles"]
180
+
181
+ # Write mask records
182
+ curr_anns = []
183
+ for idx in range(len(mask_data["segmentations"])):
184
+ ann = {
185
+ "segmentation": mask_data["segmentations"][idx],
186
+ "area": area_from_rle(mask_data["rles"][idx]),
187
+ "bbox": box_xyxy_to_xywh(mask_data["boxes"][idx]).tolist(),
188
+ "predicted_iou": mask_data["iou_preds"][idx].item(),
189
+ "point_coords": [mask_data["points"][idx].tolist()],
190
+ "stability_score": mask_data["stability_score"][idx].item(),
191
+ "crop_box": box_xyxy_to_xywh(mask_data["crop_boxes"][idx]).tolist(),
192
+ }
193
+ curr_anns.append(ann)
194
+
195
+ return curr_anns
196
+
197
+ def _generate_masks(self, image: np.ndarray) -> MaskData:
198
+ orig_size = image.shape[:2]
199
+ crop_boxes, layer_idxs = generate_crop_boxes(
200
+ orig_size, self.crop_n_layers, self.crop_overlap_ratio
201
+ )
202
+
203
+ # Iterate over image crops
204
+ data = MaskData()
205
+ for crop_box, layer_idx in zip(crop_boxes, layer_idxs):
206
+ crop_data = self._process_crop(image, crop_box, layer_idx, orig_size)
207
+ data.cat(crop_data)
208
+
209
+ # Remove duplicate masks between crops
210
+ if len(crop_boxes) > 1:
211
+ # Prefer masks from smaller crops
212
+ scores = 1 / box_area(data["crop_boxes"])
213
+ scores = scores.to(data["boxes"].device)
214
+ keep_by_nms = batched_nms(
215
+ data["boxes"].float(),
216
+ scores,
217
+ torch.zeros_like(data["boxes"][:, 0]), # categories
218
+ iou_threshold=self.crop_nms_thresh,
219
+ )
220
+ data.filter(keep_by_nms)
221
+
222
+ data.to_numpy()
223
+ return data
224
+
225
+ def _process_crop(
226
+ self,
227
+ image: np.ndarray,
228
+ crop_box: List[int],
229
+ crop_layer_idx: int,
230
+ orig_size: Tuple[int, ...],
231
+ ) -> MaskData:
232
+ # Crop the image and calculate embeddings
233
+ x0, y0, x1, y1 = crop_box
234
+ cropped_im = image[y0:y1, x0:x1, :]
235
+ cropped_im_size = cropped_im.shape[:2]
236
+ self.predictor.set_image(cropped_im)
237
+
238
+ # Get points for this crop
239
+ points_scale = np.array(cropped_im_size)[None, ::-1]
240
+ points_for_image = self.point_grids[crop_layer_idx] * points_scale
241
+
242
+ # Generate masks for this crop in batches
243
+ data = MaskData()
244
+ for (points,) in batch_iterator(self.points_per_batch, points_for_image):
245
+ batch_data = self._process_batch(points, cropped_im_size, crop_box, orig_size)
246
+ data.cat(batch_data)
247
+ del batch_data
248
+ self.predictor.reset_image()
249
+
250
+ # Remove duplicates within this crop.
251
+ keep_by_nms = batched_nms(
252
+ data["boxes"].float(),
253
+ data["iou_preds"],
254
+ torch.zeros_like(data["boxes"][:, 0]), # categories
255
+ iou_threshold=self.box_nms_thresh,
256
+ )
257
+ data.filter(keep_by_nms)
258
+
259
+ # Return to the original image frame
260
+ data["boxes"] = uncrop_boxes_xyxy(data["boxes"], crop_box)
261
+ data["points"] = uncrop_points(data["points"], crop_box)
262
+ data["crop_boxes"] = torch.tensor([crop_box for _ in range(len(data["rles"]))])
263
+
264
+ return data
265
+
266
+ def _process_batch(
267
+ self,
268
+ points: np.ndarray,
269
+ im_size: Tuple[int, ...],
270
+ crop_box: List[int],
271
+ orig_size: Tuple[int, ...],
272
+ ) -> MaskData:
273
+ orig_h, orig_w = orig_size
274
+
275
+ # Run model on this batch
276
+ transformed_points = self.predictor.transform.apply_coords(points, im_size)
277
+ in_points = torch.as_tensor(transformed_points, device=self.predictor.device)
278
+ in_labels = torch.ones(in_points.shape[0], dtype=torch.int, device=in_points.device)
279
+ masks, iou_preds, _ = self.predictor.predict_torch(
280
+ in_points[:, None, :],
281
+ in_labels[:, None],
282
+ multimask_output=True,
283
+ return_logits=True,
284
+ )
285
+
286
+ # Serialize predictions and store in MaskData
287
+ data = MaskData(
288
+ masks=masks.flatten(0, 1),
289
+ iou_preds=iou_preds.flatten(0, 1),
290
+ points=torch.as_tensor(points.repeat(masks.shape[1], axis=0)),
291
+ )
292
+ del masks
293
+
294
+ # Filter by predicted IoU
295
+ if self.pred_iou_thresh > 0.0:
296
+ keep_mask = data["iou_preds"] > self.pred_iou_thresh
297
+ data.filter(keep_mask)
298
+
299
+ # Calculate stability score
300
+ data["stability_score"] = calculate_stability_score(
301
+ data["masks"], self.predictor.model.mask_threshold, self.stability_score_offset
302
+ )
303
+ if self.stability_score_thresh > 0.0:
304
+ keep_mask = data["stability_score"] >= self.stability_score_thresh
305
+ data.filter(keep_mask)
306
+
307
+ # Threshold masks and calculate boxes
308
+ data["masks"] = data["masks"] > self.predictor.model.mask_threshold
309
+ data["boxes"] = batched_mask_to_box(data["masks"])
310
+
311
+ # Filter boxes that touch crop boundaries
312
+ keep_mask = ~is_box_near_crop_edge(data["boxes"], crop_box, [0, 0, orig_w, orig_h])
313
+ if not torch.all(keep_mask):
314
+ data.filter(keep_mask)
315
+
316
+ # Compress to RLE
317
+ data["masks"] = uncrop_masks(data["masks"], crop_box, orig_h, orig_w)
318
+ data["rles"] = mask_to_rle_pytorch(data["masks"])
319
+ del data["masks"]
320
+
321
+ return data
322
+
323
+ @staticmethod
324
+ def postprocess_small_regions(
325
+ mask_data: MaskData, min_area: int, nms_thresh: float
326
+ ) -> MaskData:
327
+ """
328
+ Removes small disconnected regions and holes in masks, then reruns
329
+ box NMS to remove any new duplicates.
330
+
331
+ Edits mask_data in place.
332
+
333
+ Requires open-cv as a dependency.
334
+ """
335
+ if len(mask_data["rles"]) == 0:
336
+ return mask_data
337
+
338
+ # Filter small disconnected regions and holes
339
+ new_masks = []
340
+ scores = []
341
+ for rle in mask_data["rles"]:
342
+ mask = rle_to_mask(rle)
343
+
344
+ mask, changed = remove_small_regions(mask, min_area, mode="holes")
345
+ unchanged = not changed
346
+ mask, changed = remove_small_regions(mask, min_area, mode="islands")
347
+ unchanged = unchanged and not changed
348
+
349
+ new_masks.append(torch.as_tensor(mask).unsqueeze(0))
350
+ # Give score=0 to changed masks and score=1 to unchanged masks
351
+ # so NMS will prefer ones that didn't need postprocessing
352
+ scores.append(float(unchanged))
353
+
354
+ # Recalculate boxes and remove any new duplicates
355
+ masks = torch.cat(new_masks, dim=0)
356
+ boxes = batched_mask_to_box(masks)
357
+ keep_by_nms = batched_nms(
358
+ boxes.float(),
359
+ torch.as_tensor(scores),
360
+ torch.zeros_like(boxes[:, 0]), # categories
361
+ iou_threshold=nms_thresh,
362
+ )
363
+
364
+ # Only recalculate RLEs for masks that have changed
365
+ for i_mask in keep_by_nms:
366
+ if scores[i_mask] == 0.0:
367
+ mask_torch = masks[i_mask].unsqueeze(0)
368
+ mask_data["rles"][i_mask] = mask_to_rle_pytorch(mask_torch)[0]
369
+ mask_data["boxes"][i_mask] = boxes[i_mask] # update res directly
370
+ mask_data.filter(keep_by_nms)
371
+
372
+ return mask_data
app/mobile_sam/build_sam.py ADDED
@@ -0,0 +1,159 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Copyright (c) Meta Platforms, Inc. and affiliates.
2
+ # All rights reserved.
3
+
4
+ # This source code is licensed under the license found in the
5
+ # LICENSE file in the root directory of this source tree.
6
+
7
+ import torch
8
+
9
+ from functools import partial
10
+
11
+ from .modeling import ImageEncoderViT, MaskDecoder, PromptEncoder, Sam, TwoWayTransformer, TinyViT
12
+
13
+
14
+ def build_sam_vit_h(checkpoint=None):
15
+ return _build_sam(
16
+ encoder_embed_dim=1280,
17
+ encoder_depth=32,
18
+ encoder_num_heads=16,
19
+ encoder_global_attn_indexes=[7, 15, 23, 31],
20
+ checkpoint=checkpoint,
21
+ )
22
+
23
+
24
+ build_sam = build_sam_vit_h
25
+
26
+
27
+ def build_sam_vit_l(checkpoint=None):
28
+ return _build_sam(
29
+ encoder_embed_dim=1024,
30
+ encoder_depth=24,
31
+ encoder_num_heads=16,
32
+ encoder_global_attn_indexes=[5, 11, 17, 23],
33
+ checkpoint=checkpoint,
34
+ )
35
+
36
+
37
+ def build_sam_vit_b(checkpoint=None):
38
+ return _build_sam(
39
+ encoder_embed_dim=768,
40
+ encoder_depth=12,
41
+ encoder_num_heads=12,
42
+ encoder_global_attn_indexes=[2, 5, 8, 11],
43
+ checkpoint=checkpoint,
44
+ )
45
+
46
+
47
+ def build_sam_vit_t(checkpoint=None):
48
+ prompt_embed_dim = 256
49
+ image_size = 1024
50
+ vit_patch_size = 16
51
+ image_embedding_size = image_size // vit_patch_size
52
+ mobile_sam = Sam(
53
+ image_encoder=TinyViT(img_size=1024, in_chans=3, num_classes=1000,
54
+ embed_dims=[64, 128, 160, 320],
55
+ depths=[2, 2, 6, 2],
56
+ num_heads=[2, 4, 5, 10],
57
+ window_sizes=[7, 7, 14, 7],
58
+ mlp_ratio=4.,
59
+ drop_rate=0.,
60
+ drop_path_rate=0.0,
61
+ use_checkpoint=False,
62
+ mbconv_expand_ratio=4.0,
63
+ local_conv_size=3,
64
+ layer_lr_decay=0.8
65
+ ),
66
+ prompt_encoder=PromptEncoder(
67
+ embed_dim=prompt_embed_dim,
68
+ image_embedding_size=(image_embedding_size, image_embedding_size),
69
+ input_image_size=(image_size, image_size),
70
+ mask_in_chans=16,
71
+ ),
72
+ mask_decoder=MaskDecoder(
73
+ num_multimask_outputs=3,
74
+ transformer=TwoWayTransformer(
75
+ depth=2,
76
+ embedding_dim=prompt_embed_dim,
77
+ mlp_dim=2048,
78
+ num_heads=8,
79
+ ),
80
+ transformer_dim=prompt_embed_dim,
81
+ iou_head_depth=3,
82
+ iou_head_hidden_dim=256,
83
+ ),
84
+ pixel_mean=[123.675, 116.28, 103.53],
85
+ pixel_std=[58.395, 57.12, 57.375],
86
+ )
87
+
88
+ mobile_sam.eval()
89
+ if checkpoint is not None:
90
+ with open(checkpoint, "rb") as f:
91
+ state_dict = torch.load(f)
92
+ mobile_sam.load_state_dict(state_dict)
93
+ return mobile_sam
94
+
95
+
96
+ sam_model_registry = {
97
+ "default": build_sam_vit_h,
98
+ "vit_h": build_sam_vit_h,
99
+ "vit_l": build_sam_vit_l,
100
+ "vit_b": build_sam_vit_b,
101
+ "vit_t": build_sam_vit_t,
102
+ }
103
+
104
+
105
+ def _build_sam(
106
+ encoder_embed_dim,
107
+ encoder_depth,
108
+ encoder_num_heads,
109
+ encoder_global_attn_indexes,
110
+ checkpoint=None,
111
+ ):
112
+ prompt_embed_dim = 256
113
+ image_size = 1024
114
+ vit_patch_size = 16
115
+ image_embedding_size = image_size // vit_patch_size
116
+ sam = Sam(
117
+ image_encoder=ImageEncoderViT(
118
+ depth=encoder_depth,
119
+ embed_dim=encoder_embed_dim,
120
+ img_size=image_size,
121
+ mlp_ratio=4,
122
+ norm_layer=partial(torch.nn.LayerNorm, eps=1e-6),
123
+ num_heads=encoder_num_heads,
124
+ patch_size=vit_patch_size,
125
+ qkv_bias=True,
126
+ use_rel_pos=True,
127
+ global_attn_indexes=encoder_global_attn_indexes,
128
+ window_size=14,
129
+ out_chans=prompt_embed_dim,
130
+ ),
131
+ prompt_encoder=PromptEncoder(
132
+ embed_dim=prompt_embed_dim,
133
+ image_embedding_size=(image_embedding_size, image_embedding_size),
134
+ input_image_size=(image_size, image_size),
135
+ mask_in_chans=16,
136
+ ),
137
+ mask_decoder=MaskDecoder(
138
+ num_multimask_outputs=3,
139
+ transformer=TwoWayTransformer(
140
+ depth=2,
141
+ embedding_dim=prompt_embed_dim,
142
+ mlp_dim=2048,
143
+ num_heads=8,
144
+ ),
145
+ transformer_dim=prompt_embed_dim,
146
+ iou_head_depth=3,
147
+ iou_head_hidden_dim=256,
148
+ ),
149
+ pixel_mean=[123.675, 116.28, 103.53],
150
+ pixel_std=[58.395, 57.12, 57.375],
151
+ )
152
+ sam.eval()
153
+ if checkpoint is not None:
154
+ with open(checkpoint, "rb") as f:
155
+ state_dict = torch.load(f)
156
+ sam.load_state_dict(state_dict)
157
+ return sam
158
+
159
+
app/mobile_sam/modeling/__init__.py ADDED
@@ -0,0 +1,12 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Copyright (c) Meta Platforms, Inc. and affiliates.
2
+ # All rights reserved.
3
+
4
+ # This source code is licensed under the license found in the
5
+ # LICENSE file in the root directory of this source tree.
6
+
7
+ from .sam import Sam
8
+ from .image_encoder import ImageEncoderViT
9
+ from .mask_decoder import MaskDecoder
10
+ from .prompt_encoder import PromptEncoder
11
+ from .transformer import TwoWayTransformer
12
+ from .tiny_vit_sam import TinyViT
app/mobile_sam/modeling/common.py ADDED
@@ -0,0 +1,43 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Copyright (c) Meta Platforms, Inc. and affiliates.
2
+ # All rights reserved.
3
+
4
+ # This source code is licensed under the license found in the
5
+ # LICENSE file in the root directory of this source tree.
6
+
7
+ import torch
8
+ import torch.nn as nn
9
+
10
+ from typing import Type
11
+
12
+
13
+ class MLPBlock(nn.Module):
14
+ def __init__(
15
+ self,
16
+ embedding_dim: int,
17
+ mlp_dim: int,
18
+ act: Type[nn.Module] = nn.GELU,
19
+ ) -> None:
20
+ super().__init__()
21
+ self.lin1 = nn.Linear(embedding_dim, mlp_dim)
22
+ self.lin2 = nn.Linear(mlp_dim, embedding_dim)
23
+ self.act = act()
24
+
25
+ def forward(self, x: torch.Tensor) -> torch.Tensor:
26
+ return self.lin2(self.act(self.lin1(x)))
27
+
28
+
29
+ # From https://github.com/facebookresearch/detectron2/blob/main/detectron2/layers/batch_norm.py # noqa
30
+ # Itself from https://github.com/facebookresearch/ConvNeXt/blob/d1fa8f6fef0a165b27399986cc2bdacc92777e40/models/convnext.py#L119 # noqa
31
+ class LayerNorm2d(nn.Module):
32
+ def __init__(self, num_channels: int, eps: float = 1e-6) -> None:
33
+ super().__init__()
34
+ self.weight = nn.Parameter(torch.ones(num_channels))
35
+ self.bias = nn.Parameter(torch.zeros(num_channels))
36
+ self.eps = eps
37
+
38
+ def forward(self, x: torch.Tensor) -> torch.Tensor:
39
+ u = x.mean(1, keepdim=True)
40
+ s = (x - u).pow(2).mean(1, keepdim=True)
41
+ x = (x - u) / torch.sqrt(s + self.eps)
42
+ x = self.weight[:, None, None] * x + self.bias[:, None, None]
43
+ return x
app/mobile_sam/modeling/image_encoder.py ADDED
@@ -0,0 +1,395 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Copyright (c) Meta Platforms, Inc. and affiliates.
2
+ # All rights reserved.
3
+
4
+ # This source code is licensed under the license found in the
5
+ # LICENSE file in the root directory of this source tree.
6
+
7
+ import torch
8
+ import torch.nn as nn
9
+ import torch.nn.functional as F
10
+
11
+ from typing import Optional, Tuple, Type
12
+
13
+ from .common import LayerNorm2d, MLPBlock
14
+
15
+
16
+ # This class and its supporting functions below lightly adapted from the ViTDet backbone available at: https://github.com/facebookresearch/detectron2/blob/main/detectron2/modeling/backbone/vit.py # noqa
17
+ class ImageEncoderViT(nn.Module):
18
+ def __init__(
19
+ self,
20
+ img_size: int = 1024,
21
+ patch_size: int = 16,
22
+ in_chans: int = 3,
23
+ embed_dim: int = 768,
24
+ depth: int = 12,
25
+ num_heads: int = 12,
26
+ mlp_ratio: float = 4.0,
27
+ out_chans: int = 256,
28
+ qkv_bias: bool = True,
29
+ norm_layer: Type[nn.Module] = nn.LayerNorm,
30
+ act_layer: Type[nn.Module] = nn.GELU,
31
+ use_abs_pos: bool = True,
32
+ use_rel_pos: bool = False,
33
+ rel_pos_zero_init: bool = True,
34
+ window_size: int = 0,
35
+ global_attn_indexes: Tuple[int, ...] = (),
36
+ ) -> None:
37
+ """
38
+ Args:
39
+ img_size (int): Input image size.
40
+ patch_size (int): Patch size.
41
+ in_chans (int): Number of input image channels.
42
+ embed_dim (int): Patch embedding dimension.
43
+ depth (int): Depth of ViT.
44
+ num_heads (int): Number of attention heads in each ViT block.
45
+ mlp_ratio (float): Ratio of mlp hidden dim to embedding dim.
46
+ qkv_bias (bool): If True, add a learnable bias to query, key, value.
47
+ norm_layer (nn.Module): Normalization layer.
48
+ act_layer (nn.Module): Activation layer.
49
+ use_abs_pos (bool): If True, use absolute positional embeddings.
50
+ use_rel_pos (bool): If True, add relative positional embeddings to the attention map.
51
+ rel_pos_zero_init (bool): If True, zero initialize relative positional parameters.
52
+ window_size (int): Window size for window attention blocks.
53
+ global_attn_indexes (list): Indexes for blocks using global attention.
54
+ """
55
+ super().__init__()
56
+ self.img_size = img_size
57
+
58
+ self.patch_embed = PatchEmbed(
59
+ kernel_size=(patch_size, patch_size),
60
+ stride=(patch_size, patch_size),
61
+ in_chans=in_chans,
62
+ embed_dim=embed_dim,
63
+ )
64
+
65
+ self.pos_embed: Optional[nn.Parameter] = None
66
+ if use_abs_pos:
67
+ # Initialize absolute positional embedding with pretrain image size.
68
+ self.pos_embed = nn.Parameter(
69
+ torch.zeros(1, img_size // patch_size, img_size // patch_size, embed_dim)
70
+ )
71
+
72
+ self.blocks = nn.ModuleList()
73
+ for i in range(depth):
74
+ block = Block(
75
+ dim=embed_dim,
76
+ num_heads=num_heads,
77
+ mlp_ratio=mlp_ratio,
78
+ qkv_bias=qkv_bias,
79
+ norm_layer=norm_layer,
80
+ act_layer=act_layer,
81
+ use_rel_pos=use_rel_pos,
82
+ rel_pos_zero_init=rel_pos_zero_init,
83
+ window_size=window_size if i not in global_attn_indexes else 0,
84
+ input_size=(img_size // patch_size, img_size // patch_size),
85
+ )
86
+ self.blocks.append(block)
87
+
88
+ self.neck = nn.Sequential(
89
+ nn.Conv2d(
90
+ embed_dim,
91
+ out_chans,
92
+ kernel_size=1,
93
+ bias=False,
94
+ ),
95
+ LayerNorm2d(out_chans),
96
+ nn.Conv2d(
97
+ out_chans,
98
+ out_chans,
99
+ kernel_size=3,
100
+ padding=1,
101
+ bias=False,
102
+ ),
103
+ LayerNorm2d(out_chans),
104
+ )
105
+
106
+ def forward(self, x: torch.Tensor) -> torch.Tensor:
107
+ x = self.patch_embed(x)
108
+ if self.pos_embed is not None:
109
+ x = x + self.pos_embed
110
+
111
+ for blk in self.blocks:
112
+ x = blk(x)
113
+
114
+ x = self.neck(x.permute(0, 3, 1, 2))
115
+
116
+ return x
117
+
118
+
119
+ class Block(nn.Module):
120
+ """Transformer blocks with support of window attention and residual propagation blocks"""
121
+
122
+ def __init__(
123
+ self,
124
+ dim: int,
125
+ num_heads: int,
126
+ mlp_ratio: float = 4.0,
127
+ qkv_bias: bool = True,
128
+ norm_layer: Type[nn.Module] = nn.LayerNorm,
129
+ act_layer: Type[nn.Module] = nn.GELU,
130
+ use_rel_pos: bool = False,
131
+ rel_pos_zero_init: bool = True,
132
+ window_size: int = 0,
133
+ input_size: Optional[Tuple[int, int]] = None,
134
+ ) -> None:
135
+ """
136
+ Args:
137
+ dim (int): Number of input channels.
138
+ num_heads (int): Number of attention heads in each ViT block.
139
+ mlp_ratio (float): Ratio of mlp hidden dim to embedding dim.
140
+ qkv_bias (bool): If True, add a learnable bias to query, key, value.
141
+ norm_layer (nn.Module): Normalization layer.
142
+ act_layer (nn.Module): Activation layer.
143
+ use_rel_pos (bool): If True, add relative positional embeddings to the attention map.
144
+ rel_pos_zero_init (bool): If True, zero initialize relative positional parameters.
145
+ window_size (int): Window size for window attention blocks. If it equals 0, then
146
+ use global attention.
147
+ input_size (tuple(int, int) or None): Input resolution for calculating the relative
148
+ positional parameter size.
149
+ """
150
+ super().__init__()
151
+ self.norm1 = norm_layer(dim)
152
+ self.attn = Attention(
153
+ dim,
154
+ num_heads=num_heads,
155
+ qkv_bias=qkv_bias,
156
+ use_rel_pos=use_rel_pos,
157
+ rel_pos_zero_init=rel_pos_zero_init,
158
+ input_size=input_size if window_size == 0 else (window_size, window_size),
159
+ )
160
+
161
+ self.norm2 = norm_layer(dim)
162
+ self.mlp = MLPBlock(embedding_dim=dim, mlp_dim=int(dim * mlp_ratio), act=act_layer)
163
+
164
+ self.window_size = window_size
165
+
166
+ def forward(self, x: torch.Tensor) -> torch.Tensor:
167
+ shortcut = x
168
+ x = self.norm1(x)
169
+ # Window partition
170
+ if self.window_size > 0:
171
+ H, W = x.shape[1], x.shape[2]
172
+ x, pad_hw = window_partition(x, self.window_size)
173
+
174
+ x = self.attn(x)
175
+ # Reverse window partition
176
+ if self.window_size > 0:
177
+ x = window_unpartition(x, self.window_size, pad_hw, (H, W))
178
+
179
+ x = shortcut + x
180
+ x = x + self.mlp(self.norm2(x))
181
+
182
+ return x
183
+
184
+
185
+ class Attention(nn.Module):
186
+ """Multi-head Attention block with relative position embeddings."""
187
+
188
+ def __init__(
189
+ self,
190
+ dim: int,
191
+ num_heads: int = 8,
192
+ qkv_bias: bool = True,
193
+ use_rel_pos: bool = False,
194
+ rel_pos_zero_init: bool = True,
195
+ input_size: Optional[Tuple[int, int]] = None,
196
+ ) -> None:
197
+ """
198
+ Args:
199
+ dim (int): Number of input channels.
200
+ num_heads (int): Number of attention heads.
201
+ qkv_bias (bool): If True, add a learnable bias to query, key, value.
202
+ rel_pos (bool): If True, add relative positional embeddings to the attention map.
203
+ rel_pos_zero_init (bool): If True, zero initialize relative positional parameters.
204
+ input_size (tuple(int, int) or None): Input resolution for calculating the relative
205
+ positional parameter size.
206
+ """
207
+ super().__init__()
208
+ self.num_heads = num_heads
209
+ head_dim = dim // num_heads
210
+ self.scale = head_dim**-0.5
211
+
212
+ self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
213
+ self.proj = nn.Linear(dim, dim)
214
+
215
+ self.use_rel_pos = use_rel_pos
216
+ if self.use_rel_pos:
217
+ assert (
218
+ input_size is not None
219
+ ), "Input size must be provided if using relative positional encoding."
220
+ # initialize relative positional embeddings
221
+ self.rel_pos_h = nn.Parameter(torch.zeros(2 * input_size[0] - 1, head_dim))
222
+ self.rel_pos_w = nn.Parameter(torch.zeros(2 * input_size[1] - 1, head_dim))
223
+
224
+ def forward(self, x: torch.Tensor) -> torch.Tensor:
225
+ B, H, W, _ = x.shape
226
+ # qkv with shape (3, B, nHead, H * W, C)
227
+ qkv = self.qkv(x).reshape(B, H * W, 3, self.num_heads, -1).permute(2, 0, 3, 1, 4)
228
+ # q, k, v with shape (B * nHead, H * W, C)
229
+ q, k, v = qkv.reshape(3, B * self.num_heads, H * W, -1).unbind(0)
230
+
231
+ attn = (q * self.scale) @ k.transpose(-2, -1)
232
+
233
+ if self.use_rel_pos:
234
+ attn = add_decomposed_rel_pos(attn, q, self.rel_pos_h, self.rel_pos_w, (H, W), (H, W))
235
+
236
+ attn = attn.softmax(dim=-1)
237
+ x = (attn @ v).view(B, self.num_heads, H, W, -1).permute(0, 2, 3, 1, 4).reshape(B, H, W, -1)
238
+ x = self.proj(x)
239
+
240
+ return x
241
+
242
+
243
+ def window_partition(x: torch.Tensor, window_size: int) -> Tuple[torch.Tensor, Tuple[int, int]]:
244
+ """
245
+ Partition into non-overlapping windows with padding if needed.
246
+ Args:
247
+ x (tensor): input tokens with [B, H, W, C].
248
+ window_size (int): window size.
249
+
250
+ Returns:
251
+ windows: windows after partition with [B * num_windows, window_size, window_size, C].
252
+ (Hp, Wp): padded height and width before partition
253
+ """
254
+ B, H, W, C = x.shape
255
+
256
+ pad_h = (window_size - H % window_size) % window_size
257
+ pad_w = (window_size - W % window_size) % window_size
258
+ if pad_h > 0 or pad_w > 0:
259
+ x = F.pad(x, (0, 0, 0, pad_w, 0, pad_h))
260
+ Hp, Wp = H + pad_h, W + pad_w
261
+
262
+ x = x.view(B, Hp // window_size, window_size, Wp // window_size, window_size, C)
263
+ windows = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(-1, window_size, window_size, C)
264
+ return windows, (Hp, Wp)
265
+
266
+
267
+ def window_unpartition(
268
+ windows: torch.Tensor, window_size: int, pad_hw: Tuple[int, int], hw: Tuple[int, int]
269
+ ) -> torch.Tensor:
270
+ """
271
+ Window unpartition into original sequences and removing padding.
272
+ Args:
273
+ windows (tensor): input tokens with [B * num_windows, window_size, window_size, C].
274
+ window_size (int): window size.
275
+ pad_hw (Tuple): padded height and width (Hp, Wp).
276
+ hw (Tuple): original height and width (H, W) before padding.
277
+
278
+ Returns:
279
+ x: unpartitioned sequences with [B, H, W, C].
280
+ """
281
+ Hp, Wp = pad_hw
282
+ H, W = hw
283
+ B = windows.shape[0] // (Hp * Wp // window_size // window_size)
284
+ x = windows.view(B, Hp // window_size, Wp // window_size, window_size, window_size, -1)
285
+ x = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(B, Hp, Wp, -1)
286
+
287
+ if Hp > H or Wp > W:
288
+ x = x[:, :H, :W, :].contiguous()
289
+ return x
290
+
291
+
292
+ def get_rel_pos(q_size: int, k_size: int, rel_pos: torch.Tensor) -> torch.Tensor:
293
+ """
294
+ Get relative positional embeddings according to the relative positions of
295
+ query and key sizes.
296
+ Args:
297
+ q_size (int): size of query q.
298
+ k_size (int): size of key k.
299
+ rel_pos (Tensor): relative position embeddings (L, C).
300
+
301
+ Returns:
302
+ Extracted positional embeddings according to relative positions.
303
+ """
304
+ max_rel_dist = int(2 * max(q_size, k_size) - 1)
305
+ # Interpolate rel pos if needed.
306
+ if rel_pos.shape[0] != max_rel_dist:
307
+ # Interpolate rel pos.
308
+ rel_pos_resized = F.interpolate(
309
+ rel_pos.reshape(1, rel_pos.shape[0], -1).permute(0, 2, 1),
310
+ size=max_rel_dist,
311
+ mode="linear",
312
+ )
313
+ rel_pos_resized = rel_pos_resized.reshape(-1, max_rel_dist).permute(1, 0)
314
+ else:
315
+ rel_pos_resized = rel_pos
316
+
317
+ # Scale the coords with short length if shapes for q and k are different.
318
+ q_coords = torch.arange(q_size)[:, None] * max(k_size / q_size, 1.0)
319
+ k_coords = torch.arange(k_size)[None, :] * max(q_size / k_size, 1.0)
320
+ relative_coords = (q_coords - k_coords) + (k_size - 1) * max(q_size / k_size, 1.0)
321
+
322
+ return rel_pos_resized[relative_coords.long()]
323
+
324
+
325
+ def add_decomposed_rel_pos(
326
+ attn: torch.Tensor,
327
+ q: torch.Tensor,
328
+ rel_pos_h: torch.Tensor,
329
+ rel_pos_w: torch.Tensor,
330
+ q_size: Tuple[int, int],
331
+ k_size: Tuple[int, int],
332
+ ) -> torch.Tensor:
333
+ """
334
+ Calculate decomposed Relative Positional Embeddings from :paper:`mvitv2`.
335
+ https://github.com/facebookresearch/mvit/blob/19786631e330df9f3622e5402b4a419a263a2c80/mvit/models/attention.py # noqa B950
336
+ Args:
337
+ attn (Tensor): attention map.
338
+ q (Tensor): query q in the attention layer with shape (B, q_h * q_w, C).
339
+ rel_pos_h (Tensor): relative position embeddings (Lh, C) for height axis.
340
+ rel_pos_w (Tensor): relative position embeddings (Lw, C) for width axis.
341
+ q_size (Tuple): spatial sequence size of query q with (q_h, q_w).
342
+ k_size (Tuple): spatial sequence size of key k with (k_h, k_w).
343
+
344
+ Returns:
345
+ attn (Tensor): attention map with added relative positional embeddings.
346
+ """
347
+ q_h, q_w = q_size
348
+ k_h, k_w = k_size
349
+ Rh = get_rel_pos(q_h, k_h, rel_pos_h)
350
+ Rw = get_rel_pos(q_w, k_w, rel_pos_w)
351
+
352
+ B, _, dim = q.shape
353
+ r_q = q.reshape(B, q_h, q_w, dim)
354
+ rel_h = torch.einsum("bhwc,hkc->bhwk", r_q, Rh)
355
+ rel_w = torch.einsum("bhwc,wkc->bhwk", r_q, Rw)
356
+
357
+ attn = (
358
+ attn.view(B, q_h, q_w, k_h, k_w) + rel_h[:, :, :, :, None] + rel_w[:, :, :, None, :]
359
+ ).view(B, q_h * q_w, k_h * k_w)
360
+
361
+ return attn
362
+
363
+
364
+ class PatchEmbed(nn.Module):
365
+ """
366
+ Image to Patch Embedding.
367
+ """
368
+
369
+ def __init__(
370
+ self,
371
+ kernel_size: Tuple[int, int] = (16, 16),
372
+ stride: Tuple[int, int] = (16, 16),
373
+ padding: Tuple[int, int] = (0, 0),
374
+ in_chans: int = 3,
375
+ embed_dim: int = 768,
376
+ ) -> None:
377
+ """
378
+ Args:
379
+ kernel_size (Tuple): kernel size of the projection layer.
380
+ stride (Tuple): stride of the projection layer.
381
+ padding (Tuple): padding size of the projection layer.
382
+ in_chans (int): Number of input image channels.
383
+ embed_dim (int): Patch embedding dimension.
384
+ """
385
+ super().__init__()
386
+
387
+ self.proj = nn.Conv2d(
388
+ in_chans, embed_dim, kernel_size=kernel_size, stride=stride, padding=padding
389
+ )
390
+
391
+ def forward(self, x: torch.Tensor) -> torch.Tensor:
392
+ x = self.proj(x)
393
+ # B C H W -> B H W C
394
+ x = x.permute(0, 2, 3, 1)
395
+ return x
app/mobile_sam/modeling/mask_decoder.py ADDED
@@ -0,0 +1,176 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Copyright (c) Meta Platforms, Inc. and affiliates.
2
+ # All rights reserved.
3
+
4
+ # This source code is licensed under the license found in the
5
+ # LICENSE file in the root directory of this source tree.
6
+
7
+ import torch
8
+ from torch import nn
9
+ from torch.nn import functional as F
10
+
11
+ from typing import List, Tuple, Type
12
+
13
+ from .common import LayerNorm2d
14
+
15
+
16
+ class MaskDecoder(nn.Module):
17
+ def __init__(
18
+ self,
19
+ *,
20
+ transformer_dim: int,
21
+ transformer: nn.Module,
22
+ num_multimask_outputs: int = 3,
23
+ activation: Type[nn.Module] = nn.GELU,
24
+ iou_head_depth: int = 3,
25
+ iou_head_hidden_dim: int = 256,
26
+ ) -> None:
27
+ """
28
+ Predicts masks given an image and prompt embeddings, using a
29
+ transformer architecture.
30
+
31
+ Arguments:
32
+ transformer_dim (int): the channel dimension of the transformer
33
+ transformer (nn.Module): the transformer used to predict masks
34
+ num_multimask_outputs (int): the number of masks to predict
35
+ when disambiguating masks
36
+ activation (nn.Module): the type of activation to use when
37
+ upscaling masks
38
+ iou_head_depth (int): the depth of the MLP used to predict
39
+ mask quality
40
+ iou_head_hidden_dim (int): the hidden dimension of the MLP
41
+ used to predict mask quality
42
+ """
43
+ super().__init__()
44
+ self.transformer_dim = transformer_dim
45
+ self.transformer = transformer
46
+
47
+ self.num_multimask_outputs = num_multimask_outputs
48
+
49
+ self.iou_token = nn.Embedding(1, transformer_dim)
50
+ self.num_mask_tokens = num_multimask_outputs + 1
51
+ self.mask_tokens = nn.Embedding(self.num_mask_tokens, transformer_dim)
52
+
53
+ self.output_upscaling = nn.Sequential(
54
+ nn.ConvTranspose2d(transformer_dim, transformer_dim // 4, kernel_size=2, stride=2),
55
+ LayerNorm2d(transformer_dim // 4),
56
+ activation(),
57
+ nn.ConvTranspose2d(transformer_dim // 4, transformer_dim // 8, kernel_size=2, stride=2),
58
+ activation(),
59
+ )
60
+ self.output_hypernetworks_mlps = nn.ModuleList(
61
+ [
62
+ MLP(transformer_dim, transformer_dim, transformer_dim // 8, 3)
63
+ for i in range(self.num_mask_tokens)
64
+ ]
65
+ )
66
+
67
+ self.iou_prediction_head = MLP(
68
+ transformer_dim, iou_head_hidden_dim, self.num_mask_tokens, iou_head_depth
69
+ )
70
+
71
+ def forward(
72
+ self,
73
+ image_embeddings: torch.Tensor,
74
+ image_pe: torch.Tensor,
75
+ sparse_prompt_embeddings: torch.Tensor,
76
+ dense_prompt_embeddings: torch.Tensor,
77
+ multimask_output: bool,
78
+ ) -> Tuple[torch.Tensor, torch.Tensor]:
79
+ """
80
+ Predict masks given image and prompt embeddings.
81
+
82
+ Arguments:
83
+ image_embeddings (torch.Tensor): the embeddings from the image encoder
84
+ image_pe (torch.Tensor): positional encoding with the shape of image_embeddings
85
+ sparse_prompt_embeddings (torch.Tensor): the embeddings of the points and boxes
86
+ dense_prompt_embeddings (torch.Tensor): the embeddings of the mask inputs
87
+ multimask_output (bool): Whether to return multiple masks or a single
88
+ mask.
89
+
90
+ Returns:
91
+ torch.Tensor: batched predicted masks
92
+ torch.Tensor: batched predictions of mask quality
93
+ """
94
+ masks, iou_pred = self.predict_masks(
95
+ image_embeddings=image_embeddings,
96
+ image_pe=image_pe,
97
+ sparse_prompt_embeddings=sparse_prompt_embeddings,
98
+ dense_prompt_embeddings=dense_prompt_embeddings,
99
+ )
100
+
101
+ # Select the correct mask or masks for output
102
+ if multimask_output:
103
+ mask_slice = slice(1, None)
104
+ else:
105
+ mask_slice = slice(0, 1)
106
+ masks = masks[:, mask_slice, :, :]
107
+ iou_pred = iou_pred[:, mask_slice]
108
+
109
+ # Prepare output
110
+ return masks, iou_pred
111
+
112
+ def predict_masks(
113
+ self,
114
+ image_embeddings: torch.Tensor,
115
+ image_pe: torch.Tensor,
116
+ sparse_prompt_embeddings: torch.Tensor,
117
+ dense_prompt_embeddings: torch.Tensor,
118
+ ) -> Tuple[torch.Tensor, torch.Tensor]:
119
+ """Predicts masks. See 'forward' for more details."""
120
+ # Concatenate output tokens
121
+ output_tokens = torch.cat([self.iou_token.weight, self.mask_tokens.weight], dim=0)
122
+ output_tokens = output_tokens.unsqueeze(0).expand(sparse_prompt_embeddings.size(0), -1, -1)
123
+ tokens = torch.cat((output_tokens, sparse_prompt_embeddings), dim=1)
124
+
125
+ # Expand per-image data in batch direction to be per-mask
126
+ src = torch.repeat_interleave(image_embeddings, tokens.shape[0], dim=0)
127
+ src = src + dense_prompt_embeddings
128
+ pos_src = torch.repeat_interleave(image_pe, tokens.shape[0], dim=0)
129
+ b, c, h, w = src.shape
130
+
131
+ # Run the transformer
132
+ hs, src = self.transformer(src, pos_src, tokens)
133
+ iou_token_out = hs[:, 0, :]
134
+ mask_tokens_out = hs[:, 1 : (1 + self.num_mask_tokens), :]
135
+
136
+ # Upscale mask embeddings and predict masks using the mask tokens
137
+ src = src.transpose(1, 2).view(b, c, h, w)
138
+ upscaled_embedding = self.output_upscaling(src)
139
+ hyper_in_list: List[torch.Tensor] = []
140
+ for i in range(self.num_mask_tokens):
141
+ hyper_in_list.append(self.output_hypernetworks_mlps[i](mask_tokens_out[:, i, :]))
142
+ hyper_in = torch.stack(hyper_in_list, dim=1)
143
+ b, c, h, w = upscaled_embedding.shape
144
+ masks = (hyper_in @ upscaled_embedding.view(b, c, h * w)).view(b, -1, h, w)
145
+
146
+ # Generate mask quality predictions
147
+ iou_pred = self.iou_prediction_head(iou_token_out)
148
+
149
+ return masks, iou_pred
150
+
151
+
152
+ # Lightly adapted from
153
+ # https://github.com/facebookresearch/MaskFormer/blob/main/mask_former/modeling/transformer/transformer_predictor.py # noqa
154
+ class MLP(nn.Module):
155
+ def __init__(
156
+ self,
157
+ input_dim: int,
158
+ hidden_dim: int,
159
+ output_dim: int,
160
+ num_layers: int,
161
+ sigmoid_output: bool = False,
162
+ ) -> None:
163
+ super().__init__()
164
+ self.num_layers = num_layers
165
+ h = [hidden_dim] * (num_layers - 1)
166
+ self.layers = nn.ModuleList(
167
+ nn.Linear(n, k) for n, k in zip([input_dim] + h, h + [output_dim])
168
+ )
169
+ self.sigmoid_output = sigmoid_output
170
+
171
+ def forward(self, x):
172
+ for i, layer in enumerate(self.layers):
173
+ x = F.relu(layer(x)) if i < self.num_layers - 1 else layer(x)
174
+ if self.sigmoid_output:
175
+ x = F.sigmoid(x)
176
+ return x
app/mobile_sam/modeling/prompt_encoder.py ADDED
@@ -0,0 +1,214 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Copyright (c) Meta Platforms, Inc. and affiliates.
2
+ # All rights reserved.
3
+
4
+ # This source code is licensed under the license found in the
5
+ # LICENSE file in the root directory of this source tree.
6
+
7
+ import numpy as np
8
+ import torch
9
+ from torch import nn
10
+
11
+ from typing import Any, Optional, Tuple, Type
12
+
13
+ from .common import LayerNorm2d
14
+
15
+
16
+ class PromptEncoder(nn.Module):
17
+ def __init__(
18
+ self,
19
+ embed_dim: int,
20
+ image_embedding_size: Tuple[int, int],
21
+ input_image_size: Tuple[int, int],
22
+ mask_in_chans: int,
23
+ activation: Type[nn.Module] = nn.GELU,
24
+ ) -> None:
25
+ """
26
+ Encodes prompts for input to SAM's mask decoder.
27
+
28
+ Arguments:
29
+ embed_dim (int): The prompts' embedding dimension
30
+ image_embedding_size (tuple(int, int)): The spatial size of the
31
+ image embedding, as (H, W).
32
+ input_image_size (int): The padded size of the image as input
33
+ to the image encoder, as (H, W).
34
+ mask_in_chans (int): The number of hidden channels used for
35
+ encoding input masks.
36
+ activation (nn.Module): The activation to use when encoding
37
+ input masks.
38
+ """
39
+ super().__init__()
40
+ self.embed_dim = embed_dim
41
+ self.input_image_size = input_image_size
42
+ self.image_embedding_size = image_embedding_size
43
+ self.pe_layer = PositionEmbeddingRandom(embed_dim // 2)
44
+
45
+ self.num_point_embeddings: int = 4 # pos/neg point + 2 box corners
46
+ point_embeddings = [nn.Embedding(1, embed_dim) for i in range(self.num_point_embeddings)]
47
+ self.point_embeddings = nn.ModuleList(point_embeddings)
48
+ self.not_a_point_embed = nn.Embedding(1, embed_dim)
49
+
50
+ self.mask_input_size = (4 * image_embedding_size[0], 4 * image_embedding_size[1])
51
+ self.mask_downscaling = nn.Sequential(
52
+ nn.Conv2d(1, mask_in_chans // 4, kernel_size=2, stride=2),
53
+ LayerNorm2d(mask_in_chans // 4),
54
+ activation(),
55
+ nn.Conv2d(mask_in_chans // 4, mask_in_chans, kernel_size=2, stride=2),
56
+ LayerNorm2d(mask_in_chans),
57
+ activation(),
58
+ nn.Conv2d(mask_in_chans, embed_dim, kernel_size=1),
59
+ )
60
+ self.no_mask_embed = nn.Embedding(1, embed_dim)
61
+
62
+ def get_dense_pe(self) -> torch.Tensor:
63
+ """
64
+ Returns the positional encoding used to encode point prompts,
65
+ applied to a dense set of points the shape of the image encoding.
66
+
67
+ Returns:
68
+ torch.Tensor: Positional encoding with shape
69
+ 1x(embed_dim)x(embedding_h)x(embedding_w)
70
+ """
71
+ return self.pe_layer(self.image_embedding_size).unsqueeze(0)
72
+
73
+ def _embed_points(
74
+ self,
75
+ points: torch.Tensor,
76
+ labels: torch.Tensor,
77
+ pad: bool,
78
+ ) -> torch.Tensor:
79
+ """Embeds point prompts."""
80
+ points = points + 0.5 # Shift to center of pixel
81
+ if pad:
82
+ padding_point = torch.zeros((points.shape[0], 1, 2), device=points.device)
83
+ padding_label = -torch.ones((labels.shape[0], 1), device=labels.device)
84
+ points = torch.cat([points, padding_point], dim=1)
85
+ labels = torch.cat([labels, padding_label], dim=1)
86
+ point_embedding = self.pe_layer.forward_with_coords(points, self.input_image_size)
87
+ point_embedding[labels == -1] = 0.0
88
+ point_embedding[labels == -1] += self.not_a_point_embed.weight
89
+ point_embedding[labels == 0] += self.point_embeddings[0].weight
90
+ point_embedding[labels == 1] += self.point_embeddings[1].weight
91
+ return point_embedding
92
+
93
+ def _embed_boxes(self, boxes: torch.Tensor) -> torch.Tensor:
94
+ """Embeds box prompts."""
95
+ boxes = boxes + 0.5 # Shift to center of pixel
96
+ coords = boxes.reshape(-1, 2, 2)
97
+ corner_embedding = self.pe_layer.forward_with_coords(coords, self.input_image_size)
98
+ corner_embedding[:, 0, :] += self.point_embeddings[2].weight
99
+ corner_embedding[:, 1, :] += self.point_embeddings[3].weight
100
+ return corner_embedding
101
+
102
+ def _embed_masks(self, masks: torch.Tensor) -> torch.Tensor:
103
+ """Embeds mask inputs."""
104
+ mask_embedding = self.mask_downscaling(masks)
105
+ return mask_embedding
106
+
107
+ def _get_batch_size(
108
+ self,
109
+ points: Optional[Tuple[torch.Tensor, torch.Tensor]],
110
+ boxes: Optional[torch.Tensor],
111
+ masks: Optional[torch.Tensor],
112
+ ) -> int:
113
+ """
114
+ Gets the batch size of the output given the batch size of the input prompts.
115
+ """
116
+ if points is not None:
117
+ return points[0].shape[0]
118
+ elif boxes is not None:
119
+ return boxes.shape[0]
120
+ elif masks is not None:
121
+ return masks.shape[0]
122
+ else:
123
+ return 1
124
+
125
+ def _get_device(self) -> torch.device:
126
+ return self.point_embeddings[0].weight.device
127
+
128
+ def forward(
129
+ self,
130
+ points: Optional[Tuple[torch.Tensor, torch.Tensor]],
131
+ boxes: Optional[torch.Tensor],
132
+ masks: Optional[torch.Tensor],
133
+ ) -> Tuple[torch.Tensor, torch.Tensor]:
134
+ """
135
+ Embeds different types of prompts, returning both sparse and dense
136
+ embeddings.
137
+
138
+ Arguments:
139
+ points (tuple(torch.Tensor, torch.Tensor) or none): point coordinates
140
+ and labels to embed.
141
+ boxes (torch.Tensor or none): boxes to embed
142
+ masks (torch.Tensor or none): masks to embed
143
+
144
+ Returns:
145
+ torch.Tensor: sparse embeddings for the points and boxes, with shape
146
+ BxNx(embed_dim), where N is determined by the number of input points
147
+ and boxes.
148
+ torch.Tensor: dense embeddings for the masks, in the shape
149
+ Bx(embed_dim)x(embed_H)x(embed_W)
150
+ """
151
+ bs = self._get_batch_size(points, boxes, masks)
152
+ sparse_embeddings = torch.empty((bs, 0, self.embed_dim), device=self._get_device())
153
+ if points is not None:
154
+ coords, labels = points
155
+ point_embeddings = self._embed_points(coords, labels, pad=(boxes is None))
156
+ sparse_embeddings = torch.cat([sparse_embeddings, point_embeddings], dim=1)
157
+ if boxes is not None:
158
+ box_embeddings = self._embed_boxes(boxes)
159
+ sparse_embeddings = torch.cat([sparse_embeddings, box_embeddings], dim=1)
160
+
161
+ if masks is not None:
162
+ dense_embeddings = self._embed_masks(masks)
163
+ else:
164
+ dense_embeddings = self.no_mask_embed.weight.reshape(1, -1, 1, 1).expand(
165
+ bs, -1, self.image_embedding_size[0], self.image_embedding_size[1]
166
+ )
167
+
168
+ return sparse_embeddings, dense_embeddings
169
+
170
+
171
+ class PositionEmbeddingRandom(nn.Module):
172
+ """
173
+ Positional encoding using random spatial frequencies.
174
+ """
175
+
176
+ def __init__(self, num_pos_feats: int = 64, scale: Optional[float] = None) -> None:
177
+ super().__init__()
178
+ if scale is None or scale <= 0.0:
179
+ scale = 1.0
180
+ self.register_buffer(
181
+ "positional_encoding_gaussian_matrix",
182
+ scale * torch.randn((2, num_pos_feats)),
183
+ )
184
+
185
+ def _pe_encoding(self, coords: torch.Tensor) -> torch.Tensor:
186
+ """Positionally encode points that are normalized to [0,1]."""
187
+ # assuming coords are in [0, 1]^2 square and have d_1 x ... x d_n x 2 shape
188
+ coords = 2 * coords - 1
189
+ coords = coords @ self.positional_encoding_gaussian_matrix
190
+ coords = 2 * np.pi * coords
191
+ # outputs d_1 x ... x d_n x C shape
192
+ return torch.cat([torch.sin(coords), torch.cos(coords)], dim=-1)
193
+
194
+ def forward(self, size: Tuple[int, int]) -> torch.Tensor:
195
+ """Generate positional encoding for a grid of the specified size."""
196
+ h, w = size
197
+ device: Any = self.positional_encoding_gaussian_matrix.device
198
+ grid = torch.ones((h, w), device=device, dtype=torch.float32)
199
+ y_embed = grid.cumsum(dim=0) - 0.5
200
+ x_embed = grid.cumsum(dim=1) - 0.5
201
+ y_embed = y_embed / h
202
+ x_embed = x_embed / w
203
+
204
+ pe = self._pe_encoding(torch.stack([x_embed, y_embed], dim=-1))
205
+ return pe.permute(2, 0, 1) # C x H x W
206
+
207
+ def forward_with_coords(
208
+ self, coords_input: torch.Tensor, image_size: Tuple[int, int]
209
+ ) -> torch.Tensor:
210
+ """Positionally encode points that are not normalized to [0,1]."""
211
+ coords = coords_input.clone()
212
+ coords[:, :, 0] = coords[:, :, 0] / image_size[1]
213
+ coords[:, :, 1] = coords[:, :, 1] / image_size[0]
214
+ return self._pe_encoding(coords.to(torch.float)) # B x N x C
app/mobile_sam/modeling/sam.py ADDED
@@ -0,0 +1,182 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Copyright (c) Meta Platforms, Inc. and affiliates.
2
+ # All rights reserved.
3
+
4
+ # This source code is licensed under the license found in the
5
+ # LICENSE file in the root directory of this source tree.
6
+
7
+ import torch
8
+ from torch import nn
9
+ from torch.nn import functional as F
10
+
11
+ from typing import Any, Dict, List, Tuple, Union
12
+
13
+ from .tiny_vit_sam import TinyViT
14
+ from .image_encoder import ImageEncoderViT
15
+ from .mask_decoder import MaskDecoder
16
+ from .prompt_encoder import PromptEncoder
17
+
18
+
19
+ class Sam(nn.Module):
20
+ mask_threshold: float = 0.0
21
+ image_format: str = "RGB"
22
+
23
+ def __init__(
24
+ self,
25
+ image_encoder: Union[ImageEncoderViT, TinyViT],
26
+ prompt_encoder: PromptEncoder,
27
+ mask_decoder: MaskDecoder,
28
+ pixel_mean: List[float] = [123.675, 116.28, 103.53],
29
+ pixel_std: List[float] = [58.395, 57.12, 57.375],
30
+ ) -> None:
31
+ """
32
+ SAM predicts object masks from an image and input prompts.
33
+
34
+ Arguments:
35
+ image_encoder (ImageEncoderViT): The backbone used to encode the
36
+ image into image embeddings that allow for efficient mask prediction.
37
+ prompt_encoder (PromptEncoder): Encodes various types of input prompts.
38
+ mask_decoder (MaskDecoder): Predicts masks from the image embeddings
39
+ and encoded prompts.
40
+ pixel_mean (list(float)): Mean values for normalizing pixels in the input image.
41
+ pixel_std (list(float)): Std values for normalizing pixels in the input image.
42
+ """
43
+ super().__init__()
44
+ self.image_encoder = image_encoder
45
+ self.prompt_encoder = prompt_encoder
46
+ self.mask_decoder = mask_decoder
47
+ self.register_buffer("pixel_mean", torch.Tensor(pixel_mean).view(-1, 1, 1), False)
48
+ self.register_buffer("pixel_std", torch.Tensor(pixel_std).view(-1, 1, 1), False)
49
+
50
+ @property
51
+ def device(self) -> Any:
52
+ return self.pixel_mean.device
53
+
54
+ @torch.no_grad()
55
+ def forward(
56
+ self,
57
+ batched_input: List[Dict[str, Any]],
58
+ multimask_output: bool,
59
+ ) -> List[Dict[str, torch.Tensor]]:
60
+ """
61
+ Predicts masks end-to-end from provided images and prompts.
62
+ If prompts are not known in advance, using SamPredictor is
63
+ recommended over calling the model directly.
64
+
65
+ Arguments:
66
+ batched_input (list(dict)): A list over input images, each a
67
+ dictionary with the following keys. A prompt key can be
68
+ excluded if it is not present.
69
+ 'image': The image as a torch tensor in 3xHxW format,
70
+ already transformed for input to the model.
71
+ 'original_size': (tuple(int, int)) The original size of
72
+ the image before transformation, as (H, W).
73
+ 'point_coords': (torch.Tensor) Batched point prompts for
74
+ this image, with shape BxNx2. Already transformed to the
75
+ input frame of the model.
76
+ 'point_labels': (torch.Tensor) Batched labels for point prompts,
77
+ with shape BxN.
78
+ 'boxes': (torch.Tensor) Batched box inputs, with shape Bx4.
79
+ Already transformed to the input frame of the model.
80
+ 'mask_inputs': (torch.Tensor) Batched mask inputs to the model,
81
+ in the form Bx1xHxW.
82
+ multimask_output (bool): Whether the model should predict multiple
83
+ disambiguating masks, or return a single mask.
84
+
85
+ Returns:
86
+ (list(dict)): A list over input images, where each element is
87
+ as dictionary with the following keys.
88
+ 'masks': (torch.Tensor) Batched binary mask predictions,
89
+ with shape BxCxHxW, where B is the number of input prompts,
90
+ C is determined by multimask_output, and (H, W) is the
91
+ original size of the image.
92
+ 'iou_predictions': (torch.Tensor) The model's predictions
93
+ of mask quality, in shape BxC.
94
+ 'low_res_logits': (torch.Tensor) Low resolution logits with
95
+ shape BxCxHxW, where H=W=256. Can be passed as mask input
96
+ to subsequent iterations of prediction.
97
+ """
98
+ input_images = torch.stack([self.preprocess(x["image"]) for x in batched_input], dim=0)
99
+ image_embeddings = self.image_encoder(input_images)
100
+
101
+ outputs = []
102
+ for image_record, curr_embedding in zip(batched_input, image_embeddings):
103
+ if "point_coords" in image_record:
104
+ points = (image_record["point_coords"], image_record["point_labels"])
105
+ else:
106
+ points = None
107
+ sparse_embeddings, dense_embeddings = self.prompt_encoder(
108
+ points=points,
109
+ boxes=image_record.get("boxes", None),
110
+ masks=image_record.get("mask_inputs", None),
111
+ )
112
+ low_res_masks, iou_predictions = self.mask_decoder(
113
+ image_embeddings=curr_embedding.unsqueeze(0),
114
+ image_pe=self.prompt_encoder.get_dense_pe(),
115
+ sparse_prompt_embeddings=sparse_embeddings,
116
+ dense_prompt_embeddings=dense_embeddings,
117
+ multimask_output=multimask_output,
118
+ )
119
+ masks = self.postprocess_masks(
120
+ low_res_masks,
121
+ input_size=image_record["image"].shape[-2:],
122
+ original_size=image_record["original_size"],
123
+ )
124
+ masks = masks > self.mask_threshold
125
+ outputs.append(
126
+ {
127
+ "masks": masks,
128
+ "iou_predictions": iou_predictions,
129
+ "low_res_logits": low_res_masks,
130
+ }
131
+ )
132
+ return outputs
133
+
134
+ def postprocess_masks(
135
+ self,
136
+ masks: torch.Tensor,
137
+ input_size: Tuple[int, ...],
138
+ original_size: Tuple[int, ...],
139
+ ) -> torch.Tensor:
140
+ """
141
+ Remove padding and upscale masks to the original image size.
142
+
143
+ Arguments:
144
+ masks (torch.Tensor): Batched masks from the mask_decoder,
145
+ in BxCxHxW format.
146
+ input_size (tuple(int, int)): The size of the image input to the
147
+ model, in (H, W) format. Used to remove padding.
148
+ original_size (tuple(int, int)): The original size of the image
149
+ before resizing for input to the model, in (H, W) format.
150
+
151
+ Returns:
152
+ (torch.Tensor): Batched masks in BxCxHxW format, where (H, W)
153
+ is given by original_size.
154
+ """
155
+ masks = F.interpolate(
156
+ masks,
157
+ (self.image_encoder.img_size, self.image_encoder.img_size),
158
+ mode="bilinear",
159
+ align_corners=False,
160
+ )
161
+ masks = masks[..., : input_size[0], : input_size[1]]
162
+ masks = F.interpolate(masks, original_size, mode="bilinear", align_corners=False)
163
+ return masks
164
+
165
+ def preprocess(self, x: torch.Tensor) -> torch.Tensor:
166
+ """Normalize pixel values and pad to a square input."""
167
+ # Normalize colors
168
+ x = (x - self.pixel_mean) / self.pixel_std
169
+
170
+ # Pad
171
+ h, w = x.shape[-2:]
172
+ padh = self.image_encoder.img_size - h
173
+ padw = self.image_encoder.img_size - w
174
+ x = F.pad(x, (0, padw, 0, padh))
175
+ return x
176
+
177
+ def preprocess_mask(self, x: torch.Tensor) -> torch.Tensor:
178
+ h, w = x.shape[-2:]
179
+ padh = self.image_encoder.img_size - h
180
+ padw = self.image_encoder.img_size - w
181
+ x = F.pad(x, (0, padw, 0, padh))
182
+ return x
app/mobile_sam/modeling/tiny_vit_sam.py ADDED
@@ -0,0 +1,718 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # --------------------------------------------------------
2
+ # TinyViT Model Architecture
3
+ # Copyright (c) 2022 Microsoft
4
+ # Adapted from LeViT and Swin Transformer
5
+ # LeViT: (https://github.com/facebookresearch/levit)
6
+ # Swin: (https://github.com/microsoft/swin-transformer)
7
+ # Build the TinyViT Model
8
+ # --------------------------------------------------------
9
+
10
+ import itertools
11
+ import torch
12
+ import torch.nn as nn
13
+ import torch.nn.functional as F
14
+ import torch.utils.checkpoint as checkpoint
15
+ from timm.models.layers import DropPath as TimmDropPath,\
16
+ to_2tuple, trunc_normal_
17
+ from timm.models.registry import register_model
18
+ from typing import Tuple
19
+
20
+
21
+ class Conv2d_BN(torch.nn.Sequential):
22
+ def __init__(self, a, b, ks=1, stride=1, pad=0, dilation=1,
23
+ groups=1, bn_weight_init=1):
24
+ super().__init__()
25
+ self.add_module('c', torch.nn.Conv2d(
26
+ a, b, ks, stride, pad, dilation, groups, bias=False))
27
+ bn = torch.nn.BatchNorm2d(b)
28
+ torch.nn.init.constant_(bn.weight, bn_weight_init)
29
+ torch.nn.init.constant_(bn.bias, 0)
30
+ self.add_module('bn', bn)
31
+
32
+ @torch.no_grad()
33
+ def fuse(self):
34
+ c, bn = self._modules.values()
35
+ w = bn.weight / (bn.running_var + bn.eps)**0.5
36
+ w = c.weight * w[:, None, None, None]
37
+ b = bn.bias - bn.running_mean * bn.weight / \
38
+ (bn.running_var + bn.eps)**0.5
39
+ m = torch.nn.Conv2d(w.size(1) * self.c.groups, w.size(
40
+ 0), w.shape[2:], stride=self.c.stride, padding=self.c.padding, dilation=self.c.dilation, groups=self.c.groups)
41
+ m.weight.data.copy_(w)
42
+ m.bias.data.copy_(b)
43
+ return m
44
+
45
+
46
+ class DropPath(TimmDropPath):
47
+ def __init__(self, drop_prob=None):
48
+ super().__init__(drop_prob=drop_prob)
49
+ self.drop_prob = drop_prob
50
+
51
+ def __repr__(self):
52
+ msg = super().__repr__()
53
+ msg += f'(drop_prob={self.drop_prob})'
54
+ return msg
55
+
56
+
57
+ class PatchEmbed(nn.Module):
58
+ def __init__(self, in_chans, embed_dim, resolution, activation):
59
+ super().__init__()
60
+ img_size: Tuple[int, int] = to_2tuple(resolution)
61
+ self.patches_resolution = (img_size[0] // 4, img_size[1] // 4)
62
+ self.num_patches = self.patches_resolution[0] * \
63
+ self.patches_resolution[1]
64
+ self.in_chans = in_chans
65
+ self.embed_dim = embed_dim
66
+ n = embed_dim
67
+ self.seq = nn.Sequential(
68
+ Conv2d_BN(in_chans, n // 2, 3, 2, 1),
69
+ activation(),
70
+ Conv2d_BN(n // 2, n, 3, 2, 1),
71
+ )
72
+
73
+ def forward(self, x):
74
+ return self.seq(x)
75
+
76
+
77
+ class MBConv(nn.Module):
78
+ def __init__(self, in_chans, out_chans, expand_ratio,
79
+ activation, drop_path):
80
+ super().__init__()
81
+ self.in_chans = in_chans
82
+ self.hidden_chans = int(in_chans * expand_ratio)
83
+ self.out_chans = out_chans
84
+
85
+ self.conv1 = Conv2d_BN(in_chans, self.hidden_chans, ks=1)
86
+ self.act1 = activation()
87
+
88
+ self.conv2 = Conv2d_BN(self.hidden_chans, self.hidden_chans,
89
+ ks=3, stride=1, pad=1, groups=self.hidden_chans)
90
+ self.act2 = activation()
91
+
92
+ self.conv3 = Conv2d_BN(
93
+ self.hidden_chans, out_chans, ks=1, bn_weight_init=0.0)
94
+ self.act3 = activation()
95
+
96
+ self.drop_path = DropPath(
97
+ drop_path) if drop_path > 0. else nn.Identity()
98
+
99
+ def forward(self, x):
100
+ shortcut = x
101
+
102
+ x = self.conv1(x)
103
+ x = self.act1(x)
104
+
105
+ x = self.conv2(x)
106
+ x = self.act2(x)
107
+
108
+ x = self.conv3(x)
109
+
110
+ x = self.drop_path(x)
111
+
112
+ x += shortcut
113
+ x = self.act3(x)
114
+
115
+ return x
116
+
117
+
118
+ class PatchMerging(nn.Module):
119
+ def __init__(self, input_resolution, dim, out_dim, activation):
120
+ super().__init__()
121
+
122
+ self.input_resolution = input_resolution
123
+ self.dim = dim
124
+ self.out_dim = out_dim
125
+ self.act = activation()
126
+ self.conv1 = Conv2d_BN(dim, out_dim, 1, 1, 0)
127
+ stride_c=2
128
+ if(out_dim==320 or out_dim==448 or out_dim==576):
129
+ stride_c=1
130
+ self.conv2 = Conv2d_BN(out_dim, out_dim, 3, stride_c, 1, groups=out_dim)
131
+ self.conv3 = Conv2d_BN(out_dim, out_dim, 1, 1, 0)
132
+
133
+ def forward(self, x):
134
+ if x.ndim == 3:
135
+ H, W = self.input_resolution
136
+ B = len(x)
137
+ # (B, C, H, W)
138
+ x = x.view(B, H, W, -1).permute(0, 3, 1, 2)
139
+
140
+ x = self.conv1(x)
141
+ x = self.act(x)
142
+
143
+ x = self.conv2(x)
144
+ x = self.act(x)
145
+ x = self.conv3(x)
146
+ x = x.flatten(2).transpose(1, 2)
147
+ return x
148
+
149
+
150
+ class ConvLayer(nn.Module):
151
+ def __init__(self, dim, input_resolution, depth,
152
+ activation,
153
+ drop_path=0., downsample=None, use_checkpoint=False,
154
+ out_dim=None,
155
+ conv_expand_ratio=4.,
156
+ ):
157
+
158
+ super().__init__()
159
+ self.dim = dim
160
+ self.input_resolution = input_resolution
161
+ self.depth = depth
162
+ self.use_checkpoint = use_checkpoint
163
+
164
+ # build blocks
165
+ self.blocks = nn.ModuleList([
166
+ MBConv(dim, dim, conv_expand_ratio, activation,
167
+ drop_path[i] if isinstance(drop_path, list) else drop_path,
168
+ )
169
+ for i in range(depth)])
170
+
171
+ # patch merging layer
172
+ if downsample is not None:
173
+ self.downsample = downsample(
174
+ input_resolution, dim=dim, out_dim=out_dim, activation=activation)
175
+ else:
176
+ self.downsample = None
177
+
178
+ def forward(self, x):
179
+ for blk in self.blocks:
180
+ if self.use_checkpoint:
181
+ x = checkpoint.checkpoint(blk, x)
182
+ else:
183
+ x = blk(x)
184
+ if self.downsample is not None:
185
+ x = self.downsample(x)
186
+ return x
187
+
188
+
189
+ class Mlp(nn.Module):
190
+ def __init__(self, in_features, hidden_features=None,
191
+ out_features=None, act_layer=nn.GELU, drop=0.):
192
+ super().__init__()
193
+ out_features = out_features or in_features
194
+ hidden_features = hidden_features or in_features
195
+ self.norm = nn.LayerNorm(in_features)
196
+ self.fc1 = nn.Linear(in_features, hidden_features)
197
+ self.fc2 = nn.Linear(hidden_features, out_features)
198
+ self.act = act_layer()
199
+ self.drop = nn.Dropout(drop)
200
+
201
+ def forward(self, x):
202
+ x = self.norm(x)
203
+
204
+ x = self.fc1(x)
205
+ x = self.act(x)
206
+ x = self.drop(x)
207
+ x = self.fc2(x)
208
+ x = self.drop(x)
209
+ return x
210
+
211
+
212
+ class Attention(torch.nn.Module):
213
+ def __init__(self, dim, key_dim, num_heads=8,
214
+ attn_ratio=4,
215
+ resolution=(14, 14),
216
+ ):
217
+ super().__init__()
218
+ # (h, w)
219
+ assert isinstance(resolution, tuple) and len(resolution) == 2
220
+ self.num_heads = num_heads
221
+ self.scale = key_dim ** -0.5
222
+ self.key_dim = key_dim
223
+ self.nh_kd = nh_kd = key_dim * num_heads
224
+ self.d = int(attn_ratio * key_dim)
225
+ self.dh = int(attn_ratio * key_dim) * num_heads
226
+ self.attn_ratio = attn_ratio
227
+ h = self.dh + nh_kd * 2
228
+
229
+ self.norm = nn.LayerNorm(dim)
230
+ self.qkv = nn.Linear(dim, h)
231
+ self.proj = nn.Linear(self.dh, dim)
232
+
233
+ points = list(itertools.product(
234
+ range(resolution[0]), range(resolution[1])))
235
+ N = len(points)
236
+ attention_offsets = {}
237
+ idxs = []
238
+ for p1 in points:
239
+ for p2 in points:
240
+ offset = (abs(p1[0] - p2[0]), abs(p1[1] - p2[1]))
241
+ if offset not in attention_offsets:
242
+ attention_offsets[offset] = len(attention_offsets)
243
+ idxs.append(attention_offsets[offset])
244
+ self.attention_biases = torch.nn.Parameter(
245
+ torch.zeros(num_heads, len(attention_offsets)))
246
+ self.register_buffer('attention_bias_idxs',
247
+ torch.LongTensor(idxs).view(N, N),
248
+ persistent=False)
249
+
250
+ @torch.no_grad()
251
+ def train(self, mode=True):
252
+ super().train(mode)
253
+ if mode and hasattr(self, 'ab'):
254
+ del self.ab
255
+ else:
256
+ self.register_buffer('ab',
257
+ self.attention_biases[:, self.attention_bias_idxs],
258
+ persistent=False)
259
+
260
+ def forward(self, x): # x (B,N,C)
261
+ B, N, _ = x.shape
262
+
263
+ # Normalization
264
+ x = self.norm(x)
265
+
266
+ qkv = self.qkv(x)
267
+ # (B, N, num_heads, d)
268
+ q, k, v = qkv.view(B, N, self.num_heads, -
269
+ 1).split([self.key_dim, self.key_dim, self.d], dim=3)
270
+ # (B, num_heads, N, d)
271
+ q = q.permute(0, 2, 1, 3)
272
+ k = k.permute(0, 2, 1, 3)
273
+ v = v.permute(0, 2, 1, 3)
274
+
275
+ attn = (
276
+ (q @ k.transpose(-2, -1)) * self.scale
277
+ +
278
+ (self.attention_biases[:, self.attention_bias_idxs]
279
+ if self.training else self.ab)
280
+ )
281
+ attn = attn.softmax(dim=-1)
282
+ x = (attn @ v).transpose(1, 2).reshape(B, N, self.dh)
283
+ x = self.proj(x)
284
+ return x
285
+
286
+
287
+ class TinyViTBlock(nn.Module):
288
+ r""" TinyViT Block.
289
+
290
+ Args:
291
+ dim (int): Number of input channels.
292
+ input_resolution (tuple[int, int]): Input resolution.
293
+ num_heads (int): Number of attention heads.
294
+ window_size (int): Window size.
295
+ mlp_ratio (float): Ratio of mlp hidden dim to embedding dim.
296
+ drop (float, optional): Dropout rate. Default: 0.0
297
+ drop_path (float, optional): Stochastic depth rate. Default: 0.0
298
+ local_conv_size (int): the kernel size of the convolution between
299
+ Attention and MLP. Default: 3
300
+ activation: the activation function. Default: nn.GELU
301
+ """
302
+
303
+ def __init__(self, dim, input_resolution, num_heads, window_size=7,
304
+ mlp_ratio=4., drop=0., drop_path=0.,
305
+ local_conv_size=3,
306
+ activation=nn.GELU,
307
+ ):
308
+ super().__init__()
309
+ self.dim = dim
310
+ self.input_resolution = input_resolution
311
+ self.num_heads = num_heads
312
+ assert window_size > 0, 'window_size must be greater than 0'
313
+ self.window_size = window_size
314
+ self.mlp_ratio = mlp_ratio
315
+
316
+ self.drop_path = DropPath(
317
+ drop_path) if drop_path > 0. else nn.Identity()
318
+
319
+ assert dim % num_heads == 0, 'dim must be divisible by num_heads'
320
+ head_dim = dim // num_heads
321
+
322
+ window_resolution = (window_size, window_size)
323
+ self.attn = Attention(dim, head_dim, num_heads,
324
+ attn_ratio=1, resolution=window_resolution)
325
+
326
+ mlp_hidden_dim = int(dim * mlp_ratio)
327
+ mlp_activation = activation
328
+ self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim,
329
+ act_layer=mlp_activation, drop=drop)
330
+
331
+ pad = local_conv_size // 2
332
+ self.local_conv = Conv2d_BN(
333
+ dim, dim, ks=local_conv_size, stride=1, pad=pad, groups=dim)
334
+
335
+ def forward(self, x):
336
+ H, W = self.input_resolution
337
+ B, L, C = x.shape
338
+ assert L == H * W, "input feature has wrong size"
339
+ res_x = x
340
+ if H == self.window_size and W == self.window_size:
341
+ x = self.attn(x)
342
+ else:
343
+ x = x.view(B, H, W, C)
344
+ pad_b = (self.window_size - H %
345
+ self.window_size) % self.window_size
346
+ pad_r = (self.window_size - W %
347
+ self.window_size) % self.window_size
348
+ padding = pad_b > 0 or pad_r > 0
349
+
350
+ if padding:
351
+ x = F.pad(x, (0, 0, 0, pad_r, 0, pad_b))
352
+
353
+ pH, pW = H + pad_b, W + pad_r
354
+ nH = pH // self.window_size
355
+ nW = pW // self.window_size
356
+ # window partition
357
+ x = x.view(B, nH, self.window_size, nW, self.window_size, C).transpose(2, 3).reshape(
358
+ B * nH * nW, self.window_size * self.window_size, C)
359
+ x = self.attn(x)
360
+ # window reverse
361
+ x = x.view(B, nH, nW, self.window_size, self.window_size,
362
+ C).transpose(2, 3).reshape(B, pH, pW, C)
363
+
364
+ if padding:
365
+ x = x[:, :H, :W].contiguous()
366
+
367
+ x = x.view(B, L, C)
368
+
369
+ x = res_x + self.drop_path(x)
370
+
371
+ x = x.transpose(1, 2).reshape(B, C, H, W)
372
+ x = self.local_conv(x)
373
+ x = x.view(B, C, L).transpose(1, 2)
374
+
375
+ x = x + self.drop_path(self.mlp(x))
376
+ return x
377
+
378
+ def extra_repr(self) -> str:
379
+ return f"dim={self.dim}, input_resolution={self.input_resolution}, num_heads={self.num_heads}, " \
380
+ f"window_size={self.window_size}, mlp_ratio={self.mlp_ratio}"
381
+
382
+
383
+ class BasicLayer(nn.Module):
384
+ """ A basic TinyViT layer for one stage.
385
+
386
+ Args:
387
+ dim (int): Number of input channels.
388
+ input_resolution (tuple[int]): Input resolution.
389
+ depth (int): Number of blocks.
390
+ num_heads (int): Number of attention heads.
391
+ window_size (int): Local window size.
392
+ mlp_ratio (float): Ratio of mlp hidden dim to embedding dim.
393
+ drop (float, optional): Dropout rate. Default: 0.0
394
+ drop_path (float | tuple[float], optional): Stochastic depth rate. Default: 0.0
395
+ downsample (nn.Module | None, optional): Downsample layer at the end of the layer. Default: None
396
+ use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False.
397
+ local_conv_size: the kernel size of the depthwise convolution between attention and MLP. Default: 3
398
+ activation: the activation function. Default: nn.GELU
399
+ out_dim: the output dimension of the layer. Default: dim
400
+ """
401
+
402
+ def __init__(self, dim, input_resolution, depth, num_heads, window_size,
403
+ mlp_ratio=4., drop=0.,
404
+ drop_path=0., downsample=None, use_checkpoint=False,
405
+ local_conv_size=3,
406
+ activation=nn.GELU,
407
+ out_dim=None,
408
+ ):
409
+
410
+ super().__init__()
411
+ self.dim = dim
412
+ self.input_resolution = input_resolution
413
+ self.depth = depth
414
+ self.use_checkpoint = use_checkpoint
415
+
416
+ # build blocks
417
+ self.blocks = nn.ModuleList([
418
+ TinyViTBlock(dim=dim, input_resolution=input_resolution,
419
+ num_heads=num_heads, window_size=window_size,
420
+ mlp_ratio=mlp_ratio,
421
+ drop=drop,
422
+ drop_path=drop_path[i] if isinstance(
423
+ drop_path, list) else drop_path,
424
+ local_conv_size=local_conv_size,
425
+ activation=activation,
426
+ )
427
+ for i in range(depth)])
428
+
429
+ # patch merging layer
430
+ if downsample is not None:
431
+ self.downsample = downsample(
432
+ input_resolution, dim=dim, out_dim=out_dim, activation=activation)
433
+ else:
434
+ self.downsample = None
435
+
436
+ def forward(self, x):
437
+ for blk in self.blocks:
438
+ if self.use_checkpoint:
439
+ x = checkpoint.checkpoint(blk, x)
440
+ else:
441
+ x = blk(x)
442
+ if self.downsample is not None:
443
+ x = self.downsample(x)
444
+ return x
445
+
446
+ def extra_repr(self) -> str:
447
+ return f"dim={self.dim}, input_resolution={self.input_resolution}, depth={self.depth}"
448
+
449
+ class LayerNorm2d(nn.Module):
450
+ def __init__(self, num_channels: int, eps: float = 1e-6) -> None:
451
+ super().__init__()
452
+ self.weight = nn.Parameter(torch.ones(num_channels))
453
+ self.bias = nn.Parameter(torch.zeros(num_channels))
454
+ self.eps = eps
455
+
456
+ def forward(self, x: torch.Tensor) -> torch.Tensor:
457
+ u = x.mean(1, keepdim=True)
458
+ s = (x - u).pow(2).mean(1, keepdim=True)
459
+ x = (x - u) / torch.sqrt(s + self.eps)
460
+ x = self.weight[:, None, None] * x + self.bias[:, None, None]
461
+ return x
462
+ class TinyViT(nn.Module):
463
+ def __init__(self, img_size=224, in_chans=3, num_classes=1000,
464
+ embed_dims=[96, 192, 384, 768], depths=[2, 2, 6, 2],
465
+ num_heads=[3, 6, 12, 24],
466
+ window_sizes=[7, 7, 14, 7],
467
+ mlp_ratio=4.,
468
+ drop_rate=0.,
469
+ drop_path_rate=0.1,
470
+ use_checkpoint=False,
471
+ mbconv_expand_ratio=4.0,
472
+ local_conv_size=3,
473
+ layer_lr_decay=1.0,
474
+ ):
475
+ super().__init__()
476
+ self.img_size=img_size
477
+ self.num_classes = num_classes
478
+ self.depths = depths
479
+ self.num_layers = len(depths)
480
+ self.mlp_ratio = mlp_ratio
481
+
482
+ activation = nn.GELU
483
+
484
+ self.patch_embed = PatchEmbed(in_chans=in_chans,
485
+ embed_dim=embed_dims[0],
486
+ resolution=img_size,
487
+ activation=activation)
488
+
489
+ patches_resolution = self.patch_embed.patches_resolution
490
+ self.patches_resolution = patches_resolution
491
+
492
+ # stochastic depth
493
+ dpr = [x.item() for x in torch.linspace(0, drop_path_rate,
494
+ sum(depths))] # stochastic depth decay rule
495
+
496
+ # build layers
497
+ self.layers = nn.ModuleList()
498
+ for i_layer in range(self.num_layers):
499
+ kwargs = dict(dim=embed_dims[i_layer],
500
+ input_resolution=(patches_resolution[0] // (2 ** (i_layer-1 if i_layer == 3 else i_layer)),
501
+ patches_resolution[1] // (2 ** (i_layer-1 if i_layer == 3 else i_layer))),
502
+ # input_resolution=(patches_resolution[0] // (2 ** i_layer),
503
+ # patches_resolution[1] // (2 ** i_layer)),
504
+ depth=depths[i_layer],
505
+ drop_path=dpr[sum(depths[:i_layer]):sum(depths[:i_layer + 1])],
506
+ downsample=PatchMerging if (
507
+ i_layer < self.num_layers - 1) else None,
508
+ use_checkpoint=use_checkpoint,
509
+ out_dim=embed_dims[min(
510
+ i_layer + 1, len(embed_dims) - 1)],
511
+ activation=activation,
512
+ )
513
+ if i_layer == 0:
514
+ layer = ConvLayer(
515
+ conv_expand_ratio=mbconv_expand_ratio,
516
+ **kwargs,
517
+ )
518
+ else:
519
+ layer = BasicLayer(
520
+ num_heads=num_heads[i_layer],
521
+ window_size=window_sizes[i_layer],
522
+ mlp_ratio=self.mlp_ratio,
523
+ drop=drop_rate,
524
+ local_conv_size=local_conv_size,
525
+ **kwargs)
526
+ self.layers.append(layer)
527
+
528
+ # Classifier head
529
+ self.norm_head = nn.LayerNorm(embed_dims[-1])
530
+ self.head = nn.Linear(
531
+ embed_dims[-1], num_classes) if num_classes > 0 else torch.nn.Identity()
532
+
533
+ # init weights
534
+ self.apply(self._init_weights)
535
+ self.set_layer_lr_decay(layer_lr_decay)
536
+ self.neck = nn.Sequential(
537
+ nn.Conv2d(
538
+ embed_dims[-1],
539
+ 256,
540
+ kernel_size=1,
541
+ bias=False,
542
+ ),
543
+ LayerNorm2d(256),
544
+ nn.Conv2d(
545
+ 256,
546
+ 256,
547
+ kernel_size=3,
548
+ padding=1,
549
+ bias=False,
550
+ ),
551
+ LayerNorm2d(256),
552
+ )
553
+ def set_layer_lr_decay(self, layer_lr_decay):
554
+ decay_rate = layer_lr_decay
555
+
556
+ # layers -> blocks (depth)
557
+ depth = sum(self.depths)
558
+ lr_scales = [decay_rate ** (depth - i - 1) for i in range(depth)]
559
+ #print("LR SCALES:", lr_scales)
560
+
561
+ def _set_lr_scale(m, scale):
562
+ for p in m.parameters():
563
+ p.lr_scale = scale
564
+
565
+ self.patch_embed.apply(lambda x: _set_lr_scale(x, lr_scales[0]))
566
+ i = 0
567
+ for layer in self.layers:
568
+ for block in layer.blocks:
569
+ block.apply(lambda x: _set_lr_scale(x, lr_scales[i]))
570
+ i += 1
571
+ if layer.downsample is not None:
572
+ layer.downsample.apply(
573
+ lambda x: _set_lr_scale(x, lr_scales[i - 1]))
574
+ assert i == depth
575
+ for m in [self.norm_head, self.head]:
576
+ m.apply(lambda x: _set_lr_scale(x, lr_scales[-1]))
577
+
578
+ for k, p in self.named_parameters():
579
+ p.param_name = k
580
+
581
+ def _check_lr_scale(m):
582
+ for p in m.parameters():
583
+ assert hasattr(p, 'lr_scale'), p.param_name
584
+
585
+ self.apply(_check_lr_scale)
586
+
587
+ def _init_weights(self, m):
588
+ if isinstance(m, nn.Linear):
589
+ trunc_normal_(m.weight, std=.02)
590
+ if isinstance(m, nn.Linear) and m.bias is not None:
591
+ nn.init.constant_(m.bias, 0)
592
+ elif isinstance(m, nn.LayerNorm):
593
+ nn.init.constant_(m.bias, 0)
594
+ nn.init.constant_(m.weight, 1.0)
595
+
596
+ @torch.jit.ignore
597
+ def no_weight_decay_keywords(self):
598
+ return {'attention_biases'}
599
+
600
+ def forward_features(self, x):
601
+ # x: (N, C, H, W)
602
+ x = self.patch_embed(x)
603
+
604
+ x = self.layers[0](x)
605
+ start_i = 1
606
+
607
+ for i in range(start_i, len(self.layers)):
608
+ layer = self.layers[i]
609
+ x = layer(x)
610
+ B,_,C=x.size()
611
+ x = x.view(B, 64, 64, C)
612
+ x=x.permute(0, 3, 1, 2)
613
+ x=self.neck(x)
614
+ return x
615
+
616
+ def forward(self, x):
617
+ x = self.forward_features(x)
618
+ #x = self.norm_head(x)
619
+ #x = self.head(x)
620
+ return x
621
+
622
+
623
+ _checkpoint_url_format = \
624
+ 'https://github.com/wkcn/TinyViT-model-zoo/releases/download/checkpoints/{}.pth'
625
+ _provided_checkpoints = {
626
+ 'tiny_vit_5m_224': 'tiny_vit_5m_22kto1k_distill',
627
+ 'tiny_vit_11m_224': 'tiny_vit_11m_22kto1k_distill',
628
+ 'tiny_vit_21m_224': 'tiny_vit_21m_22kto1k_distill',
629
+ 'tiny_vit_21m_384': 'tiny_vit_21m_22kto1k_384_distill',
630
+ 'tiny_vit_21m_512': 'tiny_vit_21m_22kto1k_512_distill',
631
+ }
632
+
633
+
634
+ def register_tiny_vit_model(fn):
635
+ '''Register a TinyViT model
636
+ It is a wrapper of `register_model` with loading the pretrained checkpoint.
637
+ '''
638
+ def fn_wrapper(pretrained=False, **kwargs):
639
+ model = fn()
640
+ if pretrained:
641
+ model_name = fn.__name__
642
+ assert model_name in _provided_checkpoints, \
643
+ f'Sorry that the checkpoint `{model_name}` is not provided yet.'
644
+ url = _checkpoint_url_format.format(
645
+ _provided_checkpoints[model_name])
646
+ checkpoint = torch.hub.load_state_dict_from_url(
647
+ url=url,
648
+ map_location='cpu', check_hash=False,
649
+ )
650
+ model.load_state_dict(checkpoint['model'])
651
+
652
+ return model
653
+
654
+ # rename the name of fn_wrapper
655
+ fn_wrapper.__name__ = fn.__name__
656
+ return register_model(fn_wrapper)
657
+
658
+
659
+ @register_tiny_vit_model
660
+ def tiny_vit_5m_224(pretrained=False, num_classes=1000, drop_path_rate=0.0):
661
+ return TinyViT(
662
+ num_classes=num_classes,
663
+ embed_dims=[64, 128, 160, 320],
664
+ depths=[2, 2, 6, 2],
665
+ num_heads=[2, 4, 5, 10],
666
+ window_sizes=[7, 7, 14, 7],
667
+ drop_path_rate=drop_path_rate,
668
+ )
669
+
670
+
671
+ @register_tiny_vit_model
672
+ def tiny_vit_11m_224(pretrained=False, num_classes=1000, drop_path_rate=0.1):
673
+ return TinyViT(
674
+ num_classes=num_classes,
675
+ embed_dims=[64, 128, 256, 448],
676
+ depths=[2, 2, 6, 2],
677
+ num_heads=[2, 4, 8, 14],
678
+ window_sizes=[7, 7, 14, 7],
679
+ drop_path_rate=drop_path_rate,
680
+ )
681
+
682
+
683
+ @register_tiny_vit_model
684
+ def tiny_vit_21m_224(pretrained=False, num_classes=1000, drop_path_rate=0.2):
685
+ return TinyViT(
686
+ num_classes=num_classes,
687
+ embed_dims=[96, 192, 384, 576],
688
+ depths=[2, 2, 6, 2],
689
+ num_heads=[3, 6, 12, 18],
690
+ window_sizes=[7, 7, 14, 7],
691
+ drop_path_rate=drop_path_rate,
692
+ )
693
+
694
+
695
+ @register_tiny_vit_model
696
+ def tiny_vit_21m_384(pretrained=False, num_classes=1000, drop_path_rate=0.1):
697
+ return TinyViT(
698
+ img_size=384,
699
+ num_classes=num_classes,
700
+ embed_dims=[96, 192, 384, 576],
701
+ depths=[2, 2, 6, 2],
702
+ num_heads=[3, 6, 12, 18],
703
+ window_sizes=[12, 12, 24, 12],
704
+ drop_path_rate=drop_path_rate,
705
+ )
706
+
707
+
708
+ @register_tiny_vit_model
709
+ def tiny_vit_21m_512(pretrained=False, num_classes=1000, drop_path_rate=0.1):
710
+ return TinyViT(
711
+ img_size=512,
712
+ num_classes=num_classes,
713
+ embed_dims=[96, 192, 384, 576],
714
+ depths=[2, 2, 6, 2],
715
+ num_heads=[3, 6, 12, 18],
716
+ window_sizes=[16, 16, 32, 16],
717
+ drop_path_rate=drop_path_rate,
718
+ )
app/mobile_sam/modeling/transformer.py ADDED
@@ -0,0 +1,240 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Copyright (c) Meta Platforms, Inc. and affiliates.
2
+ # All rights reserved.
3
+
4
+ # This source code is licensed under the license found in the
5
+ # LICENSE file in the root directory of this source tree.
6
+
7
+ import torch
8
+ from torch import Tensor, nn
9
+
10
+ import math
11
+ from typing import Tuple, Type
12
+
13
+ from .common import MLPBlock
14
+
15
+
16
+ class TwoWayTransformer(nn.Module):
17
+ def __init__(
18
+ self,
19
+ depth: int,
20
+ embedding_dim: int,
21
+ num_heads: int,
22
+ mlp_dim: int,
23
+ activation: Type[nn.Module] = nn.ReLU,
24
+ attention_downsample_rate: int = 2,
25
+ ) -> None:
26
+ """
27
+ A transformer decoder that attends to an input image using
28
+ queries whose positional embedding is supplied.
29
+
30
+ Args:
31
+ depth (int): number of layers in the transformer
32
+ embedding_dim (int): the channel dimension for the input embeddings
33
+ num_heads (int): the number of heads for multihead attention. Must
34
+ divide embedding_dim
35
+ mlp_dim (int): the channel dimension internal to the MLP block
36
+ activation (nn.Module): the activation to use in the MLP block
37
+ """
38
+ super().__init__()
39
+ self.depth = depth
40
+ self.embedding_dim = embedding_dim
41
+ self.num_heads = num_heads
42
+ self.mlp_dim = mlp_dim
43
+ self.layers = nn.ModuleList()
44
+
45
+ for i in range(depth):
46
+ self.layers.append(
47
+ TwoWayAttentionBlock(
48
+ embedding_dim=embedding_dim,
49
+ num_heads=num_heads,
50
+ mlp_dim=mlp_dim,
51
+ activation=activation,
52
+ attention_downsample_rate=attention_downsample_rate,
53
+ skip_first_layer_pe=(i == 0),
54
+ )
55
+ )
56
+
57
+ self.final_attn_token_to_image = Attention(
58
+ embedding_dim, num_heads, downsample_rate=attention_downsample_rate
59
+ )
60
+ self.norm_final_attn = nn.LayerNorm(embedding_dim)
61
+
62
+ def forward(
63
+ self,
64
+ image_embedding: Tensor,
65
+ image_pe: Tensor,
66
+ point_embedding: Tensor,
67
+ ) -> Tuple[Tensor, Tensor]:
68
+ """
69
+ Args:
70
+ image_embedding (torch.Tensor): image to attend to. Should be shape
71
+ B x embedding_dim x h x w for any h and w.
72
+ image_pe (torch.Tensor): the positional encoding to add to the image. Must
73
+ have the same shape as image_embedding.
74
+ point_embedding (torch.Tensor): the embedding to add to the query points.
75
+ Must have shape B x N_points x embedding_dim for any N_points.
76
+
77
+ Returns:
78
+ torch.Tensor: the processed point_embedding
79
+ torch.Tensor: the processed image_embedding
80
+ """
81
+ # BxCxHxW -> BxHWxC == B x N_image_tokens x C
82
+ bs, c, h, w = image_embedding.shape
83
+ image_embedding = image_embedding.flatten(2).permute(0, 2, 1)
84
+ image_pe = image_pe.flatten(2).permute(0, 2, 1)
85
+
86
+ # Prepare queries
87
+ queries = point_embedding
88
+ keys = image_embedding
89
+
90
+ # Apply transformer blocks and final layernorm
91
+ for layer in self.layers:
92
+ queries, keys = layer(
93
+ queries=queries,
94
+ keys=keys,
95
+ query_pe=point_embedding,
96
+ key_pe=image_pe,
97
+ )
98
+
99
+ # Apply the final attention layer from the points to the image
100
+ q = queries + point_embedding
101
+ k = keys + image_pe
102
+ attn_out = self.final_attn_token_to_image(q=q, k=k, v=keys)
103
+ queries = queries + attn_out
104
+ queries = self.norm_final_attn(queries)
105
+
106
+ return queries, keys
107
+
108
+
109
+ class TwoWayAttentionBlock(nn.Module):
110
+ def __init__(
111
+ self,
112
+ embedding_dim: int,
113
+ num_heads: int,
114
+ mlp_dim: int = 2048,
115
+ activation: Type[nn.Module] = nn.ReLU,
116
+ attention_downsample_rate: int = 2,
117
+ skip_first_layer_pe: bool = False,
118
+ ) -> None:
119
+ """
120
+ A transformer block with four layers: (1) self-attention of sparse
121
+ inputs, (2) cross attention of sparse inputs to dense inputs, (3) mlp
122
+ block on sparse inputs, and (4) cross attention of dense inputs to sparse
123
+ inputs.
124
+
125
+ Arguments:
126
+ embedding_dim (int): the channel dimension of the embeddings
127
+ num_heads (int): the number of heads in the attention layers
128
+ mlp_dim (int): the hidden dimension of the mlp block
129
+ activation (nn.Module): the activation of the mlp block
130
+ skip_first_layer_pe (bool): skip the PE on the first layer
131
+ """
132
+ super().__init__()
133
+ self.self_attn = Attention(embedding_dim, num_heads)
134
+ self.norm1 = nn.LayerNorm(embedding_dim)
135
+
136
+ self.cross_attn_token_to_image = Attention(
137
+ embedding_dim, num_heads, downsample_rate=attention_downsample_rate
138
+ )
139
+ self.norm2 = nn.LayerNorm(embedding_dim)
140
+
141
+ self.mlp = MLPBlock(embedding_dim, mlp_dim, activation)
142
+ self.norm3 = nn.LayerNorm(embedding_dim)
143
+
144
+ self.norm4 = nn.LayerNorm(embedding_dim)
145
+ self.cross_attn_image_to_token = Attention(
146
+ embedding_dim, num_heads, downsample_rate=attention_downsample_rate
147
+ )
148
+
149
+ self.skip_first_layer_pe = skip_first_layer_pe
150
+
151
+ def forward(
152
+ self, queries: Tensor, keys: Tensor, query_pe: Tensor, key_pe: Tensor
153
+ ) -> Tuple[Tensor, Tensor]:
154
+ # Self attention block
155
+ if self.skip_first_layer_pe:
156
+ queries = self.self_attn(q=queries, k=queries, v=queries)
157
+ else:
158
+ q = queries + query_pe
159
+ attn_out = self.self_attn(q=q, k=q, v=queries)
160
+ queries = queries + attn_out
161
+ queries = self.norm1(queries)
162
+
163
+ # Cross attention block, tokens attending to image embedding
164
+ q = queries + query_pe
165
+ k = keys + key_pe
166
+ attn_out = self.cross_attn_token_to_image(q=q, k=k, v=keys)
167
+ queries = queries + attn_out
168
+ queries = self.norm2(queries)
169
+
170
+ # MLP block
171
+ mlp_out = self.mlp(queries)
172
+ queries = queries + mlp_out
173
+ queries = self.norm3(queries)
174
+
175
+ # Cross attention block, image embedding attending to tokens
176
+ q = queries + query_pe
177
+ k = keys + key_pe
178
+ attn_out = self.cross_attn_image_to_token(q=k, k=q, v=queries)
179
+ keys = keys + attn_out
180
+ keys = self.norm4(keys)
181
+
182
+ return queries, keys
183
+
184
+
185
+ class Attention(nn.Module):
186
+ """
187
+ An attention layer that allows for downscaling the size of the embedding
188
+ after projection to queries, keys, and values.
189
+ """
190
+
191
+ def __init__(
192
+ self,
193
+ embedding_dim: int,
194
+ num_heads: int,
195
+ downsample_rate: int = 1,
196
+ ) -> None:
197
+ super().__init__()
198
+ self.embedding_dim = embedding_dim
199
+ self.internal_dim = embedding_dim // downsample_rate
200
+ self.num_heads = num_heads
201
+ assert self.internal_dim % num_heads == 0, "num_heads must divide embedding_dim."
202
+
203
+ self.q_proj = nn.Linear(embedding_dim, self.internal_dim)
204
+ self.k_proj = nn.Linear(embedding_dim, self.internal_dim)
205
+ self.v_proj = nn.Linear(embedding_dim, self.internal_dim)
206
+ self.out_proj = nn.Linear(self.internal_dim, embedding_dim)
207
+
208
+ def _separate_heads(self, x: Tensor, num_heads: int) -> Tensor:
209
+ b, n, c = x.shape
210
+ x = x.reshape(b, n, num_heads, c // num_heads)
211
+ return x.transpose(1, 2) # B x N_heads x N_tokens x C_per_head
212
+
213
+ def _recombine_heads(self, x: Tensor) -> Tensor:
214
+ b, n_heads, n_tokens, c_per_head = x.shape
215
+ x = x.transpose(1, 2)
216
+ return x.reshape(b, n_tokens, n_heads * c_per_head) # B x N_tokens x C
217
+
218
+ def forward(self, q: Tensor, k: Tensor, v: Tensor) -> Tensor:
219
+ # Input projections
220
+ q = self.q_proj(q)
221
+ k = self.k_proj(k)
222
+ v = self.v_proj(v)
223
+
224
+ # Separate into heads
225
+ q = self._separate_heads(q, self.num_heads)
226
+ k = self._separate_heads(k, self.num_heads)
227
+ v = self._separate_heads(v, self.num_heads)
228
+
229
+ # Attention
230
+ _, _, _, c_per_head = q.shape
231
+ attn = q @ k.permute(0, 1, 3, 2) # B x N_heads x N_tokens x N_tokens
232
+ attn = attn / math.sqrt(c_per_head)
233
+ attn = torch.softmax(attn, dim=-1)
234
+
235
+ # Get output
236
+ out = attn @ v
237
+ out = self._recombine_heads(out)
238
+ out = self.out_proj(out)
239
+
240
+ return out
app/mobile_sam/predictor.py ADDED
@@ -0,0 +1,281 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Copyright (c) Meta Platforms, Inc. and affiliates.
2
+ # All rights reserved.
3
+
4
+ # This source code is licensed under the license found in the
5
+ # LICENSE file in the root directory of this source tree.
6
+
7
+ import numpy as np
8
+ import torch
9
+
10
+ from .modeling import Sam
11
+
12
+ from typing import Optional, Tuple
13
+
14
+ from .utils.transforms import ResizeLongestSide
15
+
16
+
17
+ class SamPredictor:
18
+ def __init__(
19
+ self,
20
+ sam_model: Sam,
21
+ ) -> None:
22
+ """
23
+ Uses SAM to calculate the image embedding for an image, and then
24
+ allow repeated, efficient mask prediction given prompts.
25
+
26
+ Arguments:
27
+ sam_model (Sam): The model to use for mask prediction.
28
+ """
29
+ super().__init__()
30
+ self.model = sam_model
31
+ self.transform = ResizeLongestSide(sam_model.image_encoder.img_size)
32
+ self.reset_image()
33
+
34
+ def set_image(
35
+ self,
36
+ image: np.ndarray,
37
+ image_format: str = "RGB",
38
+ ) -> None:
39
+ """
40
+ Calculates the image embeddings for the provided image, allowing
41
+ masks to be predicted with the 'predict' method.
42
+
43
+ Arguments:
44
+ image (np.ndarray): The image for calculating masks. Expects an
45
+ image in HWC uint8 format, with pixel values in [0, 255].
46
+ image_format (str): The color format of the image, in ['RGB', 'BGR'].
47
+ """
48
+ assert image_format in [
49
+ "RGB",
50
+ "BGR",
51
+ ], f"image_format must be in ['RGB', 'BGR'], is {image_format}."
52
+ if image_format != self.model.image_format:
53
+ image = image[..., ::-1]
54
+
55
+ # Transform the image to the form expected by the model
56
+ input_image = self.transform.apply_image(image)
57
+ input_image_torch = torch.as_tensor(input_image, device=self.device)
58
+ input_image_torch = input_image_torch.permute(2, 0, 1).contiguous()[None, :, :, :]
59
+
60
+ self.set_torch_image(input_image_torch, image.shape[:2])
61
+
62
+ def get_mask(self, mask: np.ndarray) -> torch.Tensor:
63
+ mask = self.transform.apply_image(mask)
64
+ mask_torch = torch.as_tensor(mask, device=self.device)
65
+ mask_torch = mask_torch[None, None, :, :]
66
+ mask_torch = self.model.preprocess_mask(mask_torch)
67
+ return mask_torch
68
+
69
+ @torch.no_grad()
70
+ def set_torch_image(
71
+ self,
72
+ transformed_image: torch.Tensor,
73
+ original_image_size: Tuple[int, ...],
74
+ ) -> None:
75
+ """
76
+ Calculates the image embeddings for the provided image, allowing
77
+ masks to be predicted with the 'predict' method. Expects the input
78
+ image to be already transformed to the format expected by the model.
79
+
80
+ Arguments:
81
+ transformed_image (torch.Tensor): The input image, with shape
82
+ 1x3xHxW, which has been transformed with ResizeLongestSide.
83
+ original_image_size (tuple(int, int)): The size of the image
84
+ before transformation, in (H, W) format.
85
+ """
86
+ assert (
87
+ len(transformed_image.shape) == 4
88
+ and transformed_image.shape[1] == 3
89
+ and max(*transformed_image.shape[2:]) == self.model.image_encoder.img_size
90
+ ), f"set_torch_image input must be BCHW with long side {self.model.image_encoder.img_size}."
91
+ self.reset_image()
92
+
93
+ self.original_size = original_image_size
94
+ self.input_size = tuple(transformed_image.shape[-2:])
95
+ #import pdb; pdb.set_trace()
96
+ input_image = self.model.preprocess(transformed_image)
97
+ self.features = self.model.image_encoder(input_image)
98
+ self.is_image_set = True
99
+
100
+ def predict(
101
+ self,
102
+ point_coords: Optional[np.ndarray] = None,
103
+ point_labels: Optional[np.ndarray] = None,
104
+ box: Optional[np.ndarray] = None,
105
+ mask_input: Optional[np.ndarray] = None,
106
+ multimask_output: bool = True,
107
+ return_logits: bool = False,
108
+ return_numpy: bool = True,
109
+ ) -> Tuple[np.ndarray | torch.Tensor, np.ndarray | torch.Tensor, np.ndarray | torch.Tensor]:
110
+ """
111
+ Predict masks for the given input prompts, using the currently set image.
112
+
113
+ Arguments:
114
+ point_coords (np.ndarray or None): A Nx2 array of point prompts to the
115
+ model. Each point is in (X,Y) in pixels.
116
+ point_labels (np.ndarray or None): A length N array of labels for the
117
+ point prompts. 1 indicates a foreground point and 0 indicates a
118
+ background point.
119
+ box (np.ndarray or None): A length 4 array given a box prompt to the
120
+ model, in XYXY format.
121
+ mask_input (np.ndarray): A low resolution mask input to the model, typically
122
+ coming from a previous prediction iteration. Has form 1xHxW, where
123
+ for SAM, H=W=256.
124
+ multimask_output (bool): If true, the model will return three masks.
125
+ For ambiguous input prompts (such as a single click), this will often
126
+ produce better masks than a single prediction. If only a single
127
+ mask is needed, the model's predicted quality score can be used
128
+ to select the best mask. For non-ambiguous prompts, such as multiple
129
+ input prompts, multimask_output=False can give better results.
130
+ return_logits (bool): If true, returns un-thresholded masks logits
131
+ instead of a binary mask.
132
+
133
+ Returns:
134
+ (np.ndarray): The output masks in CxHxW format, where C is the
135
+ number of masks, and (H, W) is the original image size.
136
+ (np.ndarray): An array of length C containing the model's
137
+ predictions for the quality of each mask.
138
+ (np.ndarray): An array of shape CxHxW, where C is the number
139
+ of masks and H=W=256. These low resolution logits can be passed to
140
+ a subsequent iteration as mask input.
141
+ """
142
+ if not self.is_image_set:
143
+ raise RuntimeError("An image must be set with .set_image(...) before mask prediction.")
144
+
145
+ # Transform input prompts
146
+ coords_torch, labels_torch, box_torch, mask_input_torch = None, None, None, None
147
+ if point_coords is not None:
148
+ assert (
149
+ point_labels is not None
150
+ ), "point_labels must be supplied if point_coords is supplied."
151
+ point_coords = self.transform.apply_coords(point_coords, self.original_size)
152
+ coords_torch = torch.as_tensor(point_coords, dtype=torch.float, device=self.device)
153
+ labels_torch = torch.as_tensor(point_labels, dtype=torch.int, device=self.device)
154
+ coords_torch, labels_torch = coords_torch[None, :, :], labels_torch[None, :]
155
+ if box is not None:
156
+ box = self.transform.apply_boxes(box, self.original_size)
157
+ box_torch = torch.as_tensor(box, dtype=torch.float, device=self.device)
158
+ box_torch = box_torch[None, :]
159
+ if mask_input is not None:
160
+ mask_input_torch = torch.as_tensor(mask_input, dtype=torch.float, device=self.device)
161
+ mask_input_torch = mask_input_torch[None, :, :, :]
162
+
163
+ masks, iou_predictions, low_res_masks = self.predict_torch(
164
+ coords_torch,
165
+ labels_torch,
166
+ box_torch,
167
+ mask_input_torch,
168
+ multimask_output,
169
+ return_logits=return_logits,
170
+ )
171
+
172
+ if return_numpy:
173
+ masks_np = masks[0].detach().cpu().numpy()
174
+ iou_predictions_np = iou_predictions[0].detach().cpu().numpy()
175
+ low_res_masks_np = low_res_masks[0].detach().cpu().numpy()
176
+ return masks_np, iou_predictions_np, low_res_masks_np
177
+ else:
178
+ return masks[0], iou_predictions[0], low_res_masks[0]
179
+
180
+ @torch.no_grad()
181
+ def predict_torch(
182
+ self,
183
+ point_coords: Optional[torch.Tensor],
184
+ point_labels: Optional[torch.Tensor],
185
+ boxes: Optional[torch.Tensor] = None,
186
+ mask_input: Optional[torch.Tensor] = None,
187
+ multimask_output: bool = True,
188
+ return_logits: bool = False,
189
+ ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
190
+ """
191
+ Predict masks for the given input prompts, using the currently set image.
192
+ Input prompts are batched torch tensors and are expected to already be
193
+ transformed to the input frame using ResizeLongestSide.
194
+
195
+ Arguments:
196
+ point_coords (torch.Tensor or None): A BxNx2 array of point prompts to the
197
+ model. Each point is in (X,Y) in pixels.
198
+ point_labels (torch.Tensor or None): A BxN array of labels for the
199
+ point prompts. 1 indicates a foreground point and 0 indicates a
200
+ background point.
201
+ boxes (np.ndarray or None): A Bx4 array given a box prompt to the
202
+ model, in XYXY format.
203
+ mask_input (np.ndarray): A low resolution mask input to the model, typically
204
+ coming from a previous prediction iteration. Has form Bx1xHxW, where
205
+ for SAM, H=W=256. Masks returned by a previous iteration of the
206
+ predict method do not need further transformation.
207
+ multimask_output (bool): If true, the model will return three masks.
208
+ For ambiguous input prompts (such as a single click), this will often
209
+ produce better masks than a single prediction. If only a single
210
+ mask is needed, the model's predicted quality score can be used
211
+ to select the best mask. For non-ambiguous prompts, such as multiple
212
+ input prompts, multimask_output=False can give better results.
213
+ return_logits (bool): If true, returns un-thresholded masks logits
214
+ instead of a binary mask.
215
+
216
+ Returns:
217
+ (torch.Tensor): The output masks in BxCxHxW format, where C is the
218
+ number of masks, and (H, W) is the original image size.
219
+ (torch.Tensor): An array of shape BxC containing the model's
220
+ predictions for the quality of each mask.
221
+ (torch.Tensor): An array of shape BxCxHxW, where C is the number
222
+ of masks and H=W=256. These low res logits can be passed to
223
+ a subsequent iteration as mask input.
224
+ """
225
+ if not self.is_image_set:
226
+ raise RuntimeError("An image must be set with .set_image(...) before mask prediction.")
227
+
228
+ if point_coords is not None:
229
+ points = (point_coords, point_labels)
230
+ else:
231
+ points = None
232
+
233
+ # Embed prompts
234
+ sparse_embeddings, dense_embeddings = self.model.prompt_encoder(
235
+ points=points,
236
+ boxes=boxes,
237
+ masks=mask_input,
238
+ )
239
+
240
+ # Predict masks
241
+ low_res_masks, iou_predictions = self.model.mask_decoder(
242
+ image_embeddings=self.features,
243
+ image_pe=self.model.prompt_encoder.get_dense_pe(),
244
+ sparse_prompt_embeddings=sparse_embeddings,
245
+ dense_prompt_embeddings=dense_embeddings,
246
+ multimask_output=multimask_output,
247
+ )
248
+
249
+ # Upscale the masks to the original image resolution
250
+ masks = self.model.postprocess_masks(low_res_masks, self.input_size, self.original_size)
251
+
252
+ if not return_logits:
253
+ masks = masks > self.model.mask_threshold
254
+
255
+ return masks, iou_predictions, low_res_masks
256
+
257
+ def get_image_embedding(self) -> torch.Tensor:
258
+ """
259
+ Returns the image embeddings for the currently set image, with
260
+ shape 1xCxHxW, where C is the embedding dimension and (H,W) are
261
+ the embedding spatial dimension of SAM (typically C=256, H=W=64).
262
+ """
263
+ if not self.is_image_set:
264
+ raise RuntimeError(
265
+ "An image must be set with .set_image(...) to generate an embedding."
266
+ )
267
+ assert self.features is not None, "Features must exist if an image has been set."
268
+ return self.features
269
+
270
+ @property
271
+ def device(self) -> torch.device:
272
+ return self.model.device
273
+
274
+ def reset_image(self) -> None:
275
+ """Resets the currently set image."""
276
+ self.is_image_set = False
277
+ self.features = None
278
+ self.orig_h = None
279
+ self.orig_w = None
280
+ self.input_h = None
281
+ self.input_w = None
app/mobile_sam/utils/__init__.py ADDED
@@ -0,0 +1,7 @@
 
 
 
 
 
 
 
 
1
+ # Copyright (c) Meta Platforms, Inc. and affiliates.
2
+ # All rights reserved.
3
+
4
+ # This source code is licensed under the license found in the
5
+ # LICENSE file in the root directory of this source tree.
6
+
7
+ from .amg import batched_mask_to_box
app/mobile_sam/utils/amg.py ADDED
@@ -0,0 +1,347 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Copyright (c) Meta Platforms, Inc. and affiliates.
2
+ # All rights reserved.
3
+
4
+ # This source code is licensed under the license found in the
5
+ # LICENSE file in the root directory of this source tree.
6
+
7
+ import numpy as np
8
+ import torch
9
+
10
+ import math
11
+ from copy import deepcopy
12
+ from itertools import product
13
+ from typing import Any, Dict, Generator, ItemsView, List, Tuple
14
+
15
+
16
+ class MaskData:
17
+ """
18
+ A structure for storing masks and their related data in batched format.
19
+ Implements basic filtering and concatenation.
20
+ """
21
+
22
+ def __init__(self, **kwargs) -> None:
23
+ for v in kwargs.values():
24
+ assert isinstance(
25
+ v, (list, np.ndarray, torch.Tensor)
26
+ ), "MaskData only supports list, numpy arrays, and torch tensors."
27
+ self._stats = dict(**kwargs)
28
+
29
+ def __setitem__(self, key: str, item: Any) -> None:
30
+ assert isinstance(
31
+ item, (list, np.ndarray, torch.Tensor)
32
+ ), "MaskData only supports list, numpy arrays, and torch tensors."
33
+ self._stats[key] = item
34
+
35
+ def __delitem__(self, key: str) -> None:
36
+ del self._stats[key]
37
+
38
+ def __getitem__(self, key: str) -> Any:
39
+ return self._stats[key]
40
+
41
+ def items(self) -> ItemsView[str, Any]:
42
+ return self._stats.items()
43
+
44
+ def filter(self, keep: torch.Tensor) -> None:
45
+ for k, v in self._stats.items():
46
+ if v is None:
47
+ self._stats[k] = None
48
+ elif isinstance(v, torch.Tensor):
49
+ self._stats[k] = v[torch.as_tensor(keep, device=v.device)]
50
+ elif isinstance(v, np.ndarray):
51
+ self._stats[k] = v[keep.detach().cpu().numpy()]
52
+ elif isinstance(v, list) and keep.dtype == torch.bool:
53
+ self._stats[k] = [a for i, a in enumerate(v) if keep[i]]
54
+ elif isinstance(v, list):
55
+ self._stats[k] = [v[i] for i in keep]
56
+ else:
57
+ raise TypeError(f"MaskData key {k} has an unsupported type {type(v)}.")
58
+
59
+ def cat(self, new_stats: "MaskData") -> None:
60
+ for k, v in new_stats.items():
61
+ if k not in self._stats or self._stats[k] is None:
62
+ self._stats[k] = deepcopy(v)
63
+ elif isinstance(v, torch.Tensor):
64
+ self._stats[k] = torch.cat([self._stats[k], v], dim=0)
65
+ elif isinstance(v, np.ndarray):
66
+ self._stats[k] = np.concatenate([self._stats[k], v], axis=0)
67
+ elif isinstance(v, list):
68
+ self._stats[k] = self._stats[k] + deepcopy(v)
69
+ else:
70
+ raise TypeError(f"MaskData key {k} has an unsupported type {type(v)}.")
71
+
72
+ def to_numpy(self) -> None:
73
+ for k, v in self._stats.items():
74
+ if isinstance(v, torch.Tensor):
75
+ self._stats[k] = v.detach().cpu().numpy()
76
+
77
+
78
+ def is_box_near_crop_edge(
79
+ boxes: torch.Tensor, crop_box: List[int], orig_box: List[int], atol: float = 20.0
80
+ ) -> torch.Tensor:
81
+ """Filter masks at the edge of a crop, but not at the edge of the original image."""
82
+ crop_box_torch = torch.as_tensor(crop_box, dtype=torch.float, device=boxes.device)
83
+ orig_box_torch = torch.as_tensor(orig_box, dtype=torch.float, device=boxes.device)
84
+ boxes = uncrop_boxes_xyxy(boxes, crop_box).float()
85
+ near_crop_edge = torch.isclose(boxes, crop_box_torch[None, :], atol=atol, rtol=0)
86
+ near_image_edge = torch.isclose(boxes, orig_box_torch[None, :], atol=atol, rtol=0)
87
+ near_crop_edge = torch.logical_and(near_crop_edge, ~near_image_edge)
88
+ return torch.any(near_crop_edge, dim=1)
89
+
90
+
91
+ def box_xyxy_to_xywh(box_xyxy: torch.Tensor) -> torch.Tensor:
92
+ box_xywh = deepcopy(box_xyxy)
93
+ box_xywh[2] = box_xywh[2] - box_xywh[0]
94
+ box_xywh[3] = box_xywh[3] - box_xywh[1]
95
+ return box_xywh
96
+
97
+
98
+ def batch_iterator(batch_size: int, *args) -> Generator[List[Any], None, None]:
99
+ assert len(args) > 0 and all(
100
+ len(a) == len(args[0]) for a in args
101
+ ), "Batched iteration must have inputs of all the same size."
102
+ n_batches = len(args[0]) // batch_size + int(len(args[0]) % batch_size != 0)
103
+ for b in range(n_batches):
104
+ yield [arg[b * batch_size : (b + 1) * batch_size] for arg in args]
105
+
106
+
107
+ def mask_to_rle_pytorch(tensor: torch.Tensor) -> List[Dict[str, Any]]:
108
+ """
109
+ Encodes masks to an uncompressed RLE, in the format expected by
110
+ pycoco tools.
111
+ """
112
+ # Put in fortran order and flatten h,w
113
+ b, h, w = tensor.shape
114
+ tensor = tensor.permute(0, 2, 1).flatten(1)
115
+
116
+ # Compute change indices
117
+ diff = tensor[:, 1:] ^ tensor[:, :-1]
118
+ change_indices = diff.nonzero()
119
+
120
+ # Encode run length
121
+ out = []
122
+ for i in range(b):
123
+ cur_idxs = change_indices[change_indices[:, 0] == i, 1]
124
+ cur_idxs = torch.cat(
125
+ [
126
+ torch.tensor([0], dtype=cur_idxs.dtype, device=cur_idxs.device),
127
+ cur_idxs + 1,
128
+ torch.tensor([h * w], dtype=cur_idxs.dtype, device=cur_idxs.device),
129
+ ]
130
+ )
131
+ btw_idxs = cur_idxs[1:] - cur_idxs[:-1]
132
+ counts = [] if tensor[i, 0] == 0 else [0]
133
+ counts.extend(btw_idxs.detach().cpu().tolist())
134
+ out.append({"size": [h, w], "counts": counts})
135
+ return out
136
+
137
+
138
+ def rle_to_mask(rle: Dict[str, Any]) -> np.ndarray:
139
+ """Compute a binary mask from an uncompressed RLE."""
140
+ h, w = rle["size"]
141
+ mask = np.empty(h * w, dtype=bool)
142
+ idx = 0
143
+ parity = False
144
+ for count in rle["counts"]:
145
+ mask[idx : idx + count] = parity
146
+ idx += count
147
+ parity ^= True
148
+ mask = mask.reshape(w, h)
149
+ return mask.transpose() # Put in C order
150
+
151
+
152
+ def area_from_rle(rle: Dict[str, Any]) -> int:
153
+ return sum(rle["counts"][1::2])
154
+
155
+
156
+ def calculate_stability_score(
157
+ masks: torch.Tensor, mask_threshold: float, threshold_offset: float
158
+ ) -> torch.Tensor:
159
+ """
160
+ Computes the stability score for a batch of masks. The stability
161
+ score is the IoU between the binary masks obtained by thresholding
162
+ the predicted mask logits at high and low values.
163
+ """
164
+ # One mask is always contained inside the other.
165
+ # Save memory by preventing unnecessary cast to torch.int64
166
+ intersections = (
167
+ (masks > (mask_threshold + threshold_offset))
168
+ .sum(-1, dtype=torch.int16)
169
+ .sum(-1, dtype=torch.int32)
170
+ )
171
+ unions = (
172
+ (masks > (mask_threshold - threshold_offset))
173
+ .sum(-1, dtype=torch.int16)
174
+ .sum(-1, dtype=torch.int32)
175
+ )
176
+ return intersections / unions
177
+
178
+
179
+ def build_point_grid(n_per_side: int) -> np.ndarray:
180
+ """Generates a 2D grid of points evenly spaced in [0,1]x[0,1]."""
181
+ offset = 1 / (2 * n_per_side)
182
+ points_one_side = np.linspace(offset, 1 - offset, n_per_side)
183
+ points_x = np.tile(points_one_side[None, :], (n_per_side, 1))
184
+ points_y = np.tile(points_one_side[:, None], (1, n_per_side))
185
+ points = np.stack([points_x, points_y], axis=-1).reshape(-1, 2)
186
+ return points
187
+
188
+
189
+ def build_all_layer_point_grids(
190
+ n_per_side: int, n_layers: int, scale_per_layer: int
191
+ ) -> List[np.ndarray]:
192
+ """Generates point grids for all crop layers."""
193
+ points_by_layer = []
194
+ for i in range(n_layers + 1):
195
+ n_points = int(n_per_side / (scale_per_layer**i))
196
+ points_by_layer.append(build_point_grid(n_points))
197
+ return points_by_layer
198
+
199
+
200
+ def generate_crop_boxes(
201
+ im_size: Tuple[int, ...], n_layers: int, overlap_ratio: float
202
+ ) -> Tuple[List[List[int]], List[int]]:
203
+ """
204
+ Generates a list of crop boxes of different sizes. Each layer
205
+ has (2**i)**2 boxes for the ith layer.
206
+ """
207
+ crop_boxes, layer_idxs = [], []
208
+ im_h, im_w = im_size
209
+ short_side = min(im_h, im_w)
210
+
211
+ # Original image
212
+ crop_boxes.append([0, 0, im_w, im_h])
213
+ layer_idxs.append(0)
214
+
215
+ def crop_len(orig_len, n_crops, overlap):
216
+ return int(math.ceil((overlap * (n_crops - 1) + orig_len) / n_crops))
217
+
218
+ for i_layer in range(n_layers):
219
+ n_crops_per_side = 2 ** (i_layer + 1)
220
+ overlap = int(overlap_ratio * short_side * (2 / n_crops_per_side))
221
+
222
+ crop_w = crop_len(im_w, n_crops_per_side, overlap)
223
+ crop_h = crop_len(im_h, n_crops_per_side, overlap)
224
+
225
+ crop_box_x0 = [int((crop_w - overlap) * i) for i in range(n_crops_per_side)]
226
+ crop_box_y0 = [int((crop_h - overlap) * i) for i in range(n_crops_per_side)]
227
+
228
+ # Crops in XYWH format
229
+ for x0, y0 in product(crop_box_x0, crop_box_y0):
230
+ box = [x0, y0, min(x0 + crop_w, im_w), min(y0 + crop_h, im_h)]
231
+ crop_boxes.append(box)
232
+ layer_idxs.append(i_layer + 1)
233
+
234
+ return crop_boxes, layer_idxs
235
+
236
+
237
+ def uncrop_boxes_xyxy(boxes: torch.Tensor, crop_box: List[int]) -> torch.Tensor:
238
+ x0, y0, _, _ = crop_box
239
+ offset = torch.tensor([[x0, y0, x0, y0]], device=boxes.device)
240
+ # Check if boxes has a channel dimension
241
+ if len(boxes.shape) == 3:
242
+ offset = offset.unsqueeze(1)
243
+ return boxes + offset
244
+
245
+
246
+ def uncrop_points(points: torch.Tensor, crop_box: List[int]) -> torch.Tensor:
247
+ x0, y0, _, _ = crop_box
248
+ offset = torch.tensor([[x0, y0]], device=points.device)
249
+ # Check if points has a channel dimension
250
+ if len(points.shape) == 3:
251
+ offset = offset.unsqueeze(1)
252
+ return points + offset
253
+
254
+
255
+ def uncrop_masks(
256
+ masks: torch.Tensor, crop_box: List[int], orig_h: int, orig_w: int
257
+ ) -> torch.Tensor:
258
+ x0, y0, x1, y1 = crop_box
259
+ if x0 == 0 and y0 == 0 and x1 == orig_w and y1 == orig_h:
260
+ return masks
261
+ # Coordinate transform masks
262
+ pad_x, pad_y = orig_w - (x1 - x0), orig_h - (y1 - y0)
263
+ pad = (x0, pad_x - x0, y0, pad_y - y0)
264
+ return torch.nn.functional.pad(masks, pad, value=0)
265
+
266
+
267
+ def remove_small_regions(
268
+ mask: np.ndarray, area_thresh: float, mode: str
269
+ ) -> Tuple[np.ndarray, bool]:
270
+ """
271
+ Removes small disconnected regions and holes in a mask. Returns the
272
+ mask and an indicator of if the mask has been modified.
273
+ """
274
+ import cv2 # type: ignore
275
+
276
+ assert mode in ["holes", "islands"]
277
+ correct_holes = mode == "holes"
278
+ working_mask = (correct_holes ^ mask).astype(np.uint8)
279
+ n_labels, regions, stats, _ = cv2.connectedComponentsWithStats(working_mask, 8)
280
+ sizes = stats[:, -1][1:] # Row 0 is background label
281
+ small_regions = [i + 1 for i, s in enumerate(sizes) if s < area_thresh]
282
+ if len(small_regions) == 0:
283
+ return mask, False
284
+ fill_labels = [0] + small_regions
285
+ if not correct_holes:
286
+ fill_labels = [i for i in range(n_labels) if i not in fill_labels]
287
+ # If every region is below threshold, keep largest
288
+ if len(fill_labels) == 0:
289
+ fill_labels = [int(np.argmax(sizes)) + 1]
290
+ mask = np.isin(regions, fill_labels)
291
+ return mask, True
292
+
293
+
294
+ def coco_encode_rle(uncompressed_rle: Dict[str, Any]) -> Dict[str, Any]:
295
+ from pycocotools import mask as mask_utils # type: ignore
296
+
297
+ h, w = uncompressed_rle["size"]
298
+ rle = mask_utils.frPyObjects(uncompressed_rle, h, w)
299
+ rle["counts"] = rle["counts"].decode("utf-8") # Necessary to serialize with json
300
+ return rle
301
+
302
+
303
+ def batched_mask_to_box(masks: torch.Tensor) -> torch.Tensor:
304
+ """
305
+ Calculates boxes in XYXY format around masks. Return [0,0,0,0] for
306
+ an empty mask. For input shape C1xC2x...xHxW, the output shape is C1xC2x...x4.
307
+ """
308
+ # torch.max below raises an error on empty inputs, just skip in this case
309
+ if torch.numel(masks) == 0:
310
+ return torch.tensor([])
311
+ # return torch.zeros(*masks.shape[:-2], 4, device=masks.device)
312
+
313
+ # Normalize shape to CxHxW
314
+ shape = masks.shape
315
+ h, w = shape[-2:]
316
+ if len(shape) > 2:
317
+ masks = masks.flatten(0, -3)
318
+ else:
319
+ masks = masks.unsqueeze(0)
320
+
321
+ # Get top and bottom edges
322
+ in_height, _ = torch.max(masks, dim=-1)
323
+ in_height_coords = in_height * torch.arange(h, device=in_height.device)[None, :]
324
+ bottom_edges, _ = torch.max(in_height_coords.int(), dim=-1)
325
+ in_height_coords = in_height_coords + h * (~in_height)
326
+ top_edges, _ = torch.min(in_height_coords.int(), dim=-1)
327
+
328
+ # Get left and right edges
329
+ in_width, _ = torch.max(masks, dim=-2)
330
+ in_width_coords = in_width * torch.arange(w, device=in_width.device)[None, :]
331
+ right_edges, _ = torch.max(in_width_coords.int(), dim=-1)
332
+ in_width_coords = in_width_coords + w * (~in_width)
333
+ left_edges, _ = torch.min(in_width_coords.int(), dim=-1)
334
+
335
+ # If the mask is empty the right edge will be to the left of the left edge.
336
+ # Replace these boxes with [0, 0, 0, 0]
337
+ empty_filter = (right_edges < left_edges) | (bottom_edges < top_edges)
338
+ out = torch.stack([left_edges, top_edges, right_edges, bottom_edges], dim=-1)
339
+ out = out * (~empty_filter).unsqueeze(-1)
340
+
341
+ # Return to original shape
342
+ if len(shape) > 2:
343
+ out = out.reshape(*shape[:-2], 4)
344
+ else:
345
+ out = out[0]
346
+
347
+ return out
app/mobile_sam/utils/onnx.py ADDED
@@ -0,0 +1,144 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Copyright (c) Meta Platforms, Inc. and affiliates.
2
+ # All rights reserved.
3
+
4
+ # This source code is licensed under the license found in the
5
+ # LICENSE file in the root directory of this source tree.
6
+
7
+ import torch
8
+ import torch.nn as nn
9
+ from torch.nn import functional as F
10
+
11
+ from typing import Tuple
12
+
13
+ from ..modeling import Sam
14
+ from .amg import calculate_stability_score
15
+
16
+
17
+ class SamOnnxModel(nn.Module):
18
+ """
19
+ This model should not be called directly, but is used in ONNX export.
20
+ It combines the prompt encoder, mask decoder, and mask postprocessing of Sam,
21
+ with some functions modified to enable model tracing. Also supports extra
22
+ options controlling what information. See the ONNX export script for details.
23
+ """
24
+
25
+ def __init__(
26
+ self,
27
+ model: Sam,
28
+ return_single_mask: bool,
29
+ use_stability_score: bool = False,
30
+ return_extra_metrics: bool = False,
31
+ ) -> None:
32
+ super().__init__()
33
+ self.mask_decoder = model.mask_decoder
34
+ self.model = model
35
+ self.img_size = model.image_encoder.img_size
36
+ self.return_single_mask = return_single_mask
37
+ self.use_stability_score = use_stability_score
38
+ self.stability_score_offset = 1.0
39
+ self.return_extra_metrics = return_extra_metrics
40
+
41
+ @staticmethod
42
+ def resize_longest_image_size(
43
+ input_image_size: torch.Tensor, longest_side: int
44
+ ) -> torch.Tensor:
45
+ input_image_size = input_image_size.to(torch.float32)
46
+ scale = longest_side / torch.max(input_image_size)
47
+ transformed_size = scale * input_image_size
48
+ transformed_size = torch.floor(transformed_size + 0.5).to(torch.int64)
49
+ return transformed_size
50
+
51
+ def _embed_points(self, point_coords: torch.Tensor, point_labels: torch.Tensor) -> torch.Tensor:
52
+ point_coords = point_coords + 0.5
53
+ point_coords = point_coords / self.img_size
54
+ point_embedding = self.model.prompt_encoder.pe_layer._pe_encoding(point_coords)
55
+ point_labels = point_labels.unsqueeze(-1).expand_as(point_embedding)
56
+
57
+ point_embedding = point_embedding * (point_labels != -1)
58
+ point_embedding = point_embedding + self.model.prompt_encoder.not_a_point_embed.weight * (
59
+ point_labels == -1
60
+ )
61
+
62
+ for i in range(self.model.prompt_encoder.num_point_embeddings):
63
+ point_embedding = point_embedding + self.model.prompt_encoder.point_embeddings[
64
+ i
65
+ ].weight * (point_labels == i)
66
+
67
+ return point_embedding
68
+
69
+ def _embed_masks(self, input_mask: torch.Tensor, has_mask_input: torch.Tensor) -> torch.Tensor:
70
+ mask_embedding = has_mask_input * self.model.prompt_encoder.mask_downscaling(input_mask)
71
+ mask_embedding = mask_embedding + (
72
+ 1 - has_mask_input
73
+ ) * self.model.prompt_encoder.no_mask_embed.weight.reshape(1, -1, 1, 1)
74
+ return mask_embedding
75
+
76
+ def mask_postprocessing(self, masks: torch.Tensor, orig_im_size: torch.Tensor) -> torch.Tensor:
77
+ masks = F.interpolate(
78
+ masks,
79
+ size=(self.img_size, self.img_size),
80
+ mode="bilinear",
81
+ align_corners=False,
82
+ )
83
+
84
+ prepadded_size = self.resize_longest_image_size(orig_im_size, self.img_size).to(torch.int64)
85
+ masks = masks[..., : prepadded_size[0], : prepadded_size[1]] # type: ignore
86
+
87
+ orig_im_size = orig_im_size.to(torch.int64)
88
+ h, w = orig_im_size[0], orig_im_size[1]
89
+ masks = F.interpolate(masks, size=(h, w), mode="bilinear", align_corners=False)
90
+ return masks
91
+
92
+ def select_masks(
93
+ self, masks: torch.Tensor, iou_preds: torch.Tensor, num_points: int
94
+ ) -> Tuple[torch.Tensor, torch.Tensor]:
95
+ # Determine if we should return the multiclick mask or not from the number of points.
96
+ # The reweighting is used to avoid control flow.
97
+ score_reweight = torch.tensor(
98
+ [[1000] + [0] * (self.model.mask_decoder.num_mask_tokens - 1)]
99
+ ).to(iou_preds.device)
100
+ score = iou_preds + (num_points - 2.5) * score_reweight
101
+ best_idx = torch.argmax(score, dim=1)
102
+ masks = masks[torch.arange(masks.shape[0]), best_idx, :, :].unsqueeze(1)
103
+ iou_preds = iou_preds[torch.arange(masks.shape[0]), best_idx].unsqueeze(1)
104
+
105
+ return masks, iou_preds
106
+
107
+ @torch.no_grad()
108
+ def forward(
109
+ self,
110
+ image_embeddings: torch.Tensor,
111
+ point_coords: torch.Tensor,
112
+ point_labels: torch.Tensor,
113
+ mask_input: torch.Tensor,
114
+ has_mask_input: torch.Tensor,
115
+ orig_im_size: torch.Tensor,
116
+ ):
117
+ sparse_embedding = self._embed_points(point_coords, point_labels)
118
+ dense_embedding = self._embed_masks(mask_input, has_mask_input)
119
+
120
+ masks, scores = self.model.mask_decoder.predict_masks(
121
+ image_embeddings=image_embeddings,
122
+ image_pe=self.model.prompt_encoder.get_dense_pe(),
123
+ sparse_prompt_embeddings=sparse_embedding,
124
+ dense_prompt_embeddings=dense_embedding,
125
+ )
126
+
127
+ if self.use_stability_score:
128
+ scores = calculate_stability_score(
129
+ masks, self.model.mask_threshold, self.stability_score_offset
130
+ )
131
+
132
+ if self.return_single_mask:
133
+ masks, scores = self.select_masks(masks, scores, point_coords.shape[1])
134
+
135
+ upscaled_masks = self.mask_postprocessing(masks, orig_im_size)
136
+
137
+ if self.return_extra_metrics:
138
+ stability_scores = calculate_stability_score(
139
+ upscaled_masks, self.model.mask_threshold, self.stability_score_offset
140
+ )
141
+ areas = (upscaled_masks > self.model.mask_threshold).sum(-1).sum(-1)
142
+ return upscaled_masks, scores, stability_scores, areas, masks
143
+
144
+ return upscaled_masks, scores, masks
app/mobile_sam/utils/transforms.py ADDED
@@ -0,0 +1,102 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Copyright (c) Meta Platforms, Inc. and affiliates.
2
+ # All rights reserved.
3
+
4
+ # This source code is licensed under the license found in the
5
+ # LICENSE file in the root directory of this source tree.
6
+
7
+ import numpy as np
8
+ import torch
9
+ from torch.nn import functional as F
10
+ from torchvision.transforms.functional import resize, to_pil_image # type: ignore
11
+
12
+ from copy import deepcopy
13
+ from typing import Tuple
14
+
15
+
16
+ class ResizeLongestSide:
17
+ """
18
+ Resizes images to the longest side 'target_length', as well as provides
19
+ methods for resizing coordinates and boxes. Provides methods for
20
+ transforming both numpy array and batched torch tensors.
21
+ """
22
+
23
+ def __init__(self, target_length: int) -> None:
24
+ self.target_length = target_length
25
+
26
+ def apply_image(self, image: np.ndarray) -> np.ndarray:
27
+ """
28
+ Expects a numpy array with shape HxWxC in uint8 format.
29
+ """
30
+ target_size = self.get_preprocess_shape(image.shape[0], image.shape[1], self.target_length)
31
+ return np.array(resize(to_pil_image(image), target_size))
32
+
33
+ def apply_coords(self, coords: np.ndarray, original_size: Tuple[int, ...]) -> np.ndarray:
34
+ """
35
+ Expects a numpy array of length 2 in the final dimension. Requires the
36
+ original image size in (H, W) format.
37
+ """
38
+ old_h, old_w = original_size
39
+ new_h, new_w = self.get_preprocess_shape(
40
+ original_size[0], original_size[1], self.target_length
41
+ )
42
+ coords = deepcopy(coords).astype(float)
43
+ coords[..., 0] = coords[..., 0] * (new_w / old_w)
44
+ coords[..., 1] = coords[..., 1] * (new_h / old_h)
45
+ return coords
46
+
47
+ def apply_boxes(self, boxes: np.ndarray, original_size: Tuple[int, ...]) -> np.ndarray:
48
+ """
49
+ Expects a numpy array shape Bx4. Requires the original image size
50
+ in (H, W) format.
51
+ """
52
+ boxes = self.apply_coords(boxes.reshape(-1, 2, 2), original_size)
53
+ return boxes.reshape(-1, 4)
54
+
55
+ def apply_image_torch(self, image: torch.Tensor) -> torch.Tensor:
56
+ """
57
+ Expects batched images with shape BxCxHxW and float format. This
58
+ transformation may not exactly match apply_image. apply_image is
59
+ the transformation expected by the model.
60
+ """
61
+ # Expects an image in BCHW format. May not exactly match apply_image.
62
+ target_size = self.get_preprocess_shape(image.shape[2], image.shape[3], self.target_length)
63
+ return F.interpolate(
64
+ image, target_size, mode="bilinear", align_corners=False, antialias=True
65
+ )
66
+
67
+ def apply_coords_torch(
68
+ self, coords: torch.Tensor, original_size: Tuple[int, ...]
69
+ ) -> torch.Tensor:
70
+ """
71
+ Expects a torch tensor with length 2 in the last dimension. Requires the
72
+ original image size in (H, W) format.
73
+ """
74
+ old_h, old_w = original_size
75
+ new_h, new_w = self.get_preprocess_shape(
76
+ original_size[0], original_size[1], self.target_length
77
+ )
78
+ coords = deepcopy(coords).to(torch.float)
79
+ coords[..., 0] = coords[..., 0] * (new_w / old_w)
80
+ coords[..., 1] = coords[..., 1] * (new_h / old_h)
81
+ return coords
82
+
83
+ def apply_boxes_torch(
84
+ self, boxes: torch.Tensor, original_size: Tuple[int, ...]
85
+ ) -> torch.Tensor:
86
+ """
87
+ Expects a torch tensor with shape Bx4. Requires the original image
88
+ size in (H, W) format.
89
+ """
90
+ boxes = self.apply_coords_torch(boxes.reshape(-1, 2, 2), original_size)
91
+ return boxes.reshape(-1, 4)
92
+
93
+ @staticmethod
94
+ def get_preprocess_shape(oldh: int, oldw: int, long_side_length: int) -> Tuple[int, int]:
95
+ """
96
+ Compute the output size given input size and target long side length.
97
+ """
98
+ scale = long_side_length * 1.0 / max(oldh, oldw)
99
+ newh, neww = oldh * scale, oldw * scale
100
+ neww = int(neww + 0.5)
101
+ newh = int(newh + 0.5)
102
+ return (newh, neww)
app/model.py ADDED
@@ -0,0 +1,70 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import numpy.typing as npt
2
+ import time
3
+
4
+ # from .sam import build_sam, SamPredictor
5
+ # from .sam_hq import build_sam as build_sam_hq, SamPredictor as SamHqPredictor
6
+ # from .mobile_sam import (
7
+ # build_sam_vit_t as build_mobile_sam,
8
+ # SamPredictor as MobileSamPredictor,
9
+ # )
10
+ # from .per_sam import train, PerSAM
11
+ # from .configs import DEVICE
12
+
13
+ from app.sam import build_sam, SamPredictor
14
+ from app.sam_hq import build_sam as build_sam_hq, SamPredictor as SamHqPredictor
15
+ from app.mobile_sam import (
16
+ build_sam_vit_t as build_mobile_sam,
17
+ SamPredictor as MobileSamPredictor,
18
+ )
19
+ from app.per_sam import train, PerSAM
20
+ from app.configs import DEVICE
21
+
22
+
23
+ def build_sam_predictor(checkpoint: str | None = None):
24
+ sam = build_sam(checkpoint)
25
+ sam = sam.to(DEVICE)
26
+ return SamPredictor(sam)
27
+
28
+
29
+ def build_sam_hq_predictor(checkpoint: str | None = None):
30
+ sam = build_sam_hq(checkpoint)
31
+ sam = sam.to(DEVICE)
32
+ return SamHqPredictor(sam)
33
+
34
+
35
+ def build_mobile_sam_predictor(checkpoint: str | None = None):
36
+ sam = build_mobile_sam(checkpoint)
37
+ sam = sam.to(DEVICE)
38
+ return MobileSamPredictor(sam)
39
+
40
+
41
+ def get_multi_label_predictor(
42
+ sam: MobileSamPredictor, image: npt.NDArray, mask: npt.NDArray,
43
+ ) -> PerSAM:
44
+ start = time.perf_counter()
45
+ weights, target_feat = train(sam, [image], [mask])
46
+ print(f"training time {time.perf_counter() - start}")
47
+ per_sam_model = PerSAM(sam, target_feat, 100, 0.6, 0.2, weights)
48
+ return per_sam_model
49
+
50
+
51
+ if __name__ == "__main__":
52
+ import numpy as np
53
+ from PIL import Image
54
+ from torchvision.transforms.functional import resize
55
+ from app.transforms import ResizeLongestSide
56
+ T = ResizeLongestSide(1024)
57
+ image = Image.open("/Users/dillonlaird/code/instance_labeler/seals-labeled/img2.png").convert("RGB")
58
+ target_size = T.get_preprocess_shape(image.size[1], image.size[0], T.target_length)
59
+ image_np = np.array(resize(image, target_size))
60
+ mask = Image.open("/Users/dillonlaird/code/instance_labeler/seals-labeled/img2.seal.10.png")
61
+ target_size = T.get_preprocess_shape(mask.size[1], mask.size[0], T.target_length)
62
+ mask_np = np.array(resize(mask, target_size).convert("L"))
63
+
64
+ model = build_mobile_sam_predictor("/Users/dillonlaird/code/instance_labeler/mobile_sam.pth")
65
+ start = time.perf_counter()
66
+ per_sam_model = get_multi_label_predictor(model, image_np, mask_np)
67
+ print(f"training time {time.perf_counter() - start}")
68
+ start = time.perf_counter()
69
+ masks, bboxes, _ = per_sam_model(image_np)
70
+ print(f"prediction time {time.perf_counter() - start}")
app/per_sam/__init__.py ADDED
@@ -0,0 +1,2 @@
 
 
 
1
+ from .train import train
2
+ from .model import PerSAM
app/per_sam/loss.py ADDED
@@ -0,0 +1,45 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import numpy as np
2
+ import numpy.typing as npt
3
+
4
+ from torch.nn import functional as F
5
+ from torch import Tensor
6
+
7
+
8
+ def calculate_dice_loss(inputs: Tensor, targets: Tensor, num_masks: int = 1) -> Tensor:
9
+ inputs = inputs.sigmoid()
10
+ inputs = inputs.flatten(1)
11
+ numerator = 2 * (inputs * targets).sum(-1)
12
+ denominator = inputs.sum(-1) + targets.sum(-1)
13
+ loss = 1 - (numerator + 1) / (denominator + 1)
14
+ return loss.sum() / num_masks
15
+
16
+
17
+ def calculate_sigmoid_focal_loss(
18
+ inputs: Tensor,
19
+ targets: Tensor,
20
+ num_masks: int = 1,
21
+ alpha: float = 0.25,
22
+ gamma: float = 2,
23
+ ) -> Tensor:
24
+ prob = inputs.sigmoid()
25
+ ce_loss = F.binary_cross_entropy_with_logits(inputs, targets, reduction="none")
26
+ p_t = prob * targets + (1 - prob) * (1 - targets)
27
+ loss = ce_loss * ((1 - p_t) ** gamma)
28
+
29
+ if alpha >= 0:
30
+ alpha_t = alpha * targets + (1 - alpha) * (1 - targets)
31
+ loss = alpha_t * loss
32
+
33
+ return loss.mean(1).sum() / num_masks
34
+
35
+
36
+ def calculate_iou(mask1: npt.NDArray, mask2: npt.NDArray) -> float:
37
+ mask1 = mask1.sum(axis=2)
38
+ mask2 = mask2.sum(axis=2)
39
+
40
+ mask1 = np.where(mask1 == 128, 1, 0)
41
+ mask2 = np.where(mask2 == 128, 1, 0)
42
+ intersection = np.sum(np.logical_and(mask1, mask2))
43
+ union = np.sum(np.logical_or(mask1, mask2))
44
+ iou = intersection / union
45
+ return iou
app/per_sam/model.py ADDED
@@ -0,0 +1,204 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import numpy.typing as npt
2
+ import numpy as np
3
+ import torch
4
+ import torch.nn as nn
5
+ import torch.nn.functional as F
6
+ import cv2
7
+
8
+ from torchvision.ops.boxes import batched_nms
9
+ from app.mobile_sam import SamPredictor
10
+ from app.mobile_sam.utils import batched_mask_to_box
11
+
12
+
13
+ def point_selection(mask_sim, topk: int = 1):
14
+ # Top-1 point selection
15
+ _, h = mask_sim.shape
16
+ topk_xy = mask_sim.flatten(0).topk(topk)[1]
17
+ topk_x = (topk_xy // h).unsqueeze(0)
18
+ topk_y = topk_xy - topk_x * h
19
+ topk_xy = torch.cat((topk_y, topk_x), dim=0).permute(1, 0)
20
+ topk_label = np.array([1] * topk)
21
+ topk_xy = topk_xy.cpu().numpy()
22
+
23
+ return topk_xy, topk_label
24
+
25
+
26
+ def mask_nms(
27
+ masks: list[npt.NDArray], scores: list[float], iou_thresh: float = 0.2
28
+ ) -> tuple[list[npt.NDArray], list[float]]:
29
+ ious = np.zeros((len(masks), len(masks)))
30
+ np_masks = np.array(masks).astype(bool)
31
+ np_scores = np.array(scores)
32
+ remove_indices = set()
33
+ for i in range(len(masks)):
34
+ mask_i = np_masks[i, :, :]
35
+ intersection_sum = np.logical_and(mask_i, np_masks).sum(axis=(1, 2))
36
+ union = np.logical_or(mask_i, np_masks)
37
+ ious_i = intersection_sum / union.sum(axis=(1, 2))
38
+ ious[i, :] = ious_i
39
+
40
+ # if the mask completely overlaps another mask, take the highest
41
+ # scoring mask and remove the lower (current) one
42
+ overlap = intersection_sum >= np_masks.sum(axis=(1, 2)) * 0.90
43
+ argmax_idx = np_scores[overlap].argmax()
44
+ max_idx = np.where(overlap == True)[0][argmax_idx]
45
+ if max_idx != i:
46
+ remove_indices.add(i)
47
+
48
+ for i in range(ious.shape[0]):
49
+ ious_i = ious[i, :]
50
+ idxs = np.where(ious_i > iou_thresh)[0]
51
+ keep = idxs[np.argmax(np_scores[idxs])]
52
+ if keep != i:
53
+ remove_indices.add(i)
54
+
55
+ return [masks[i] for i in range(len(masks)) if i not in remove_indices], [
56
+ scores[i] for i in range(len(masks)) if i not in remove_indices
57
+ ]
58
+
59
+
60
+ class MaskWeights(nn.Module):
61
+ def __init__(self):
62
+ super().__init__()
63
+ self.weights = nn.Parameter(torch.ones(2, 1, requires_grad=True) / 3)
64
+
65
+
66
+ class PerSAM:
67
+ def __init__(
68
+ self,
69
+ sam: SamPredictor,
70
+ target_feat: torch.Tensor,
71
+ max_objects: int,
72
+ score_thresh: float,
73
+ nms_iou_thresh: float,
74
+ mask_weights: torch.Tensor,
75
+ ) -> None:
76
+ super().__init__()
77
+ self.sam = sam
78
+ self.weights = mask_weights
79
+ self.target_feat = target_feat
80
+ self.max_objects = max_objects
81
+ self.score_thresh = score_thresh
82
+ self.nms_iou_thresh = nms_iou_thresh
83
+
84
+ def __call__(self, x: npt.NDArray) -> tuple[npt.NDArray, npt.NDArray, npt.NDArray]:
85
+ return fast_inference(
86
+ self.sam,
87
+ x,
88
+ self.target_feat,
89
+ self.weights,
90
+ self.max_objects,
91
+ self.score_thresh,
92
+ self.nms_iou_thresh,
93
+ )
94
+
95
+
96
+ def fast_inference(
97
+ predictor: SamPredictor,
98
+ image: npt.NDArray,
99
+ target_feat: torch.Tensor,
100
+ weights: torch.Tensor,
101
+ max_objects: int,
102
+ score_thresh: float,
103
+ nms_iou_thresh: float = 0.2,
104
+ ) -> tuple[npt.NDArray, npt.NDArray, npt.NDArray]:
105
+ weights_np = weights.detach().cpu().numpy()
106
+ pred_masks = []
107
+ pred_scores = []
108
+
109
+ # Image feature encoding
110
+ predictor.set_image(image)
111
+ test_feat = predictor.features.squeeze()
112
+
113
+ # Cosine similarity
114
+ C, h, w = test_feat.shape
115
+ test_feat = test_feat / test_feat.norm(dim=0, keepdim=True)
116
+ test_feat = test_feat.reshape(C, h * w)
117
+ sim = target_feat @ test_feat
118
+
119
+ sim = sim.reshape(1, 1, h, w)
120
+ sim = F.interpolate(sim, scale_factor=4, mode="bilinear")
121
+ sim = predictor.model.postprocess_masks(
122
+ sim, input_size=predictor.input_size, original_size=predictor.original_size
123
+ ).squeeze()
124
+
125
+ for _ in range(max_objects):
126
+ # Positive location prior
127
+ topk_xy, topk_label = point_selection(sim, topk=1)
128
+
129
+ # First-step prediction
130
+ logits_high, scores, logits = predictor.predict(
131
+ point_coords=topk_xy,
132
+ point_labels=topk_label,
133
+ multimask_output=True,
134
+ return_logits=True,
135
+ return_numpy=False,
136
+ )
137
+ logits = logits.detach().cpu().numpy()
138
+
139
+ # Weighted sum three-scale masks
140
+ logits_high = logits_high * weights.unsqueeze(-1)
141
+ logit_high = logits_high.sum(0)
142
+ mask = (logit_high > 0).detach().cpu().numpy()
143
+
144
+ logits = logits * weights_np[..., None]
145
+ logit = logits.sum(0)
146
+
147
+ # Cascaded Post-refinement-1
148
+ y, x = np.nonzero(mask)
149
+ x_min = x.min()
150
+ x_max = x.max()
151
+ y_min = y.min()
152
+ y_max = y.max()
153
+ input_box = np.array([x_min, y_min, x_max, y_max])
154
+ masks, scores, logits = predictor.predict(
155
+ point_coords=topk_xy,
156
+ point_labels=topk_label,
157
+ box=input_box[None, :],
158
+ mask_input=logit[None, :, :],
159
+ multimask_output=True,
160
+ )
161
+ best_idx = np.argmax(scores)
162
+
163
+ # Cascaded Post-refinement-2
164
+ y, x = np.nonzero(masks[best_idx])
165
+ x_min = x.min()
166
+ x_max = x.max()
167
+ y_min = y.min()
168
+ y_max = y.max()
169
+ input_box = np.array([x_min, y_min, x_max, y_max])
170
+ masks, scores, logits = predictor.predict(
171
+ point_coords=topk_xy,
172
+ point_labels=topk_label,
173
+ box=input_box[None, :],
174
+ mask_input=logits[best_idx : best_idx + 1, :, :],
175
+ multimask_output=True,
176
+ return_numpy=False,
177
+ )
178
+
179
+ best_idx = np.argmax(scores.detach().cpu().numpy())
180
+ final_mask = masks[best_idx]
181
+ score = sim[topk_xy[0][1], topk_xy[0][0]].item()
182
+ final_mask_dilate = cv2.dilate(
183
+ final_mask.detach().cpu().numpy().astype(np.uint8), np.ones((5, 5), np.uint8), iterations=1
184
+ )
185
+
186
+ if score < score_thresh:
187
+ break
188
+
189
+ sim[final_mask_dilate] = 0
190
+ pred_masks.append(final_mask)
191
+ pred_scores.append(score)
192
+
193
+ pred_masks = torch.stack(pred_masks)
194
+ bboxes = batched_mask_to_box(pred_masks)
195
+ keep_by_nms = batched_nms(
196
+ bboxes.float(),
197
+ torch.as_tensor(pred_scores),
198
+ torch.zeros_like(bboxes[:, 0]),
199
+ iou_threshold=nms_iou_thresh,
200
+ )
201
+ pred_masks = pred_masks[keep_by_nms].cpu().numpy()
202
+ pred_scores = np.array(pred_scores)[keep_by_nms.cpu().numpy()]
203
+ bboxes = bboxes[keep_by_nms].int().cpu().numpy()
204
+ return pred_masks, bboxes, pred_scores
app/per_sam/train.py ADDED
@@ -0,0 +1,107 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import torch
2
+ import numpy.typing as npt
3
+ import torch.nn.functional as F
4
+
5
+ from app.configs import DEVICE
6
+ from app.mobile_sam import SamPredictor
7
+ from .model import point_selection, MaskWeights
8
+ from .loss import calculate_dice_loss, calculate_sigmoid_focal_loss
9
+
10
+
11
+ def train(
12
+ predictor: SamPredictor,
13
+ ref_images: list[npt.NDArray],
14
+ ref_masks: list[npt.NDArray],
15
+ lr: float = 1e-3,
16
+ epochs: int = 200,
17
+ ) -> tuple[torch.Tensor, torch.Tensor]:
18
+ gt_masks = []
19
+ points = []
20
+ target_feats = []
21
+ for ref_image, ref_mask in zip(ref_images, ref_masks):
22
+ gt_mask = torch.from_numpy(ref_mask)[:, :] > 0
23
+ gt_mask = gt_mask.float().unsqueeze(0).flatten(1).to(DEVICE)
24
+ gt_masks.append(gt_mask)
25
+
26
+ # Image features encoding
27
+ predictor.set_image(ref_image)
28
+ ref_mask = predictor.get_mask(ref_mask[:, :, None])
29
+ ref_feat = predictor.features.squeeze().permute(1, 2, 0)
30
+
31
+ ref_mask = F.interpolate(ref_mask, size=ref_feat.shape[0:2], mode="bilinear")
32
+ ref_mask = ref_mask.squeeze()
33
+
34
+ # Target feature extraction
35
+ target_feat = ref_feat[ref_mask > 0]
36
+ target_feat_mean = target_feat.mean(0)
37
+ target_feat_max = torch.max(target_feat, dim=0)[0]
38
+ target_feat = (target_feat_max / 2 + target_feat_mean / 2).unsqueeze(0)
39
+
40
+ # Cosine similarity
41
+ h, w, C = ref_feat.shape
42
+ target_feat = target_feat / target_feat.norm(dim=-1, keepdim=True)
43
+ target_feats.append(target_feat)
44
+ ref_feat = ref_feat / ref_feat.norm(dim=-1, keepdim=True)
45
+ ref_feat = ref_feat.permute(2, 0, 1).reshape(C, h * w)
46
+ sim = target_feat @ ref_feat
47
+
48
+ sim = sim.reshape(1, 1, h, w)
49
+ sim = F.interpolate(sim, scale_factor=4, mode="bilinear")
50
+ sim = predictor.model.postprocess_masks(
51
+ sim, input_size=predictor.input_size, original_size=predictor.original_size
52
+ ).squeeze()
53
+
54
+ # Positive location prior
55
+ topk_xy, topk_label = point_selection(sim, topk=1)
56
+ points.append((topk_xy, topk_label))
57
+
58
+ target_feat = torch.concat(target_feats, axis=0).mean(axis=0)
59
+
60
+ # Learnable mask weights
61
+ mask_weights = MaskWeights().to(DEVICE)
62
+ mask_weights.train()
63
+
64
+ optimizer = torch.optim.AdamW(mask_weights.parameters(), lr=lr, eps=1e-4)
65
+ scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(optimizer, epochs)
66
+
67
+ for _ in range(epochs):
68
+ for i in range(len(gt_masks)):
69
+ gt_mask = gt_masks[i]
70
+ topk_xy, topk_label = points[i]
71
+ # Run the decoder
72
+ (
73
+ logits_high,
74
+ _,
75
+ _,
76
+ ) = predictor.predict(
77
+ point_coords=topk_xy,
78
+ point_labels=topk_label,
79
+ multimask_output=True,
80
+ return_logits=True,
81
+ return_numpy=False,
82
+ )
83
+ logits_high = logits_high.flatten(1)
84
+
85
+ # Weighted sum three-scale masks
86
+ weights = torch.cat(
87
+ (1 - mask_weights.weights.sum(0).unsqueeze(0), mask_weights.weights),
88
+ dim=0,
89
+ )
90
+ logits_high = logits_high * weights
91
+ logits_high = logits_high.sum(0).unsqueeze(0)
92
+
93
+ dice_loss = calculate_dice_loss(logits_high, gt_mask)
94
+ focal_loss = calculate_sigmoid_focal_loss(logits_high, gt_mask, alpha=1.0)
95
+ loss = dice_loss + focal_loss
96
+
97
+ optimizer.zero_grad()
98
+ loss.backward()
99
+ optimizer.step()
100
+ scheduler.step()
101
+
102
+ # current_lr = scheduler.get_last_lr()[0]
103
+ mask_weights.eval()
104
+ weights = torch.cat(
105
+ (1 - mask_weights.weights.sum(0).unsqueeze(0), mask_weights.weights), dim=0
106
+ )
107
+ return weights, target_feat
app/sam/__init__.py ADDED
@@ -0,0 +1,2 @@
 
 
 
1
+ from .build_sam import build_sam
2
+ from .predictor import SamPredictor
app/sam/build_sam.py ADDED
@@ -0,0 +1,111 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Copyright (c) Meta Platforms, Inc. and affiliates.
2
+ # All rights reserved.
3
+
4
+ # This source code is licensed under the license found in the
5
+ # LICENSE file in the root directory of this source tree.
6
+
7
+ import torch
8
+
9
+ from functools import partial
10
+
11
+ from .image_encoder import ImageEncoderViT
12
+ from .mask_decoder import MaskDecoder
13
+ from .prompt_encoder import PromptEncoder
14
+ from .transformer import TwoWayTransformer
15
+ from .sam import Sam
16
+
17
+
18
+ def build_sam_vit_h(checkpoint=None):
19
+ return _build_sam(
20
+ encoder_embed_dim=1280,
21
+ encoder_depth=32,
22
+ encoder_num_heads=16,
23
+ encoder_global_attn_indexes=[7, 15, 23, 31],
24
+ checkpoint=checkpoint,
25
+ )
26
+
27
+
28
+ build_sam = build_sam_vit_h
29
+
30
+
31
+ def build_sam_vit_l(checkpoint=None):
32
+ return _build_sam(
33
+ encoder_embed_dim=1024,
34
+ encoder_depth=24,
35
+ encoder_num_heads=16,
36
+ encoder_global_attn_indexes=[5, 11, 17, 23],
37
+ checkpoint=checkpoint,
38
+ )
39
+
40
+
41
+ def build_sam_vit_b(checkpoint=None):
42
+ return _build_sam(
43
+ encoder_embed_dim=768,
44
+ encoder_depth=12,
45
+ encoder_num_heads=12,
46
+ encoder_global_attn_indexes=[2, 5, 8, 11],
47
+ checkpoint=checkpoint,
48
+ )
49
+
50
+
51
+ sam_model_registry = {
52
+ "default": build_sam_vit_h,
53
+ "vit_h": build_sam_vit_h,
54
+ "vit_l": build_sam_vit_l,
55
+ "vit_b": build_sam_vit_b,
56
+ }
57
+
58
+
59
+ def _build_sam(
60
+ encoder_embed_dim,
61
+ encoder_depth,
62
+ encoder_num_heads,
63
+ encoder_global_attn_indexes,
64
+ checkpoint=None,
65
+ ):
66
+ prompt_embed_dim = 256
67
+ image_size = 1024
68
+ vit_patch_size = 16
69
+ image_embedding_size = image_size // vit_patch_size
70
+ sam = Sam(
71
+ image_encoder=ImageEncoderViT(
72
+ depth=encoder_depth,
73
+ embed_dim=encoder_embed_dim,
74
+ img_size=image_size,
75
+ mlp_ratio=4,
76
+ norm_layer=partial(torch.nn.LayerNorm, eps=1e-6),
77
+ num_heads=encoder_num_heads,
78
+ patch_size=vit_patch_size,
79
+ qkv_bias=True,
80
+ use_rel_pos=True,
81
+ global_attn_indexes=encoder_global_attn_indexes,
82
+ window_size=14,
83
+ out_chans=prompt_embed_dim,
84
+ ),
85
+ prompt_encoder=PromptEncoder(
86
+ embed_dim=prompt_embed_dim,
87
+ image_embedding_size=(image_embedding_size, image_embedding_size),
88
+ input_image_size=(image_size, image_size),
89
+ mask_in_chans=16,
90
+ ),
91
+ mask_decoder=MaskDecoder(
92
+ num_multimask_outputs=3,
93
+ transformer=TwoWayTransformer(
94
+ depth=2,
95
+ embedding_dim=prompt_embed_dim,
96
+ mlp_dim=2048,
97
+ num_heads=8,
98
+ ),
99
+ transformer_dim=prompt_embed_dim,
100
+ iou_head_depth=3,
101
+ iou_head_hidden_dim=256,
102
+ ),
103
+ pixel_mean=[123.675, 116.28, 103.53],
104
+ pixel_std=[58.395, 57.12, 57.375],
105
+ )
106
+ sam.eval()
107
+ if checkpoint is not None:
108
+ with open(checkpoint, "rb") as f:
109
+ state_dict = torch.load(f)
110
+ sam.load_state_dict(state_dict)
111
+ return sam
app/sam/common.py ADDED
@@ -0,0 +1,43 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Copyright (c) Meta Platforms, Inc. and affiliates.
2
+ # All rights reserved.
3
+
4
+ # This source code is licensed under the license found in the
5
+ # LICENSE file in the root directory of this source tree.
6
+
7
+ import torch
8
+ import torch.nn as nn
9
+
10
+ from typing import Type
11
+
12
+
13
+ class MLPBlock(nn.Module):
14
+ def __init__(
15
+ self,
16
+ embedding_dim: int,
17
+ mlp_dim: int,
18
+ act: Type[nn.Module] = nn.GELU,
19
+ ) -> None:
20
+ super().__init__()
21
+ self.lin1 = nn.Linear(embedding_dim, mlp_dim)
22
+ self.lin2 = nn.Linear(mlp_dim, embedding_dim)
23
+ self.act = act()
24
+
25
+ def forward(self, x: torch.Tensor) -> torch.Tensor:
26
+ return self.lin2(self.act(self.lin1(x)))
27
+
28
+
29
+ # From https://github.com/facebookresearch/detectron2/blob/main/detectron2/layers/batch_norm.py # noqa
30
+ # Itself from https://github.com/facebookresearch/ConvNeXt/blob/d1fa8f6fef0a165b27399986cc2bdacc92777e40/models/convnext.py#L119 # noqa
31
+ class LayerNorm2d(nn.Module):
32
+ def __init__(self, num_channels: int, eps: float = 1e-6) -> None:
33
+ super().__init__()
34
+ self.weight = nn.Parameter(torch.ones(num_channels))
35
+ self.bias = nn.Parameter(torch.zeros(num_channels))
36
+ self.eps = eps
37
+
38
+ def forward(self, x: torch.Tensor) -> torch.Tensor:
39
+ u = x.mean(1, keepdim=True)
40
+ s = (x - u).pow(2).mean(1, keepdim=True)
41
+ x = (x - u) / torch.sqrt(s + self.eps)
42
+ x = self.weight[:, None, None] * x + self.bias[:, None, None]
43
+ return x
app/sam/image_encoder.py ADDED
@@ -0,0 +1,395 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Copyright (c) Meta Platforms, Inc. and affiliates.
2
+ # All rights reserved.
3
+
4
+ # This source code is licensed under the license found in the
5
+ # LICENSE file in the root directory of this source tree.
6
+
7
+ import torch
8
+ import torch.nn as nn
9
+ import torch.nn.functional as F
10
+
11
+ from typing import Optional, Tuple, Type
12
+
13
+ from .common import LayerNorm2d, MLPBlock
14
+
15
+
16
+ # This class and its supporting functions below lightly adapted from the ViTDet backbone available at: https://github.com/facebookresearch/detectron2/blob/main/detectron2/modeling/backbone/vit.py # noqa
17
+ class ImageEncoderViT(nn.Module):
18
+ def __init__(
19
+ self,
20
+ img_size: int = 1024,
21
+ patch_size: int = 16,
22
+ in_chans: int = 3,
23
+ embed_dim: int = 768,
24
+ depth: int = 12,
25
+ num_heads: int = 12,
26
+ mlp_ratio: float = 4.0,
27
+ out_chans: int = 256,
28
+ qkv_bias: bool = True,
29
+ norm_layer: Type[nn.Module] = nn.LayerNorm,
30
+ act_layer: Type[nn.Module] = nn.GELU,
31
+ use_abs_pos: bool = True,
32
+ use_rel_pos: bool = False,
33
+ rel_pos_zero_init: bool = True,
34
+ window_size: int = 0,
35
+ global_attn_indexes: Tuple[int, ...] = (),
36
+ ) -> None:
37
+ """
38
+ Args:
39
+ img_size (int): Input image size.
40
+ patch_size (int): Patch size.
41
+ in_chans (int): Number of input image channels.
42
+ embed_dim (int): Patch embedding dimension.
43
+ depth (int): Depth of ViT.
44
+ num_heads (int): Number of attention heads in each ViT block.
45
+ mlp_ratio (float): Ratio of mlp hidden dim to embedding dim.
46
+ qkv_bias (bool): If True, add a learnable bias to query, key, value.
47
+ norm_layer (nn.Module): Normalization layer.
48
+ act_layer (nn.Module): Activation layer.
49
+ use_abs_pos (bool): If True, use absolute positional embeddings.
50
+ use_rel_pos (bool): If True, add relative positional embeddings to the attention map.
51
+ rel_pos_zero_init (bool): If True, zero initialize relative positional parameters.
52
+ window_size (int): Window size for window attention blocks.
53
+ global_attn_indexes (list): Indexes for blocks using global attention.
54
+ """
55
+ super().__init__()
56
+ self.img_size = img_size
57
+
58
+ self.patch_embed = PatchEmbed(
59
+ kernel_size=(patch_size, patch_size),
60
+ stride=(patch_size, patch_size),
61
+ in_chans=in_chans,
62
+ embed_dim=embed_dim,
63
+ )
64
+
65
+ self.pos_embed: Optional[nn.Parameter] = None
66
+ if use_abs_pos:
67
+ # Initialize absolute positional embedding with pretrain image size.
68
+ self.pos_embed = nn.Parameter(
69
+ torch.zeros(1, img_size // patch_size, img_size // patch_size, embed_dim)
70
+ )
71
+
72
+ self.blocks = nn.ModuleList()
73
+ for i in range(depth):
74
+ block = Block(
75
+ dim=embed_dim,
76
+ num_heads=num_heads,
77
+ mlp_ratio=mlp_ratio,
78
+ qkv_bias=qkv_bias,
79
+ norm_layer=norm_layer,
80
+ act_layer=act_layer,
81
+ use_rel_pos=use_rel_pos,
82
+ rel_pos_zero_init=rel_pos_zero_init,
83
+ window_size=window_size if i not in global_attn_indexes else 0,
84
+ input_size=(img_size // patch_size, img_size // patch_size),
85
+ )
86
+ self.blocks.append(block)
87
+
88
+ self.neck = nn.Sequential(
89
+ nn.Conv2d(
90
+ embed_dim,
91
+ out_chans,
92
+ kernel_size=1,
93
+ bias=False,
94
+ ),
95
+ LayerNorm2d(out_chans),
96
+ nn.Conv2d(
97
+ out_chans,
98
+ out_chans,
99
+ kernel_size=3,
100
+ padding=1,
101
+ bias=False,
102
+ ),
103
+ LayerNorm2d(out_chans),
104
+ )
105
+
106
+ def forward(self, x: torch.Tensor) -> torch.Tensor:
107
+ x = self.patch_embed(x)
108
+ if self.pos_embed is not None:
109
+ x = x + self.pos_embed
110
+
111
+ for blk in self.blocks:
112
+ x = blk(x)
113
+
114
+ x = self.neck(x.permute(0, 3, 1, 2))
115
+
116
+ return x
117
+
118
+
119
+ class Block(nn.Module):
120
+ """Transformer blocks with support of window attention and residual propagation blocks"""
121
+
122
+ def __init__(
123
+ self,
124
+ dim: int,
125
+ num_heads: int,
126
+ mlp_ratio: float = 4.0,
127
+ qkv_bias: bool = True,
128
+ norm_layer: Type[nn.Module] = nn.LayerNorm,
129
+ act_layer: Type[nn.Module] = nn.GELU,
130
+ use_rel_pos: bool = False,
131
+ rel_pos_zero_init: bool = True,
132
+ window_size: int = 0,
133
+ input_size: Optional[Tuple[int, int]] = None,
134
+ ) -> None:
135
+ """
136
+ Args:
137
+ dim (int): Number of input channels.
138
+ num_heads (int): Number of attention heads in each ViT block.
139
+ mlp_ratio (float): Ratio of mlp hidden dim to embedding dim.
140
+ qkv_bias (bool): If True, add a learnable bias to query, key, value.
141
+ norm_layer (nn.Module): Normalization layer.
142
+ act_layer (nn.Module): Activation layer.
143
+ use_rel_pos (bool): If True, add relative positional embeddings to the attention map.
144
+ rel_pos_zero_init (bool): If True, zero initialize relative positional parameters.
145
+ window_size (int): Window size for window attention blocks. If it equals 0, then
146
+ use global attention.
147
+ input_size (tuple(int, int) or None): Input resolution for calculating the relative
148
+ positional parameter size.
149
+ """
150
+ super().__init__()
151
+ self.norm1 = norm_layer(dim)
152
+ self.attn = Attention(
153
+ dim,
154
+ num_heads=num_heads,
155
+ qkv_bias=qkv_bias,
156
+ use_rel_pos=use_rel_pos,
157
+ rel_pos_zero_init=rel_pos_zero_init,
158
+ input_size=input_size if window_size == 0 else (window_size, window_size),
159
+ )
160
+
161
+ self.norm2 = norm_layer(dim)
162
+ self.mlp = MLPBlock(embedding_dim=dim, mlp_dim=int(dim * mlp_ratio), act=act_layer)
163
+
164
+ self.window_size = window_size
165
+
166
+ def forward(self, x: torch.Tensor) -> torch.Tensor:
167
+ shortcut = x
168
+ x = self.norm1(x)
169
+ # Window partition
170
+ if self.window_size > 0:
171
+ H, W = x.shape[1], x.shape[2]
172
+ x, pad_hw = window_partition(x, self.window_size)
173
+
174
+ x = self.attn(x)
175
+ # Reverse window partition
176
+ if self.window_size > 0:
177
+ x = window_unpartition(x, self.window_size, pad_hw, (H, W))
178
+
179
+ x = shortcut + x
180
+ x = x + self.mlp(self.norm2(x))
181
+
182
+ return x
183
+
184
+
185
+ class Attention(nn.Module):
186
+ """Multi-head Attention block with relative position embeddings."""
187
+
188
+ def __init__(
189
+ self,
190
+ dim: int,
191
+ num_heads: int = 8,
192
+ qkv_bias: bool = True,
193
+ use_rel_pos: bool = False,
194
+ rel_pos_zero_init: bool = True,
195
+ input_size: Optional[Tuple[int, int]] = None,
196
+ ) -> None:
197
+ """
198
+ Args:
199
+ dim (int): Number of input channels.
200
+ num_heads (int): Number of attention heads.
201
+ qkv_bias (bool): If True, add a learnable bias to query, key, value.
202
+ rel_pos (bool): If True, add relative positional embeddings to the attention map.
203
+ rel_pos_zero_init (bool): If True, zero initialize relative positional parameters.
204
+ input_size (tuple(int, int) or None): Input resolution for calculating the relative
205
+ positional parameter size.
206
+ """
207
+ super().__init__()
208
+ self.num_heads = num_heads
209
+ head_dim = dim // num_heads
210
+ self.scale = head_dim**-0.5
211
+
212
+ self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
213
+ self.proj = nn.Linear(dim, dim)
214
+
215
+ self.use_rel_pos = use_rel_pos
216
+ if self.use_rel_pos:
217
+ assert (
218
+ input_size is not None
219
+ ), "Input size must be provided if using relative positional encoding."
220
+ # initialize relative positional embeddings
221
+ self.rel_pos_h = nn.Parameter(torch.zeros(2 * input_size[0] - 1, head_dim))
222
+ self.rel_pos_w = nn.Parameter(torch.zeros(2 * input_size[1] - 1, head_dim))
223
+
224
+ def forward(self, x: torch.Tensor) -> torch.Tensor:
225
+ B, H, W, _ = x.shape
226
+ # qkv with shape (3, B, nHead, H * W, C)
227
+ qkv = self.qkv(x).reshape(B, H * W, 3, self.num_heads, -1).permute(2, 0, 3, 1, 4)
228
+ # q, k, v with shape (B * nHead, H * W, C)
229
+ q, k, v = qkv.reshape(3, B * self.num_heads, H * W, -1).unbind(0)
230
+
231
+ attn = (q * self.scale) @ k.transpose(-2, -1)
232
+
233
+ if self.use_rel_pos:
234
+ attn = add_decomposed_rel_pos(attn, q, self.rel_pos_h, self.rel_pos_w, (H, W), (H, W))
235
+
236
+ attn = attn.softmax(dim=-1)
237
+ x = (attn @ v).view(B, self.num_heads, H, W, -1).permute(0, 2, 3, 1, 4).reshape(B, H, W, -1)
238
+ x = self.proj(x)
239
+
240
+ return x
241
+
242
+
243
+ def window_partition(x: torch.Tensor, window_size: int) -> Tuple[torch.Tensor, Tuple[int, int]]:
244
+ """
245
+ Partition into non-overlapping windows with padding if needed.
246
+ Args:
247
+ x (tensor): input tokens with [B, H, W, C].
248
+ window_size (int): window size.
249
+
250
+ Returns:
251
+ windows: windows after partition with [B * num_windows, window_size, window_size, C].
252
+ (Hp, Wp): padded height and width before partition
253
+ """
254
+ B, H, W, C = x.shape
255
+
256
+ pad_h = (window_size - H % window_size) % window_size
257
+ pad_w = (window_size - W % window_size) % window_size
258
+ if pad_h > 0 or pad_w > 0:
259
+ x = F.pad(x, (0, 0, 0, pad_w, 0, pad_h))
260
+ Hp, Wp = H + pad_h, W + pad_w
261
+
262
+ x = x.view(B, Hp // window_size, window_size, Wp // window_size, window_size, C)
263
+ windows = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(-1, window_size, window_size, C)
264
+ return windows, (Hp, Wp)
265
+
266
+
267
+ def window_unpartition(
268
+ windows: torch.Tensor, window_size: int, pad_hw: Tuple[int, int], hw: Tuple[int, int]
269
+ ) -> torch.Tensor:
270
+ """
271
+ Window unpartition into original sequences and removing padding.
272
+ Args:
273
+ windows (tensor): input tokens with [B * num_windows, window_size, window_size, C].
274
+ window_size (int): window size.
275
+ pad_hw (Tuple): padded height and width (Hp, Wp).
276
+ hw (Tuple): original height and width (H, W) before padding.
277
+
278
+ Returns:
279
+ x: unpartitioned sequences with [B, H, W, C].
280
+ """
281
+ Hp, Wp = pad_hw
282
+ H, W = hw
283
+ B = windows.shape[0] // (Hp * Wp // window_size // window_size)
284
+ x = windows.view(B, Hp // window_size, Wp // window_size, window_size, window_size, -1)
285
+ x = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(B, Hp, Wp, -1)
286
+
287
+ if Hp > H or Wp > W:
288
+ x = x[:, :H, :W, :].contiguous()
289
+ return x
290
+
291
+
292
+ def get_rel_pos(q_size: int, k_size: int, rel_pos: torch.Tensor) -> torch.Tensor:
293
+ """
294
+ Get relative positional embeddings according to the relative positions of
295
+ query and key sizes.
296
+ Args:
297
+ q_size (int): size of query q.
298
+ k_size (int): size of key k.
299
+ rel_pos (Tensor): relative position embeddings (L, C).
300
+
301
+ Returns:
302
+ Extracted positional embeddings according to relative positions.
303
+ """
304
+ max_rel_dist = int(2 * max(q_size, k_size) - 1)
305
+ # Interpolate rel pos if needed.
306
+ if rel_pos.shape[0] != max_rel_dist:
307
+ # Interpolate rel pos.
308
+ rel_pos_resized = F.interpolate(
309
+ rel_pos.reshape(1, rel_pos.shape[0], -1).permute(0, 2, 1),
310
+ size=max_rel_dist,
311
+ mode="linear",
312
+ )
313
+ rel_pos_resized = rel_pos_resized.reshape(-1, max_rel_dist).permute(1, 0)
314
+ else:
315
+ rel_pos_resized = rel_pos
316
+
317
+ # Scale the coords with short length if shapes for q and k are different.
318
+ q_coords = torch.arange(q_size)[:, None] * max(k_size / q_size, 1.0)
319
+ k_coords = torch.arange(k_size)[None, :] * max(q_size / k_size, 1.0)
320
+ relative_coords = (q_coords - k_coords) + (k_size - 1) * max(q_size / k_size, 1.0)
321
+
322
+ return rel_pos_resized[relative_coords.long()]
323
+
324
+
325
+ def add_decomposed_rel_pos(
326
+ attn: torch.Tensor,
327
+ q: torch.Tensor,
328
+ rel_pos_h: torch.Tensor,
329
+ rel_pos_w: torch.Tensor,
330
+ q_size: Tuple[int, int],
331
+ k_size: Tuple[int, int],
332
+ ) -> torch.Tensor:
333
+ """
334
+ Calculate decomposed Relative Positional Embeddings from :paper:`mvitv2`.
335
+ https://github.com/facebookresearch/mvit/blob/19786631e330df9f3622e5402b4a419a263a2c80/mvit/models/attention.py # noqa B950
336
+ Args:
337
+ attn (Tensor): attention map.
338
+ q (Tensor): query q in the attention layer with shape (B, q_h * q_w, C).
339
+ rel_pos_h (Tensor): relative position embeddings (Lh, C) for height axis.
340
+ rel_pos_w (Tensor): relative position embeddings (Lw, C) for width axis.
341
+ q_size (Tuple): spatial sequence size of query q with (q_h, q_w).
342
+ k_size (Tuple): spatial sequence size of key k with (k_h, k_w).
343
+
344
+ Returns:
345
+ attn (Tensor): attention map with added relative positional embeddings.
346
+ """
347
+ q_h, q_w = q_size
348
+ k_h, k_w = k_size
349
+ Rh = get_rel_pos(q_h, k_h, rel_pos_h)
350
+ Rw = get_rel_pos(q_w, k_w, rel_pos_w)
351
+
352
+ B, _, dim = q.shape
353
+ r_q = q.reshape(B, q_h, q_w, dim)
354
+ rel_h = torch.einsum("bhwc,hkc->bhwk", r_q, Rh)
355
+ rel_w = torch.einsum("bhwc,wkc->bhwk", r_q, Rw)
356
+
357
+ attn = (
358
+ attn.view(B, q_h, q_w, k_h, k_w) + rel_h[:, :, :, :, None] + rel_w[:, :, :, None, :]
359
+ ).view(B, q_h * q_w, k_h * k_w)
360
+
361
+ return attn
362
+
363
+
364
+ class PatchEmbed(nn.Module):
365
+ """
366
+ Image to Patch Embedding.
367
+ """
368
+
369
+ def __init__(
370
+ self,
371
+ kernel_size: Tuple[int, int] = (16, 16),
372
+ stride: Tuple[int, int] = (16, 16),
373
+ padding: Tuple[int, int] = (0, 0),
374
+ in_chans: int = 3,
375
+ embed_dim: int = 768,
376
+ ) -> None:
377
+ """
378
+ Args:
379
+ kernel_size (Tuple): kernel size of the projection layer.
380
+ stride (Tuple): stride of the projection layer.
381
+ padding (Tuple): padding size of the projection layer.
382
+ in_chans (int): Number of input image channels.
383
+ embed_dim (int): Patch embedding dimension.
384
+ """
385
+ super().__init__()
386
+
387
+ self.proj = nn.Conv2d(
388
+ in_chans, embed_dim, kernel_size=kernel_size, stride=stride, padding=padding
389
+ )
390
+
391
+ def forward(self, x: torch.Tensor) -> torch.Tensor:
392
+ x = self.proj(x)
393
+ # B C H W -> B H W C
394
+ x = x.permute(0, 2, 3, 1)
395
+ return x
app/sam/mask_decoder.py ADDED
@@ -0,0 +1,176 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Copyright (c) Meta Platforms, Inc. and affiliates.
2
+ # All rights reserved.
3
+
4
+ # This source code is licensed under the license found in the
5
+ # LICENSE file in the root directory of this source tree.
6
+
7
+ import torch
8
+ from torch import nn
9
+ from torch.nn import functional as F
10
+
11
+ from typing import List, Tuple, Type
12
+
13
+ from .common import LayerNorm2d
14
+
15
+
16
+ class MaskDecoder(nn.Module):
17
+ def __init__(
18
+ self,
19
+ *,
20
+ transformer_dim: int,
21
+ transformer: nn.Module,
22
+ num_multimask_outputs: int = 3,
23
+ activation: Type[nn.Module] = nn.GELU,
24
+ iou_head_depth: int = 3,
25
+ iou_head_hidden_dim: int = 256,
26
+ ) -> None:
27
+ """
28
+ Predicts masks given an image and prompt embeddings, using a
29
+ transformer architecture.
30
+
31
+ Arguments:
32
+ transformer_dim (int): the channel dimension of the transformer
33
+ transformer (nn.Module): the transformer used to predict masks
34
+ num_multimask_outputs (int): the number of masks to predict
35
+ when disambiguating masks
36
+ activation (nn.Module): the type of activation to use when
37
+ upscaling masks
38
+ iou_head_depth (int): the depth of the MLP used to predict
39
+ mask quality
40
+ iou_head_hidden_dim (int): the hidden dimension of the MLP
41
+ used to predict mask quality
42
+ """
43
+ super().__init__()
44
+ self.transformer_dim = transformer_dim
45
+ self.transformer = transformer
46
+
47
+ self.num_multimask_outputs = num_multimask_outputs
48
+
49
+ self.iou_token = nn.Embedding(1, transformer_dim)
50
+ self.num_mask_tokens = num_multimask_outputs + 1
51
+ self.mask_tokens = nn.Embedding(self.num_mask_tokens, transformer_dim)
52
+
53
+ self.output_upscaling = nn.Sequential(
54
+ nn.ConvTranspose2d(transformer_dim, transformer_dim // 4, kernel_size=2, stride=2),
55
+ LayerNorm2d(transformer_dim // 4),
56
+ activation(),
57
+ nn.ConvTranspose2d(transformer_dim // 4, transformer_dim // 8, kernel_size=2, stride=2),
58
+ activation(),
59
+ )
60
+ self.output_hypernetworks_mlps = nn.ModuleList(
61
+ [
62
+ MLP(transformer_dim, transformer_dim, transformer_dim // 8, 3)
63
+ for i in range(self.num_mask_tokens)
64
+ ]
65
+ )
66
+
67
+ self.iou_prediction_head = MLP(
68
+ transformer_dim, iou_head_hidden_dim, self.num_mask_tokens, iou_head_depth
69
+ )
70
+
71
+ def forward(
72
+ self,
73
+ image_embeddings: torch.Tensor,
74
+ image_pe: torch.Tensor,
75
+ sparse_prompt_embeddings: torch.Tensor,
76
+ dense_prompt_embeddings: torch.Tensor,
77
+ multimask_output: bool,
78
+ ) -> Tuple[torch.Tensor, torch.Tensor]:
79
+ """
80
+ Predict masks given image and prompt embeddings.
81
+
82
+ Arguments:
83
+ image_embeddings (torch.Tensor): the embeddings from the image encoder
84
+ image_pe (torch.Tensor): positional encoding with the shape of image_embeddings
85
+ sparse_prompt_embeddings (torch.Tensor): the embeddings of the points and boxes
86
+ dense_prompt_embeddings (torch.Tensor): the embeddings of the mask inputs
87
+ multimask_output (bool): Whether to return multiple masks or a single
88
+ mask.
89
+
90
+ Returns:
91
+ torch.Tensor: batched predicted masks
92
+ torch.Tensor: batched predictions of mask quality
93
+ """
94
+ masks, iou_pred = self.predict_masks(
95
+ image_embeddings=image_embeddings,
96
+ image_pe=image_pe,
97
+ sparse_prompt_embeddings=sparse_prompt_embeddings,
98
+ dense_prompt_embeddings=dense_prompt_embeddings,
99
+ )
100
+
101
+ # Select the correct mask or masks for output
102
+ if multimask_output:
103
+ mask_slice = slice(1, None)
104
+ else:
105
+ mask_slice = slice(0, 1)
106
+ masks = masks[:, mask_slice, :, :]
107
+ iou_pred = iou_pred[:, mask_slice]
108
+
109
+ # Prepare output
110
+ return masks, iou_pred
111
+
112
+ def predict_masks(
113
+ self,
114
+ image_embeddings: torch.Tensor,
115
+ image_pe: torch.Tensor,
116
+ sparse_prompt_embeddings: torch.Tensor,
117
+ dense_prompt_embeddings: torch.Tensor,
118
+ ) -> Tuple[torch.Tensor, torch.Tensor]:
119
+ """Predicts masks. See 'forward' for more details."""
120
+ # Concatenate output tokens
121
+ output_tokens = torch.cat([self.iou_token.weight, self.mask_tokens.weight], dim=0)
122
+ output_tokens = output_tokens.unsqueeze(0).expand(sparse_prompt_embeddings.size(0), -1, -1)
123
+ tokens = torch.cat((output_tokens, sparse_prompt_embeddings), dim=1)
124
+
125
+ # Expand per-image data in batch direction to be per-mask
126
+ src = torch.repeat_interleave(image_embeddings, tokens.shape[0], dim=0)
127
+ src = src + dense_prompt_embeddings
128
+ pos_src = torch.repeat_interleave(image_pe, tokens.shape[0], dim=0)
129
+ b, c, h, w = src.shape
130
+
131
+ # Run the transformer
132
+ hs, src = self.transformer(src, pos_src, tokens)
133
+ iou_token_out = hs[:, 0, :]
134
+ mask_tokens_out = hs[:, 1 : (1 + self.num_mask_tokens), :]
135
+
136
+ # Upscale mask embeddings and predict masks using the mask tokens
137
+ src = src.transpose(1, 2).view(b, c, h, w)
138
+ upscaled_embedding = self.output_upscaling(src)
139
+ hyper_in_list: List[torch.Tensor] = []
140
+ for i in range(self.num_mask_tokens):
141
+ hyper_in_list.append(self.output_hypernetworks_mlps[i](mask_tokens_out[:, i, :]))
142
+ hyper_in = torch.stack(hyper_in_list, dim=1)
143
+ b, c, h, w = upscaled_embedding.shape
144
+ masks = (hyper_in @ upscaled_embedding.view(b, c, h * w)).view(b, -1, h, w)
145
+
146
+ # Generate mask quality predictions
147
+ iou_pred = self.iou_prediction_head(iou_token_out)
148
+
149
+ return masks, iou_pred
150
+
151
+
152
+ # Lightly adapted from
153
+ # https://github.com/facebookresearch/MaskFormer/blob/main/mask_former/modeling/transformer/transformer_predictor.py # noqa
154
+ class MLP(nn.Module):
155
+ def __init__(
156
+ self,
157
+ input_dim: int,
158
+ hidden_dim: int,
159
+ output_dim: int,
160
+ num_layers: int,
161
+ sigmoid_output: bool = False,
162
+ ) -> None:
163
+ super().__init__()
164
+ self.num_layers = num_layers
165
+ h = [hidden_dim] * (num_layers - 1)
166
+ self.layers = nn.ModuleList(
167
+ nn.Linear(n, k) for n, k in zip([input_dim] + h, h + [output_dim])
168
+ )
169
+ self.sigmoid_output = sigmoid_output
170
+
171
+ def forward(self, x):
172
+ for i, layer in enumerate(self.layers):
173
+ x = F.relu(layer(x)) if i < self.num_layers - 1 else layer(x)
174
+ if self.sigmoid_output:
175
+ x = F.sigmoid(x)
176
+ return x
app/sam/predictor.py ADDED
@@ -0,0 +1,267 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Copyright (c) Meta Platforms, Inc. and affiliates.
2
+ # All rights reserved.
3
+
4
+ # This source code is licensed under the license found in the
5
+ # LICENSE file in the root directory of this source tree.
6
+
7
+ import numpy as np
8
+ import torch
9
+
10
+ from typing import Optional, Tuple
11
+ from .sam import Sam
12
+ from .transforms import ResizeLongestSide
13
+
14
+
15
+ class SamPredictor:
16
+ def __init__(
17
+ self,
18
+ sam_model: Sam,
19
+ ) -> None:
20
+ """
21
+ Uses SAM to calculate the image embedding for an image, and then
22
+ allow repeated, efficient mask prediction given prompts.
23
+
24
+ Arguments:
25
+ sam_model (Sam): The model to use for mask prediction.
26
+ """
27
+ super().__init__()
28
+ self.model = sam_model
29
+ self.transform = ResizeLongestSide(sam_model.image_encoder.img_size)
30
+ self.reset_image()
31
+
32
+ def set_image(
33
+ self,
34
+ image: np.ndarray,
35
+ image_format: str = "RGB",
36
+ ) -> None:
37
+ """
38
+ Calculates the image embeddings for the provided image, allowing
39
+ masks to be predicted with the 'predict' method.
40
+
41
+ Arguments:
42
+ image (np.ndarray): The image for calculating masks. Expects an
43
+ image in HWC uint8 format, with pixel values in [0, 255].
44
+ image_format (str): The color format of the image, in ['RGB', 'BGR'].
45
+ """
46
+ assert image_format in [
47
+ "RGB",
48
+ "BGR",
49
+ ], f"image_format must be in ['RGB', 'BGR'], is {image_format}."
50
+ if image_format != self.model.image_format:
51
+ image = image[..., ::-1]
52
+
53
+ # Transform the image to the form expected by the model
54
+ input_image = self.transform.apply_image(image)
55
+ input_image_torch = torch.as_tensor(input_image, device=self.device)
56
+ input_image_torch = input_image_torch.permute(2, 0, 1).contiguous()[None, :, :, :]
57
+
58
+ self.set_torch_image(input_image_torch, image.shape[:2])
59
+
60
+ @torch.no_grad()
61
+ def set_torch_image(
62
+ self,
63
+ transformed_image: torch.Tensor,
64
+ original_image_size: Tuple[int, ...],
65
+ ) -> None:
66
+ """
67
+ Calculates the image embeddings for the provided image, allowing
68
+ masks to be predicted with the 'predict' method. Expects the input
69
+ image to be already transformed to the format expected by the model.
70
+
71
+ Arguments:
72
+ transformed_image (torch.Tensor): The input image, with shape
73
+ 1x3xHxW, which has been transformed with ResizeLongestSide.
74
+ original_image_size (tuple(int, int)): The size of the image
75
+ before transformation, in (H, W) format.
76
+ """
77
+ assert (
78
+ len(transformed_image.shape) == 4
79
+ and transformed_image.shape[1] == 3
80
+ and max(*transformed_image.shape[2:]) == self.model.image_encoder.img_size
81
+ ), f"set_torch_image input must be BCHW with long side {self.model.image_encoder.img_size}."
82
+ self.reset_image()
83
+
84
+ self.original_size = original_image_size
85
+ self.input_size = tuple(transformed_image.shape[-2:])
86
+ input_image = self.model.preprocess(transformed_image)
87
+ self.features = self.model.image_encoder(input_image)
88
+ self.is_image_set = True
89
+
90
+ def predict(
91
+ self,
92
+ point_coords: Optional[np.ndarray] = None,
93
+ point_labels: Optional[np.ndarray] = None,
94
+ box: Optional[np.ndarray] = None,
95
+ mask_input: Optional[np.ndarray] = None,
96
+ multimask_output: bool = True,
97
+ return_logits: bool = False,
98
+ ) -> Tuple[np.ndarray, np.ndarray, np.ndarray]:
99
+ """
100
+ Predict masks for the given input prompts, using the currently set image.
101
+
102
+ Arguments:
103
+ point_coords (np.ndarray or None): A Nx2 array of point prompts to the
104
+ model. Each point is in (X,Y) in pixels.
105
+ point_labels (np.ndarray or None): A length N array of labels for the
106
+ point prompts. 1 indicates a foreground point and 0 indicates a
107
+ background point.
108
+ box (np.ndarray or None): A length 4 array given a box prompt to the
109
+ model, in XYXY format.
110
+ mask_input (np.ndarray): A low resolution mask input to the model, typically
111
+ coming from a previous prediction iteration. Has form 1xHxW, where
112
+ for SAM, H=W=256.
113
+ multimask_output (bool): If true, the model will return three masks.
114
+ For ambiguous input prompts (such as a single click), this will often
115
+ produce better masks than a single prediction. If only a single
116
+ mask is needed, the model's predicted quality score can be used
117
+ to select the best mask. For non-ambiguous prompts, such as multiple
118
+ input prompts, multimask_output=False can give better results.
119
+ return_logits (bool): If true, returns un-thresholded masks logits
120
+ instead of a binary mask.
121
+
122
+ Returns:
123
+ (np.ndarray): The output masks in CxHxW format, where C is the
124
+ number of masks, and (H, W) is the original image size.
125
+ (np.ndarray): An array of length C containing the model's
126
+ predictions for the quality of each mask.
127
+ (np.ndarray): An array of shape CxHxW, where C is the number
128
+ of masks and H=W=256. These low resolution logits can be passed to
129
+ a subsequent iteration as mask input.
130
+ """
131
+ if not self.is_image_set:
132
+ raise RuntimeError("An image must be set with .set_image(...) before mask prediction.")
133
+
134
+ # Transform input prompts
135
+ coords_torch, labels_torch, box_torch, mask_input_torch = None, None, None, None
136
+ if point_coords is not None:
137
+ assert (
138
+ point_labels is not None
139
+ ), "point_labels must be supplied if point_coords is supplied."
140
+ point_coords = self.transform.apply_coords(point_coords, self.original_size)
141
+ coords_torch = torch.as_tensor(point_coords, dtype=torch.float, device=self.device)
142
+ labels_torch = torch.as_tensor(point_labels, dtype=torch.int, device=self.device)
143
+ coords_torch, labels_torch = coords_torch[None, :, :], labels_torch[None, :]
144
+ if box is not None:
145
+ box = self.transform.apply_boxes(box, self.original_size)
146
+ box_torch = torch.as_tensor(box, dtype=torch.float, device=self.device)
147
+ box_torch = box_torch[None, :]
148
+ if mask_input is not None:
149
+ mask_input_torch = torch.as_tensor(mask_input, dtype=torch.float, device=self.device)
150
+ mask_input_torch = mask_input_torch[None, :, :, :]
151
+
152
+ masks, iou_predictions, low_res_masks = self.predict_torch(
153
+ coords_torch,
154
+ labels_torch,
155
+ box_torch,
156
+ mask_input_torch,
157
+ multimask_output,
158
+ return_logits=return_logits,
159
+ )
160
+
161
+ masks_np = masks[0].detach().cpu().numpy()
162
+ iou_predictions_np = iou_predictions[0].detach().cpu().numpy()
163
+ low_res_masks_np = low_res_masks[0].detach().cpu().numpy()
164
+ return masks_np, iou_predictions_np, low_res_masks_np
165
+
166
+ @torch.no_grad()
167
+ def predict_torch(
168
+ self,
169
+ point_coords: Optional[torch.Tensor],
170
+ point_labels: Optional[torch.Tensor],
171
+ boxes: Optional[torch.Tensor] = None,
172
+ mask_input: Optional[torch.Tensor] = None,
173
+ multimask_output: bool = True,
174
+ return_logits: bool = False,
175
+ ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
176
+ """
177
+ Predict masks for the given input prompts, using the currently set image.
178
+ Input prompts are batched torch tensors and are expected to already be
179
+ transformed to the input frame using ResizeLongestSide.
180
+
181
+ Arguments:
182
+ point_coords (torch.Tensor or None): A BxNx2 array of point prompts to the
183
+ model. Each point is in (X,Y) in pixels.
184
+ point_labels (torch.Tensor or None): A BxN array of labels for the
185
+ point prompts. 1 indicates a foreground point and 0 indicates a
186
+ background point.
187
+ boxes (np.ndarray or None): A Bx4 array given a box prompt to the
188
+ model, in XYXY format.
189
+ mask_input (np.ndarray): A low resolution mask input to the model, typically
190
+ coming from a previous prediction iteration. Has form Bx1xHxW, where
191
+ for SAM, H=W=256. Masks returned by a previous iteration of the
192
+ predict method do not need further transformation.
193
+ multimask_output (bool): If true, the model will return three masks.
194
+ For ambiguous input prompts (such as a single click), this will often
195
+ produce better masks than a single prediction. If only a single
196
+ mask is needed, the model's predicted quality score can be used
197
+ to select the best mask. For non-ambiguous prompts, such as multiple
198
+ input prompts, multimask_output=False can give better results.
199
+ return_logits (bool): If true, returns un-thresholded masks logits
200
+ instead of a binary mask.
201
+
202
+ Returns:
203
+ (torch.Tensor): The output masks in BxCxHxW format, where C is the
204
+ number of masks, and (H, W) is the original image size.
205
+ (torch.Tensor): An array of shape BxC containing the model's
206
+ predictions for the quality of each mask.
207
+ (torch.Tensor): An array of shape BxCxHxW, where C is the number
208
+ of masks and H=W=256. These low res logits can be passed to
209
+ a subsequent iteration as mask input.
210
+ """
211
+ if not self.is_image_set:
212
+ raise RuntimeError("An image must be set with .set_image(...) before mask prediction.")
213
+
214
+ if point_coords is not None:
215
+ points = (point_coords, point_labels)
216
+ else:
217
+ points = None
218
+
219
+ # Embed prompts
220
+ sparse_embeddings, dense_embeddings = self.model.prompt_encoder(
221
+ points=points,
222
+ boxes=boxes,
223
+ masks=mask_input,
224
+ )
225
+
226
+ # Predict masks
227
+ low_res_masks, iou_predictions = self.model.mask_decoder(
228
+ image_embeddings=self.features,
229
+ image_pe=self.model.prompt_encoder.get_dense_pe(),
230
+ sparse_prompt_embeddings=sparse_embeddings,
231
+ dense_prompt_embeddings=dense_embeddings,
232
+ multimask_output=multimask_output,
233
+ )
234
+
235
+ # Upscale the masks to the original image resolution
236
+ masks = self.model.postprocess_masks(low_res_masks, self.input_size, self.original_size)
237
+
238
+ if not return_logits:
239
+ masks = masks > self.model.mask_threshold
240
+
241
+ return masks, iou_predictions, low_res_masks
242
+
243
+ def get_image_embedding(self) -> torch.Tensor:
244
+ """
245
+ Returns the image embeddings for the currently set image, with
246
+ shape 1xCxHxW, where C is the embedding dimension and (H,W) are
247
+ the embedding spatial dimension of SAM (typically C=256, H=W=64).
248
+ """
249
+ if not self.is_image_set:
250
+ raise RuntimeError(
251
+ "An image must be set with .set_image(...) to generate an embedding."
252
+ )
253
+ assert self.features is not None, "Features must exist if an image has been set."
254
+ return self.features
255
+
256
+ @property
257
+ def device(self) -> torch.device:
258
+ return self.model.device
259
+
260
+ def reset_image(self) -> None:
261
+ """Resets the currently set image."""
262
+ self.is_image_set = False
263
+ self.features = None
264
+ self.orig_h = None
265
+ self.orig_w = None
266
+ self.input_h = None
267
+ self.input_w = None
app/sam/prompt_encoder.py ADDED
@@ -0,0 +1,214 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Copyright (c) Meta Platforms, Inc. and affiliates.
2
+ # All rights reserved.
3
+
4
+ # This source code is licensed under the license found in the
5
+ # LICENSE file in the root directory of this source tree.
6
+
7
+ import numpy as np
8
+ import torch
9
+ from torch import nn
10
+
11
+ from typing import Any, Optional, Tuple, Type
12
+
13
+ from .common import LayerNorm2d
14
+
15
+
16
+ class PromptEncoder(nn.Module):
17
+ def __init__(
18
+ self,
19
+ embed_dim: int,
20
+ image_embedding_size: Tuple[int, int],
21
+ input_image_size: Tuple[int, int],
22
+ mask_in_chans: int,
23
+ activation: Type[nn.Module] = nn.GELU,
24
+ ) -> None:
25
+ """
26
+ Encodes prompts for input to SAM's mask decoder.
27
+
28
+ Arguments:
29
+ embed_dim (int): The prompts' embedding dimension
30
+ image_embedding_size (tuple(int, int)): The spatial size of the
31
+ image embedding, as (H, W).
32
+ input_image_size (int): The padded size of the image as input
33
+ to the image encoder, as (H, W).
34
+ mask_in_chans (int): The number of hidden channels used for
35
+ encoding input masks.
36
+ activation (nn.Module): The activation to use when encoding
37
+ input masks.
38
+ """
39
+ super().__init__()
40
+ self.embed_dim = embed_dim
41
+ self.input_image_size = input_image_size
42
+ self.image_embedding_size = image_embedding_size
43
+ self.pe_layer = PositionEmbeddingRandom(embed_dim // 2)
44
+
45
+ self.num_point_embeddings: int = 4 # pos/neg point + 2 box corners
46
+ point_embeddings = [nn.Embedding(1, embed_dim) for i in range(self.num_point_embeddings)]
47
+ self.point_embeddings = nn.ModuleList(point_embeddings)
48
+ self.not_a_point_embed = nn.Embedding(1, embed_dim)
49
+
50
+ self.mask_input_size = (4 * image_embedding_size[0], 4 * image_embedding_size[1])
51
+ self.mask_downscaling = nn.Sequential(
52
+ nn.Conv2d(1, mask_in_chans // 4, kernel_size=2, stride=2),
53
+ LayerNorm2d(mask_in_chans // 4),
54
+ activation(),
55
+ nn.Conv2d(mask_in_chans // 4, mask_in_chans, kernel_size=2, stride=2),
56
+ LayerNorm2d(mask_in_chans),
57
+ activation(),
58
+ nn.Conv2d(mask_in_chans, embed_dim, kernel_size=1),
59
+ )
60
+ self.no_mask_embed = nn.Embedding(1, embed_dim)
61
+
62
+ def get_dense_pe(self) -> torch.Tensor:
63
+ """
64
+ Returns the positional encoding used to encode point prompts,
65
+ applied to a dense set of points the shape of the image encoding.
66
+
67
+ Returns:
68
+ torch.Tensor: Positional encoding with shape
69
+ 1x(embed_dim)x(embedding_h)x(embedding_w)
70
+ """
71
+ return self.pe_layer(self.image_embedding_size).unsqueeze(0)
72
+
73
+ def _embed_points(
74
+ self,
75
+ points: torch.Tensor,
76
+ labels: torch.Tensor,
77
+ pad: bool,
78
+ ) -> torch.Tensor:
79
+ """Embeds point prompts."""
80
+ points = points + 0.5 # Shift to center of pixel
81
+ if pad:
82
+ padding_point = torch.zeros((points.shape[0], 1, 2), device=points.device)
83
+ padding_label = -torch.ones((labels.shape[0], 1), device=labels.device)
84
+ points = torch.cat([points, padding_point], dim=1)
85
+ labels = torch.cat([labels, padding_label], dim=1)
86
+ point_embedding = self.pe_layer.forward_with_coords(points, self.input_image_size)
87
+ point_embedding[labels == -1] = 0.0
88
+ point_embedding[labels == -1] += self.not_a_point_embed.weight
89
+ point_embedding[labels == 0] += self.point_embeddings[0].weight
90
+ point_embedding[labels == 1] += self.point_embeddings[1].weight
91
+ return point_embedding
92
+
93
+ def _embed_boxes(self, boxes: torch.Tensor) -> torch.Tensor:
94
+ """Embeds box prompts."""
95
+ boxes = boxes + 0.5 # Shift to center of pixel
96
+ coords = boxes.reshape(-1, 2, 2)
97
+ corner_embedding = self.pe_layer.forward_with_coords(coords, self.input_image_size)
98
+ corner_embedding[:, 0, :] += self.point_embeddings[2].weight
99
+ corner_embedding[:, 1, :] += self.point_embeddings[3].weight
100
+ return corner_embedding
101
+
102
+ def _embed_masks(self, masks: torch.Tensor) -> torch.Tensor:
103
+ """Embeds mask inputs."""
104
+ mask_embedding = self.mask_downscaling(masks)
105
+ return mask_embedding
106
+
107
+ def _get_batch_size(
108
+ self,
109
+ points: Optional[Tuple[torch.Tensor, torch.Tensor]],
110
+ boxes: Optional[torch.Tensor],
111
+ masks: Optional[torch.Tensor],
112
+ ) -> int:
113
+ """
114
+ Gets the batch size of the output given the batch size of the input prompts.
115
+ """
116
+ if points is not None:
117
+ return points[0].shape[0]
118
+ elif boxes is not None:
119
+ return boxes.shape[0]
120
+ elif masks is not None:
121
+ return masks.shape[0]
122
+ else:
123
+ return 1
124
+
125
+ def _get_device(self) -> torch.device:
126
+ return self.point_embeddings[0].weight.device
127
+
128
+ def forward(
129
+ self,
130
+ points: Optional[Tuple[torch.Tensor, torch.Tensor]],
131
+ boxes: Optional[torch.Tensor],
132
+ masks: Optional[torch.Tensor],
133
+ ) -> Tuple[torch.Tensor, torch.Tensor]:
134
+ """
135
+ Embeds different types of prompts, returning both sparse and dense
136
+ embeddings.
137
+
138
+ Arguments:
139
+ points (tuple(torch.Tensor, torch.Tensor) or none): point coordinates
140
+ and labels to embed.
141
+ boxes (torch.Tensor or none): boxes to embed
142
+ masks (torch.Tensor or none): masks to embed
143
+
144
+ Returns:
145
+ torch.Tensor: sparse embeddings for the points and boxes, with shape
146
+ BxNx(embed_dim), where N is determined by the number of input points
147
+ and boxes.
148
+ torch.Tensor: dense embeddings for the masks, in the shape
149
+ Bx(embed_dim)x(embed_H)x(embed_W)
150
+ """
151
+ bs = self._get_batch_size(points, boxes, masks)
152
+ sparse_embeddings = torch.empty((bs, 0, self.embed_dim), device=self._get_device())
153
+ if points is not None:
154
+ coords, labels = points
155
+ point_embeddings = self._embed_points(coords, labels, pad=(boxes is None))
156
+ sparse_embeddings = torch.cat([sparse_embeddings, point_embeddings], dim=1)
157
+ if boxes is not None:
158
+ box_embeddings = self._embed_boxes(boxes)
159
+ sparse_embeddings = torch.cat([sparse_embeddings, box_embeddings], dim=1)
160
+
161
+ if masks is not None:
162
+ dense_embeddings = self._embed_masks(masks)
163
+ else:
164
+ dense_embeddings = self.no_mask_embed.weight.reshape(1, -1, 1, 1).expand(
165
+ bs, -1, self.image_embedding_size[0], self.image_embedding_size[1]
166
+ )
167
+
168
+ return sparse_embeddings, dense_embeddings
169
+
170
+
171
+ class PositionEmbeddingRandom(nn.Module):
172
+ """
173
+ Positional encoding using random spatial frequencies.
174
+ """
175
+
176
+ def __init__(self, num_pos_feats: int = 64, scale: Optional[float] = None) -> None:
177
+ super().__init__()
178
+ if scale is None or scale <= 0.0:
179
+ scale = 1.0
180
+ self.register_buffer(
181
+ "positional_encoding_gaussian_matrix",
182
+ scale * torch.randn((2, num_pos_feats)),
183
+ )
184
+
185
+ def _pe_encoding(self, coords: torch.Tensor) -> torch.Tensor:
186
+ """Positionally encode points that are normalized to [0,1]."""
187
+ # assuming coords are in [0, 1]^2 square and have d_1 x ... x d_n x 2 shape
188
+ coords = 2 * coords - 1
189
+ coords = coords @ self.positional_encoding_gaussian_matrix
190
+ coords = 2 * np.pi * coords
191
+ # outputs d_1 x ... x d_n x C shape
192
+ return torch.cat([torch.sin(coords), torch.cos(coords)], dim=-1)
193
+
194
+ def forward(self, size: Tuple[int, int]) -> torch.Tensor:
195
+ """Generate positional encoding for a grid of the specified size."""
196
+ h, w = size
197
+ device: Any = self.positional_encoding_gaussian_matrix.device
198
+ grid = torch.ones((h, w), device=device, dtype=torch.float32)
199
+ y_embed = grid.cumsum(dim=0) - 0.5
200
+ x_embed = grid.cumsum(dim=1) - 0.5
201
+ y_embed = y_embed / h
202
+ x_embed = x_embed / w
203
+
204
+ pe = self._pe_encoding(torch.stack([x_embed, y_embed], dim=-1))
205
+ return pe.permute(2, 0, 1) # C x H x W
206
+
207
+ def forward_with_coords(
208
+ self, coords_input: torch.Tensor, image_size: Tuple[int, int]
209
+ ) -> torch.Tensor:
210
+ """Positionally encode points that are not normalized to [0,1]."""
211
+ coords = coords_input.clone()
212
+ coords[:, :, 0] = coords[:, :, 0] / image_size[1]
213
+ coords[:, :, 1] = coords[:, :, 1] / image_size[0]
214
+ return self._pe_encoding(coords.to(torch.float)) # B x N x C
app/sam/sam.py ADDED
@@ -0,0 +1,174 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Copyright (c) Meta Platforms, Inc. and affiliates.
2
+ # All rights reserved.
3
+
4
+ # This source code is licensed under the license found in the
5
+ # LICENSE file in the root directory of this source tree.
6
+
7
+ import torch
8
+ from torch import nn
9
+ from torch.nn import functional as F
10
+
11
+ from typing import Any, Dict, List, Tuple
12
+
13
+ from .image_encoder import ImageEncoderViT
14
+ from .mask_decoder import MaskDecoder
15
+ from .prompt_encoder import PromptEncoder
16
+
17
+
18
+ class Sam(nn.Module):
19
+ mask_threshold: float = 0.0
20
+ image_format: str = "RGB"
21
+
22
+ def __init__(
23
+ self,
24
+ image_encoder: ImageEncoderViT,
25
+ prompt_encoder: PromptEncoder,
26
+ mask_decoder: MaskDecoder,
27
+ pixel_mean: List[float] = [123.675, 116.28, 103.53],
28
+ pixel_std: List[float] = [58.395, 57.12, 57.375],
29
+ ) -> None:
30
+ """
31
+ SAM predicts object masks from an image and input prompts.
32
+
33
+ Arguments:
34
+ image_encoder (ImageEncoderViT): The backbone used to encode the
35
+ image into image embeddings that allow for efficient mask prediction.
36
+ prompt_encoder (PromptEncoder): Encodes various types of input prompts.
37
+ mask_decoder (MaskDecoder): Predicts masks from the image embeddings
38
+ and encoded prompts.
39
+ pixel_mean (list(float)): Mean values for normalizing pixels in the input image.
40
+ pixel_std (list(float)): Std values for normalizing pixels in the input image.
41
+ """
42
+ super().__init__()
43
+ self.image_encoder = image_encoder
44
+ self.prompt_encoder = prompt_encoder
45
+ self.mask_decoder = mask_decoder
46
+ self.register_buffer("pixel_mean", torch.Tensor(pixel_mean).view(-1, 1, 1), False)
47
+ self.register_buffer("pixel_std", torch.Tensor(pixel_std).view(-1, 1, 1), False)
48
+
49
+ @property
50
+ def device(self) -> Any:
51
+ return self.pixel_mean.device
52
+
53
+ @torch.no_grad()
54
+ def forward(
55
+ self,
56
+ batched_input: List[Dict[str, Any]],
57
+ multimask_output: bool,
58
+ ) -> List[Dict[str, torch.Tensor]]:
59
+ """
60
+ Predicts masks end-to-end from provided images and prompts.
61
+ If prompts are not known in advance, using SamPredictor is
62
+ recommended over calling the model directly.
63
+
64
+ Arguments:
65
+ batched_input (list(dict)): A list over input images, each a
66
+ dictionary with the following keys. A prompt key can be
67
+ excluded if it is not present.
68
+ 'image': The image as a torch tensor in 3xHxW format,
69
+ already transformed for input to the model.
70
+ 'original_size': (tuple(int, int)) The original size of
71
+ the image before transformation, as (H, W).
72
+ 'point_coords': (torch.Tensor) Batched point prompts for
73
+ this image, with shape BxNx2. Already transformed to the
74
+ input frame of the model.
75
+ 'point_labels': (torch.Tensor) Batched labels for point prompts,
76
+ with shape BxN.
77
+ 'boxes': (torch.Tensor) Batched box inputs, with shape Bx4.
78
+ Already transformed to the input frame of the model.
79
+ 'mask_inputs': (torch.Tensor) Batched mask inputs to the model,
80
+ in the form Bx1xHxW.
81
+ multimask_output (bool): Whether the model should predict multiple
82
+ disambiguating masks, or return a single mask.
83
+
84
+ Returns:
85
+ (list(dict)): A list over input images, where each element is
86
+ as dictionary with the following keys.
87
+ 'masks': (torch.Tensor) Batched binary mask predictions,
88
+ with shape BxCxHxW, where B is the number of input prompts,
89
+ C is determined by multimask_output, and (H, W) is the
90
+ original size of the image.
91
+ 'iou_predictions': (torch.Tensor) The model's predictions
92
+ of mask quality, in shape BxC.
93
+ 'low_res_logits': (torch.Tensor) Low resolution logits with
94
+ shape BxCxHxW, where H=W=256. Can be passed as mask input
95
+ to subsequent iterations of prediction.
96
+ """
97
+ input_images = torch.stack([self.preprocess(x["image"]) for x in batched_input], dim=0)
98
+ image_embeddings = self.image_encoder(input_images)
99
+
100
+ outputs = []
101
+ for image_record, curr_embedding in zip(batched_input, image_embeddings):
102
+ if "point_coords" in image_record:
103
+ points = (image_record["point_coords"], image_record["point_labels"])
104
+ else:
105
+ points = None
106
+ sparse_embeddings, dense_embeddings = self.prompt_encoder(
107
+ points=points,
108
+ boxes=image_record.get("boxes", None),
109
+ masks=image_record.get("mask_inputs", None),
110
+ )
111
+ low_res_masks, iou_predictions = self.mask_decoder(
112
+ image_embeddings=curr_embedding.unsqueeze(0),
113
+ image_pe=self.prompt_encoder.get_dense_pe(),
114
+ sparse_prompt_embeddings=sparse_embeddings,
115
+ dense_prompt_embeddings=dense_embeddings,
116
+ multimask_output=multimask_output,
117
+ )
118
+ masks = self.postprocess_masks(
119
+ low_res_masks,
120
+ input_size=image_record["image"].shape[-2:],
121
+ original_size=image_record["original_size"],
122
+ )
123
+ masks = masks > self.mask_threshold
124
+ outputs.append(
125
+ {
126
+ "masks": masks,
127
+ "iou_predictions": iou_predictions,
128
+ "low_res_logits": low_res_masks,
129
+ }
130
+ )
131
+ return outputs
132
+
133
+ def postprocess_masks(
134
+ self,
135
+ masks: torch.Tensor,
136
+ input_size: Tuple[int, ...],
137
+ original_size: Tuple[int, ...],
138
+ ) -> torch.Tensor:
139
+ """
140
+ Remove padding and upscale masks to the original image size.
141
+
142
+ Arguments:
143
+ masks (torch.Tensor): Batched masks from the mask_decoder,
144
+ in BxCxHxW format.
145
+ input_size (tuple(int, int)): The size of the image input to the
146
+ model, in (H, W) format. Used to remove padding.
147
+ original_size (tuple(int, int)): The original size of the image
148
+ before resizing for input to the model, in (H, W) format.
149
+
150
+ Returns:
151
+ (torch.Tensor): Batched masks in BxCxHxW format, where (H, W)
152
+ is given by original_size.
153
+ """
154
+ masks = F.interpolate(
155
+ masks,
156
+ (self.image_encoder.img_size, self.image_encoder.img_size),
157
+ mode="bilinear",
158
+ align_corners=False,
159
+ )
160
+ masks = masks[..., : input_size[0], : input_size[1]]
161
+ masks = F.interpolate(masks, original_size, mode="bilinear", align_corners=False)
162
+ return masks
163
+
164
+ def preprocess(self, x: torch.Tensor) -> torch.Tensor:
165
+ """Normalize pixel values and pad to a square input."""
166
+ # Normalize colors
167
+ x = (x - self.pixel_mean) / self.pixel_std
168
+
169
+ # Pad
170
+ h, w = x.shape[-2:]
171
+ padh = self.image_encoder.img_size - h
172
+ padw = self.image_encoder.img_size - w
173
+ x = F.pad(x, (0, padw, 0, padh))
174
+ return x
app/sam/transformer.py ADDED
@@ -0,0 +1,240 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Copyright (c) Meta Platforms, Inc. and affiliates.
2
+ # All rights reserved.
3
+
4
+ # This source code is licensed under the license found in the
5
+ # LICENSE file in the root directory of this source tree.
6
+
7
+ import torch
8
+ from torch import Tensor, nn
9
+
10
+ import math
11
+ from typing import Tuple, Type
12
+
13
+ from .common import MLPBlock
14
+
15
+
16
+ class TwoWayTransformer(nn.Module):
17
+ def __init__(
18
+ self,
19
+ depth: int,
20
+ embedding_dim: int,
21
+ num_heads: int,
22
+ mlp_dim: int,
23
+ activation: Type[nn.Module] = nn.ReLU,
24
+ attention_downsample_rate: int = 2,
25
+ ) -> None:
26
+ """
27
+ A transformer decoder that attends to an input image using
28
+ queries whose positional embedding is supplied.
29
+
30
+ Args:
31
+ depth (int): number of layers in the transformer
32
+ embedding_dim (int): the channel dimension for the input embeddings
33
+ num_heads (int): the number of heads for multihead attention. Must
34
+ divide embedding_dim
35
+ mlp_dim (int): the channel dimension internal to the MLP block
36
+ activation (nn.Module): the activation to use in the MLP block
37
+ """
38
+ super().__init__()
39
+ self.depth = depth
40
+ self.embedding_dim = embedding_dim
41
+ self.num_heads = num_heads
42
+ self.mlp_dim = mlp_dim
43
+ self.layers = nn.ModuleList()
44
+
45
+ for i in range(depth):
46
+ self.layers.append(
47
+ TwoWayAttentionBlock(
48
+ embedding_dim=embedding_dim,
49
+ num_heads=num_heads,
50
+ mlp_dim=mlp_dim,
51
+ activation=activation,
52
+ attention_downsample_rate=attention_downsample_rate,
53
+ skip_first_layer_pe=(i == 0),
54
+ )
55
+ )
56
+
57
+ self.final_attn_token_to_image = Attention(
58
+ embedding_dim, num_heads, downsample_rate=attention_downsample_rate
59
+ )
60
+ self.norm_final_attn = nn.LayerNorm(embedding_dim)
61
+
62
+ def forward(
63
+ self,
64
+ image_embedding: Tensor,
65
+ image_pe: Tensor,
66
+ point_embedding: Tensor,
67
+ ) -> Tuple[Tensor, Tensor]:
68
+ """
69
+ Args:
70
+ image_embedding (torch.Tensor): image to attend to. Should be shape
71
+ B x embedding_dim x h x w for any h and w.
72
+ image_pe (torch.Tensor): the positional encoding to add to the image. Must
73
+ have the same shape as image_embedding.
74
+ point_embedding (torch.Tensor): the embedding to add to the query points.
75
+ Must have shape B x N_points x embedding_dim for any N_points.
76
+
77
+ Returns:
78
+ torch.Tensor: the processed point_embedding
79
+ torch.Tensor: the processed image_embedding
80
+ """
81
+ # BxCxHxW -> BxHWxC == B x N_image_tokens x C
82
+ bs, c, h, w = image_embedding.shape
83
+ image_embedding = image_embedding.flatten(2).permute(0, 2, 1)
84
+ image_pe = image_pe.flatten(2).permute(0, 2, 1)
85
+
86
+ # Prepare queries
87
+ queries = point_embedding
88
+ keys = image_embedding
89
+
90
+ # Apply transformer blocks and final layernorm
91
+ for layer in self.layers:
92
+ queries, keys = layer(
93
+ queries=queries,
94
+ keys=keys,
95
+ query_pe=point_embedding,
96
+ key_pe=image_pe,
97
+ )
98
+
99
+ # Apply the final attention layer from the points to the image
100
+ q = queries + point_embedding
101
+ k = keys + image_pe
102
+ attn_out = self.final_attn_token_to_image(q=q, k=k, v=keys)
103
+ queries = queries + attn_out
104
+ queries = self.norm_final_attn(queries)
105
+
106
+ return queries, keys
107
+
108
+
109
+ class TwoWayAttentionBlock(nn.Module):
110
+ def __init__(
111
+ self,
112
+ embedding_dim: int,
113
+ num_heads: int,
114
+ mlp_dim: int = 2048,
115
+ activation: Type[nn.Module] = nn.ReLU,
116
+ attention_downsample_rate: int = 2,
117
+ skip_first_layer_pe: bool = False,
118
+ ) -> None:
119
+ """
120
+ A transformer block with four layers: (1) self-attention of sparse
121
+ inputs, (2) cross attention of sparse inputs to dense inputs, (3) mlp
122
+ block on sparse inputs, and (4) cross attention of dense inputs to sparse
123
+ inputs.
124
+
125
+ Arguments:
126
+ embedding_dim (int): the channel dimension of the embeddings
127
+ num_heads (int): the number of heads in the attention layers
128
+ mlp_dim (int): the hidden dimension of the mlp block
129
+ activation (nn.Module): the activation of the mlp block
130
+ skip_first_layer_pe (bool): skip the PE on the first layer
131
+ """
132
+ super().__init__()
133
+ self.self_attn = Attention(embedding_dim, num_heads)
134
+ self.norm1 = nn.LayerNorm(embedding_dim)
135
+
136
+ self.cross_attn_token_to_image = Attention(
137
+ embedding_dim, num_heads, downsample_rate=attention_downsample_rate
138
+ )
139
+ self.norm2 = nn.LayerNorm(embedding_dim)
140
+
141
+ self.mlp = MLPBlock(embedding_dim, mlp_dim, activation)
142
+ self.norm3 = nn.LayerNorm(embedding_dim)
143
+
144
+ self.norm4 = nn.LayerNorm(embedding_dim)
145
+ self.cross_attn_image_to_token = Attention(
146
+ embedding_dim, num_heads, downsample_rate=attention_downsample_rate
147
+ )
148
+
149
+ self.skip_first_layer_pe = skip_first_layer_pe
150
+
151
+ def forward(
152
+ self, queries: Tensor, keys: Tensor, query_pe: Tensor, key_pe: Tensor
153
+ ) -> Tuple[Tensor, Tensor]:
154
+ # Self attention block
155
+ if self.skip_first_layer_pe:
156
+ queries = self.self_attn(q=queries, k=queries, v=queries)
157
+ else:
158
+ q = queries + query_pe
159
+ attn_out = self.self_attn(q=q, k=q, v=queries)
160
+ queries = queries + attn_out
161
+ queries = self.norm1(queries)
162
+
163
+ # Cross attention block, tokens attending to image embedding
164
+ q = queries + query_pe
165
+ k = keys + key_pe
166
+ attn_out = self.cross_attn_token_to_image(q=q, k=k, v=keys)
167
+ queries = queries + attn_out
168
+ queries = self.norm2(queries)
169
+
170
+ # MLP block
171
+ mlp_out = self.mlp(queries)
172
+ queries = queries + mlp_out
173
+ queries = self.norm3(queries)
174
+
175
+ # Cross attention block, image embedding attending to tokens
176
+ q = queries + query_pe
177
+ k = keys + key_pe
178
+ attn_out = self.cross_attn_image_to_token(q=k, k=q, v=queries)
179
+ keys = keys + attn_out
180
+ keys = self.norm4(keys)
181
+
182
+ return queries, keys
183
+
184
+
185
+ class Attention(nn.Module):
186
+ """
187
+ An attention layer that allows for downscaling the size of the embedding
188
+ after projection to queries, keys, and values.
189
+ """
190
+
191
+ def __init__(
192
+ self,
193
+ embedding_dim: int,
194
+ num_heads: int,
195
+ downsample_rate: int = 1,
196
+ ) -> None:
197
+ super().__init__()
198
+ self.embedding_dim = embedding_dim
199
+ self.internal_dim = embedding_dim // downsample_rate
200
+ self.num_heads = num_heads
201
+ assert self.internal_dim % num_heads == 0, "num_heads must divide embedding_dim."
202
+
203
+ self.q_proj = nn.Linear(embedding_dim, self.internal_dim)
204
+ self.k_proj = nn.Linear(embedding_dim, self.internal_dim)
205
+ self.v_proj = nn.Linear(embedding_dim, self.internal_dim)
206
+ self.out_proj = nn.Linear(self.internal_dim, embedding_dim)
207
+
208
+ def _separate_heads(self, x: Tensor, num_heads: int) -> Tensor:
209
+ b, n, c = x.shape
210
+ x = x.reshape(b, n, num_heads, c // num_heads)
211
+ return x.transpose(1, 2) # B x N_heads x N_tokens x C_per_head
212
+
213
+ def _recombine_heads(self, x: Tensor) -> Tensor:
214
+ b, n_heads, n_tokens, c_per_head = x.shape
215
+ x = x.transpose(1, 2)
216
+ return x.reshape(b, n_tokens, n_heads * c_per_head) # B x N_tokens x C
217
+
218
+ def forward(self, q: Tensor, k: Tensor, v: Tensor) -> Tensor:
219
+ # Input projections
220
+ q = self.q_proj(q)
221
+ k = self.k_proj(k)
222
+ v = self.v_proj(v)
223
+
224
+ # Separate into heads
225
+ q = self._separate_heads(q, self.num_heads)
226
+ k = self._separate_heads(k, self.num_heads)
227
+ v = self._separate_heads(v, self.num_heads)
228
+
229
+ # Attention
230
+ _, _, _, c_per_head = q.shape
231
+ attn = q @ k.permute(0, 1, 3, 2) # B x N_heads x N_tokens x N_tokens
232
+ attn = attn / math.sqrt(c_per_head)
233
+ attn = torch.softmax(attn, dim=-1)
234
+
235
+ # Get output
236
+ out = attn @ v
237
+ out = self._recombine_heads(out)
238
+ out = self.out_proj(out)
239
+
240
+ return out
app/sam/transforms.py ADDED
@@ -0,0 +1,102 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Copyright (c) Meta Platforms, Inc. and affiliates.
2
+ # All rights reserved.
3
+
4
+ # This source code is licensed under the license found in the
5
+ # LICENSE file in the root directory of this source tree.
6
+
7
+ import numpy as np
8
+ import torch
9
+ from torch.nn import functional as F
10
+ from torchvision.transforms.functional import resize, to_pil_image # type: ignore
11
+
12
+ from copy import deepcopy
13
+ from typing import Tuple
14
+
15
+
16
+ class ResizeLongestSide:
17
+ """
18
+ Resizes images to the longest side 'target_length', as well as provides
19
+ methods for resizing coordinates and boxes. Provides methods for
20
+ transforming both numpy array and batched torch tensors.
21
+ """
22
+
23
+ def __init__(self, target_length: int) -> None:
24
+ self.target_length = target_length
25
+
26
+ def apply_image(self, image: np.ndarray) -> np.ndarray:
27
+ """
28
+ Expects a numpy array with shape HxWxC in uint8 format.
29
+ """
30
+ target_size = self.get_preprocess_shape(image.shape[0], image.shape[1], self.target_length)
31
+ return np.array(resize(to_pil_image(image), target_size))
32
+
33
+ def apply_coords(self, coords: np.ndarray, original_size: Tuple[int, ...]) -> np.ndarray:
34
+ """
35
+ Expects a numpy array of length 2 in the final dimension. Requires the
36
+ original image size in (H, W) format.
37
+ """
38
+ old_h, old_w = original_size
39
+ new_h, new_w = self.get_preprocess_shape(
40
+ original_size[0], original_size[1], self.target_length
41
+ )
42
+ coords = deepcopy(coords).astype(float)
43
+ coords[..., 0] = coords[..., 0] * (new_w / old_w)
44
+ coords[..., 1] = coords[..., 1] * (new_h / old_h)
45
+ return coords
46
+
47
+ def apply_boxes(self, boxes: np.ndarray, original_size: Tuple[int, ...]) -> np.ndarray:
48
+ """
49
+ Expects a numpy array shape Bx4. Requires the original image size
50
+ in (H, W) format.
51
+ """
52
+ boxes = self.apply_coords(boxes.reshape(-1, 2, 2), original_size)
53
+ return boxes.reshape(-1, 4)
54
+
55
+ def apply_image_torch(self, image: torch.Tensor) -> torch.Tensor:
56
+ """
57
+ Expects batched images with shape BxCxHxW and float format. This
58
+ transformation may not exactly match apply_image. apply_image is
59
+ the transformation expected by the model.
60
+ """
61
+ # Expects an image in BCHW format. May not exactly match apply_image.
62
+ target_size = self.get_preprocess_shape(image.shape[2], image.shape[3], self.target_length)
63
+ return F.interpolate(
64
+ image, target_size, mode="bilinear", align_corners=False, antialias=True
65
+ )
66
+
67
+ def apply_coords_torch(
68
+ self, coords: torch.Tensor, original_size: Tuple[int, ...]
69
+ ) -> torch.Tensor:
70
+ """
71
+ Expects a torch tensor with length 2 in the last dimension. Requires the
72
+ original image size in (H, W) format.
73
+ """
74
+ old_h, old_w = original_size
75
+ new_h, new_w = self.get_preprocess_shape(
76
+ original_size[0], original_size[1], self.target_length
77
+ )
78
+ coords = deepcopy(coords).to(torch.float)
79
+ coords[..., 0] = coords[..., 0] * (new_w / old_w)
80
+ coords[..., 1] = coords[..., 1] * (new_h / old_h)
81
+ return coords
82
+
83
+ def apply_boxes_torch(
84
+ self, boxes: torch.Tensor, original_size: Tuple[int, ...]
85
+ ) -> torch.Tensor:
86
+ """
87
+ Expects a torch tensor with shape Bx4. Requires the original image
88
+ size in (H, W) format.
89
+ """
90
+ boxes = self.apply_coords_torch(boxes.reshape(-1, 2, 2), original_size)
91
+ return boxes.reshape(-1, 4)
92
+
93
+ @staticmethod
94
+ def get_preprocess_shape(oldh: int, oldw: int, long_side_length: int) -> Tuple[int, int]:
95
+ """
96
+ Compute the output size given input size and target long side length.
97
+ """
98
+ scale = long_side_length * 1.0 / max(oldh, oldw)
99
+ newh, neww = oldh * scale, oldw * scale
100
+ neww = int(neww + 0.5)
101
+ newh = int(newh + 0.5)
102
+ return (newh, neww)
app/sam_hq ADDED
@@ -0,0 +1 @@
 
 
1
+ Subproject commit 759948f24f5524e5946bc274b4086ff1cc13b676
app/transforms.py ADDED
@@ -0,0 +1,102 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Copyright (c) Meta Platforms, Inc. and affiliates.
2
+ # All rights reserved.
3
+
4
+ # This source code is licensed under the license found in the
5
+ # LICENSE file in the root directory of this source tree.
6
+
7
+ import numpy as np
8
+ import torch
9
+ from torch.nn import functional as F
10
+ from torchvision.transforms.functional import resize, to_pil_image # type: ignore
11
+
12
+ from copy import deepcopy
13
+ from typing import Tuple
14
+
15
+
16
+ class ResizeLongestSide:
17
+ """
18
+ Resizes images to the longest side 'target_length', as well as provides
19
+ methods for resizing coordinates and boxes. Provides methods for
20
+ transforming both numpy array and batched torch tensors.
21
+ """
22
+
23
+ def __init__(self, target_length: int) -> None:
24
+ self.target_length = target_length
25
+
26
+ def apply_image(self, image: np.ndarray) -> np.ndarray:
27
+ """
28
+ Expects a numpy array with shape HxWxC in uint8 format.
29
+ """
30
+ target_size = self.get_preprocess_shape(image.shape[0], image.shape[1], self.target_length)
31
+ return np.array(resize(to_pil_image(image), target_size))
32
+
33
+ def apply_coords(self, coords: np.ndarray, original_size: Tuple[int, ...]) -> np.ndarray:
34
+ """
35
+ Expects a numpy array of length 2 in the final dimension. Requires the
36
+ original image size in (H, W) format.
37
+ """
38
+ old_h, old_w = original_size
39
+ new_h, new_w = self.get_preprocess_shape(
40
+ original_size[0], original_size[1], self.target_length
41
+ )
42
+ coords = deepcopy(coords).astype(float)
43
+ coords[..., 0] = coords[..., 0] * (new_w / old_w)
44
+ coords[..., 1] = coords[..., 1] * (new_h / old_h)
45
+ return coords
46
+
47
+ def apply_boxes(self, boxes: np.ndarray, original_size: Tuple[int, ...]) -> np.ndarray:
48
+ """
49
+ Expects a numpy array shape Bx4. Requires the original image size
50
+ in (H, W) format.
51
+ """
52
+ boxes = self.apply_coords(boxes.reshape(-1, 2, 2), original_size)
53
+ return boxes.reshape(-1, 4)
54
+
55
+ def apply_image_torch(self, image: torch.Tensor) -> torch.Tensor:
56
+ """
57
+ Expects batched images with shape BxCxHxW and float format. This
58
+ transformation may not exactly match apply_image. apply_image is
59
+ the transformation expected by the model.
60
+ """
61
+ # Expects an image in BCHW format. May not exactly match apply_image.
62
+ target_size = self.get_preprocess_shape(image.shape[2], image.shape[3], self.target_length)
63
+ return F.interpolate(
64
+ image, target_size, mode="bilinear", align_corners=False, antialias=True
65
+ )
66
+
67
+ def apply_coords_torch(
68
+ self, coords: torch.Tensor, original_size: Tuple[int, ...]
69
+ ) -> torch.Tensor:
70
+ """
71
+ Expects a torch tensor with length 2 in the last dimension. Requires the
72
+ original image size in (H, W) format.
73
+ """
74
+ old_h, old_w = original_size
75
+ new_h, new_w = self.get_preprocess_shape(
76
+ original_size[0], original_size[1], self.target_length
77
+ )
78
+ coords = deepcopy(coords).to(torch.float)
79
+ coords[..., 0] = coords[..., 0] * (new_w / old_w)
80
+ coords[..., 1] = coords[..., 1] * (new_h / old_h)
81
+ return coords
82
+
83
+ def apply_boxes_torch(
84
+ self, boxes: torch.Tensor, original_size: Tuple[int, ...]
85
+ ) -> torch.Tensor:
86
+ """
87
+ Expects a torch tensor with shape Bx4. Requires the original image
88
+ size in (H, W) format.
89
+ """
90
+ boxes = self.apply_coords_torch(boxes.reshape(-1, 2, 2), original_size)
91
+ return boxes.reshape(-1, 4)
92
+
93
+ @staticmethod
94
+ def get_preprocess_shape(oldh: int, oldw: int, long_side_length: int) -> Tuple[int, int]:
95
+ """
96
+ Compute the output size given input size and target long side length.
97
+ """
98
+ scale = long_side_length * 1.0 / max(oldh, oldw)
99
+ newh, neww = oldh * scale, oldw * scale
100
+ neww = int(neww + 0.5)
101
+ newh = int(newh + 0.5)
102
+ return (newh, neww)
instance-labeler ADDED
@@ -0,0 +1 @@
 
 
1
+ Subproject commit f4dde8b7122daba0a4cf869dc78d450fe80ffe81
main.py ADDED
@@ -0,0 +1,5 @@
 
 
 
 
 
 
1
+ import uvicorn
2
+
3
+
4
+ if __name__ == "__main__":
5
+ uvicorn.run("app.main:app", host="0.0.0.0", port=7860, reload=True)
mobile_sam.pth ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:6dbb90523a35330fedd7f1d3dfc66f995213d81b29a5ca8108dbcdd4e37d6c2f
3
+ size 40728226
requirements.txt ADDED
@@ -0,0 +1,8 @@
 
 
 
 
 
 
 
 
 
1
+ torch
2
+ torchvision
3
+ opencv-python
4
+ numpy
5
+ Pillow
6
+ fastapi
7
+ uvicorn
8
+ timm