Dramb commited on
Commit
3f894e0
·
verified ·
1 Parent(s): 9c91372

Update inference.py

Browse files
Files changed (1) hide show
  1. inference.py +13 -61
inference.py CHANGED
@@ -1,66 +1,18 @@
1
- import base64
2
- import io
3
- import torch
4
- import numpy as np
5
  from PIL import Image
6
- from skimage import transform
7
- from sam2.build_sam import build_sam2
8
- from sam2.sam2_image_predictor import SAM2ImagePredictor
9
-
10
- # Инициализация модели один раз
11
- device = "cuda" if torch.cuda.is_available() else "cpu"
12
- torch.autocast(device_type=device, dtype=torch.bfloat16).__enter__()
13
- if torch.cuda.is_available() and torch.cuda.get_device_properties(0).major >= 8:
14
- torch.backends.cuda.matmul.allow_tf32 = True
15
- torch.backends.cudnn.allow_tf32 = True
16
-
17
- # Загрузи путь к своей модели
18
- medsam2_model = build_sam2('sam2_hiera_t', 'MedSAM2_pretrain_10ep_b1_AMD-SD_sam2_hiera_t.pth', device=device)
19
- predictor = SAM2ImagePredictor(medsam2_model)
20
-
21
- # --- Вспомогательные функции ---
22
- def decode_image(base64_str):
23
- img_bytes = base64.b64decode(base64_str)
24
- img = Image.open(io.BytesIO(img_bytes)).convert("RGB")
25
- return np.array(img)
26
-
27
- def encode_mask_to_base64(mask_np):
28
- mask_img = Image.fromarray((mask_np * 255).astype(np.uint8))
29
- buffer = io.BytesIO()
30
- mask_img.save(buffer, format="PNG")
31
- return base64.b64encode(buffer.getvalue()).decode("utf-8")
32
-
33
- # --- Главная функция вызова инференса ---
34
- def predict(input_dict):
35
- try:
36
- image_b64 = input_dict["image"]
37
- box = input_dict["box"] # [x1, y1, x2, y2]
38
-
39
- image = decode_image(image_b64)
40
-
41
- # Resize image to 1024x1024 for MedSAM2
42
- img_3c = np.repeat(image[:, :, None], 3, axis=-1) if len(image.shape) == 2 else image
43
- img_1024 = transform.resize(img_3c, (1024, 1024), order=3, preserve_range=True, anti_aliasing=True).astype(np.uint8)
44
-
45
- # Normalize box
46
- box_np = np.array(box)
47
- box_1024 = box_np / np.array([img_3c.shape[1], img_3c.shape[0], img_3c.shape[1], img_3c.shape[0]]) * 1024
48
 
49
- with torch.inference_mode(), torch.autocast(device, dtype=torch.bfloat16):
50
- predictor.set_image(img_1024)
51
- masks, _, _ = predictor.predict(
52
- point_coords=None,
53
- point_labels=None,
54
- box=np.array([box_1024]),
55
- multimask_output=False,
56
- )
57
 
58
- mask = masks[0] # (H, W) np.uint8
 
59
 
60
- return {
61
- "mask": encode_mask_to_base64(mask.astype(np.uint8)),
62
- "shape": mask.shape,
63
- }
64
 
65
- except Exception as e:
66
- return {"error": str(e)}
 
 
 
 
 
1
  from PIL import Image
2
+ import numpy as np
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3
 
4
+ class PreTrainedPipeline:
5
+ def __init__(self, path=""):
6
+ # здесь можно подгрузить модель
7
+ print("Model initialized")
 
 
 
 
8
 
9
+ def __call__(self, inputs):
10
+ image = inputs.get("image")
11
 
12
+ if isinstance(image, str):
13
+ from base64 import b64decode
14
+ from io import BytesIO
15
+ image = Image.open(BytesIO(b64decode(image)))
16
 
17
+ # Заглушка: возвращаем просто shape
18
+ return {"msg": "Got image", "size": image.size}