Spaces:
Runtime error
Runtime error
fix the resizing issue during concurrent visiting
Browse files- app.py +16 -2
- segment_anything/onnx/predictor_onnx.py +12 -16
- segment_anything/predictor.py +17 -20
app.py
CHANGED
@@ -109,6 +109,8 @@ def reset(session_state):
|
|
109 |
session_state['ori_image'] = None
|
110 |
session_state['image_with_prompt'] = None
|
111 |
session_state['feature'] = None
|
|
|
|
|
112 |
return None, None, None, session_state
|
113 |
|
114 |
|
@@ -119,6 +121,8 @@ def reset_all(session_state):
|
|
119 |
session_state['ori_image'] = None
|
120 |
session_state['image_with_prompt'] = None
|
121 |
session_state['feature'] = None
|
|
|
|
|
122 |
return None, None, None, None, None, None, session_state
|
123 |
|
124 |
|
@@ -149,7 +153,7 @@ def on_image_upload(
|
|
149 |
session_state['image_with_prompt'] = copy.deepcopy(image)
|
150 |
print("Image changed")
|
151 |
nd_image = np.array(image)
|
152 |
-
session_state['feature'] = predictor.set_image(nd_image)
|
153 |
|
154 |
return image, None, None, session_state
|
155 |
|
@@ -190,12 +194,16 @@ def segment_with_points(
|
|
190 |
fill=point_color,
|
191 |
)
|
192 |
image = session_state['image_with_prompt']
|
|
|
|
|
193 |
|
194 |
if ENABLE_ONNX:
|
195 |
coord_np = np.array(session_state['coord_list'])[None]
|
196 |
label_np = np.array(session_state['label_list'])[None]
|
197 |
masks, scores, _ = predictor.predict(
|
198 |
features=session_state['feature'],
|
|
|
|
|
199 |
point_coords=coord_np,
|
200 |
point_labels=label_np,
|
201 |
)
|
@@ -206,6 +214,8 @@ def segment_with_points(
|
|
206 |
label_np = np.array(session_state['label_list'])
|
207 |
masks, scores, logits = predictor.predict(
|
208 |
features=session_state['feature'],
|
|
|
|
|
209 |
point_coords=coord_np,
|
210 |
point_labels=label_np,
|
211 |
num_multimask_outputs=4,
|
@@ -233,7 +243,7 @@ def segment_with_points(
|
|
233 |
binary_mask = np.where(annotations[0] > 0.5, 255, 0).astype(np.uint8)
|
234 |
mask = Image.fromarray(binary_mask)
|
235 |
binary_mask = np.expand_dims(binary_mask, axis=2)
|
236 |
-
crop = Image.fromarray(np.concatenate((
|
237 |
return seg, mask, crop, session_state
|
238 |
|
239 |
|
@@ -282,6 +292,8 @@ def segment_with_box(
|
|
282 |
point_labels = np.array([2, 3])[None]
|
283 |
masks, _, _ = predictor.predict(
|
284 |
features=session_state['feature'],
|
|
|
|
|
285 |
point_coords=point_coords,
|
286 |
point_labels=point_labels,
|
287 |
)
|
@@ -289,6 +301,8 @@ def segment_with_box(
|
|
289 |
else:
|
290 |
masks, scores, _ = predictor.predict(
|
291 |
features=session_state['feature'],
|
|
|
|
|
292 |
box=box_np,
|
293 |
num_multimask_outputs=1,
|
294 |
)
|
|
|
109 |
session_state['ori_image'] = None
|
110 |
session_state['image_with_prompt'] = None
|
111 |
session_state['feature'] = None
|
112 |
+
session_state['input_size'] = None
|
113 |
+
session_state['original_size'] = None
|
114 |
return None, None, None, session_state
|
115 |
|
116 |
|
|
|
121 |
session_state['ori_image'] = None
|
122 |
session_state['image_with_prompt'] = None
|
123 |
session_state['feature'] = None
|
124 |
+
session_state['input_size'] = None
|
125 |
+
session_state['original_size'] = None
|
126 |
return None, None, None, None, None, None, session_state
|
127 |
|
128 |
|
|
|
153 |
session_state['image_with_prompt'] = copy.deepcopy(image)
|
154 |
print("Image changed")
|
155 |
nd_image = np.array(image)
|
156 |
+
session_state['feature'], session_state['input_size'], session_state['original_size'] = predictor.set_image(nd_image)
|
157 |
|
158 |
return image, None, None, session_state
|
159 |
|
|
|
194 |
fill=point_color,
|
195 |
)
|
196 |
image = session_state['image_with_prompt']
|
197 |
+
print(f"image: {image.size}")
|
198 |
+
nd_image = np.array(session_state['ori_image'])
|
199 |
|
200 |
if ENABLE_ONNX:
|
201 |
coord_np = np.array(session_state['coord_list'])[None]
|
202 |
label_np = np.array(session_state['label_list'])[None]
|
203 |
masks, scores, _ = predictor.predict(
|
204 |
features=session_state['feature'],
|
205 |
+
input_size=session_state['input_size'],
|
206 |
+
original_size=session_state['original_size'],
|
207 |
point_coords=coord_np,
|
208 |
point_labels=label_np,
|
209 |
)
|
|
|
214 |
label_np = np.array(session_state['label_list'])
|
215 |
masks, scores, logits = predictor.predict(
|
216 |
features=session_state['feature'],
|
217 |
+
input_size=session_state['input_size'],
|
218 |
+
original_size=session_state['original_size'],
|
219 |
point_coords=coord_np,
|
220 |
point_labels=label_np,
|
221 |
num_multimask_outputs=4,
|
|
|
243 |
binary_mask = np.where(annotations[0] > 0.5, 255, 0).astype(np.uint8)
|
244 |
mask = Image.fromarray(binary_mask)
|
245 |
binary_mask = np.expand_dims(binary_mask, axis=2)
|
246 |
+
crop = Image.fromarray(np.concatenate((nd_image, binary_mask), axis=2), "RGBA")
|
247 |
return seg, mask, crop, session_state
|
248 |
|
249 |
|
|
|
292 |
point_labels = np.array([2, 3])[None]
|
293 |
masks, _, _ = predictor.predict(
|
294 |
features=session_state['feature'],
|
295 |
+
input_size=session_state['input_size'],
|
296 |
+
original_size=session_state['original_size'],
|
297 |
point_coords=point_coords,
|
298 |
point_labels=point_labels,
|
299 |
)
|
|
|
301 |
else:
|
302 |
masks, scores, _ = predictor.predict(
|
303 |
features=session_state['feature'],
|
304 |
+
input_size=session_state['input_size'],
|
305 |
+
original_size=session_state['original_size'],
|
306 |
box=box_np,
|
307 |
num_multimask_outputs=1,
|
308 |
)
|
segment_anything/onnx/predictor_onnx.py
CHANGED
@@ -53,34 +53,30 @@ class SamPredictorONNX:
|
|
53 |
input_image = self.transform.apply_image(image)
|
54 |
input_image = input_image.transpose(2, 0, 1)[None, :, :, :]
|
55 |
self.reset_image()
|
56 |
-
|
57 |
-
|
58 |
input_image = self.preprocess(input_image).astype(np.float32)
|
59 |
outputs = self.encoder.run(None, {'image': input_image})
|
60 |
-
|
61 |
-
self.is_image_set = True
|
62 |
|
63 |
-
return
|
64 |
|
65 |
def predict(
|
66 |
self,
|
67 |
-
features: np.ndarray
|
|
|
|
|
68 |
point_coords: Optional[np.ndarray] = None,
|
69 |
point_labels: Optional[np.ndarray] = None,
|
70 |
) -> Tuple[np.ndarray, np.ndarray, np.ndarray]:
|
71 |
-
|
72 |
-
raise RuntimeError("An image must be set with .set_image(...) before mask prediction.")
|
73 |
-
if features is None:
|
74 |
-
features = self.features
|
75 |
-
|
76 |
-
point_coords = self.transform.apply_coords(point_coords, self.original_size)
|
77 |
outputs = self.decoder.run(None, {
|
78 |
'image_embeddings': features,
|
79 |
'point_coords': point_coords.astype(np.float32),
|
80 |
'point_labels': point_labels.astype(np.float32)
|
81 |
})
|
82 |
scores, low_res_masks = outputs[0], outputs[1]
|
83 |
-
masks = self.postprocess_masks(low_res_masks)
|
84 |
masks = masks > self.mask_threshold
|
85 |
|
86 |
return masks, scores, low_res_masks
|
@@ -102,10 +98,10 @@ class SamPredictorONNX:
|
|
102 |
x = np.pad(x, ((0, 0), (0, 0), (0, padh), (0, padw)), mode='constant', constant_values=0)
|
103 |
return x
|
104 |
|
105 |
-
def postprocess_masks(self, mask: np.ndarray):
|
106 |
mask = mask.squeeze(0).transpose(1, 2, 0)
|
107 |
mask = cv2.resize(mask, (self.img_size, self.img_size), interpolation=cv2.INTER_LINEAR)
|
108 |
-
mask = mask[:
|
109 |
-
mask = cv2.resize(mask, (
|
110 |
mask = mask.transpose(2, 0, 1)[None, :, :, :]
|
111 |
return mask
|
|
|
53 |
input_image = self.transform.apply_image(image)
|
54 |
input_image = input_image.transpose(2, 0, 1)[None, :, :, :]
|
55 |
self.reset_image()
|
56 |
+
original_size = image.shape[:2]
|
57 |
+
input_size = tuple(input_image.shape[-2:])
|
58 |
input_image = self.preprocess(input_image).astype(np.float32)
|
59 |
outputs = self.encoder.run(None, {'image': input_image})
|
60 |
+
features = outputs[0]
|
|
|
61 |
|
62 |
+
return features, input_size, original_size
|
63 |
|
64 |
def predict(
|
65 |
self,
|
66 |
+
features: np.ndarray,
|
67 |
+
input_size: Tuple[int, int],
|
68 |
+
original_size: Tuple[int, int],
|
69 |
point_coords: Optional[np.ndarray] = None,
|
70 |
point_labels: Optional[np.ndarray] = None,
|
71 |
) -> Tuple[np.ndarray, np.ndarray, np.ndarray]:
|
72 |
+
point_coords = self.transform.apply_coords(point_coords, original_size)
|
|
|
|
|
|
|
|
|
|
|
73 |
outputs = self.decoder.run(None, {
|
74 |
'image_embeddings': features,
|
75 |
'point_coords': point_coords.astype(np.float32),
|
76 |
'point_labels': point_labels.astype(np.float32)
|
77 |
})
|
78 |
scores, low_res_masks = outputs[0], outputs[1]
|
79 |
+
masks = self.postprocess_masks(low_res_masks, input_size, original_size)
|
80 |
masks = masks > self.mask_threshold
|
81 |
|
82 |
return masks, scores, low_res_masks
|
|
|
98 |
x = np.pad(x, ((0, 0), (0, 0), (0, padh), (0, padw)), mode='constant', constant_values=0)
|
99 |
return x
|
100 |
|
101 |
+
def postprocess_masks(self, mask: np.ndarray, input_size: Tuple[int, int], original_size: Tuple[int, int]) -> np.ndarray:
|
102 |
mask = mask.squeeze(0).transpose(1, 2, 0)
|
103 |
mask = cv2.resize(mask, (self.img_size, self.img_size), interpolation=cv2.INTER_LINEAR)
|
104 |
+
mask = mask[:input_size[0], :input_size[1], :]
|
105 |
+
mask = cv2.resize(mask, (original_size[1], original_size[0]), interpolation=cv2.INTER_LINEAR)
|
106 |
mask = mask.transpose(2, 0, 1)[None, :, :, :]
|
107 |
return mask
|
segment_anything/predictor.py
CHANGED
@@ -59,13 +59,15 @@ class SamPredictor:
|
|
59 |
input_image_torch = torch.as_tensor(input_image, device=self.device)
|
60 |
input_image_torch = input_image_torch.permute(2, 0, 1).contiguous()[None, :, :, :]
|
61 |
|
62 |
-
|
|
|
|
|
|
|
63 |
|
64 |
@torch.no_grad()
|
65 |
def set_torch_image(
|
66 |
self,
|
67 |
transformed_image: torch.Tensor,
|
68 |
-
original_image_size: Tuple[int, ...],
|
69 |
) -> torch.Tensor:
|
70 |
"""
|
71 |
Calculates the image embeddings for the provided image, allowing
|
@@ -75,8 +77,6 @@ class SamPredictor:
|
|
75 |
Arguments:
|
76 |
transformed_image (torch.Tensor): The input image, with shape
|
77 |
1x3xHxW, which has been transformed with ResizeLongestSide.
|
78 |
-
original_image_size (tuple(int, int)): The size of the image
|
79 |
-
before transformation, in (H, W) format.
|
80 |
"""
|
81 |
assert (
|
82 |
len(transformed_image.shape) == 4
|
@@ -85,24 +85,23 @@ class SamPredictor:
|
|
85 |
), f"set_torch_image input must be BCHW with long side {self.model.image_encoder.img_size}."
|
86 |
self.reset_image()
|
87 |
|
88 |
-
self.original_size = original_image_size
|
89 |
-
self.input_size = tuple(transformed_image.shape[-2:])
|
90 |
input_image = self.model.preprocess(transformed_image)
|
91 |
-
|
92 |
-
self.is_image_set = True
|
93 |
|
94 |
-
return
|
95 |
|
96 |
def predict(
|
97 |
self,
|
98 |
-
features: torch.Tensor
|
|
|
|
|
99 |
point_coords: Optional[np.ndarray] = None,
|
100 |
point_labels: Optional[np.ndarray] = None,
|
101 |
box: Optional[np.ndarray] = None,
|
102 |
mask_input: Optional[np.ndarray] = None,
|
103 |
num_multimask_outputs: int = 3,
|
104 |
return_logits: bool = False,
|
105 |
-
use_stability_score: bool = False
|
106 |
) -> Tuple[np.ndarray, np.ndarray, np.ndarray]:
|
107 |
"""
|
108 |
Predict masks for the given input prompts, using the currently set image.
|
@@ -134,24 +133,18 @@ class SamPredictor:
|
|
134 |
of masks and H=W=256. These low resolution logits can be passed to
|
135 |
a subsequent iteration as mask input.
|
136 |
"""
|
137 |
-
if features is None and not self.is_image_set:
|
138 |
-
raise RuntimeError("An image must be set with .set_image(...) before mask prediction.")
|
139 |
-
|
140 |
-
if features is None:
|
141 |
-
features = self.features
|
142 |
-
|
143 |
# Transform input prompts
|
144 |
coords_torch, labels_torch, box_torch, mask_input_torch = None, None, None, None
|
145 |
if point_coords is not None:
|
146 |
assert (
|
147 |
point_labels is not None
|
148 |
), "point_labels must be supplied if point_coords is supplied."
|
149 |
-
point_coords = self.transform.apply_coords(point_coords,
|
150 |
coords_torch = torch.as_tensor(point_coords, dtype=torch.float, device=self.device)
|
151 |
labels_torch = torch.as_tensor(point_labels, dtype=torch.int, device=self.device)
|
152 |
coords_torch, labels_torch = coords_torch[None, :, :], labels_torch[None, :]
|
153 |
if box is not None:
|
154 |
-
box = self.transform.apply_boxes(box,
|
155 |
box_torch = torch.as_tensor(box, dtype=torch.float, device=self.device)
|
156 |
box_torch = box_torch[None, :]
|
157 |
if mask_input is not None:
|
@@ -160,6 +153,8 @@ class SamPredictor:
|
|
160 |
|
161 |
masks, iou_predictions, low_res_masks = self.predict_torch(
|
162 |
features,
|
|
|
|
|
163 |
coords_torch,
|
164 |
labels_torch,
|
165 |
box_torch,
|
@@ -178,6 +173,8 @@ class SamPredictor:
|
|
178 |
def predict_torch(
|
179 |
self,
|
180 |
features: torch.Tensor,
|
|
|
|
|
181 |
point_coords: Optional[torch.Tensor],
|
182 |
point_labels: Optional[torch.Tensor],
|
183 |
boxes: Optional[torch.Tensor] = None,
|
@@ -249,7 +246,7 @@ class SamPredictor:
|
|
249 |
)
|
250 |
|
251 |
# Upscale the masks to the original image resolution
|
252 |
-
masks = self.model.postprocess_masks(low_res_masks,
|
253 |
|
254 |
if not return_logits:
|
255 |
masks = masks > self.model.mask_threshold
|
|
|
59 |
input_image_torch = torch.as_tensor(input_image, device=self.device)
|
60 |
input_image_torch = input_image_torch.permute(2, 0, 1).contiguous()[None, :, :, :]
|
61 |
|
62 |
+
input_size = tuple(input_image_torch.shape[-2:])
|
63 |
+
original_size = image.shape[:2]
|
64 |
+
|
65 |
+
return self.set_torch_image(input_image_torch), input_size, original_size
|
66 |
|
67 |
@torch.no_grad()
|
68 |
def set_torch_image(
|
69 |
self,
|
70 |
transformed_image: torch.Tensor,
|
|
|
71 |
) -> torch.Tensor:
|
72 |
"""
|
73 |
Calculates the image embeddings for the provided image, allowing
|
|
|
77 |
Arguments:
|
78 |
transformed_image (torch.Tensor): The input image, with shape
|
79 |
1x3xHxW, which has been transformed with ResizeLongestSide.
|
|
|
|
|
80 |
"""
|
81 |
assert (
|
82 |
len(transformed_image.shape) == 4
|
|
|
85 |
), f"set_torch_image input must be BCHW with long side {self.model.image_encoder.img_size}."
|
86 |
self.reset_image()
|
87 |
|
|
|
|
|
88 |
input_image = self.model.preprocess(transformed_image)
|
89 |
+
features = self.model.image_encoder(input_image)
|
|
|
90 |
|
91 |
+
return features
|
92 |
|
93 |
def predict(
|
94 |
self,
|
95 |
+
features: torch.Tensor,
|
96 |
+
input_size: Tuple[int, int],
|
97 |
+
original_size: Tuple[int, int],
|
98 |
point_coords: Optional[np.ndarray] = None,
|
99 |
point_labels: Optional[np.ndarray] = None,
|
100 |
box: Optional[np.ndarray] = None,
|
101 |
mask_input: Optional[np.ndarray] = None,
|
102 |
num_multimask_outputs: int = 3,
|
103 |
return_logits: bool = False,
|
104 |
+
use_stability_score: bool = False,
|
105 |
) -> Tuple[np.ndarray, np.ndarray, np.ndarray]:
|
106 |
"""
|
107 |
Predict masks for the given input prompts, using the currently set image.
|
|
|
133 |
of masks and H=W=256. These low resolution logits can be passed to
|
134 |
a subsequent iteration as mask input.
|
135 |
"""
|
|
|
|
|
|
|
|
|
|
|
|
|
136 |
# Transform input prompts
|
137 |
coords_torch, labels_torch, box_torch, mask_input_torch = None, None, None, None
|
138 |
if point_coords is not None:
|
139 |
assert (
|
140 |
point_labels is not None
|
141 |
), "point_labels must be supplied if point_coords is supplied."
|
142 |
+
point_coords = self.transform.apply_coords(point_coords, original_size)
|
143 |
coords_torch = torch.as_tensor(point_coords, dtype=torch.float, device=self.device)
|
144 |
labels_torch = torch.as_tensor(point_labels, dtype=torch.int, device=self.device)
|
145 |
coords_torch, labels_torch = coords_torch[None, :, :], labels_torch[None, :]
|
146 |
if box is not None:
|
147 |
+
box = self.transform.apply_boxes(box, original_size)
|
148 |
box_torch = torch.as_tensor(box, dtype=torch.float, device=self.device)
|
149 |
box_torch = box_torch[None, :]
|
150 |
if mask_input is not None:
|
|
|
153 |
|
154 |
masks, iou_predictions, low_res_masks = self.predict_torch(
|
155 |
features,
|
156 |
+
input_size,
|
157 |
+
original_size,
|
158 |
coords_torch,
|
159 |
labels_torch,
|
160 |
box_torch,
|
|
|
173 |
def predict_torch(
|
174 |
self,
|
175 |
features: torch.Tensor,
|
176 |
+
input_size: Tuple[int, int],
|
177 |
+
original_size: Tuple[int, int],
|
178 |
point_coords: Optional[torch.Tensor],
|
179 |
point_labels: Optional[torch.Tensor],
|
180 |
boxes: Optional[torch.Tensor] = None,
|
|
|
246 |
)
|
247 |
|
248 |
# Upscale the masks to the original image resolution
|
249 |
+
masks = self.model.postprocess_masks(low_res_masks, input_size, original_size)
|
250 |
|
251 |
if not return_logits:
|
252 |
masks = masks > self.model.mask_threshold
|