-
Notifications
You must be signed in to change notification settings - Fork 0
/
morph_gan.py
226 lines (190 loc) · 8.28 KB
/
morph_gan.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
from stylegan_layers import G_mapping,G_synthesis
import torch
import torch.optim as optim
import torch.nn as nn
from collections import OrderedDict
import torch.nn.functional as F
from PIL import Image
from torchvision import transforms
import torchvision
from torchvision import models
from torchvision.utils import save_image
import numpy as np
from math import log10
import matplotlib.pyplot as plt
from tqdm import tqdm
from perceptual_model import VGG16_perceptual
from SSIM import SSIM
from mlxtend.image import extract_face_landmarks
import scipy.ndimage
import pdb
import PIL
#Load the pre-trained model
def loss_function(syn_img, img, img_p, MSE_loss, upsample, perceptual):
#UpSample synthesized image to match the input size of VGG-16 input.
#Extract mid level features for real and synthesized image and find the MSE loss between them for perceptual loss.
#Find MSE loss between the real and synthesized images of actual size
syn_img_p = upsample(syn_img)
syn0, syn1, syn2, syn3 = perceptual(syn_img_p)
r0, r1, r2, r3 = perceptual(img_p)
mse = MSE_loss(syn_img,img)
per_loss = 0
per_loss += MSE_loss(syn0,r0)
per_loss += MSE_loss(syn1,r1)
per_loss += MSE_loss(syn2,r2)
per_loss += MSE_loss(syn3,r3)
return mse, per_loss
def PSNR(mse, flag = 0):
#flag = 0 if a single image is used and 1 if loss for a batch of images is to be calculated
if flag == 0:
psnr = 10 * log10(1 / mse.item())
return psnr
def processImage(img, landmark, output_size=1024, transform_size=4096, enable_padding = True):
lm = np.array(landmark)
lm_chin = lm[0 : 17] # left-right
lm_eyebrow_left = lm[17 : 22] # left-right
lm_eyebrow_right = lm[22 : 27] # left-right
lm_nose = lm[27 : 31] # top-down
lm_nostrils = lm[31 : 36] # top-down
lm_eye_left = lm[36 : 42] # left-clockwise
lm_eye_right = lm[42 : 48] # left-clockwise
lm_mouth_outer = lm[48 : 60] # left-clockwise
lm_mouth_inner = lm[60 : 68] # left-clockwise
# Calculate auxiliary vectors.
eye_left = np.mean(lm_eye_left, axis=0)
eye_right = np.mean(lm_eye_right, axis=0)
eye_avg = (eye_left + eye_right) * 0.5
eye_to_eye = eye_right - eye_left
mouth_left = lm_mouth_outer[0]
mouth_right = lm_mouth_outer[6]
mouth_avg = (mouth_left + mouth_right) * 0.5
eye_to_mouth = mouth_avg - eye_avg
# Choose oriented crop rectangle.
x = eye_to_eye - np.flipud(eye_to_mouth) * [-1, 1]
x /= np.hypot(*x)
x *= max(np.hypot(*eye_to_eye) * 2.0, np.hypot(*eye_to_mouth) * 1.8)
y = np.flipud(x) * [-1, 1]
c = eye_avg + eye_to_mouth * 0.1
quad = np.stack([c - x - y, c - x + y, c + x + y, c + x - y])
qsize = np.hypot(*x) * 2
# Shrink.
shrink = int(np.floor(qsize / output_size * 0.5))
if shrink > 1:
rsize = (int(np.rint(float(img.size[0]) / shrink)), int(np.rint(float(img.size[1]) / shrink)))
img = img.resize(rsize, PIL.Image.ANTIALIAS)
quad /= shrink
qsize /= shrink
# Crop.
border = max(int(np.rint(qsize * 0.1)), 3)
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]))))
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]))
if crop[2] - crop[0] < img.size[0] or crop[3] - crop[1] < img.size[1]:
img = img.crop(crop)
quad -= crop[0:2]
# Pad.
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]))))
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))
if enable_padding and max(pad) > border - 4:
pad = np.maximum(pad, int(np.rint(qsize * 0.3)))
img = np.pad(np.float32(img), ((pad[1], pad[3]), (pad[0], pad[2]), (0, 0)), 'reflect')
h, w, _ = img.shape
y, x, _ = np.ogrid[:h, :w, :1]
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]))
blur = qsize * 0.02
img += (scipy.ndimage.gaussian_filter(img, [blur, blur, 0]) - img) * np.clip(mask * 3.0 + 1.0, 0.0, 1.0)
img += (np.median(img, axis=(0,1)) - img) * np.clip(mask, 0.0, 1.0)
img = PIL.Image.fromarray(np.uint8(np.clip(np.rint(img), 0, 255)), 'RGB')
quad += pad[:2]
# Transform.
img = img.transform((transform_size, transform_size), PIL.Image.QUAD, (quad + 0.5).flatten(), PIL.Image.BILINEAR)
if output_size < transform_size:
img = img.resize((output_size, output_size), PIL.Image.ANTIALIAS)
return img
import imageio
def extract_landmark(path):
img = imageio.imread(path)
landmarks = extract_face_landmarks(img)
return landmarks
def embedding_function(image, g_mapping, g_synthesis):
upsample = torch.nn.Upsample(scale_factor = 256/1024, mode = 'bilinear')
img_p = image.clone()
img_p = upsample(img_p)
#Perceptual loss initialise object
perceptual = VGG16_perceptual().to(device)
#MSE loss object
MSE_loss = nn.MSELoss(reduction="mean")
#since the synthesis network expects 18 w vectors of size 1x512 thus we take latent vector of the same size
latents = torch.zeros((1,18,512), requires_grad = True, device = device)
#Optimizer to change latent code in each backward step
optimizer = optim.Adam({latents},lr=0.01,betas=(0.9,0.999),eps=1e-8)
#Loop to optimise latent vector to match the generated image to input image
loss_ = []
loss_psnr = []
count = 0
ssim_loss = SSIM()
print("Start Optimizing Latent Space")
for e in tqdm(range(500)):
optimizer.zero_grad()
syn_img = g_synthesis(latents)
syn_img = (syn_img+1.0)/2.0
ssim_loss_syn = upsample(syn_img)
ssim_loss_out = -ssim_loss(ssim_loss_syn, img_p)
mse, per_loss = loss_function(syn_img, image, img_p, MSE_loss, upsample, perceptual)
psnr = PSNR(mse, flag = 0)
loss = per_loss + mse
loss.backward()
optimizer.step()
loss_np=loss.detach().cpu().numpy()
loss_p=per_loss.detach().cpu().numpy()
loss_m=mse.detach().cpu().numpy()
loss_psnr.append(psnr)
loss_.append(loss_np)
if (e+1)%250==0 :
print("iter{}: loss -- {}, mse_loss --{}, percep_loss --{}, psnr --{}".format(e+1,loss_np,loss_m,loss_p,psnr))
count = count + 1
save_image(syn_img.clamp(0,1),"./imgs/progress{}.png".format(count))
return latents
if __name__ == "__main__":
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
img_path = './imbedimgs/Vittoria.jpeg'
with open(img_path,"rb") as f:
image1=Image.open(f)
image1=image1.convert("RGB")
landmark1 = extract_landmark(img_path)
image1 = processImage(image1, landmark1)
transform = transforms.Compose([transforms.ToTensor()])
image1 = transform(image1)
image1 = image1.unsqueeze(0)
image1 = image1.to(device)
print("image1 loaded")
#Read a sample image we want to find a latent vector for
img_path = './imbedimgs/DariaS.jpeg'
with open(img_path,"rb") as f:
image2=Image.open(f)
image2=image2.convert("RGB")
landmark2 = extract_landmark(img_path)
image2 = processImage(image2, landmark2)
transform = transforms.Compose([transforms.ToTensor()])
image2 = transform(image2)
image2 = image2.unsqueeze(0)
image2 = image2.to(device)
print("image2 loaded")
device = 'cuda:0' if torch.cuda.is_available() else 'cpu'
g_all = nn.Sequential(OrderedDict([('g_mapping', G_mapping()),
#('truncation', Truncation(avg_latent)),
('g_synthesis', G_synthesis(resolution=1024))
]))
#Load the pre-trained model
g_all.load_state_dict(torch.load('./PretrainedModel/karras2019stylegan-ffhq-1024x1024.pt', map_location=device))
g_all.eval()
g_all.to(device)
g_mapping, g_synthesis = g_all[0],g_all[1]
latent1 = embedding_function(image1, g_mapping, g_synthesis)
latent2 = embedding_function(image2, g_mapping, g_synthesis)
#Morph
for i in tqdm(range(40)):
a = (1/40)*i
w = latent1 * (1-a) + latent2 * a
syn_img = g_synthesis(w)
syn_img = (syn_img+1.0)/2.0
save_image(syn_img.clamp(0,1),"./imgs/Morphed{}.png".format(i))