Spaces:
Sleeping
Sleeping
Update app.py
Browse files
app.py
CHANGED
@@ -8,37 +8,37 @@ from transformers import AutoProcessor, CLIPSegForImageSegmentation
|
|
8 |
processor = AutoProcessor.from_pretrained("CIDAS/clipseg-rd64-refined")
|
9 |
model = CLIPSegForImageSegmentation.from_pretrained("CIDAS/clipseg-rd64-refined")
|
10 |
|
|
|
|
|
|
|
|
|
11 |
def segment_everything(image):
|
12 |
-
# Check if image is a list and extract the actual image data
|
13 |
if isinstance(image, list):
|
14 |
image = image[0]
|
15 |
|
16 |
-
# Convert numpy array to PIL Image
|
17 |
if isinstance(image, np.ndarray):
|
18 |
image = Image.fromarray(image)
|
19 |
|
20 |
-
inputs = processor(text=["object"], images=[image], padding="max_length", return_tensors="pt")
|
21 |
with torch.no_grad():
|
22 |
outputs = model(**inputs)
|
23 |
-
preds = outputs.logits.squeeze().sigmoid()
|
24 |
segmentation = (preds.numpy() * 255).astype(np.uint8)
|
25 |
return Image.fromarray(segmentation)
|
26 |
|
27 |
def segment_box(image, x1, y1, x2, y2):
|
28 |
-
# Check if image is a list and extract the actual image data
|
29 |
if isinstance(image, list):
|
30 |
image = image[0]
|
31 |
|
32 |
-
# Convert PIL Image to numpy array if necessary
|
33 |
if isinstance(image, Image.Image):
|
34 |
image = np.array(image)
|
35 |
|
36 |
x1, y1, x2, y2 = int(x1), int(y1), int(x2), int(y2)
|
37 |
cropped_image = image[y1:y2, x1:x2]
|
38 |
-
inputs = processor(text=["object"], images=[Image.fromarray(cropped_image)], padding="max_length", return_tensors="pt")
|
39 |
with torch.no_grad():
|
40 |
outputs = model(**inputs)
|
41 |
-
preds = outputs.logits.squeeze().sigmoid()
|
42 |
segmentation = np.zeros((image.shape[0], image.shape[1]), dtype=np.uint8)
|
43 |
segmentation[y1:y2, x1:x2] = (preds.numpy() * 255).astype(np.uint8)
|
44 |
return Image.fromarray(segmentation)
|
@@ -47,24 +47,19 @@ def update_image(image, segmentation):
|
|
47 |
if segmentation is None:
|
48 |
return image
|
49 |
|
50 |
-
# Check if image is a list and extract the actual image data
|
51 |
if isinstance(image, list):
|
52 |
image = image[0]
|
53 |
|
54 |
-
# Ensure image is in the correct format (PIL Image)
|
55 |
if isinstance(image, np.ndarray):
|
56 |
image_pil = Image.fromarray(image)
|
57 |
else:
|
58 |
image_pil = image
|
59 |
|
60 |
-
# Convert segmentation to RGBA
|
61 |
seg_pil = Image.fromarray(segmentation).convert('RGBA')
|
62 |
|
63 |
-
# Resize segmentation to match input image if necessary
|
64 |
if image_pil.size != seg_pil.size:
|
65 |
seg_pil = seg_pil.resize(image_pil.size, Image.NEAREST)
|
66 |
|
67 |
-
# Blend images
|
68 |
blended = Image.blend(image_pil.convert('RGBA'), seg_pil, 0.5)
|
69 |
|
70 |
return np.array(blended)
|
@@ -101,4 +96,4 @@ with gr.Blocks() as demo:
|
|
101 |
outputs=[output_image]
|
102 |
)
|
103 |
|
104 |
-
demo.launch()
|
|
|
8 |
processor = AutoProcessor.from_pretrained("CIDAS/clipseg-rd64-refined")
|
9 |
model = CLIPSegForImageSegmentation.from_pretrained("CIDAS/clipseg-rd64-refined")
|
10 |
|
11 |
+
# Ensure that the model uses GPU if available
|
12 |
+
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
|
13 |
+
model.to(device)
|
14 |
+
|
15 |
def segment_everything(image):
|
|
|
16 |
if isinstance(image, list):
|
17 |
image = image[0]
|
18 |
|
|
|
19 |
if isinstance(image, np.ndarray):
|
20 |
image = Image.fromarray(image)
|
21 |
|
22 |
+
inputs = processor(text=["object"], images=[image], padding="max_length", return_tensors="pt").to(device)
|
23 |
with torch.no_grad():
|
24 |
outputs = model(**inputs)
|
25 |
+
preds = outputs.logits.squeeze().sigmoid().cpu()
|
26 |
segmentation = (preds.numpy() * 255).astype(np.uint8)
|
27 |
return Image.fromarray(segmentation)
|
28 |
|
29 |
def segment_box(image, x1, y1, x2, y2):
|
|
|
30 |
if isinstance(image, list):
|
31 |
image = image[0]
|
32 |
|
|
|
33 |
if isinstance(image, Image.Image):
|
34 |
image = np.array(image)
|
35 |
|
36 |
x1, y1, x2, y2 = int(x1), int(y1), int(x2), int(y2)
|
37 |
cropped_image = image[y1:y2, x1:x2]
|
38 |
+
inputs = processor(text=["object"], images=[Image.fromarray(cropped_image)], padding="max_length", return_tensors="pt").to(device)
|
39 |
with torch.no_grad():
|
40 |
outputs = model(**inputs)
|
41 |
+
preds = outputs.logits.squeeze().sigmoid().cpu()
|
42 |
segmentation = np.zeros((image.shape[0], image.shape[1]), dtype=np.uint8)
|
43 |
segmentation[y1:y2, x1:x2] = (preds.numpy() * 255).astype(np.uint8)
|
44 |
return Image.fromarray(segmentation)
|
|
|
47 |
if segmentation is None:
|
48 |
return image
|
49 |
|
|
|
50 |
if isinstance(image, list):
|
51 |
image = image[0]
|
52 |
|
|
|
53 |
if isinstance(image, np.ndarray):
|
54 |
image_pil = Image.fromarray(image)
|
55 |
else:
|
56 |
image_pil = image
|
57 |
|
|
|
58 |
seg_pil = Image.fromarray(segmentation).convert('RGBA')
|
59 |
|
|
|
60 |
if image_pil.size != seg_pil.size:
|
61 |
seg_pil = seg_pil.resize(image_pil.size, Image.NEAREST)
|
62 |
|
|
|
63 |
blended = Image.blend(image_pil.convert('RGBA'), seg_pil, 0.5)
|
64 |
|
65 |
return np.array(blended)
|
|
|
96 |
outputs=[output_image]
|
97 |
)
|
98 |
|
99 |
+
demo.launch()
|