Spaces:
Running
Running
Split file
Browse files- app.py +7 -163
- dualstylegan.py +167 -0
app.py
CHANGED
|
@@ -5,27 +5,10 @@ from __future__ import annotations
|
|
| 5 |
import argparse
|
| 6 |
import os
|
| 7 |
import pathlib
|
| 8 |
-
import sys
|
| 9 |
-
from typing import Callable
|
| 10 |
|
| 11 |
-
import dlib
|
| 12 |
import gradio as gr
|
| 13 |
-
import huggingface_hub
|
| 14 |
-
import numpy as np
|
| 15 |
-
import PIL.Image
|
| 16 |
-
import torch
|
| 17 |
-
import torch.nn as nn
|
| 18 |
-
import torchvision.transforms as T
|
| 19 |
|
| 20 |
-
|
| 21 |
-
os.system("sed -i '10,17d' DualStyleGAN/model/stylegan/op/fused_act.py")
|
| 22 |
-
os.system("sed -i '10,17d' DualStyleGAN/model/stylegan/op/upfirdn2d.py")
|
| 23 |
-
|
| 24 |
-
sys.path.insert(0, 'DualStyleGAN')
|
| 25 |
-
|
| 26 |
-
from model.dualstylegan import DualStyleGAN
|
| 27 |
-
from model.encoder.align_all_parallel import align_face
|
| 28 |
-
from model.encoder.psp import pSp
|
| 29 |
|
| 30 |
TOKEN = os.environ['TOKEN']
|
| 31 |
MODEL_REPO = 'hysts/DualStyleGAN'
|
|
@@ -43,146 +26,6 @@ def parse_args() -> argparse.Namespace:
|
|
| 43 |
return parser.parse_args()
|
| 44 |
|
| 45 |
|
| 46 |
-
class App:
|
| 47 |
-
|
| 48 |
-
def __init__(self, device: torch.device):
|
| 49 |
-
self.device = device
|
| 50 |
-
self.landmark_model = self._create_dlib_landmark_model()
|
| 51 |
-
self.encoder = self._load_encoder()
|
| 52 |
-
self.transform = self._create_transform()
|
| 53 |
-
|
| 54 |
-
self.style_types = [
|
| 55 |
-
'cartoon',
|
| 56 |
-
'caricature',
|
| 57 |
-
'anime',
|
| 58 |
-
'arcane',
|
| 59 |
-
'comic',
|
| 60 |
-
'pixar',
|
| 61 |
-
'slamdunk',
|
| 62 |
-
]
|
| 63 |
-
self.generator_dict = {
|
| 64 |
-
style_type: self._load_generator(style_type)
|
| 65 |
-
for style_type in self.style_types
|
| 66 |
-
}
|
| 67 |
-
self.exstyle_dict = {
|
| 68 |
-
style_type: self._load_exstylecode(style_type)
|
| 69 |
-
for style_type in self.style_types
|
| 70 |
-
}
|
| 71 |
-
|
| 72 |
-
@staticmethod
|
| 73 |
-
def _create_dlib_landmark_model():
|
| 74 |
-
path = huggingface_hub.hf_hub_download(
|
| 75 |
-
'hysts/dlib_face_landmark_model',
|
| 76 |
-
'shape_predictor_68_face_landmarks.dat',
|
| 77 |
-
use_auth_token=TOKEN)
|
| 78 |
-
return dlib.shape_predictor(path)
|
| 79 |
-
|
| 80 |
-
def _load_encoder(self) -> nn.Module:
|
| 81 |
-
ckpt_path = huggingface_hub.hf_hub_download(MODEL_REPO,
|
| 82 |
-
'models/encoder.pt',
|
| 83 |
-
use_auth_token=TOKEN)
|
| 84 |
-
ckpt = torch.load(ckpt_path, map_location='cpu')
|
| 85 |
-
opts = ckpt['opts']
|
| 86 |
-
opts['device'] = self.device.type
|
| 87 |
-
opts['checkpoint_path'] = ckpt_path
|
| 88 |
-
opts = argparse.Namespace(**opts)
|
| 89 |
-
model = pSp(opts)
|
| 90 |
-
model.to(self.device)
|
| 91 |
-
model.eval()
|
| 92 |
-
return model
|
| 93 |
-
|
| 94 |
-
@staticmethod
|
| 95 |
-
def _create_transform() -> Callable:
|
| 96 |
-
transform = T.Compose([
|
| 97 |
-
T.Resize(256),
|
| 98 |
-
T.CenterCrop(256),
|
| 99 |
-
T.ToTensor(),
|
| 100 |
-
T.Normalize([0.5, 0.5, 0.5], [0.5, 0.5, 0.5]),
|
| 101 |
-
])
|
| 102 |
-
return transform
|
| 103 |
-
|
| 104 |
-
def _load_generator(self, style_type: str) -> nn.Module:
|
| 105 |
-
model = DualStyleGAN(1024, 512, 8, 2, res_index=6)
|
| 106 |
-
ckpt_path = huggingface_hub.hf_hub_download(
|
| 107 |
-
MODEL_REPO,
|
| 108 |
-
f'models/{style_type}/generator.pt',
|
| 109 |
-
use_auth_token=TOKEN)
|
| 110 |
-
ckpt = torch.load(ckpt_path, map_location='cpu')
|
| 111 |
-
model.load_state_dict(ckpt['g_ema'])
|
| 112 |
-
model.to(self.device)
|
| 113 |
-
model.eval()
|
| 114 |
-
return model
|
| 115 |
-
|
| 116 |
-
@staticmethod
|
| 117 |
-
def _load_exstylecode(style_type: str) -> dict[str, np.ndarray]:
|
| 118 |
-
if style_type in ['cartoon', 'caricature', 'anime']:
|
| 119 |
-
filename = 'refined_exstyle_code.npy'
|
| 120 |
-
else:
|
| 121 |
-
filename = 'exstyle_code.npy'
|
| 122 |
-
path = huggingface_hub.hf_hub_download(
|
| 123 |
-
MODEL_REPO,
|
| 124 |
-
f'models/{style_type}/{filename}',
|
| 125 |
-
use_auth_token=TOKEN)
|
| 126 |
-
exstyles = np.load(path, allow_pickle=True).item()
|
| 127 |
-
return exstyles
|
| 128 |
-
|
| 129 |
-
def detect_and_align_face(self, image) -> np.ndarray:
|
| 130 |
-
image = align_face(filepath=image.name, predictor=self.landmark_model)
|
| 131 |
-
return image
|
| 132 |
-
|
| 133 |
-
@staticmethod
|
| 134 |
-
def denormalize(tensor: torch.Tensor) -> torch.Tensor:
|
| 135 |
-
return torch.clamp((tensor + 1) / 2 * 255, 0, 255).to(torch.uint8)
|
| 136 |
-
|
| 137 |
-
def postprocess(self, tensor: torch.Tensor) -> np.ndarray:
|
| 138 |
-
tensor = self.denormalize(tensor)
|
| 139 |
-
return tensor.cpu().numpy().transpose(1, 2, 0)
|
| 140 |
-
|
| 141 |
-
@torch.inference_mode()
|
| 142 |
-
def reconstruct_face(self,
|
| 143 |
-
image: np.ndarray) -> tuple[np.ndarray, torch.Tensor]:
|
| 144 |
-
image = PIL.Image.fromarray(image)
|
| 145 |
-
input_data = self.transform(image).unsqueeze(0).to(self.device)
|
| 146 |
-
img_rec, instyle = self.encoder(input_data,
|
| 147 |
-
randomize_noise=False,
|
| 148 |
-
return_latents=True,
|
| 149 |
-
z_plus_latent=True,
|
| 150 |
-
return_z_plus_latent=True,
|
| 151 |
-
resize=False)
|
| 152 |
-
img_rec = torch.clamp(img_rec.detach(), -1, 1)
|
| 153 |
-
img_rec = self.postprocess(img_rec[0])
|
| 154 |
-
return img_rec, instyle
|
| 155 |
-
|
| 156 |
-
@torch.inference_mode()
|
| 157 |
-
def generate(self, style_type: str, style_id: int, structure_weight: float,
|
| 158 |
-
color_weight: float, structure_only: bool,
|
| 159 |
-
instyle: torch.Tensor) -> np.ndarray:
|
| 160 |
-
generator = self.generator_dict[style_type]
|
| 161 |
-
exstyles = self.exstyle_dict[style_type]
|
| 162 |
-
|
| 163 |
-
style_id = int(style_id)
|
| 164 |
-
stylename = list(exstyles.keys())[style_id]
|
| 165 |
-
|
| 166 |
-
latent = torch.tensor(exstyles[stylename]).to(self.device)
|
| 167 |
-
if structure_only:
|
| 168 |
-
latent[0, 7:18] = instyle[0, 7:18]
|
| 169 |
-
exstyle = generator.generator.style(
|
| 170 |
-
latent.reshape(latent.shape[0] * latent.shape[1],
|
| 171 |
-
latent.shape[2])).reshape(latent.shape)
|
| 172 |
-
|
| 173 |
-
img_gen, _ = generator([instyle],
|
| 174 |
-
exstyle,
|
| 175 |
-
z_plus_latent=True,
|
| 176 |
-
truncation=0.7,
|
| 177 |
-
truncation_latent=0,
|
| 178 |
-
use_res=True,
|
| 179 |
-
interp_weights=[structure_weight] * 7 +
|
| 180 |
-
[color_weight] * 11)
|
| 181 |
-
img_gen = torch.clamp(img_gen.detach(), -1, 1)
|
| 182 |
-
img_gen = self.postprocess(img_gen[0])
|
| 183 |
-
return img_gen
|
| 184 |
-
|
| 185 |
-
|
| 186 |
def get_style_image_url(style_name: str) -> str:
|
| 187 |
base_url = 'https://raw.githubusercontent.com/williamyang1991/DualStyleGAN/main/doc_images'
|
| 188 |
filenames = {
|
|
@@ -240,7 +83,7 @@ def set_example_weights(example: list) -> list[dict]:
|
|
| 240 |
|
| 241 |
def main():
|
| 242 |
args = parse_args()
|
| 243 |
-
|
| 244 |
|
| 245 |
css = '''
|
| 246 |
h1#title {
|
|
@@ -304,7 +147,8 @@ img#style-image {
|
|
| 304 |
''')
|
| 305 |
with gr.Row():
|
| 306 |
with gr.Column():
|
| 307 |
-
style_type = gr.Radio(
|
|
|
|
| 308 |
text = get_style_image_markdown_text('cartoon')
|
| 309 |
style_image = gr.Markdown(value=text)
|
| 310 |
style_index = gr.Slider(0,
|
|
@@ -366,10 +210,10 @@ img#style-image {
|
|
| 366 |
'<center><img src="https://visitor-badge.glitch.me/badge?page_id=gradio-blocks.dualstylegan" alt="visitor badge"/></center>'
|
| 367 |
)
|
| 368 |
|
| 369 |
-
detect_button.click(fn=
|
| 370 |
inputs=input_image,
|
| 371 |
outputs=aligned_face)
|
| 372 |
-
reconstruct_button.click(fn=
|
| 373 |
inputs=aligned_face,
|
| 374 |
outputs=[reconstructed_face, instyle])
|
| 375 |
style_type.change(fn=update_slider,
|
|
@@ -378,7 +222,7 @@ img#style-image {
|
|
| 378 |
style_type.change(fn=update_style_image,
|
| 379 |
inputs=style_type,
|
| 380 |
outputs=style_image)
|
| 381 |
-
generate_button.click(fn=
|
| 382 |
inputs=[
|
| 383 |
style_type,
|
| 384 |
style_index,
|
|
|
|
| 5 |
import argparse
|
| 6 |
import os
|
| 7 |
import pathlib
|
|
|
|
|
|
|
| 8 |
|
|
|
|
| 9 |
import gradio as gr
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 10 |
|
| 11 |
+
from dualstylegan import Model
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 12 |
|
| 13 |
TOKEN = os.environ['TOKEN']
|
| 14 |
MODEL_REPO = 'hysts/DualStyleGAN'
|
|
|
|
| 26 |
return parser.parse_args()
|
| 27 |
|
| 28 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 29 |
def get_style_image_url(style_name: str) -> str:
|
| 30 |
base_url = 'https://raw.githubusercontent.com/williamyang1991/DualStyleGAN/main/doc_images'
|
| 31 |
filenames = {
|
|
|
|
| 83 |
|
| 84 |
def main():
|
| 85 |
args = parse_args()
|
| 86 |
+
model = Model(device=args.device)
|
| 87 |
|
| 88 |
css = '''
|
| 89 |
h1#title {
|
|
|
|
| 147 |
''')
|
| 148 |
with gr.Row():
|
| 149 |
with gr.Column():
|
| 150 |
+
style_type = gr.Radio(model.style_types,
|
| 151 |
+
label='Style Type')
|
| 152 |
text = get_style_image_markdown_text('cartoon')
|
| 153 |
style_image = gr.Markdown(value=text)
|
| 154 |
style_index = gr.Slider(0,
|
|
|
|
| 210 |
'<center><img src="https://visitor-badge.glitch.me/badge?page_id=gradio-blocks.dualstylegan" alt="visitor badge"/></center>'
|
| 211 |
)
|
| 212 |
|
| 213 |
+
detect_button.click(fn=model.detect_and_align_face,
|
| 214 |
inputs=input_image,
|
| 215 |
outputs=aligned_face)
|
| 216 |
+
reconstruct_button.click(fn=model.reconstruct_face,
|
| 217 |
inputs=aligned_face,
|
| 218 |
outputs=[reconstructed_face, instyle])
|
| 219 |
style_type.change(fn=update_slider,
|
|
|
|
| 222 |
style_type.change(fn=update_style_image,
|
| 223 |
inputs=style_type,
|
| 224 |
outputs=style_image)
|
| 225 |
+
generate_button.click(fn=model.generate,
|
| 226 |
inputs=[
|
| 227 |
style_type,
|
| 228 |
style_index,
|
dualstylegan.py
ADDED
|
@@ -0,0 +1,167 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
from __future__ import annotations
|
| 2 |
+
|
| 3 |
+
import argparse
|
| 4 |
+
import os
|
| 5 |
+
import sys
|
| 6 |
+
from typing import Callable, Union
|
| 7 |
+
|
| 8 |
+
import dlib
|
| 9 |
+
import huggingface_hub
|
| 10 |
+
import numpy as np
|
| 11 |
+
import PIL.Image
|
| 12 |
+
import torch
|
| 13 |
+
import torch.nn as nn
|
| 14 |
+
import torchvision.transforms as T
|
| 15 |
+
|
| 16 |
+
if os.environ.get('SYSTEM') == 'spaces':
|
| 17 |
+
os.system("sed -i '10,17d' DualStyleGAN/model/stylegan/op/fused_act.py")
|
| 18 |
+
os.system("sed -i '10,17d' DualStyleGAN/model/stylegan/op/upfirdn2d.py")
|
| 19 |
+
|
| 20 |
+
sys.path.insert(0, 'DualStyleGAN')
|
| 21 |
+
|
| 22 |
+
from model.dualstylegan import DualStyleGAN
|
| 23 |
+
from model.encoder.align_all_parallel import align_face
|
| 24 |
+
from model.encoder.psp import pSp
|
| 25 |
+
|
| 26 |
+
TOKEN = os.environ['TOKEN']
|
| 27 |
+
MODEL_REPO = 'hysts/DualStyleGAN'
|
| 28 |
+
|
| 29 |
+
|
| 30 |
+
class Model:
|
| 31 |
+
|
| 32 |
+
def __init__(self, device: Union[torch.device, str]):
|
| 33 |
+
self.device = torch.device(device)
|
| 34 |
+
self.landmark_model = self._create_dlib_landmark_model()
|
| 35 |
+
self.encoder = self._load_encoder()
|
| 36 |
+
self.transform = self._create_transform()
|
| 37 |
+
|
| 38 |
+
self.style_types = [
|
| 39 |
+
'cartoon',
|
| 40 |
+
'caricature',
|
| 41 |
+
'anime',
|
| 42 |
+
'arcane',
|
| 43 |
+
'comic',
|
| 44 |
+
'pixar',
|
| 45 |
+
'slamdunk',
|
| 46 |
+
]
|
| 47 |
+
self.generator_dict = {
|
| 48 |
+
style_type: self._load_generator(style_type)
|
| 49 |
+
for style_type in self.style_types
|
| 50 |
+
}
|
| 51 |
+
self.exstyle_dict = {
|
| 52 |
+
style_type: self._load_exstylecode(style_type)
|
| 53 |
+
for style_type in self.style_types
|
| 54 |
+
}
|
| 55 |
+
|
| 56 |
+
@staticmethod
|
| 57 |
+
def _create_dlib_landmark_model():
|
| 58 |
+
path = huggingface_hub.hf_hub_download(
|
| 59 |
+
'hysts/dlib_face_landmark_model',
|
| 60 |
+
'shape_predictor_68_face_landmarks.dat',
|
| 61 |
+
use_auth_token=TOKEN)
|
| 62 |
+
return dlib.shape_predictor(path)
|
| 63 |
+
|
| 64 |
+
def _load_encoder(self) -> nn.Module:
|
| 65 |
+
ckpt_path = huggingface_hub.hf_hub_download(MODEL_REPO,
|
| 66 |
+
'models/encoder.pt',
|
| 67 |
+
use_auth_token=TOKEN)
|
| 68 |
+
ckpt = torch.load(ckpt_path, map_location='cpu')
|
| 69 |
+
opts = ckpt['opts']
|
| 70 |
+
opts['device'] = self.device.type
|
| 71 |
+
opts['checkpoint_path'] = ckpt_path
|
| 72 |
+
opts = argparse.Namespace(**opts)
|
| 73 |
+
model = pSp(opts)
|
| 74 |
+
model.to(self.device)
|
| 75 |
+
model.eval()
|
| 76 |
+
return model
|
| 77 |
+
|
| 78 |
+
@staticmethod
|
| 79 |
+
def _create_transform() -> Callable:
|
| 80 |
+
transform = T.Compose([
|
| 81 |
+
T.Resize(256),
|
| 82 |
+
T.CenterCrop(256),
|
| 83 |
+
T.ToTensor(),
|
| 84 |
+
T.Normalize([0.5, 0.5, 0.5], [0.5, 0.5, 0.5]),
|
| 85 |
+
])
|
| 86 |
+
return transform
|
| 87 |
+
|
| 88 |
+
def _load_generator(self, style_type: str) -> nn.Module:
|
| 89 |
+
model = DualStyleGAN(1024, 512, 8, 2, res_index=6)
|
| 90 |
+
ckpt_path = huggingface_hub.hf_hub_download(
|
| 91 |
+
MODEL_REPO,
|
| 92 |
+
f'models/{style_type}/generator.pt',
|
| 93 |
+
use_auth_token=TOKEN)
|
| 94 |
+
ckpt = torch.load(ckpt_path, map_location='cpu')
|
| 95 |
+
model.load_state_dict(ckpt['g_ema'])
|
| 96 |
+
model.to(self.device)
|
| 97 |
+
model.eval()
|
| 98 |
+
return model
|
| 99 |
+
|
| 100 |
+
@staticmethod
|
| 101 |
+
def _load_exstylecode(style_type: str) -> dict[str, np.ndarray]:
|
| 102 |
+
if style_type in ['cartoon', 'caricature', 'anime']:
|
| 103 |
+
filename = 'refined_exstyle_code.npy'
|
| 104 |
+
else:
|
| 105 |
+
filename = 'exstyle_code.npy'
|
| 106 |
+
path = huggingface_hub.hf_hub_download(
|
| 107 |
+
MODEL_REPO,
|
| 108 |
+
f'models/{style_type}/{filename}',
|
| 109 |
+
use_auth_token=TOKEN)
|
| 110 |
+
exstyles = np.load(path, allow_pickle=True).item()
|
| 111 |
+
return exstyles
|
| 112 |
+
|
| 113 |
+
def detect_and_align_face(self, image) -> np.ndarray:
|
| 114 |
+
image = align_face(filepath=image.name, predictor=self.landmark_model)
|
| 115 |
+
return image
|
| 116 |
+
|
| 117 |
+
@staticmethod
|
| 118 |
+
def denormalize(tensor: torch.Tensor) -> torch.Tensor:
|
| 119 |
+
return torch.clamp((tensor + 1) / 2 * 255, 0, 255).to(torch.uint8)
|
| 120 |
+
|
| 121 |
+
def postprocess(self, tensor: torch.Tensor) -> np.ndarray:
|
| 122 |
+
tensor = self.denormalize(tensor)
|
| 123 |
+
return tensor.cpu().numpy().transpose(1, 2, 0)
|
| 124 |
+
|
| 125 |
+
@torch.inference_mode()
|
| 126 |
+
def reconstruct_face(self,
|
| 127 |
+
image: np.ndarray) -> tuple[np.ndarray, torch.Tensor]:
|
| 128 |
+
image = PIL.Image.fromarray(image)
|
| 129 |
+
input_data = self.transform(image).unsqueeze(0).to(self.device)
|
| 130 |
+
img_rec, instyle = self.encoder(input_data,
|
| 131 |
+
randomize_noise=False,
|
| 132 |
+
return_latents=True,
|
| 133 |
+
z_plus_latent=True,
|
| 134 |
+
return_z_plus_latent=True,
|
| 135 |
+
resize=False)
|
| 136 |
+
img_rec = torch.clamp(img_rec.detach(), -1, 1)
|
| 137 |
+
img_rec = self.postprocess(img_rec[0])
|
| 138 |
+
return img_rec, instyle
|
| 139 |
+
|
| 140 |
+
@torch.inference_mode()
|
| 141 |
+
def generate(self, style_type: str, style_id: int, structure_weight: float,
|
| 142 |
+
color_weight: float, structure_only: bool,
|
| 143 |
+
instyle: torch.Tensor) -> np.ndarray:
|
| 144 |
+
generator = self.generator_dict[style_type]
|
| 145 |
+
exstyles = self.exstyle_dict[style_type]
|
| 146 |
+
|
| 147 |
+
style_id = int(style_id)
|
| 148 |
+
stylename = list(exstyles.keys())[style_id]
|
| 149 |
+
|
| 150 |
+
latent = torch.tensor(exstyles[stylename]).to(self.device)
|
| 151 |
+
if structure_only:
|
| 152 |
+
latent[0, 7:18] = instyle[0, 7:18]
|
| 153 |
+
exstyle = generator.generator.style(
|
| 154 |
+
latent.reshape(latent.shape[0] * latent.shape[1],
|
| 155 |
+
latent.shape[2])).reshape(latent.shape)
|
| 156 |
+
|
| 157 |
+
img_gen, _ = generator([instyle],
|
| 158 |
+
exstyle,
|
| 159 |
+
z_plus_latent=True,
|
| 160 |
+
truncation=0.7,
|
| 161 |
+
truncation_latent=0,
|
| 162 |
+
use_res=True,
|
| 163 |
+
interp_weights=[structure_weight] * 7 +
|
| 164 |
+
[color_weight] * 11)
|
| 165 |
+
img_gen = torch.clamp(img_gen.detach(), -1, 1)
|
| 166 |
+
img_gen = self.postprocess(img_gen[0])
|
| 167 |
+
return img_gen
|