Spaces:
Running
on
Zero
Running
on
Zero
Delete app.py
Browse files
app.py
DELETED
@@ -1,415 +0,0 @@
|
|
1 |
-
import os
|
2 |
-
import gradio as gr
|
3 |
-
import json
|
4 |
-
import logging
|
5 |
-
import torch
|
6 |
-
from PIL import Image
|
7 |
-
import spaces
|
8 |
-
from diffusers import StableDiffusionXLPipeline, StableDiffusionXLImg2ImgPipeline
|
9 |
-
from diffusers.utils import load_image
|
10 |
-
from huggingface_hub import hf_hub_download, HfFileSystem, ModelCard
|
11 |
-
import copy
|
12 |
-
import random
|
13 |
-
import time
|
14 |
-
import re
|
15 |
-
|
16 |
-
# Load LoRAs from JSON file
|
17 |
-
with open('loras.json', 'r') as f:
|
18 |
-
loras = json.load(f)
|
19 |
-
|
20 |
-
# Initialize the base model for SDXL
|
21 |
-
dtype = torch.float16 if torch.cuda.is_available() else torch.float32
|
22 |
-
device = "cuda" if torch.cuda.is_available() else "cpu"
|
23 |
-
base_model = "stabilityai/stable-diffusion-xl-base-1.0"
|
24 |
-
|
25 |
-
# Load SDXL pipelines
|
26 |
-
pipe = StableDiffusionXLPipeline.from_pretrained(
|
27 |
-
base_model,
|
28 |
-
torch_dtype=dtype,
|
29 |
-
use_safetensors=True
|
30 |
-
).to(device)
|
31 |
-
|
32 |
-
pipe_i2i = StableDiffusionXLImg2ImgPipeline.from_pretrained(
|
33 |
-
base_model,
|
34 |
-
torch_dtype=dtype,
|
35 |
-
use_safetensors=True
|
36 |
-
).to(device)
|
37 |
-
|
38 |
-
MAX_SEED = 2**32 - 1
|
39 |
-
|
40 |
-
# Custom SDXL generation function for live preview
|
41 |
-
@torch.inference_mode()
|
42 |
-
def generate_sdxl_images(
|
43 |
-
pipe,
|
44 |
-
prompt: str,
|
45 |
-
height: int = 1024,
|
46 |
-
width: int = 1024,
|
47 |
-
num_inference_steps: int = 50,
|
48 |
-
guidance_scale: float = 7.5,
|
49 |
-
generator: Optional[torch.Generator] = None,
|
50 |
-
output_type: str = "pil",
|
51 |
-
):
|
52 |
-
# Encode prompt
|
53 |
-
prompt_embeds, negative_prompt_embeds, pooled_prompt_embeds, negative_pooled_prompt_embeds = pipe.encode_prompt(
|
54 |
-
prompt=prompt,
|
55 |
-
num_images_per_prompt=1,
|
56 |
-
do_classifier_free_guidance=True,
|
57 |
-
)
|
58 |
-
# Prepare latents
|
59 |
-
latents = pipe.prepare_latents(
|
60 |
-
batch_size=1,
|
61 |
-
num_channels_latents=pipe.unet.config.in_channels,
|
62 |
-
height=height,
|
63 |
-
width=width,
|
64 |
-
dtype=prompt_embeds.dtype,
|
65 |
-
device=pipe.device,
|
66 |
-
generator=generator,
|
67 |
-
)
|
68 |
-
# Prepare timesteps
|
69 |
-
pipe.scheduler.set_timesteps(num_inference_steps, device=pipe.device)
|
70 |
-
timesteps = pipe.scheduler.timesteps
|
71 |
-
# Prepare guidance
|
72 |
-
do_classifier_free_guidance = guidance_scale > 1.0
|
73 |
-
if do_classifier_free_guidance:
|
74 |
-
prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
|
75 |
-
pooled_prompt_embeds = torch.cat([negative_pooled_prompt_embeds, pooled_prompt_embeds])
|
76 |
-
# Denoising loop
|
77 |
-
for i, t in enumerate(timesteps):
|
78 |
-
# Expand latents for guidance
|
79 |
-
latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
|
80 |
-
# Predict noise
|
81 |
-
noise_pred = pipe.unet(
|
82 |
-
latent_model_input,
|
83 |
-
t,
|
84 |
-
encoder_hidden_states=prompt_embeds,
|
85 |
-
added_cond_kwargs={"text_embeds": pooled_prompt_embeds},
|
86 |
-
).sample
|
87 |
-
# Perform guidance
|
88 |
-
if do_classifier_free_guidance:
|
89 |
-
noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
|
90 |
-
noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
|
91 |
-
# Step scheduler
|
92 |
-
latents = pipe.scheduler.step(noise_pred, t, latents).prev_sample
|
93 |
-
# Decode latents to image every step
|
94 |
-
image = pipe.vae.decode(latents / pipe.vae.config.scaling_factor, return_dict=False)[0]
|
95 |
-
yield pipe.image_processor.postprocess(image, output_type=output_type)[0]
|
96 |
-
# Final image
|
97 |
-
image = pipe.vae.decode(latents / pipe.vae.config.scaling_factor, return_dict=False)[0]
|
98 |
-
yield pipe.image_processor.postprocess(image, output_type=output_type)[0]
|
99 |
-
|
100 |
-
class calculateDuration:
|
101 |
-
def __init__(self, activity_name=""):
|
102 |
-
self.activity_name = activity_name
|
103 |
-
|
104 |
-
def __enter__(self):
|
105 |
-
self.start_time = time.time()
|
106 |
-
return self
|
107 |
-
|
108 |
-
def __exit__(self, exc_type, exc_value, traceback):
|
109 |
-
self.end_time = time.time()
|
110 |
-
self.elapsed_time = self.end_time - self.start_time
|
111 |
-
if self.activity_name:
|
112 |
-
print(f"Elapsed time for {self.activity_name}: {self.elapsed_time:.6f} seconds")
|
113 |
-
else:
|
114 |
-
print(f"Elapsed time: {self.elapsed_time:.6f} seconds")
|
115 |
-
|
116 |
-
def update_selection(evt: gr.SelectData, width, height):
|
117 |
-
selected_lora = loras[evt.index]
|
118 |
-
new_placeholder = f"Type a prompt for {selected_lora['title']}"
|
119 |
-
lora_repo = selected_lora["repo"]
|
120 |
-
updated_text = f"### Selected: [{lora_repo}](https://huggingface.co/{lora_repo}) ✨"
|
121 |
-
if "aspect" in selected_lora:
|
122 |
-
if selected_lora["aspect"] == "portrait":
|
123 |
-
width = 768
|
124 |
-
height = 1024
|
125 |
-
elif selected_lora["aspect"] == "landscape":
|
126 |
-
width = 1024
|
127 |
-
height = 768
|
128 |
-
else:
|
129 |
-
width = 1024
|
130 |
-
height = 1024
|
131 |
-
return (
|
132 |
-
gr.update(placeholder=new_placeholder),
|
133 |
-
updated_text,
|
134 |
-
evt.index,
|
135 |
-
width,
|
136 |
-
height,
|
137 |
-
)
|
138 |
-
|
139 |
-
@spaces.GPU(duration=70)
|
140 |
-
def generate_image(prompt_mash, steps, seed, cfg_scale, width, height, progress):
|
141 |
-
pipe.to("cuda")
|
142 |
-
generator = torch.Generator(device="cuda").manual_seed(seed)
|
143 |
-
with calculateDuration("Generating image"):
|
144 |
-
for img in generate_sdxl_images(
|
145 |
-
pipe,
|
146 |
-
prompt=prompt_mash,
|
147 |
-
num_inference_steps=steps,
|
148 |
-
guidance_scale=cfg_scale,
|
149 |
-
width=width,
|
150 |
-
height=height,
|
151 |
-
generator=generator,
|
152 |
-
output_type="pil",
|
153 |
-
):
|
154 |
-
yield img
|
155 |
-
|
156 |
-
def generate_image_to_image(prompt_mash, image_input_path, image_strength, steps, cfg_scale, width, height, seed):
|
157 |
-
generator = torch.Generator(device="cuda").manual_seed(seed)
|
158 |
-
pipe_i2i.to("cuda")
|
159 |
-
image_input = load_image(image_input_path)
|
160 |
-
final_image = pipe_i2i(
|
161 |
-
prompt=prompt_mash,
|
162 |
-
image=image_input,
|
163 |
-
strength=image_strength,
|
164 |
-
num_inference_steps=steps,
|
165 |
-
guidance_scale=cfg_scale,
|
166 |
-
width=width,
|
167 |
-
height=height,
|
168 |
-
generator=generator,
|
169 |
-
output_type="pil",
|
170 |
-
).images[0]
|
171 |
-
return final_image
|
172 |
-
|
173 |
-
@spaces.GPU(duration=70)
|
174 |
-
def run_lora(prompt, image_input, image_strength, cfg_scale, steps, selected_index, randomize_seed, seed, width, height, lora_scale, progress=gr.Progress(track_tqdm=True)):
|
175 |
-
if selected_index is None:
|
176 |
-
raise gr.Error("You must select a LoRA before proceeding.")
|
177 |
-
selected_lora = loras[selected_index]
|
178 |
-
lora_path = selected_lora["repo"]
|
179 |
-
trigger_word = selected_lora["trigger_word"]
|
180 |
-
if trigger_word:
|
181 |
-
if "trigger_position" in selected_lora and selected_lora["trigger_position"] == "prepend":
|
182 |
-
prompt_mash = f"{trigger_word} {prompt}"
|
183 |
-
else:
|
184 |
-
prompt_mash = f"{prompt} {trigger_word}"
|
185 |
-
else:
|
186 |
-
prompt_mash = prompt
|
187 |
-
|
188 |
-
# Unload previous LoRA weights
|
189 |
-
with calculateDuration("Unloading LoRA"):
|
190 |
-
pipe.unload_lora_weights()
|
191 |
-
pipe_i2i.unload_lora_weights()
|
192 |
-
|
193 |
-
# Load LoRA weights and set adapter scale
|
194 |
-
with calculateDuration(f"Loading LoRA weights for {selected_lora['title']}"):
|
195 |
-
weight_name = selected_lora.get("weights", None)
|
196 |
-
adapter_name = "lora"
|
197 |
-
pipe.load_lora_weights(lora_path, weight_name=weight_name, adapter_name=adapter_name)
|
198 |
-
pipe.set_adapters([adapter_name], [lora_scale])
|
199 |
-
pipe_i2i.load_lora_weights(lora_path, weight_name=weight_name, adapter_name=adapter_name)
|
200 |
-
pipe_i2i.set_adapters([adapter_name], [lora_scale])
|
201 |
-
|
202 |
-
# Set random seed
|
203 |
-
with calculateDuration("Randomizing seed"):
|
204 |
-
if randomize_seed:
|
205 |
-
seed = random.randint(0, MAX_SEED)
|
206 |
-
|
207 |
-
if image_input is not None:
|
208 |
-
final_image = generate_image_to_image(prompt_mash, image_input, image_strength, steps, cfg_scale, width, height, seed)
|
209 |
-
yield final_image, seed, gr.update(visible=False)
|
210 |
-
else:
|
211 |
-
image_generator = generate_image(prompt_mash, steps, seed, cfg_scale, width, height, progress)
|
212 |
-
final_image = None
|
213 |
-
step_counter = 0
|
214 |
-
for image in image_generator:
|
215 |
-
step_counter += 1
|
216 |
-
final_image = image
|
217 |
-
progress_bar = f'<div class="progress-container"><div class="progress-bar" style="--current: {step_counter}; --total: {steps};"></div></div>'
|
218 |
-
yield image, seed, gr.update(value=progress_bar, visible=True)
|
219 |
-
yield final_image, seed, gr.update(value=progress_bar, visible=False)
|
220 |
-
|
221 |
-
def get_huggingface_safetensors(link):
|
222 |
-
split_link = link.split("/")
|
223 |
-
if len(split_link) != 2:
|
224 |
-
raise Exception("Invalid Hugging Face repository link format.")
|
225 |
-
|
226 |
-
# Load model card
|
227 |
-
model_card = ModelCard.load(link)
|
228 |
-
base_model = model_card.data.get("base_model")
|
229 |
-
print(base_model)
|
230 |
-
|
231 |
-
# Validate model type for SDXL
|
232 |
-
if base_model != "stabilityai/stable-diffusion-xl-base-1.0":
|
233 |
-
raise Exception("Not an SDXL LoRA!")
|
234 |
-
|
235 |
-
# Extract image and trigger word
|
236 |
-
image_path = model_card.data.get("widget", [{}])[0].get("output", {}).get("url", None)
|
237 |
-
trigger_word = model_card.data.get("instance_prompt", "")
|
238 |
-
image_url = f"https://huggingface.co/{link}/resolve/main/{image_path}" if image_path else None
|
239 |
-
|
240 |
-
# Initialize Hugging Face file system
|
241 |
-
fs = HfFileSystem()
|
242 |
-
try:
|
243 |
-
list_of_files = fs.ls(link, detail=False)
|
244 |
-
safetensors_name = None
|
245 |
-
highest_trained_file = None
|
246 |
-
highest_steps = -1
|
247 |
-
last_safetensors_file = None
|
248 |
-
step_pattern = re.compile(r"_0{3,}\d+") # Detects step count `_000...`
|
249 |
-
|
250 |
-
for file in list_of_files:
|
251 |
-
filename = file.split("/")[-1]
|
252 |
-
if filename.endswith(".safetensors"):
|
253 |
-
last_safetensors_file = filename
|
254 |
-
match = step_pattern.search(filename)
|
255 |
-
if not match:
|
256 |
-
safetensors_name = filename
|
257 |
-
break
|
258 |
-
else:
|
259 |
-
steps = int(match.group().lstrip("_"))
|
260 |
-
if steps > highest_steps:
|
261 |
-
highest_trained_file = filename
|
262 |
-
highest_steps = steps
|
263 |
-
if not image_url and filename.lower().endswith((".jpg", ".jpeg", ".png", ".webp")):
|
264 |
-
image_url = f"https://huggingface.co/{link}/resolve/main/{filename}"
|
265 |
-
|
266 |
-
if not safetensors_name:
|
267 |
-
safetensors_name = highest_trained_file if highest_trained_file else last_safetensors_file
|
268 |
-
if not safetensors_name:
|
269 |
-
raise Exception("No valid *.safetensors file found in the repository.")
|
270 |
-
except Exception as e:
|
271 |
-
print(e)
|
272 |
-
raise Exception("You didn't include a valid Hugging Face repository with a *.safetensors LoRA")
|
273 |
-
|
274 |
-
return split_link[1], link, safetensors_name, trigger_word, image_url
|
275 |
-
|
276 |
-
def check_custom_model(link):
|
277 |
-
if link.startswith("https://"):
|
278 |
-
if link.startswith("https://huggingface.co") or link.startswith("https://www.huggingface.co"):
|
279 |
-
link_split = link.split("huggingface.co/")
|
280 |
-
return get_huggingface_safetensors(link_split[1])
|
281 |
-
else:
|
282 |
-
return get_huggingface_safetensors(link)
|
283 |
-
|
284 |
-
def add_custom_lora(custom_lora):
|
285 |
-
global loras
|
286 |
-
if custom_lora:
|
287 |
-
try:
|
288 |
-
title, repo, path, trigger_word, image = check_custom_model(custom_lora)
|
289 |
-
print(f"Loaded custom LoRA: {repo}")
|
290 |
-
card = f'''
|
291 |
-
<div class="custom_lora_card">
|
292 |
-
<span>Loaded custom LoRA:</span>
|
293 |
-
<div class="card_internal">
|
294 |
-
<img src="{image}" />
|
295 |
-
<div>
|
296 |
-
<h3>{title}</h3>
|
297 |
-
<small>{"Using: <code><b>"+trigger_word+"</code></b> as the trigger word" if trigger_word else "No trigger word found. If there's a trigger word, include it in your prompt"}<br></small>
|
298 |
-
</div>
|
299 |
-
</div>
|
300 |
-
</div>
|
301 |
-
'''
|
302 |
-
existing_item_index = next((index for (index, item) in enumerate(loras) if item['repo'] == repo), None)
|
303 |
-
if not existing_item_index:
|
304 |
-
new_item = {
|
305 |
-
"image": image,
|
306 |
-
"title": title,
|
307 |
-
"repo": repo,
|
308 |
-
"weights": path,
|
309 |
-
"trigger_word": trigger_word
|
310 |
-
}
|
311 |
-
print(new_item)
|
312 |
-
existing_item_index = len(loras)
|
313 |
-
loras.append(new_item)
|
314 |
-
return gr.update(visible=True, value=card), gr.update(visible=True), gr.Gallery(selected_index=None), f"Custom: {path}", existing_item_index, trigger_word
|
315 |
-
except Exception as e:
|
316 |
-
gr.Warning(f"Invalid LoRA: either you entered an invalid link, or a non-SDXL LoRA")
|
317 |
-
return gr.update(visible=True, value=f"Invalid LoRA: either you entered an invalid link, a non-SDXL LoRA"), gr.update(visible=True), gr.update(), "", None, ""
|
318 |
-
else:
|
319 |
-
return gr.update(visible=False), gr.update(visible=False), gr.update(), "", None, ""
|
320 |
-
|
321 |
-
def remove_custom_lora():
|
322 |
-
return gr.update(visible=False), gr.update(visible=False), gr.update(), "", None, ""
|
323 |
-
|
324 |
-
run_lora.zerogpu = True
|
325 |
-
|
326 |
-
css = '''
|
327 |
-
#gen_btn{height: 100%}
|
328 |
-
#gen_column{align-self: stretch}
|
329 |
-
#title{text-align: center}
|
330 |
-
#title h1{font-size: 3em; display:inline-flex; align-items:center}
|
331 |
-
#title img{width: 100px; margin-right: 0.5em}
|
332 |
-
#gallery .grid-wrap{height: 10vh}
|
333 |
-
#lora_list{background: var(--block-background-fill);padding: 0 1em .3em; font-size: 90%}
|
334 |
-
.card_internal{display: flex;height: 100px;margin-top: .5em}
|
335 |
-
.card_internal img{margin-right: 1em}
|
336 |
-
.styler{--form-gap-width: 0px !important}
|
337 |
-
#progress{height:30px}
|
338 |
-
#progress .generating{display:none}
|
339 |
-
.progress-container {width: 100%;height: 30px;background-color: #f0f0f0;border-radius: 15px;overflow: hidden;margin-bottom: 20px}
|
340 |
-
.progress-bar {height: 100%;background-color: #4f46e5;width: calc(var(--current) / var(--total) * 100%);transition: width 0.5s ease-in-out}
|
341 |
-
'''
|
342 |
-
font = [gr.themes.GoogleFont("Source Sans Pro"), "Arial", "sans-serif"]
|
343 |
-
with gr.Blocks(theme=gr.themes.Soft(font=font), css=css, delete_cache=(60, 60)) as app:
|
344 |
-
title = gr.HTML(
|
345 |
-
"""<h1>SDXL LoRA DLC</h1>""",
|
346 |
-
elem_id="title",
|
347 |
-
)
|
348 |
-
selected_index = gr.State(None)
|
349 |
-
with gr.Row():
|
350 |
-
with gr.Column(scale=3):
|
351 |
-
prompt = gr.Textbox(label="Prompt", lines=1, placeholder="Type a prompt after selecting a LoRA")
|
352 |
-
with gr.Column(scale=1, elem_id="gen_column"):
|
353 |
-
generate_button = gr.Button("Generate", variant="primary", elem_id="gen_btn")
|
354 |
-
with gr.Row():
|
355 |
-
with gr.Column():
|
356 |
-
selected_info = gr.Markdown("")
|
357 |
-
gallery = gr.Gallery(
|
358 |
-
[(item["image"], item["title"]) for item in loras],
|
359 |
-
label="LoRA Gallery",
|
360 |
-
allow_preview=False,
|
361 |
-
columns=3,
|
362 |
-
elem_id="gallery",
|
363 |
-
show_share_button=False
|
364 |
-
)
|
365 |
-
with gr.Group():
|
366 |
-
custom_lora = gr.Textbox(label="Custom LoRA", info="LoRA Hugging Face path", placeholder="username/sdxl-lora-model")
|
367 |
-
gr.Markdown("[Check the list of SDXL LoRAs](https://huggingface.co/models?other=base_model:stabilityai/stable-diffusion-xl-base-1.0)", elem_id="lora_list")
|
368 |
-
custom_lora_info = gr.HTML(visible=False)
|
369 |
-
custom_lora_button = gr.Button("Remove custom LoRA", visible=False)
|
370 |
-
with gr.Column():
|
371 |
-
progress_bar = gr.Markdown(elem_id="progress", visible=False)
|
372 |
-
result = gr.Image(label="Generated Image")
|
373 |
-
|
374 |
-
with gr.Row():
|
375 |
-
with gr.Accordion("Advanced Settings", open=False):
|
376 |
-
with gr.Row():
|
377 |
-
input_image = gr.Image(label="Input image", type="filepath")
|
378 |
-
image_strength = gr.Slider(label="Denoise Strength", info="Lower means more image influence", minimum=0.1, maximum=1.0, step=0.01, value=0.75)
|
379 |
-
with gr.Column():
|
380 |
-
with gr.Row():
|
381 |
-
cfg_scale = gr.Slider(label="CFG Scale", minimum=1, maximum=20, step=0.5, value=7.5)
|
382 |
-
steps = gr.Slider(label="Steps", minimum=1, maximum=50, step=1, value=30)
|
383 |
-
|
384 |
-
with gr.Row():
|
385 |
-
width = gr.Slider(label="Width", minimum=256, maximum=1536, step=64, value=1024)
|
386 |
-
height = gr.Slider(label="Height", minimum=256, maximum=1536, step=64, value=1024)
|
387 |
-
|
388 |
-
with gr.Row():
|
389 |
-
randomize_seed = gr.Checkbox(True, label="Randomize seed")
|
390 |
-
seed = gr.Slider(label="Seed", minimum=0, maximum=MAX_SEED, step=1, value=0, randomize=True)
|
391 |
-
lora_scale = gr.Slider(label="LoRA Scale", minimum=0, maximum=3, step=0.01, value=1.0)
|
392 |
-
|
393 |
-
gallery.select(
|
394 |
-
update_selection,
|
395 |
-
inputs=[width, height],
|
396 |
-
outputs=[prompt, selected_info, selected_index, width, height]
|
397 |
-
)
|
398 |
-
custom_lora.input(
|
399 |
-
add_custom_lora,
|
400 |
-
inputs=[custom_lora],
|
401 |
-
outputs=[custom_lora_info, custom_lora_button, gallery, selected_info, selected_index, prompt]
|
402 |
-
)
|
403 |
-
custom_lora_button.click(
|
404 |
-
remove_custom_lora,
|
405 |
-
outputs=[custom_lora_info, custom_lora_button, gallery, selected_info, selected_index, custom_lora]
|
406 |
-
)
|
407 |
-
gr.on(
|
408 |
-
triggers=[generate_button.click, prompt.submit],
|
409 |
-
fn=run_lora,
|
410 |
-
inputs=[prompt, input_image, image_strength, cfg_scale, steps, selected_index, randomize_seed, seed, width, height, lora_scale],
|
411 |
-
outputs=[result, seed, progress_bar]
|
412 |
-
)
|
413 |
-
|
414 |
-
app.queue()
|
415 |
-
app.launch()
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|