Spaces:
Running
on
T4
Running
on
T4
Maitreyapatel
commited on
Commit
•
964ba1e
1
Parent(s):
594c562
changes
Browse files
app.py
CHANGED
@@ -18,7 +18,7 @@ import cv2
|
|
18 |
import numpy as np
|
19 |
import torch
|
20 |
|
21 |
-
|
22 |
PRETRAINED_MODEL_NAME_OR_PATH = "./checkpoints/"
|
23 |
|
24 |
|
@@ -37,8 +37,13 @@ def get_image_grid(images: List[Image.Image]) -> Image:
|
|
37 |
|
38 |
class AttributionModel:
|
39 |
def __init__(self):
|
|
|
|
|
|
|
|
|
40 |
self.pipe = StableDiffusionPipeline.from_pretrained('stabilityai/stable-diffusion-2')#, safety_checker=None, torch_dtype=torch.float16)
|
41 |
-
|
|
|
42 |
self.resize_transform = transforms.Resize(512, interpolation=transforms.InterpolationMode.BILINEAR)
|
43 |
self.vae = AutoencoderKL.from_pretrained(
|
44 |
'stabilityai/stable-diffusion-2', subfolder="vae"
|
@@ -55,9 +60,10 @@ class AttributionModel:
|
|
55 |
self.mapping_network.load_state_dict(torch.load(os.path.join(PRETRAINED_MODEL_NAME_OR_PATH, 'mapping_network.pth')))
|
56 |
self.decoding_network.load_state_dict(torch.load(os.path.join(PRETRAINED_MODEL_NAME_OR_PATH, 'decoding_network.pth')))
|
57 |
|
58 |
-
|
59 |
-
|
60 |
-
|
|
|
61 |
|
62 |
self.test_norm = transforms.Compose(
|
63 |
[
|
@@ -65,21 +71,26 @@ class AttributionModel:
|
|
65 |
]
|
66 |
)
|
67 |
|
68 |
-
def infer(self, prompt, negative, guidance_scale):
|
69 |
-
images = []
|
70 |
with torch.no_grad():
|
71 |
-
out_latents = self.pipe([prompt], output_type="latent", num_inference_steps=
|
72 |
-
|
73 |
-
|
74 |
-
|
75 |
-
|
76 |
-
|
|
|
|
|
|
|
|
|
77 |
|
78 |
def inference_without_attribution(self, latents):
|
79 |
latents = 1 / 0.18215 * latents
|
80 |
with torch.no_grad():
|
81 |
image = self.pipe.vae.decode(latents).sample
|
82 |
image = image.clamp(-1,1)
|
|
|
|
|
83 |
return image
|
84 |
|
85 |
def get_phis(self, phi_dimension, batch_size ,eps = 1e-8):
|
@@ -111,8 +122,29 @@ class AttributionModel:
|
|
111 |
|
112 |
|
113 |
attribution_model = AttributionModel()
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
114 |
|
115 |
with gr.Blocks() as demo:
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
116 |
with gr.Row(elem_id="prompt-container").style(mobile_collapse=False, equal_height=True):
|
117 |
with gr.Column():
|
118 |
text = gr.Textbox(
|
@@ -137,20 +169,52 @@ with gr.Blocks() as demo:
|
|
137 |
rounded=(True, False, False, True),
|
138 |
container=False,
|
139 |
)
|
140 |
-
btn = gr.Button("Generate image").style(full_width=False)
|
141 |
-
|
142 |
-
with gr.Row():
|
143 |
-
img_output_simple = gr.Image()
|
144 |
-
img_output_attribute = gr.Image()
|
145 |
-
img_output_diff = gr.Image()
|
146 |
-
|
147 |
|
148 |
with gr.Row():
|
|
|
149 |
guidance_scale = gr.Slider(
|
150 |
label="Guidance Scale", minimum=0, maximum=10, value=9, step=0.1
|
151 |
)
|
152 |
-
btn.click(attribution_model.infer, inputs=[text, negative, guidance_scale], outputs=[img_output_simple, img_output_attribute, img_output_diff], postprocess=False)
|
153 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
154 |
|
155 |
-
if __name__=="__main__":
|
156 |
-
demo.
|
|
|
18 |
import numpy as np
|
19 |
import torch
|
20 |
|
21 |
+
|
22 |
PRETRAINED_MODEL_NAME_OR_PATH = "./checkpoints/"
|
23 |
|
24 |
|
|
|
37 |
|
38 |
class AttributionModel:
|
39 |
def __init__(self):
|
40 |
+
is_cuda = False
|
41 |
+
if torch.cuda.is_available():
|
42 |
+
is_cuda = True
|
43 |
+
|
44 |
self.pipe = StableDiffusionPipeline.from_pretrained('stabilityai/stable-diffusion-2')#, safety_checker=None, torch_dtype=torch.float16)
|
45 |
+
if is_cuda:
|
46 |
+
self.pipe = self.pipe.to("cuda")
|
47 |
self.resize_transform = transforms.Resize(512, interpolation=transforms.InterpolationMode.BILINEAR)
|
48 |
self.vae = AutoencoderKL.from_pretrained(
|
49 |
'stabilityai/stable-diffusion-2', subfolder="vae"
|
|
|
60 |
self.mapping_network.load_state_dict(torch.load(os.path.join(PRETRAINED_MODEL_NAME_OR_PATH, 'mapping_network.pth')))
|
61 |
self.decoding_network.load_state_dict(torch.load(os.path.join(PRETRAINED_MODEL_NAME_OR_PATH, 'decoding_network.pth')))
|
62 |
|
63 |
+
if is_cuda:
|
64 |
+
self.vae = self.vae.to("cuda")
|
65 |
+
self.mapping_network = self.mapping_network.to("cuda")
|
66 |
+
self.decoding_network = self.decoding_network.to("cuda")
|
67 |
|
68 |
self.test_norm = transforms.Compose(
|
69 |
[
|
|
|
71 |
]
|
72 |
)
|
73 |
|
74 |
+
def infer(self, prompt, negative, steps, guidance_scale):
|
|
|
75 |
with torch.no_grad():
|
76 |
+
out_latents = self.pipe([prompt], negative_prompt=[negative], output_type="latent", num_inference_steps=steps, guidance_scale=guidance_scale).images
|
77 |
+
image_attr = self.inference_with_attribution(out_latents)
|
78 |
+
image_attr_pil = self.pipe.numpy_to_pil(image_attr[0])
|
79 |
+
|
80 |
+
image_org = self.inference_without_attribution(out_latents)
|
81 |
+
image_org_pil = self.pipe.numpy_to_pil(image_org[0])
|
82 |
+
|
83 |
+
image_diff_pil = self.pipe.numpy_to_pil(image_attr[0] - image_org[0])
|
84 |
+
|
85 |
+
return image_org_pil[0], image_attr_pil[0], image_diff_pil[0]
|
86 |
|
87 |
def inference_without_attribution(self, latents):
|
88 |
latents = 1 / 0.18215 * latents
|
89 |
with torch.no_grad():
|
90 |
image = self.pipe.vae.decode(latents).sample
|
91 |
image = image.clamp(-1,1)
|
92 |
+
image = (image / 2 + 0.5).clamp(0, 1)
|
93 |
+
image = image.cpu().permute(0, 2, 3, 1).float().numpy()
|
94 |
return image
|
95 |
|
96 |
def get_phis(self, phi_dimension, batch_size ,eps = 1e-8):
|
|
|
122 |
|
123 |
|
124 |
attribution_model = AttributionModel()
|
125 |
+
def get_images(prompt, negative, steps, guidence_scale):
|
126 |
+
x1, x2, x3 = attribution_model.infer(prompt, negative, steps, guidence_scale)
|
127 |
+
return [x1, x2, x3]
|
128 |
+
|
129 |
+
|
130 |
+
image_examples = [
|
131 |
+
["A pikachu fine dining with a view to the Eiffel Tower", "low quality", 50, 10],
|
132 |
+
["A mecha robot in a favela in expressionist style", "low quality, 3d, photorealistic", 50, 10]
|
133 |
+
]
|
134 |
|
135 |
with gr.Blocks() as demo:
|
136 |
+
gr.Markdown(
|
137 |
+
"""<h1 style="text-align: center;"><b>WOUAF:
|
138 |
+
Weight Modulation for User Attribution and Fingerprinting in Text-to-Image Diffusion Models</b> <br> <a href="https://energy-based-model.github.io/Compositional-Visual-Generation-with-Composable-Diffusion
|
139 |
+
-Models/">Project Page</a></h1>""")
|
140 |
+
|
141 |
+
gr.Markdown(
|
142 |
+
"""<h3>Demo: Text-to-Image (Stable diffusion 2) with random user attribution</h3>
|
143 |
+
WOUAF can be applied to other applications such as In-painting, Image-editing, Image Super-Resolution etc.
|
144 |
+
<br>More details at: <a href="https://arxiv.org/abs/2306.04744">Paper</a>
|
145 |
+
"""
|
146 |
+
)
|
147 |
+
|
148 |
with gr.Row(elem_id="prompt-container").style(mobile_collapse=False, equal_height=True):
|
149 |
with gr.Column():
|
150 |
text = gr.Textbox(
|
|
|
169 |
rounded=(True, False, False, True),
|
170 |
container=False,
|
171 |
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
172 |
|
173 |
with gr.Row():
|
174 |
+
steps = gr.Slider(label="Steps", minimum=1, maximum=50, value=45, step=1)
|
175 |
guidance_scale = gr.Slider(
|
176 |
label="Guidance Scale", minimum=0, maximum=10, value=9, step=0.1
|
177 |
)
|
|
|
178 |
|
179 |
+
with gr.Row():
|
180 |
+
btn = gr.Button(value="Generate Image", full_width=False)
|
181 |
+
|
182 |
+
with gr.Row():
|
183 |
+
im_2 = gr.Image(type="pil", label="without attribution")
|
184 |
+
im_3 = gr.Image(type="pil", label="**with** attribution")
|
185 |
+
im_4 = gr.Image(type="pil", label="pixel-wise difference")
|
186 |
+
|
187 |
+
|
188 |
+
btn.click(get_images, inputs=[text, negative, steps, guidance_scale], outputs=[im_2, im_3, im_4])
|
189 |
+
|
190 |
+
gr.Examples(
|
191 |
+
examples=image_examples,
|
192 |
+
inputs=[text, negative, steps, guidance_scale],
|
193 |
+
outputs=[im_2, im_3, im_4],
|
194 |
+
fn=get_images,
|
195 |
+
cache_examples=True,
|
196 |
+
)
|
197 |
+
|
198 |
+
gr.HTML(
|
199 |
+
"""
|
200 |
+
<div class="footer">
|
201 |
+
<p>Pre-trained model by <a href="https://huggingface.co/stabilityai" style="text-decoration: underline;" target="_blank">StabilityAI</a>
|
202 |
+
</p>
|
203 |
+
<p>
|
204 |
+
Fine-tuned by authors for research purpose.
|
205 |
+
</p>
|
206 |
+
</div>
|
207 |
+
"""
|
208 |
+
)
|
209 |
+
with gr.Accordion(label="Ethics & Privacy", open=False):
|
210 |
+
gr.HTML(
|
211 |
+
"""<div class="acknowledgments">
|
212 |
+
<p><h4>Privacy</h4>
|
213 |
+
We do not collect any images or key data. This demo is designed with sole purpose of fun and reducing misuse of AI.
|
214 |
+
<p><h4>Biases and content acknowledgment</h4>
|
215 |
+
This model will have the same biases as Stable Diffusion V2.1 </div>
|
216 |
+
"""
|
217 |
+
)
|
218 |
|
219 |
+
if __name__ == "__main__":
|
220 |
+
demo.launch()
|
test.py
DELETED
@@ -1,220 +0,0 @@
|
|
1 |
-
import gradio as gr
|
2 |
-
from PIL import Image
|
3 |
-
|
4 |
-
import torch
|
5 |
-
import re
|
6 |
-
import os
|
7 |
-
import requests
|
8 |
-
|
9 |
-
from customization import customize_vae_decoder
|
10 |
-
from diffusers import AutoencoderKL, DDPMScheduler, StableDiffusionPipeline, UNet2DConditionModel, DDIMScheduler, EulerDiscreteScheduler
|
11 |
-
from torchvision import transforms
|
12 |
-
from attribution import MappingNetwork
|
13 |
-
|
14 |
-
import math
|
15 |
-
from typing import List
|
16 |
-
from PIL import Image
|
17 |
-
import cv2
|
18 |
-
import numpy as np
|
19 |
-
import torch
|
20 |
-
|
21 |
-
|
22 |
-
PRETRAINED_MODEL_NAME_OR_PATH = "./checkpoints/"
|
23 |
-
|
24 |
-
|
25 |
-
def get_image_grid(images: List[Image.Image]) -> Image:
|
26 |
-
num_images = len(images)
|
27 |
-
cols = 3#int(math.ceil(math.sqrt(num_images)))
|
28 |
-
rows = 1#int(math.ceil(num_images / cols))
|
29 |
-
width, height = images[0].size
|
30 |
-
grid_image = Image.new('RGB', (cols * width, rows * height))
|
31 |
-
for i, img in enumerate(images):
|
32 |
-
x = i % cols
|
33 |
-
y = i // cols
|
34 |
-
grid_image.paste(img, (x * width, y * height))
|
35 |
-
return grid_image
|
36 |
-
|
37 |
-
|
38 |
-
class AttributionModel:
|
39 |
-
def __init__(self):
|
40 |
-
is_cuda = False
|
41 |
-
if torch.cuda.is_available():
|
42 |
-
is_cuda = True
|
43 |
-
|
44 |
-
self.pipe = StableDiffusionPipeline.from_pretrained('stabilityai/stable-diffusion-2')#, safety_checker=None, torch_dtype=torch.float16)
|
45 |
-
if is_cuda:
|
46 |
-
self.pipe = self.pipe.to("cuda")
|
47 |
-
self.resize_transform = transforms.Resize(512, interpolation=transforms.InterpolationMode.BILINEAR)
|
48 |
-
self.vae = AutoencoderKL.from_pretrained(
|
49 |
-
'stabilityai/stable-diffusion-2', subfolder="vae"
|
50 |
-
)
|
51 |
-
self.vae = customize_vae_decoder(self.vae, 128, "qkv", "all", False, 1.0)
|
52 |
-
|
53 |
-
self.mapping_network = MappingNetwork(32, 0, 128, None, num_layers=2, w_avg_beta=None, normalization = False).to("cuda")
|
54 |
-
|
55 |
-
from torchvision.models import resnet50, ResNet50_Weights
|
56 |
-
self.decoding_network = resnet50(weights=ResNet50_Weights.IMAGENET1K_V2)
|
57 |
-
self.decoding_network.fc = torch.nn.Linear(2048,32)
|
58 |
-
|
59 |
-
self.vae.decoder.load_state_dict(torch.load(os.path.join(PRETRAINED_MODEL_NAME_OR_PATH, 'vae_decoder.pth')))
|
60 |
-
self.mapping_network.load_state_dict(torch.load(os.path.join(PRETRAINED_MODEL_NAME_OR_PATH, 'mapping_network.pth')))
|
61 |
-
self.decoding_network.load_state_dict(torch.load(os.path.join(PRETRAINED_MODEL_NAME_OR_PATH, 'decoding_network.pth')))
|
62 |
-
|
63 |
-
if is_cuda:
|
64 |
-
self.vae = self.vae.to("cuda")
|
65 |
-
self.mapping_network = self.mapping_network.to("cuda")
|
66 |
-
self.decoding_network = self.decoding_network.to("cuda")
|
67 |
-
|
68 |
-
self.test_norm = transforms.Compose(
|
69 |
-
[
|
70 |
-
transforms.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225)),
|
71 |
-
]
|
72 |
-
)
|
73 |
-
|
74 |
-
def infer(self, prompt, negative, steps, guidance_scale):
|
75 |
-
with torch.no_grad():
|
76 |
-
out_latents = self.pipe([prompt], negative_prompt=[negative], output_type="latent", num_inference_steps=steps, guidance_scale=guidance_scale).images
|
77 |
-
image_attr = self.inference_with_attribution(out_latents)
|
78 |
-
image_attr_pil = self.pipe.numpy_to_pil(image_attr[0])
|
79 |
-
|
80 |
-
image_org = self.inference_without_attribution(out_latents)
|
81 |
-
image_org_pil = self.pipe.numpy_to_pil(image_org[0])
|
82 |
-
|
83 |
-
image_diff_pil = self.pipe.numpy_to_pil(image_attr[0] - image_org[0])
|
84 |
-
|
85 |
-
return image_org_pil[0], image_attr_pil[0], image_diff_pil[0]
|
86 |
-
|
87 |
-
def inference_without_attribution(self, latents):
|
88 |
-
latents = 1 / 0.18215 * latents
|
89 |
-
with torch.no_grad():
|
90 |
-
image = self.pipe.vae.decode(latents).sample
|
91 |
-
image = image.clamp(-1,1)
|
92 |
-
image = (image / 2 + 0.5).clamp(0, 1)
|
93 |
-
image = image.cpu().permute(0, 2, 3, 1).float().numpy()
|
94 |
-
return image
|
95 |
-
|
96 |
-
def get_phis(self, phi_dimension, batch_size ,eps = 1e-8):
|
97 |
-
phi_length = phi_dimension
|
98 |
-
b = batch_size
|
99 |
-
phi = torch.empty(b,phi_length).uniform_(0,1)
|
100 |
-
return torch.bernoulli(phi) + eps
|
101 |
-
|
102 |
-
|
103 |
-
def inference_with_attribution(self, latents, key=None):
|
104 |
-
if key==None:
|
105 |
-
key = self.get_phis(32, 1)
|
106 |
-
|
107 |
-
latents = 1 / 0.18215 * latents
|
108 |
-
with torch.no_grad():
|
109 |
-
image = self.vae.decode(latents, self.mapping_network(key.cuda())).sample
|
110 |
-
image = image.clamp(-1,1)
|
111 |
-
image = (image / 2 + 0.5).clamp(0, 1)
|
112 |
-
image = image.cpu().permute(0, 2, 3, 1).float().numpy()
|
113 |
-
return image
|
114 |
-
|
115 |
-
def postprocess(self, image):
|
116 |
-
image = self.resize_transform(image)
|
117 |
-
return image
|
118 |
-
|
119 |
-
def detect_key(self, image):
|
120 |
-
reconstructed_keys = self.decoding_network(self.test_norm((image / 2 + 0.5).clamp(0, 1)))
|
121 |
-
return reconstructed_keys
|
122 |
-
|
123 |
-
|
124 |
-
attribution_model = AttributionModel()
|
125 |
-
def get_images(prompt, negative, steps, guidence_scale):
|
126 |
-
x1, x2, x3 = attribution_model.infer(prompt, negative, steps, guidence_scale)
|
127 |
-
return [x1, x2, x3]
|
128 |
-
|
129 |
-
|
130 |
-
image_examples = [
|
131 |
-
["A pikachu fine dining with a view to the Eiffel Tower", "low quality", 50, 10],
|
132 |
-
["A mecha robot in a favela in expressionist style", "low quality, 3d, photorealistic", 50, 10]
|
133 |
-
]
|
134 |
-
|
135 |
-
with gr.Blocks() as demo:
|
136 |
-
gr.Markdown(
|
137 |
-
"""<h1 style="text-align: center;"><b>WOUAF:
|
138 |
-
Weight Modulation for User Attribution and Fingerprinting in Text-to-Image Diffusion Models</b> <br> <a href="https://energy-based-model.github.io/Compositional-Visual-Generation-with-Composable-Diffusion
|
139 |
-
-Models/">Project Page</a></h1>""")
|
140 |
-
|
141 |
-
gr.Markdown(
|
142 |
-
"""<h3>Demo: Text-to-Image (Stable diffusion 2) with random user attribution</h3>
|
143 |
-
WOUAF can be applied to other applications such as In-painting, Image-editing, Image Super-Resolution etc.
|
144 |
-
<br>More details at: <a href="https://arxiv.org/abs/2306.04744">Paper</a>
|
145 |
-
"""
|
146 |
-
)
|
147 |
-
|
148 |
-
with gr.Row(elem_id="prompt-container").style(mobile_collapse=False, equal_height=True):
|
149 |
-
with gr.Column():
|
150 |
-
text = gr.Textbox(
|
151 |
-
label="Enter your prompt",
|
152 |
-
show_label=False,
|
153 |
-
max_lines=1,
|
154 |
-
placeholder="Enter your prompt",
|
155 |
-
elem_id="prompt-text-input",
|
156 |
-
).style(
|
157 |
-
border=(True, False, True, True),
|
158 |
-
rounded=(True, False, False, True),
|
159 |
-
container=False,
|
160 |
-
)
|
161 |
-
negative = gr.Textbox(
|
162 |
-
label="Enter your negative prompt",
|
163 |
-
show_label=False,
|
164 |
-
max_lines=1,
|
165 |
-
placeholder="Enter a negative prompt",
|
166 |
-
elem_id="negative-prompt-text-input",
|
167 |
-
).style(
|
168 |
-
border=(True, False, True, True),
|
169 |
-
rounded=(True, False, False, True),
|
170 |
-
container=False,
|
171 |
-
)
|
172 |
-
|
173 |
-
with gr.Row():
|
174 |
-
steps = gr.Slider(label="Steps", minimum=1, maximum=50, value=45, step=1)
|
175 |
-
guidance_scale = gr.Slider(
|
176 |
-
label="Guidance Scale", minimum=0, maximum=10, value=9, step=0.1
|
177 |
-
)
|
178 |
-
|
179 |
-
with gr.Row():
|
180 |
-
btn = gr.Button(value="Generate Image", full_width=False)
|
181 |
-
|
182 |
-
with gr.Row():
|
183 |
-
im_2 = gr.Image(type="pil", label="without attribution")
|
184 |
-
im_3 = gr.Image(type="pil", label="**with** attribution")
|
185 |
-
im_4 = gr.Image(type="pil", label="pixel-wise difference")
|
186 |
-
|
187 |
-
|
188 |
-
btn.click(get_images, inputs=[text, negative, steps, guidance_scale], outputs=[im_2, im_3, im_4])
|
189 |
-
|
190 |
-
gr.Examples(
|
191 |
-
examples=image_examples,
|
192 |
-
inputs=[text, negative, steps, guidance_scale],
|
193 |
-
outputs=[im_2, im_3, im_4],
|
194 |
-
fn=get_images,
|
195 |
-
cache_examples=True,
|
196 |
-
)
|
197 |
-
|
198 |
-
gr.HTML(
|
199 |
-
"""
|
200 |
-
<div class="footer">
|
201 |
-
<p>Pre-trained model by <a href="https://huggingface.co/stabilityai" style="text-decoration: underline;" target="_blank">StabilityAI</a>
|
202 |
-
</p>
|
203 |
-
<p>
|
204 |
-
Fine-tuned by authors for research purpose.
|
205 |
-
</p>
|
206 |
-
</div>
|
207 |
-
"""
|
208 |
-
)
|
209 |
-
with gr.Accordion(label="Ethics & Privacy", open=False):
|
210 |
-
gr.HTML(
|
211 |
-
"""<div class="acknowledgments">
|
212 |
-
<p><h4>Privacy</h4>
|
213 |
-
We do not collect any images or key data. This demo is designed with sole purpose of fun and reducing misuse of AI.
|
214 |
-
<p><h4>Biases and content acknowledgment</h4>
|
215 |
-
This model will have the same biases as Stable Diffusion V2.1 </div>
|
216 |
-
"""
|
217 |
-
)
|
218 |
-
|
219 |
-
if __name__ == "__main__":
|
220 |
-
demo.launch()
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|