Spaces:
Sleeping
Sleeping
Hamam
commited on
Update app.py
Browse files
app.py
CHANGED
@@ -1,15 +1,13 @@
|
|
1 |
-
|
2 |
-
|
3 |
import numpy as np
|
|
|
4 |
import torch
|
5 |
import torchvision.transforms as T
|
6 |
-
import
|
7 |
-
from model.u2net import U2NET
|
8 |
-
from io import BytesIO
|
9 |
|
10 |
-
#
|
11 |
-
|
12 |
-
DEFAULT_IMAGE_PATH = "image (4).png" # Path to your default image
|
13 |
|
14 |
# Initialize the U2NET model
|
15 |
u2net = U2NET(in_ch=3, out_ch=1)
|
@@ -20,7 +18,7 @@ def load_model(model, model_path, device):
|
|
20 |
return model
|
21 |
|
22 |
# Load the model onto the specified device
|
23 |
-
u2net = load_model(model=u2net, model_path="u2net.pth", device="cpu")
|
24 |
|
25 |
# Mean and std for normalization
|
26 |
mean = torch.tensor([0.485, 0.456, 0.406])
|
@@ -44,14 +42,6 @@ def prepare_single_image(image, resize, transforms, device):
|
|
44 |
image_batch = image_trans.unsqueeze(0).to(device) # Add batch dimension
|
45 |
return image_batch
|
46 |
|
47 |
-
def denorm_image(image_tensor):
|
48 |
-
"""Denormalize and convert tensor to numpy image."""
|
49 |
-
image_tensor = image_tensor.cpu().clone()
|
50 |
-
image_tensor = image_tensor * std[:, None, None] + mean[:, None, None]
|
51 |
-
image_tensor = torch.clamp(image_tensor * 255., min=0., max=255.)
|
52 |
-
image_tensor = image_tensor.permute(1, 2, 0).numpy().astype(np.uint8)
|
53 |
-
return image_tensor
|
54 |
-
|
55 |
def prepare_prediction(model, image_batch):
|
56 |
model.eval()
|
57 |
with torch.no_grad():
|
@@ -67,93 +57,55 @@ def normPRED(predicted_map):
|
|
67 |
|
68 |
def apply_mask(image, mask):
|
69 |
"""Apply the mask to the original image and return the result with transparent background."""
|
70 |
-
# Remove the extra dimension if present
|
71 |
mask = np.squeeze(mask)
|
72 |
-
|
73 |
-
# Normalize and convert the mask to uint8
|
74 |
mask = normPRED(mask)
|
75 |
mask = (mask * 255).astype(np.uint8)
|
76 |
-
|
77 |
-
# Convert the mask to a PIL image
|
78 |
mask_image = Image.fromarray(mask, mode='L') # 'L' mode for grayscale
|
79 |
-
|
80 |
-
# Open the original image and resize it
|
81 |
original_image = image.convert("RGB")
|
82 |
original_image = original_image.resize(resize_shape, resample=Image.BILINEAR)
|
83 |
-
|
84 |
-
# Convert original image to RGBA
|
85 |
original_image_rgba = original_image.convert("RGBA")
|
86 |
-
|
87 |
-
# Create a new image with transparency
|
88 |
transparent_background = Image.new("RGBA", original_image_rgba.size, (0, 0, 0, 0))
|
89 |
-
|
90 |
-
# Apply the mask to create an image with alpha channel
|
91 |
masked_image = Image.composite(original_image_rgba, transparent_background, mask_image)
|
92 |
-
|
93 |
return masked_image
|
94 |
|
95 |
-
|
96 |
-
|
97 |
-
|
98 |
-
|
99 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
100 |
|
|
|
|
|
|
|
|
|
101 |
image_batch = prepare_single_image(image, resize_shape, transforms, "cpu")
|
102 |
prediction_u2net = prepare_prediction(u2net, image_batch)
|
103 |
masked_image = apply_mask(image, prediction_u2net)
|
104 |
-
return masked_image
|
105 |
|
106 |
-
|
107 |
-
|
108 |
-
|
109 |
-
|
110 |
-
|
111 |
-
|
112 |
-
|
113 |
-
st.
|
114 |
-
|
115 |
-
|
116 |
-
|
117 |
-
|
118 |
-
|
119 |
-
|
120 |
-
|
121 |
-
|
122 |
-
|
123 |
-
|
124 |
-
file_name="segmented_image.png",
|
125 |
-
mime="image/png"
|
126 |
-
)
|
127 |
-
|
128 |
-
# Define the pages
|
129 |
-
def page_one():
|
130 |
-
"""Page for image segmentation."""
|
131 |
-
st.title("Image Segmentation with U2NET")
|
132 |
-
st.write("Upload an image to segment it using the U2NET model. The background of the segmented output will be transparent.")
|
133 |
-
|
134 |
-
uploaded_file = st.file_uploader("Choose an image...", type=["jpg", "jpeg", "png"])
|
135 |
-
|
136 |
-
# Determine image processing
|
137 |
-
if uploaded_file is not None:
|
138 |
-
if uploaded_file.size > MAX_FILE_SIZE:
|
139 |
-
st.error("The uploaded file is too large. Please upload an image smaller than 5MB.")
|
140 |
-
else:
|
141 |
-
fix_image(upload=uploaded_file)
|
142 |
-
else:
|
143 |
-
fix_image() # Use default image if none uploaded
|
144 |
-
|
145 |
-
def page_two():
|
146 |
-
"""Page for other code."""
|
147 |
-
st.title("Other Feature")
|
148 |
-
st.write("This page is for the second feature you want to implement.")
|
149 |
-
# Add other code or features here
|
150 |
-
|
151 |
-
# Sidebar navigation
|
152 |
-
st.sidebar.title("Navigation")
|
153 |
-
page = st.sidebar.radio("Go to", ("Image Segmentation", "Other Feature"))
|
154 |
-
|
155 |
-
# Page selection logic
|
156 |
-
if page == "Image Segmentation":
|
157 |
-
page_one()
|
158 |
-
elif page == "Other Feature":
|
159 |
-
page_two()
|
|
|
1 |
+
%%file app.py
|
2 |
+
import streamlit as st
|
3 |
import numpy as np
|
4 |
+
from PIL import Image
|
5 |
import torch
|
6 |
import torchvision.transforms as T
|
7 |
+
import io
|
|
|
|
|
8 |
|
9 |
+
# Assuming you have the U2NET model defined somewhere
|
10 |
+
from model.u2net import U2NET # Replace with your actual import path
|
|
|
11 |
|
12 |
# Initialize the U2NET model
|
13 |
u2net = U2NET(in_ch=3, out_ch=1)
|
|
|
18 |
return model
|
19 |
|
20 |
# Load the model onto the specified device
|
21 |
+
u2net = load_model(model=u2net, model_path="/content/u2net.pth", device="cpu")
|
22 |
|
23 |
# Mean and std for normalization
|
24 |
mean = torch.tensor([0.485, 0.456, 0.406])
|
|
|
42 |
image_batch = image_trans.unsqueeze(0).to(device) # Add batch dimension
|
43 |
return image_batch
|
44 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
45 |
def prepare_prediction(model, image_batch):
|
46 |
model.eval()
|
47 |
with torch.no_grad():
|
|
|
57 |
|
58 |
def apply_mask(image, mask):
|
59 |
"""Apply the mask to the original image and return the result with transparent background."""
|
|
|
60 |
mask = np.squeeze(mask)
|
|
|
|
|
61 |
mask = normPRED(mask)
|
62 |
mask = (mask * 255).astype(np.uint8)
|
|
|
|
|
63 |
mask_image = Image.fromarray(mask, mode='L') # 'L' mode for grayscale
|
|
|
|
|
64 |
original_image = image.convert("RGB")
|
65 |
original_image = original_image.resize(resize_shape, resample=Image.BILINEAR)
|
|
|
|
|
66 |
original_image_rgba = original_image.convert("RGBA")
|
|
|
|
|
67 |
transparent_background = Image.new("RGBA", original_image_rgba.size, (0, 0, 0, 0))
|
|
|
|
|
68 |
masked_image = Image.composite(original_image_rgba, transparent_background, mask_image)
|
|
|
69 |
return masked_image
|
70 |
|
71 |
+
# Streamlit app setup
|
72 |
+
st.title("Image Segmentation with U2NET")
|
73 |
+
|
74 |
+
# Sidebar for file upload and controls
|
75 |
+
st.sidebar.title("Controls :gear:")
|
76 |
+
uploaded_file = st.sidebar.file_uploader("Choose an image...", type=["png", "jpg", "jpeg"])
|
77 |
+
|
78 |
+
# Function to handle image and segmentation display
|
79 |
+
def fix_image(upload=None):
|
80 |
+
if upload is None:
|
81 |
+
st.write("Please upload an image.")
|
82 |
+
return
|
83 |
+
|
84 |
+
image = Image.open(upload)
|
85 |
|
86 |
+
# Display the original image
|
87 |
+
st.image(image, caption="Uploaded Image", use_column_width=True)
|
88 |
+
|
89 |
+
# Prepare image for segmentation
|
90 |
image_batch = prepare_single_image(image, resize_shape, transforms, "cpu")
|
91 |
prediction_u2net = prepare_prediction(u2net, image_batch)
|
92 |
masked_image = apply_mask(image, prediction_u2net)
|
|
|
93 |
|
94 |
+
# Display segmented image
|
95 |
+
st.image(masked_image, caption='Segmented Image', use_column_width=True)
|
96 |
+
|
97 |
+
# Provide download option for segmented image
|
98 |
+
buf = io.BytesIO()
|
99 |
+
masked_image.save(buf, format='PNG')
|
100 |
+
byte_im = buf.getvalue()
|
101 |
+
st.sidebar.markdown('### Download Segmented Image')
|
102 |
+
st.sidebar.download_button(
|
103 |
+
label="Download Segmented Image",
|
104 |
+
data=byte_im,
|
105 |
+
file_name="segmented_image.png",
|
106 |
+
mime="image/png"
|
107 |
+
)
|
108 |
+
|
109 |
+
# Handle image processing based on user input
|
110 |
+
if uploaded_file is not None:
|
111 |
+
fix_image(upload=uploaded_file)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|