Create app.py
Browse files
app.py
ADDED
@@ -0,0 +1,94 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import matplotlib.pyplot as plt
|
2 |
+
import numpy as np
|
3 |
+
from PIL import Image
|
4 |
+
from skimage.transform import resize
|
5 |
+
import tensorflow as tf
|
6 |
+
from tensorflow.keras.models import load_model
|
7 |
+
|
8 |
+
import gradio as gr
|
9 |
+
import os
|
10 |
+
|
11 |
+
REPO_ID = "amosfang/segmentation_u_net"
|
12 |
+
|
13 |
+
def pil_image_as_numpy_array(pilimg):
|
14 |
+
img_array = tf.keras.utils.img_to_array(pilimg)
|
15 |
+
return img_array
|
16 |
+
|
17 |
+
def resize_image(image, input_shape=(224, 224, 3)):
|
18 |
+
# Convert to NumPy array and normalize
|
19 |
+
image_array = pil_image_as_numpy_array(image)
|
20 |
+
image = image_array.astype(np.float32) / 255.
|
21 |
+
|
22 |
+
# Resize the image to 224x224
|
23 |
+
image_resized = resize(image, input_shape, anti_aliasing=True)
|
24 |
+
|
25 |
+
return image_resized
|
26 |
+
|
27 |
+
def load_model():
|
28 |
+
model_dir = snapshot_download(REPO_ID)
|
29 |
+
# saved_model_dir = os.path.join(download_dir, "saved_model")
|
30 |
+
unet_model = load_model(model_dir)
|
31 |
+
return unet_model
|
32 |
+
|
33 |
+
def ensemble_predict(X_array):
|
34 |
+
#
|
35 |
+
# Call the predict methods of the unet_model and the vgg16_unet_model
|
36 |
+
# to retrieve their predictions.
|
37 |
+
#
|
38 |
+
# Sum the two predictions together and return their results.
|
39 |
+
# You can also consider multiplying a different weight on
|
40 |
+
# one or both of the models to improve performance
|
41 |
+
|
42 |
+
X_array = np.expand_dims(X_array, axis=0)
|
43 |
+
|
44 |
+
unet_model = load_model('REPO_ID/train_2024-02-14 11-20-17/base_u_net.0098-acc-0.75-val_acc-0.74-loss-0.79.h5')
|
45 |
+
vgg16_model = load_model('REPO_ID/vgg16_u_net.0092-acc-0.74-val_acc-0.74-loss-0.82.h5')
|
46 |
+
resnet50_model = load_model('REPO_ID/resnet50_u_net.0095-acc-0.79-val_acc-0.76-loss-0.72.h5')
|
47 |
+
|
48 |
+
pred_y_unet = unet_model.predict(X_array)
|
49 |
+
pred_y_vgg16 = vgg16_model.predict(X_array)
|
50 |
+
pred_y_resnet50 = resnet50_model.predict(X_array)
|
51 |
+
|
52 |
+
return (pred_y_unet + pred_y_vgg16 + pred_y_resnet50) / 3
|
53 |
+
|
54 |
+
def get_predictions(y_prediction_encoded):
|
55 |
+
|
56 |
+
# Convert predictions to categorical indices
|
57 |
+
predicted_label_indices = np.argmax(y_prediction_encoded, axis=-1) + 1
|
58 |
+
|
59 |
+
return predicted_label_indices
|
60 |
+
|
61 |
+
def predict(image):
|
62 |
+
sample_image_resized = resize_image(image, input_shape)
|
63 |
+
y_pred = ensemble_predict(sample_image_resized)
|
64 |
+
y_pred = get_predictions(y_pred).squeeze()
|
65 |
+
|
66 |
+
# Create a figure without saving it to a file
|
67 |
+
fig, ax = plt.subplots()
|
68 |
+
cax = ax.imshow(y_pred, cmap='viridis', vmin=1, vmax=7)
|
69 |
+
|
70 |
+
# Convert the figure to a PIL Image
|
71 |
+
image_buffer = io.BytesIO()
|
72 |
+
plt.savefig(image_buffer, format='png')
|
73 |
+
image_buffer.seek(0)
|
74 |
+
image_pil = Image.open(image_buffer)
|
75 |
+
|
76 |
+
# Close the figure to release resources
|
77 |
+
plt.close(fig)
|
78 |
+
|
79 |
+
return image_pil
|
80 |
+
|
81 |
+
# Specify paths to example images
|
82 |
+
sample_images = [['989953_sat.jpg'], ['999380_sat.jpg'], ['988205_sat.jpg']]
|
83 |
+
|
84 |
+
# Launch Gradio Interface
|
85 |
+
gr.Interface(
|
86 |
+
predict,
|
87 |
+
title='Land Cover Segmentation',
|
88 |
+
inputs=[gr.Image()],
|
89 |
+
outputs=[gr.Image()],
|
90 |
+
examples=sample_images
|
91 |
+
).launch(debug=True, share=True)
|
92 |
+
|
93 |
+
# Launch the interface
|
94 |
+
iface.launch(share=True)
|