Spaces:
Sleeping
Sleeping
Update app.py
Browse files
app.py
CHANGED
@@ -9,7 +9,78 @@ import io
|
|
9 |
import zipfile
|
10 |
|
11 |
# Assuming you have these functions defined elsewhere
|
12 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
13 |
|
14 |
def extract_tiles(map_file, model, tile_size=512, overlap=0, batch_size=4):
|
15 |
tiles = []
|
|
|
9 |
import zipfile
|
10 |
|
11 |
# Assuming you have these functions defined elsewhere
|
12 |
+
import torch
|
13 |
+
import numpy as np
|
14 |
+
from PIL import Image
|
15 |
+
import albumentations as albu
|
16 |
+
import segmentation_models_pytorch as smp
|
17 |
+
from albumentations.pytorch.transforms import ToTensorV2
|
18 |
+
|
19 |
+
|
20 |
+
|
21 |
+
DEVICE = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
22 |
+
ENCODER = 'se_resnext50_32x4d'
|
23 |
+
ENCODER_WEIGHTS = 'imagenet'
|
24 |
+
|
25 |
+
# Load and prepare the model
|
26 |
+
best_model = torch.load('deeplabv3+ v15.pth', map_location=DEVICE)
|
27 |
+
best_model.eval().float()
|
28 |
+
|
29 |
+
def to_tensor(x, **kwargs):
|
30 |
+
return x.astype('float32')#.transpose(2, 0, 1)
|
31 |
+
|
32 |
+
# Preprocessing
|
33 |
+
preprocessing_fn = smp.encoders.get_preprocessing_fn(ENCODER, ENCODER_WEIGHTS)
|
34 |
+
|
35 |
+
def get_preprocessing():
|
36 |
+
_transform = [
|
37 |
+
albu.Resize(512, 512),
|
38 |
+
albu.Lambda(image=preprocessing_fn),
|
39 |
+
albu.Lambda(image=to_tensor, mask=to_tensor),
|
40 |
+
ToTensorV2(),
|
41 |
+
#albu.Normalize(mean=MEAN,std=STD)
|
42 |
+
|
43 |
+
]
|
44 |
+
return albu.Compose(_transform)
|
45 |
+
|
46 |
+
|
47 |
+
preprocess = get_preprocessing()
|
48 |
+
|
49 |
+
@torch.no_grad()
|
50 |
+
def process_and_predict(image, model):
|
51 |
+
# Convert PIL Image to numpy array if necessary
|
52 |
+
if isinstance(image, Image.Image):
|
53 |
+
image = np.array(image)
|
54 |
+
|
55 |
+
# Ensure image is 3-channel
|
56 |
+
if image.ndim == 2:
|
57 |
+
image = np.stack([image] * 3, axis=-1)
|
58 |
+
elif image.shape[2] == 4:
|
59 |
+
image = image[:, :, :3]
|
60 |
+
|
61 |
+
# Apply preprocessing
|
62 |
+
preprocessed = preprocess(image=image)['image']
|
63 |
+
#preprocessed=torch.tensor(preprocessed)
|
64 |
+
# Add batch dimension and move to device
|
65 |
+
input_tensor = preprocessed.unsqueeze(0).to(DEVICE)
|
66 |
+
|
67 |
+
print(input_tensor.shape)
|
68 |
+
# Predict
|
69 |
+
mask = model(input_tensor)
|
70 |
+
mask = torch.sigmoid(mask)
|
71 |
+
mask = (mask > 0.6).float()
|
72 |
+
|
73 |
+
# Convert to PIL Image
|
74 |
+
mask_image = Image.fromarray((mask.squeeze().cpu().numpy() * 255).astype(np.uint8))
|
75 |
+
|
76 |
+
return mask_image
|
77 |
+
|
78 |
+
#example
|
79 |
+
def main(image_path):
|
80 |
+
image = Image.open(image_path)
|
81 |
+
mask = process_and_predict(image, best_model)
|
82 |
+
return mask
|
83 |
+
|
84 |
|
85 |
def extract_tiles(map_file, model, tile_size=512, overlap=0, batch_size=4):
|
86 |
tiles = []
|