Spaces:
Runtime error
Runtime error
Create app.py
Browse files
app.py
ADDED
@@ -0,0 +1,154 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
"""Stylegan-nada-ailanta.ipynb
|
2 |
+
Automatically generated by Colab.
|
3 |
+
Original file is located at
|
4 |
+
https://colab.research.google.com/drive/1ysq4Y2sv7WTE0sW-n5W_HSgE28vaUDNE
|
5 |
+
# Проект "CLIP-Guided Domain Adaptation of Image Generators"
|
6 |
+
Данный проект представляет собой имплементацию подхода StyleGAN-NADA, предложенного в статье [StyleGAN-NADA: CLIP-Guided Domain Adaptation of Image Generators](https://arxiv.org/pdf/2108.00946).
|
7 |
+
Представленный ниже функционал предназначен для визуализации реализованного проекта и включает в себя:
|
8 |
+
- Сдвиг генератора по текстовому промпту
|
9 |
+
- Генерация примеров
|
10 |
+
- Генерация примеров из готовых пресетов
|
11 |
+
- Веб-демо
|
12 |
+
- Стилизация изображения из файла
|
13 |
+
## 1. Установка
|
14 |
+
"""
|
15 |
+
|
16 |
+
# @title
|
17 |
+
# Импорт нужных библиотек
|
18 |
+
import os
|
19 |
+
import sys
|
20 |
+
from tqdm import tqdm
|
21 |
+
import torch
|
22 |
+
import torch.nn as nn
|
23 |
+
import torch.optim as optim
|
24 |
+
from torchvision import transforms
|
25 |
+
from torchvision.utils import save_image
|
26 |
+
from PIL import Image
|
27 |
+
import numpy as np
|
28 |
+
import matplotlib.pyplot as plt
|
29 |
+
|
30 |
+
# Настройка устройства
|
31 |
+
device = "cuda" if torch.cuda.is_available() else "cpu"
|
32 |
+
|
33 |
+
import os
|
34 |
+
import subprocess
|
35 |
+
|
36 |
+
if not os.path.exists("stylegan2-pytorch"):
|
37 |
+
subprocess.run(["git", "clone", "https://github.com/rosinality/stylegan2-pytorch.git"])
|
38 |
+
os.chdir("stylegan2-pytorch")
|
39 |
+
|
40 |
+
import gdown
|
41 |
+
gdown.download('https://drive.google.com/uc?id=1EM87UquaoQmk17Q8d5kYIAHqu0dkYqdT')
|
42 |
+
gdown.download('https://drive.google.com/uc?id=1N0MZSqPRJpLfP4mFQCS14ikrVSe8vQlL')
|
43 |
+
sys.path.append(os.path.abspath("stylegan2-pytorch"))
|
44 |
+
from model import Generator
|
45 |
+
|
46 |
+
# Параметры генератора
|
47 |
+
latent_dim = 512
|
48 |
+
f_generator = Generator(size=1024, style_dim=latent_dim, n_mlp=8).to(device)
|
49 |
+
state_dict = torch.load('stylegan2-ffhq-config-f.pt', map_location=device)
|
50 |
+
f_generator.load_state_dict(state_dict['g_ema'])
|
51 |
+
f_generator.eval()
|
52 |
+
|
53 |
+
g_generator = Generator(size=1024, style_dim=latent_dim, n_mlp=8).to(device)
|
54 |
+
g_generator.load_state_dict(state_dict['g_ema'])
|
55 |
+
|
56 |
+
# Загрузка модели CLIP
|
57 |
+
import clip
|
58 |
+
clip_model, preprocess = clip.load("ViT-B/32", device=device)
|
59 |
+
|
60 |
+
latent_dim=512
|
61 |
+
batch_size=4
|
62 |
+
|
63 |
+
"""## 6. Готовые пресеты"""
|
64 |
+
|
65 |
+
# @title Загрузка пресетов
|
66 |
+
os.makedirs("/content/presets", exist_ok=True)
|
67 |
+
|
68 |
+
gdown.download('https://drive.google.com/uc?id=1trcBvlz7jeBRLNeCyNVCXE4esW25GPaZ', '/content/presets/sketch.pth')
|
69 |
+
gdown.download('https://drive.google.com/uc?id=1N4C-aTwxeOamZX2GeEElppsMv-ALKojL', '/content/presets/modigliani.pth')
|
70 |
+
gdown.download('https://drive.google.com/uc?id=1VZHEalFyEFGWIaHei98f9XPyHHvMBp6J', '/content/presets/werewolf.pth')
|
71 |
+
|
72 |
+
# @title Генерация примеров из пресета
|
73 |
+
# Загрузка генератора из файла
|
74 |
+
def load_model(file_path, latent_dim=512, size=1024):
|
75 |
+
|
76 |
+
state_dicts = torch.load(file_path, map_location=device)
|
77 |
+
|
78 |
+
# Инициализация
|
79 |
+
trained_generator = Generator(size=size, style_dim=latent_dim, n_mlp=8).to(device)
|
80 |
+
|
81 |
+
# Загрузка весов
|
82 |
+
trained_generator.load_state_dict(state_dicts)
|
83 |
+
|
84 |
+
trained_generator.eval()
|
85 |
+
|
86 |
+
return trained_generator
|
87 |
+
|
88 |
+
model_paths = {
|
89 |
+
"Photo -> Pencil Sketch": "/content/presets/sketch.pth",
|
90 |
+
"Photo -> Modigliani Painting": "/content/presets/modigliani.pth",
|
91 |
+
"Human -> Werewolf": "/content/presets/werewolf.pth"
|
92 |
+
}
|
93 |
+
|
94 |
+
|
95 |
+
"""## 8. Веб-демо"""
|
96 |
+
|
97 |
+
import gradio as gr
|
98 |
+
|
99 |
+
def get_avg_image(net):
|
100 |
+
avg_image = net(net.latent_avg.unsqueeze(0),
|
101 |
+
input_code=True,
|
102 |
+
randomize_noise=False,
|
103 |
+
return_latents=False,
|
104 |
+
average_code=True)[0]
|
105 |
+
avg_image = avg_image.to('cuda').float().detach()
|
106 |
+
return avg_image
|
107 |
+
|
108 |
+
# Функция обработки изображения
|
109 |
+
def process_image(image):
|
110 |
+
# Конвертация в объект PIL
|
111 |
+
image = Image.fromarray(image)
|
112 |
+
|
113 |
+
# Изменение размера до 256x256
|
114 |
+
image = image.resize((256, 256))
|
115 |
+
|
116 |
+
input_image = transform(image).unsqueeze(0).to(device)
|
117 |
+
|
118 |
+
opts.n_iters_per_batch = 5
|
119 |
+
opts.resize_outputs = False # generate outputs at full resolution
|
120 |
+
|
121 |
+
from restyle.utils.inference_utils import run_on_batch
|
122 |
+
|
123 |
+
with torch.no_grad():
|
124 |
+
avg_image = get_avg_image(restyle_net)
|
125 |
+
result_batch, result_latents = run_on_batch(input_image, restyle_net, opts, avg_image)
|
126 |
+
|
127 |
+
inverted_latent = torch.Tensor(result_latents[0][4]).cuda().unsqueeze(0).unsqueeze(1)
|
128 |
+
|
129 |
+
with torch.no_grad():
|
130 |
+
sampled_src = f_generator(inverted_latent, input_is_latent=True)[0]
|
131 |
+
frozen_image = (sampled_src.clamp(-1, 1) + 1) / 2.0 # Нормализация к [0, 1]
|
132 |
+
frozen_image = frozen_image.permute(0, 2, 3, 1).cpu().numpy()
|
133 |
+
|
134 |
+
g_generator.eval()
|
135 |
+
|
136 |
+
sampled_src = g_generator(inverted_latent, input_is_latent=True)[0]
|
137 |
+
trained_image = (sampled_src.clamp(-1, 1) + 1) / 2.0 # Нормализация к [0, 1]
|
138 |
+
trained_image = trained_image.permute(0, 2, 3, 1).cpu().numpy()
|
139 |
+
images = []
|
140 |
+
images.append(image)
|
141 |
+
images.append(frozen_image.squeeze(0))
|
142 |
+
images.append(trained_image.squeeze(0))
|
143 |
+
return images
|
144 |
+
|
145 |
+
# Интерфейс Gradio
|
146 |
+
iface = gr.Interface(
|
147 |
+
fn=process_image, # Функция обработки
|
148 |
+
inputs=gr.Image(type="numpy"), # Поле для загрузки изображения
|
149 |
+
outputs=gr.Gallery(label="Результаты генерации", columns=2),
|
150 |
+
title="Обработка изображения",
|
151 |
+
description="Загрузите изображение"
|
152 |
+
)
|
153 |
+
|
154 |
+
iface.launch()
|