Upload 6 files
Browse files- .gitattributes +4 -0
- app.py +133 -0
- example1.jpg +3 -0
- example2.jpg +3 -0
- example3.jpg +3 -0
- example4.jpg +3 -0
- requirements.txt +8 -0
.gitattributes
CHANGED
@@ -33,3 +33,7 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
|
|
33 |
*.zip filter=lfs diff=lfs merge=lfs -text
|
34 |
*.zst filter=lfs diff=lfs merge=lfs -text
|
35 |
*tfevents* filter=lfs diff=lfs merge=lfs -text
|
|
|
|
|
|
|
|
|
|
33 |
*.zip filter=lfs diff=lfs merge=lfs -text
|
34 |
*.zst filter=lfs diff=lfs merge=lfs -text
|
35 |
*tfevents* filter=lfs diff=lfs merge=lfs -text
|
36 |
+
example1.jpg filter=lfs diff=lfs merge=lfs -text
|
37 |
+
example2.jpg filter=lfs diff=lfs merge=lfs -text
|
38 |
+
example3.jpg filter=lfs diff=lfs merge=lfs -text
|
39 |
+
example4.jpg filter=lfs diff=lfs merge=lfs -text
|
app.py
ADDED
@@ -0,0 +1,133 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
import numpy as np
|
3 |
+
from PIL import Image, ImageDraw, ImageFont
|
4 |
+
import gradio as gr
|
5 |
+
from diffusers import DiffusionPipeline
|
6 |
+
from huggingface_hub import hf_hub_download
|
7 |
+
import os
|
8 |
+
|
9 |
+
# Настройки
|
10 |
+
use_custom_weights = True
|
11 |
+
custom_weights_path = hf_hub_download(
|
12 |
+
repo_id="focuzz/depth-estimation",
|
13 |
+
filename="unet_weights.pth"
|
14 |
+
)
|
15 |
+
|
16 |
+
device = "cuda" if torch.cuda.is_available() else "cpu"
|
17 |
+
dtype = torch.float16 if device == "cuda" else torch.float32
|
18 |
+
|
19 |
+
# Загрузка пайплайна
|
20 |
+
pipe = DiffusionPipeline.from_pretrained(
|
21 |
+
"prs-eth/marigold-v1-0",
|
22 |
+
custom_pipeline="marigold_depth_estimation",
|
23 |
+
torch_dtype=dtype
|
24 |
+
).to(device)
|
25 |
+
|
26 |
+
# Загрузка дообученных весов
|
27 |
+
if use_custom_weights:
|
28 |
+
state_dict = torch.load(custom_weights_path, map_location=device)
|
29 |
+
prefix = "unet.conv_in." if any(k.startswith("unet.conv_in.") for k in state_dict) else "conv_in."
|
30 |
+
conv_in_dict = {
|
31 |
+
k.replace(prefix, ""): v
|
32 |
+
for k, v in state_dict.items()
|
33 |
+
if k.startswith(prefix)
|
34 |
+
}
|
35 |
+
pipe.unet.conv_in.load_state_dict(conv_in_dict)
|
36 |
+
print("Загружены дообученные веса conv_in из:", custom_weights_path)
|
37 |
+
|
38 |
+
# Добавление overlay-текста
|
39 |
+
def add_overlay(image: Image.Image, label: str) -> Image.Image:
|
40 |
+
image = image.copy()
|
41 |
+
draw = ImageDraw.Draw(image)
|
42 |
+
font = ImageFont.load_default()
|
43 |
+
draw.text((10, 10), label, fill="white", font=font)
|
44 |
+
return image
|
45 |
+
|
46 |
+
# Генерация галереи из примеров
|
47 |
+
TARGET_SIZE = (768, 768)
|
48 |
+
def normalize_depth(depth_np):
|
49 |
+
d = np.copy(depth_np)
|
50 |
+
d_min = np.percentile(d, 1)
|
51 |
+
d_max = np.percentile(d, 99)
|
52 |
+
d = np.clip((d - d_min) / (d_max - d_min), 0, 1)
|
53 |
+
return (d * 255).astype(np.uint8)
|
54 |
+
|
55 |
+
def generate_gallery():
|
56 |
+
example_files = ["example1.jpg", "example2.jpg", "example3.jpg", "example4.jpg"]
|
57 |
+
rgbs = []
|
58 |
+
depths_gray = []
|
59 |
+
depths_color = []
|
60 |
+
|
61 |
+
for path in example_files:
|
62 |
+
if not os.path.exists(path):
|
63 |
+
continue
|
64 |
+
|
65 |
+
rgb = Image.open(path).convert("RGB").resize(TARGET_SIZE)
|
66 |
+
|
67 |
+
with torch.no_grad():
|
68 |
+
output = pipe(
|
69 |
+
rgb,
|
70 |
+
denoising_steps=4,
|
71 |
+
ensemble_size=5,
|
72 |
+
processing_res=768,
|
73 |
+
match_input_res=True,
|
74 |
+
batch_size=0,
|
75 |
+
color_map="Spectral",
|
76 |
+
show_progress_bar=False,
|
77 |
+
)
|
78 |
+
|
79 |
+
depth_np = output.depth_np
|
80 |
+
gray_normalized = normalize_depth(depth_np)
|
81 |
+
depth_gray = Image.fromarray(gray_normalized).convert("RGB").resize(TARGET_SIZE, Image.BILINEAR)
|
82 |
+
depth_color = output.depth_colored.resize(TARGET_SIZE, Image.BILINEAR)
|
83 |
+
|
84 |
+
rgbs.append(add_overlay(rgb, "RGB"))
|
85 |
+
depths_gray.append(add_overlay(depth_gray, "Depth (gray)"))
|
86 |
+
depths_color.append(add_overlay(depth_color, "Depth (color)"))
|
87 |
+
|
88 |
+
return rgbs + depths_color + depths_gray
|
89 |
+
|
90 |
+
# Интерфейс Blocks с галереей и инференсом
|
91 |
+
with gr.Blocks() as demo:
|
92 |
+
gr.Markdown("## Генерация карт глубины")
|
93 |
+
gr.Markdown(
|
94 |
+
"Модель основана на Marigold (ETH), дообучена на indoor-сценах из NYUv2. "
|
95 |
+
"Сохраняет способность обрабатывать произвольные изображения благодаря наличию оригинальных U-Net весов."
|
96 |
+
)
|
97 |
+
|
98 |
+
with gr.Row():
|
99 |
+
with gr.Column(scale=1):
|
100 |
+
input_image = gr.Image(label="Загрузите RGB изображение", type="pil")
|
101 |
+
denoise = gr.Slider(1, 50, value=4, step=1, label="Denoising Steps")
|
102 |
+
ensemble = gr.Slider(1, 10, value=5, step=1, label="Ensemble Size")
|
103 |
+
resolution = gr.Slider(256, 1024, value=768, step=64, label="Processing Resolution")
|
104 |
+
match_res = gr.Checkbox(value=True, label="Match Input Resolution")
|
105 |
+
with gr.Column(scale=1):
|
106 |
+
output_image = gr.Image(label="Карта глубины")
|
107 |
+
|
108 |
+
def predict_depth(image, denoising_steps, ensemble_size, processing_res, match_input_res):
|
109 |
+
with torch.no_grad():
|
110 |
+
output = pipe(
|
111 |
+
image,
|
112 |
+
denoising_steps=denoising_steps,
|
113 |
+
ensemble_size=ensemble_size,
|
114 |
+
processing_res=processing_res,
|
115 |
+
match_input_res=match_input_res,
|
116 |
+
batch_size=0,
|
117 |
+
color_map="Spectral",
|
118 |
+
show_progress_bar=False,
|
119 |
+
)
|
120 |
+
return output.depth_colored
|
121 |
+
|
122 |
+
submit_btn = gr.Button("Выполнить предсказание")
|
123 |
+
submit_btn.click(
|
124 |
+
predict_depth,
|
125 |
+
inputs=[input_image, denoise, ensemble, resolution, match_res],
|
126 |
+
outputs=output_image
|
127 |
+
)
|
128 |
+
|
129 |
+
gr.Markdown("### Примеры:")
|
130 |
+
gallery = gr.Gallery(label="Сравнение RGB и Depth", columns=4)
|
131 |
+
demo.load(fn=generate_gallery, outputs=gallery)
|
132 |
+
|
133 |
+
demo.launch(share=True)
|
example1.jpg
ADDED
![]() |
Git LFS Details
|
example2.jpg
ADDED
![]() |
Git LFS Details
|
example3.jpg
ADDED
![]() |
Git LFS Details
|
example4.jpg
ADDED
![]() |
Git LFS Details
|
requirements.txt
ADDED
@@ -0,0 +1,8 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
torch
|
2 |
+
diffusers>=0.25.0
|
3 |
+
transformers
|
4 |
+
accelerate
|
5 |
+
gradio
|
6 |
+
huggingface_hub
|
7 |
+
matplotlib
|
8 |
+
scipy
|