Ahsen Khaliq
commited on
Commit
·
56a97f7
1
Parent(s):
0b2237b
Create app.py
Browse files
app.py
ADDED
|
@@ -0,0 +1,223 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import os
|
| 2 |
+
os.system("git clone https://github.com/bryandlee/animegan2-pytorch")
|
| 3 |
+
|
| 4 |
+
os.system("gdown https://drive.google.com/uc?id=1WK5Mdt6mwlcsqCZMHkCUSDJxN1UyFi0-")
|
| 5 |
+
os.system("gdown https://drive.google.com/uc?id=18H3iK09_d54qEDoWIc82SyWB2xun4gjU")
|
| 6 |
+
|
| 7 |
+
import sys
|
| 8 |
+
sys.path.append("animegan2-pytorch")
|
| 9 |
+
|
| 10 |
+
import torch
|
| 11 |
+
torch.set_grad_enabled(False)
|
| 12 |
+
|
| 13 |
+
from model import Generator
|
| 14 |
+
|
| 15 |
+
device = "cpu"
|
| 16 |
+
|
| 17 |
+
model = Generator().eval().to(device)
|
| 18 |
+
model.load_state_dict(torch.load("face_paint_512_v2_0.pt"))
|
| 19 |
+
|
| 20 |
+
from PIL import Image
|
| 21 |
+
from torchvision.transforms.functional import to_tensor, to_pil_image
|
| 22 |
+
|
| 23 |
+
def face2paint(
|
| 24 |
+
img: Image.Image,
|
| 25 |
+
size: int,
|
| 26 |
+
side_by_side: bool = True,
|
| 27 |
+
) -> Image.Image:
|
| 28 |
+
|
| 29 |
+
w, h = img.size
|
| 30 |
+
s = min(w, h)
|
| 31 |
+
img = img.crop(((w - s) // 2, (h - s) // 2, (w + s) // 2, (h + s) // 2))
|
| 32 |
+
img = img.resize((size, size), Image.LANCZOS)
|
| 33 |
+
|
| 34 |
+
input = to_tensor(img).unsqueeze(0) * 2 - 1
|
| 35 |
+
output = model(input.to(device)).cpu()[0]
|
| 36 |
+
|
| 37 |
+
if side_by_side:
|
| 38 |
+
output = torch.cat([input[0], output], dim=2)
|
| 39 |
+
|
| 40 |
+
output = (output * 0.5 + 0.5).clip(0, 1)
|
| 41 |
+
|
| 42 |
+
return to_pil_image(output)
|
| 43 |
+
|
| 44 |
+
|
| 45 |
+
#@title Face Detector & FFHQ-style Alignment
|
| 46 |
+
|
| 47 |
+
# https://github.com/woctezuma/stylegan2-projecting-images
|
| 48 |
+
|
| 49 |
+
import os
|
| 50 |
+
import dlib
|
| 51 |
+
import collections
|
| 52 |
+
from typing import Union, List
|
| 53 |
+
import numpy as np
|
| 54 |
+
from PIL import Image
|
| 55 |
+
import matplotlib.pyplot as plt
|
| 56 |
+
|
| 57 |
+
|
| 58 |
+
def get_dlib_face_detector(predictor_path: str = "shape_predictor_68_face_landmarks.dat"):
|
| 59 |
+
|
| 60 |
+
if not os.path.isfile(predictor_path):
|
| 61 |
+
model_file = "shape_predictor_68_face_landmarks.dat.bz2"
|
| 62 |
+
os.system(f"wget http://dlib.net/files/{model_file}")
|
| 63 |
+
os.system(f"bzip2 -dk {model_file}")
|
| 64 |
+
|
| 65 |
+
detector = dlib.get_frontal_face_detector()
|
| 66 |
+
shape_predictor = dlib.shape_predictor(predictor_path)
|
| 67 |
+
|
| 68 |
+
def detect_face_landmarks(img: Union[Image.Image, np.ndarray]):
|
| 69 |
+
if isinstance(img, Image.Image):
|
| 70 |
+
img = np.array(img)
|
| 71 |
+
faces = []
|
| 72 |
+
dets = detector(img)
|
| 73 |
+
for d in dets:
|
| 74 |
+
shape = shape_predictor(img, d)
|
| 75 |
+
faces.append(np.array([[v.x, v.y] for v in shape.parts()]))
|
| 76 |
+
return faces
|
| 77 |
+
|
| 78 |
+
return detect_face_landmarks
|
| 79 |
+
|
| 80 |
+
|
| 81 |
+
def display_facial_landmarks(
|
| 82 |
+
img: Image,
|
| 83 |
+
landmarks: List[np.ndarray],
|
| 84 |
+
fig_size=[15, 15]
|
| 85 |
+
):
|
| 86 |
+
plot_style = dict(
|
| 87 |
+
marker='o',
|
| 88 |
+
markersize=4,
|
| 89 |
+
linestyle='-',
|
| 90 |
+
lw=2
|
| 91 |
+
)
|
| 92 |
+
pred_type = collections.namedtuple('prediction_type', ['slice', 'color'])
|
| 93 |
+
pred_types = {
|
| 94 |
+
'face': pred_type(slice(0, 17), (0.682, 0.780, 0.909, 0.5)),
|
| 95 |
+
'eyebrow1': pred_type(slice(17, 22), (1.0, 0.498, 0.055, 0.4)),
|
| 96 |
+
'eyebrow2': pred_type(slice(22, 27), (1.0, 0.498, 0.055, 0.4)),
|
| 97 |
+
'nose': pred_type(slice(27, 31), (0.345, 0.239, 0.443, 0.4)),
|
| 98 |
+
'nostril': pred_type(slice(31, 36), (0.345, 0.239, 0.443, 0.4)),
|
| 99 |
+
'eye1': pred_type(slice(36, 42), (0.596, 0.875, 0.541, 0.3)),
|
| 100 |
+
'eye2': pred_type(slice(42, 48), (0.596, 0.875, 0.541, 0.3)),
|
| 101 |
+
'lips': pred_type(slice(48, 60), (0.596, 0.875, 0.541, 0.3)),
|
| 102 |
+
'teeth': pred_type(slice(60, 68), (0.596, 0.875, 0.541, 0.4))
|
| 103 |
+
}
|
| 104 |
+
|
| 105 |
+
fig = plt.figure(figsize=fig_size)
|
| 106 |
+
ax = fig.add_subplot(1, 1, 1)
|
| 107 |
+
ax.imshow(img)
|
| 108 |
+
ax.axis('off')
|
| 109 |
+
|
| 110 |
+
for face in landmarks:
|
| 111 |
+
for pred_type in pred_types.values():
|
| 112 |
+
ax.plot(
|
| 113 |
+
face[pred_type.slice, 0],
|
| 114 |
+
face[pred_type.slice, 1],
|
| 115 |
+
color=pred_type.color, **plot_style
|
| 116 |
+
)
|
| 117 |
+
plt.show()
|
| 118 |
+
|
| 119 |
+
import PIL.Image
|
| 120 |
+
import PIL.ImageFile
|
| 121 |
+
import numpy as np
|
| 122 |
+
import scipy.ndimage
|
| 123 |
+
|
| 124 |
+
|
| 125 |
+
def align_and_crop_face(
|
| 126 |
+
img: Image.Image,
|
| 127 |
+
landmarks: np.ndarray,
|
| 128 |
+
expand: float = 1.0,
|
| 129 |
+
output_size: int = 1024,
|
| 130 |
+
transform_size: int = 4096,
|
| 131 |
+
enable_padding: bool = True,
|
| 132 |
+
):
|
| 133 |
+
# Parse landmarks.
|
| 134 |
+
# pylint: disable=unused-variable
|
| 135 |
+
lm = landmarks
|
| 136 |
+
lm_chin = lm[0 : 17] # left-right
|
| 137 |
+
lm_eyebrow_left = lm[17 : 22] # left-right
|
| 138 |
+
lm_eyebrow_right = lm[22 : 27] # left-right
|
| 139 |
+
lm_nose = lm[27 : 31] # top-down
|
| 140 |
+
lm_nostrils = lm[31 : 36] # top-down
|
| 141 |
+
lm_eye_left = lm[36 : 42] # left-clockwise
|
| 142 |
+
lm_eye_right = lm[42 : 48] # left-clockwise
|
| 143 |
+
lm_mouth_outer = lm[48 : 60] # left-clockwise
|
| 144 |
+
lm_mouth_inner = lm[60 : 68] # left-clockwise
|
| 145 |
+
|
| 146 |
+
# Calculate auxiliary vectors.
|
| 147 |
+
eye_left = np.mean(lm_eye_left, axis=0)
|
| 148 |
+
eye_right = np.mean(lm_eye_right, axis=0)
|
| 149 |
+
eye_avg = (eye_left + eye_right) * 0.5
|
| 150 |
+
eye_to_eye = eye_right - eye_left
|
| 151 |
+
mouth_left = lm_mouth_outer[0]
|
| 152 |
+
mouth_right = lm_mouth_outer[6]
|
| 153 |
+
mouth_avg = (mouth_left + mouth_right) * 0.5
|
| 154 |
+
eye_to_mouth = mouth_avg - eye_avg
|
| 155 |
+
|
| 156 |
+
# Choose oriented crop rectangle.
|
| 157 |
+
x = eye_to_eye - np.flipud(eye_to_mouth) * [-1, 1]
|
| 158 |
+
x /= np.hypot(*x)
|
| 159 |
+
x *= max(np.hypot(*eye_to_eye) * 2.0, np.hypot(*eye_to_mouth) * 1.8)
|
| 160 |
+
x *= expand
|
| 161 |
+
y = np.flipud(x) * [-1, 1]
|
| 162 |
+
c = eye_avg + eye_to_mouth * 0.1
|
| 163 |
+
quad = np.stack([c - x - y, c - x + y, c + x + y, c + x - y])
|
| 164 |
+
qsize = np.hypot(*x) * 2
|
| 165 |
+
|
| 166 |
+
# Shrink.
|
| 167 |
+
shrink = int(np.floor(qsize / output_size * 0.5))
|
| 168 |
+
if shrink > 1:
|
| 169 |
+
rsize = (int(np.rint(float(img.size[0]) / shrink)), int(np.rint(float(img.size[1]) / shrink)))
|
| 170 |
+
img = img.resize(rsize, PIL.Image.ANTIALIAS)
|
| 171 |
+
quad /= shrink
|
| 172 |
+
qsize /= shrink
|
| 173 |
+
|
| 174 |
+
# Crop.
|
| 175 |
+
border = max(int(np.rint(qsize * 0.1)), 3)
|
| 176 |
+
crop = (int(np.floor(min(quad[:,0]))), int(np.floor(min(quad[:,1]))), int(np.ceil(max(quad[:,0]))), int(np.ceil(max(quad[:,1]))))
|
| 177 |
+
crop = (max(crop[0] - border, 0), max(crop[1] - border, 0), min(crop[2] + border, img.size[0]), min(crop[3] + border, img.size[1]))
|
| 178 |
+
if crop[2] - crop[0] < img.size[0] or crop[3] - crop[1] < img.size[1]:
|
| 179 |
+
img = img.crop(crop)
|
| 180 |
+
quad -= crop[0:2]
|
| 181 |
+
|
| 182 |
+
# Pad.
|
| 183 |
+
pad = (int(np.floor(min(quad[:,0]))), int(np.floor(min(quad[:,1]))), int(np.ceil(max(quad[:,0]))), int(np.ceil(max(quad[:,1]))))
|
| 184 |
+
pad = (max(-pad[0] + border, 0), max(-pad[1] + border, 0), max(pad[2] - img.size[0] + border, 0), max(pad[3] - img.size[1] + border, 0))
|
| 185 |
+
if enable_padding and max(pad) > border - 4:
|
| 186 |
+
pad = np.maximum(pad, int(np.rint(qsize * 0.3)))
|
| 187 |
+
img = np.pad(np.float32(img), ((pad[1], pad[3]), (pad[0], pad[2]), (0, 0)), 'reflect')
|
| 188 |
+
h, w, _ = img.shape
|
| 189 |
+
y, x, _ = np.ogrid[:h, :w, :1]
|
| 190 |
+
mask = np.maximum(1.0 - np.minimum(np.float32(x) / pad[0], np.float32(w-1-x) / pad[2]), 1.0 - np.minimum(np.float32(y) / pad[1], np.float32(h-1-y) / pad[3]))
|
| 191 |
+
blur = qsize * 0.02
|
| 192 |
+
img += (scipy.ndimage.gaussian_filter(img, [blur, blur, 0]) - img) * np.clip(mask * 3.0 + 1.0, 0.0, 1.0)
|
| 193 |
+
img += (np.median(img, axis=(0,1)) - img) * np.clip(mask, 0.0, 1.0)
|
| 194 |
+
img = PIL.Image.fromarray(np.uint8(np.clip(np.rint(img), 0, 255)), 'RGB')
|
| 195 |
+
quad += pad[:2]
|
| 196 |
+
|
| 197 |
+
# Transform.
|
| 198 |
+
img = img.transform((transform_size, transform_size), PIL.Image.QUAD, (quad + 0.5).flatten(), PIL.Image.BILINEAR)
|
| 199 |
+
if output_size < transform_size:
|
| 200 |
+
img = img.resize((output_size, output_size), PIL.Image.ANTIALIAS)
|
| 201 |
+
|
| 202 |
+
return img
|
| 203 |
+
|
| 204 |
+
|
| 205 |
+
import requests
|
| 206 |
+
|
| 207 |
+
def inference(image):
|
| 208 |
+
img = image
|
| 209 |
+
face_detector = get_dlib_face_detector()
|
| 210 |
+
landmarks = face_detector(img)
|
| 211 |
+
|
| 212 |
+
display_facial_landmarks(img, landmarks, fig_size=[5, 5])
|
| 213 |
+
|
| 214 |
+
for landmark in landmarks:
|
| 215 |
+
face = align_and_crop_face(img, landmark, expand=1.3)
|
| 216 |
+
out = face2paint(face, 512)
|
| 217 |
+
|
| 218 |
+
return out
|
| 219 |
+
|
| 220 |
+
|
| 221 |
+
|
| 222 |
+
iface = gr.Interface(inference, "image", "image")
|
| 223 |
+
iface.launch()
|