Spaces:
Running
Running
Update face alignment
Browse files- app.py +9 -67
- requirements.txt +1 -0
- utils.py +114 -0
app.py
CHANGED
@@ -4,17 +4,22 @@ os.system("pip -qq install facenet_pytorch")
|
|
4 |
from facenet_pytorch import MTCNN
|
5 |
from torchvision import transforms
|
6 |
import torch, PIL
|
7 |
-
from tqdm.notebook import tqdm
|
8 |
import gradio as gr
|
9 |
import torch
|
|
|
10 |
|
11 |
device = "cuda:0" if torch.cuda.is_available() else "cpu"
|
12 |
|
13 |
image_size = 512
|
|
|
14 |
|
15 |
means = [0.5, 0.5, 0.5]
|
16 |
stds = [0.5, 0.5, 0.5]
|
17 |
|
|
|
|
|
|
|
|
|
18 |
model_path = hf_hub_download(repo_id="jjeamin/ArcaneStyleTransfer", filename="pytorch_model.bin")
|
19 |
|
20 |
if 'cuda' in device:
|
@@ -25,76 +30,14 @@ else:
|
|
25 |
style_transfer = torch.jit.load(model_path).eval().cpu()
|
26 |
t_stds = torch.tensor(stds).cpu()[:,None,None]
|
27 |
t_means = torch.tensor(means).cpu()[:,None,None]
|
28 |
-
|
29 |
-
mtcnn = MTCNN(image_size=image_size, margin=80)
|
30 |
-
|
31 |
-
def detect(img):
|
32 |
-
|
33 |
-
# Detect faces
|
34 |
-
batch_boxes, batch_probs, batch_points = mtcnn.detect(img, landmarks=True)
|
35 |
-
# Select faces
|
36 |
-
if not mtcnn.keep_all:
|
37 |
-
batch_boxes, batch_probs, batch_points = mtcnn.select_boxes(
|
38 |
-
batch_boxes, batch_probs, batch_points, img, method=mtcnn.selection_method
|
39 |
-
)
|
40 |
-
|
41 |
-
return batch_boxes, batch_points
|
42 |
-
|
43 |
-
def makeEven(_x):
|
44 |
-
return _x if (_x % 2 == 0) else _x+1
|
45 |
-
|
46 |
-
def scale(boxes, _img, max_res=1_500_000, target_face=256, fixed_ratio=0, max_upscale=2, VERBOSE=False):
|
47 |
-
|
48 |
-
x, y = _img.size
|
49 |
-
|
50 |
-
ratio = 2 #initial ratio
|
51 |
-
|
52 |
-
#scale to desired face size
|
53 |
-
if (boxes is not None):
|
54 |
-
if len(boxes)>0:
|
55 |
-
ratio = target_face/max(boxes[0][2:]-boxes[0][:2]);
|
56 |
-
ratio = min(ratio, max_upscale)
|
57 |
-
if VERBOSE: print('up by', ratio)
|
58 |
|
59 |
-
if fixed_ratio>0:
|
60 |
-
if VERBOSE: print('fixed ratio')
|
61 |
-
ratio = fixed_ratio
|
62 |
-
|
63 |
-
x*=ratio
|
64 |
-
y*=ratio
|
65 |
-
|
66 |
-
#downscale to fit into max res
|
67 |
-
res = x*y
|
68 |
-
if res > max_res:
|
69 |
-
ratio = pow(res/max_res,1/2);
|
70 |
-
if VERBOSE: print(ratio)
|
71 |
-
x=int(x/ratio)
|
72 |
-
y=int(y/ratio)
|
73 |
-
|
74 |
-
#make dimensions even, because usually NNs fail on uneven dimensions due skip connection size mismatch
|
75 |
-
x = makeEven(int(x))
|
76 |
-
y = makeEven(int(y))
|
77 |
-
|
78 |
-
size = (x, y)
|
79 |
-
|
80 |
-
return _img.resize(size)
|
81 |
-
|
82 |
-
def scale_by_face_size(_img, max_res=1_500_000, target_face=256, fix_ratio=0, max_upscale=2, VERBOSE=False):
|
83 |
-
boxes = None
|
84 |
-
boxes, _ = detect(_img)
|
85 |
-
if VERBOSE: print('boxes',boxes)
|
86 |
-
img_resized = scale(boxes, _img, max_res, target_face, fix_ratio, max_upscale, VERBOSE)
|
87 |
-
return img_resized
|
88 |
-
|
89 |
-
|
90 |
-
img_transforms = transforms.Compose([
|
91 |
-
transforms.ToTensor(),
|
92 |
-
transforms.Normalize(means, stds)])
|
93 |
-
|
94 |
def tensor2im(var):
|
95 |
return var.mul(t_stds).add(t_means).mul(255.).clamp(0,255).permute(1,2,0)
|
96 |
|
97 |
def proc_pil_img(input_image):
|
|
|
|
|
|
|
98 |
if 'cuda' in device:
|
99 |
transformed_image = img_transforms(input_image)[None,...].cuda().half()
|
100 |
else:
|
@@ -108,7 +51,6 @@ def proc_pil_img(input_image):
|
|
108 |
return output_image
|
109 |
|
110 |
def process(im):
|
111 |
-
im = scale_by_face_size(im, target_face=image_size, max_res=1_500_000, max_upscale=1)
|
112 |
res = proc_pil_img(im)
|
113 |
return res
|
114 |
|
|
|
4 |
from facenet_pytorch import MTCNN
|
5 |
from torchvision import transforms
|
6 |
import torch, PIL
|
|
|
7 |
import gradio as gr
|
8 |
import torch
|
9 |
+
from utils import align_face
|
10 |
|
11 |
device = "cuda:0" if torch.cuda.is_available() else "cpu"
|
12 |
|
13 |
image_size = 512
|
14 |
+
transform_size = 1024
|
15 |
|
16 |
means = [0.5, 0.5, 0.5]
|
17 |
stds = [0.5, 0.5, 0.5]
|
18 |
|
19 |
+
img_transforms = transforms.Compose([
|
20 |
+
transforms.ToTensor(),
|
21 |
+
transforms.Normalize(means, stds)])
|
22 |
+
|
23 |
model_path = hf_hub_download(repo_id="jjeamin/ArcaneStyleTransfer", filename="pytorch_model.bin")
|
24 |
|
25 |
if 'cuda' in device:
|
|
|
30 |
style_transfer = torch.jit.load(model_path).eval().cpu()
|
31 |
t_stds = torch.tensor(stds).cpu()[:,None,None]
|
32 |
t_means = torch.tensor(means).cpu()[:,None,None]
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
33 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
34 |
def tensor2im(var):
|
35 |
return var.mul(t_stds).add(t_means).mul(255.).clamp(0,255).permute(1,2,0)
|
36 |
|
37 |
def proc_pil_img(input_image):
|
38 |
+
# input_image = PIL
|
39 |
+
input_image = align_face(input_image, output_size=image_size, transform_size=transform_size)
|
40 |
+
|
41 |
if 'cuda' in device:
|
42 |
transformed_image = img_transforms(input_image)[None,...].cuda().half()
|
43 |
else:
|
|
|
51 |
return output_image
|
52 |
|
53 |
def process(im):
|
|
|
54 |
res = proc_pil_img(im)
|
55 |
return res
|
56 |
|
requirements.txt
CHANGED
@@ -7,3 +7,4 @@ scipy
|
|
7 |
cmake
|
8 |
onnxruntime-gpu
|
9 |
opencv-python-headless
|
|
|
|
7 |
cmake
|
8 |
onnxruntime-gpu
|
9 |
opencv-python-headless
|
10 |
+
dlib
|
utils.py
ADDED
@@ -0,0 +1,114 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import dlib
|
2 |
+
import numpy as np
|
3 |
+
import scipy
|
4 |
+
from PIL import Image
|
5 |
+
from huggingface_hub import hf_hub_download
|
6 |
+
|
7 |
+
shape_predictor_path = hf_hub_download(repo_id="jjeamin/ArcaneStyleTransfer", filename="shape_predictor_68_face_landmarks.dat")
|
8 |
+
|
9 |
+
def get_landmark(img, predictor):
|
10 |
+
"""get landmark with dlib
|
11 |
+
:return: np.array shape=(68, 2)
|
12 |
+
"""
|
13 |
+
detector = dlib.get_frontal_face_detector()
|
14 |
+
|
15 |
+
dets = detector(img, 1)
|
16 |
+
assert len(dets) > 0, "Face not detected, try another face image"
|
17 |
+
|
18 |
+
for k, d in enumerate(dets):
|
19 |
+
shape = predictor(img, d)
|
20 |
+
|
21 |
+
t = list(shape.parts())
|
22 |
+
a = []
|
23 |
+
for tt in t:
|
24 |
+
a.append([tt.x, tt.y])
|
25 |
+
lm = np.array(a)
|
26 |
+
return lm
|
27 |
+
|
28 |
+
|
29 |
+
def align_face(img, output_size=512, transform_size=1024, enable_padding=True):
|
30 |
+
|
31 |
+
"""
|
32 |
+
:param filepath: str
|
33 |
+
:return: PIL Image
|
34 |
+
"""
|
35 |
+
np_img = np.array(img)
|
36 |
+
predictor = dlib.shape_predictor(shape_predictor_path)
|
37 |
+
lm = get_landmark(np_img, predictor)
|
38 |
+
|
39 |
+
lm_chin = lm[0: 17] # left-right
|
40 |
+
lm_eyebrow_left = lm[17: 22] # left-right
|
41 |
+
lm_eyebrow_right = lm[22: 27] # left-right
|
42 |
+
lm_nose = lm[27: 31] # top-down
|
43 |
+
lm_nostrils = lm[31: 36] # top-down
|
44 |
+
lm_eye_left = lm[36: 42] # left-clockwise
|
45 |
+
lm_eye_right = lm[42: 48] # left-clockwise
|
46 |
+
lm_mouth_outer = lm[48: 60] # left-clockwise
|
47 |
+
lm_mouth_inner = lm[60: 68] # left-clockwise
|
48 |
+
|
49 |
+
# Calculate auxiliary vectors.
|
50 |
+
eye_left = np.mean(lm_eye_left, axis=0)
|
51 |
+
eye_right = np.mean(lm_eye_right, axis=0)
|
52 |
+
eye_avg = (eye_left + eye_right) * 0.5
|
53 |
+
eye_to_eye = eye_right - eye_left
|
54 |
+
mouth_left = lm_mouth_outer[0]
|
55 |
+
mouth_right = lm_mouth_outer[6]
|
56 |
+
mouth_avg = (mouth_left + mouth_right) * 0.5
|
57 |
+
eye_to_mouth = mouth_avg - eye_avg
|
58 |
+
|
59 |
+
# Choose oriented crop rectangle.
|
60 |
+
x = eye_to_eye - np.flipud(eye_to_mouth) * [-1, 1]
|
61 |
+
x /= np.hypot(*x)
|
62 |
+
x *= max(np.hypot(*eye_to_eye) * 2.0, np.hypot(*eye_to_mouth) * 1.8)
|
63 |
+
y = np.flipud(x) * [-1, 1]
|
64 |
+
c = eye_avg + eye_to_mouth * 0.1
|
65 |
+
quad = np.stack([c - x - y, c - x + y, c + x + y, c + x - y])
|
66 |
+
qsize = np.hypot(*x) * 2
|
67 |
+
|
68 |
+
# read image
|
69 |
+
transform_size = output_size
|
70 |
+
enable_padding = True
|
71 |
+
|
72 |
+
# Shrink.
|
73 |
+
shrink = int(np.floor(qsize / output_size * 0.5))
|
74 |
+
if shrink > 1:
|
75 |
+
rsize = (int(np.rint(float(img.size[0]) / shrink)), int(np.rint(float(img.size[1]) / shrink)))
|
76 |
+
img = img.resize(rsize, Image.ANTIALIAS)
|
77 |
+
quad /= shrink
|
78 |
+
qsize /= shrink
|
79 |
+
|
80 |
+
# Crop.
|
81 |
+
border = max(int(np.rint(qsize * 0.1)), 3)
|
82 |
+
crop = (int(np.floor(min(quad[:, 0]))), int(np.floor(min(quad[:, 1]))), int(np.ceil(max(quad[:, 0]))),
|
83 |
+
int(np.ceil(max(quad[:, 1]))))
|
84 |
+
crop = (max(crop[0] - border, 0), max(crop[1] - border, 0), min(crop[2] + border, img.size[0]),
|
85 |
+
min(crop[3] + border, img.size[1]))
|
86 |
+
if crop[2] - crop[0] < img.size[0] or crop[3] - crop[1] < img.size[1]:
|
87 |
+
img = img.crop(crop)
|
88 |
+
quad -= crop[0:2]
|
89 |
+
|
90 |
+
# Pad.
|
91 |
+
pad = (int(np.floor(min(quad[:, 0]))), int(np.floor(min(quad[:, 1]))), int(np.ceil(max(quad[:, 0]))),
|
92 |
+
int(np.ceil(max(quad[:, 1]))))
|
93 |
+
pad = (max(-pad[0] + border, 0), max(-pad[1] + border, 0), max(pad[2] - img.size[0] + border, 0),
|
94 |
+
max(pad[3] - img.size[1] + border, 0))
|
95 |
+
if enable_padding and max(pad) > border - 4:
|
96 |
+
pad = np.maximum(pad, int(np.rint(qsize * 0.3)))
|
97 |
+
img = np.pad(np.float32(img), ((pad[1], pad[3]), (pad[0], pad[2]), (0, 0)), 'reflect')
|
98 |
+
h, w, _ = img.shape
|
99 |
+
y, x, _ = np.ogrid[:h, :w, :1]
|
100 |
+
mask = np.maximum(1.0 - np.minimum(np.float32(x) / pad[0], np.float32(w - 1 - x) / pad[2]),
|
101 |
+
1.0 - np.minimum(np.float32(y) / pad[1], np.float32(h - 1 - y) / pad[3]))
|
102 |
+
blur = qsize * 0.02
|
103 |
+
img += (scipy.ndimage.gaussian_filter(img, [blur, blur, 0]) - img) * np.clip(mask * 3.0 + 1.0, 0.0, 1.0)
|
104 |
+
img += (np.median(img, axis=(0, 1)) - img) * np.clip(mask, 0.0, 1.0)
|
105 |
+
img = Image.fromarray(np.uint8(np.clip(np.rint(img), 0, 255)), 'RGB')
|
106 |
+
quad += pad[:2]
|
107 |
+
|
108 |
+
# Transform.
|
109 |
+
img = img.transform((transform_size, transform_size), Image.QUAD, (quad + 0.5).flatten(), Image.BILINEAR)
|
110 |
+
if output_size < transform_size:
|
111 |
+
img = img.resize((output_size, output_size), Image.ANTIALIAS)
|
112 |
+
|
113 |
+
# Return aligned image.
|
114 |
+
return img
|