-
Notifications
You must be signed in to change notification settings - Fork 140
/
AE_Model.py
208 lines (162 loc) · 7.49 KB
/
AE_Model.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
# this is model for fix netP and mask only left eye, right eye, nose and skin
# original loss
import random
import numpy as np
import os
from . import networks
# from scipy.ndimage import median_filter
import jittor as jt
from jittor import init
from jittor import nn
import jittor.transform as transform
# from pdb import set_trace as st
# import heapq
from numpy.linalg import solve
import time
class AE_Model(nn.Module):
def name(self):
return self.name
def init_loss_filter(self):
flags = (True,True,True)
def loss_filter(kl_loss,l2_image,l2_mask):
return [l for (l,f) in zip((kl_loss,l2_image,l2_mask),flags) if f]
return loss_filter
def initialize(self, opt):
# assert opt.vae_encoder == True
self.opt = opt
self.save_dir = os.path.join(opt.param, opt.name)
self.name = 'AE_Model'
# BaseModel.initialize(self, opt)
input_nc = opt.input_nc
self.output_nc = opt.output_nc
self.input_nc = input_nc
self.model_partial_name = opt.partial
##### define networks
# Generator network
netG_input_nc = input_nc
self.net_encoder = networks.define_part_encoder(model=self.model_partial_name,
input_nc = opt.input_nc,
norm=opt.norm,
latent_dim = opt.latant_dim)
self.net_decoder = networks.define_part_decoder(model=self.model_partial_name,
output_nc = opt.input_nc,
norm=opt.norm,
latent_dim = opt.latant_dim)
self.load_network(self.net_encoder, 'encoder_'+self.model_partial_name, 'latest', '')
self.load_network(self.net_decoder, 'decoder_'+self.model_partial_name+'_image', 'latest', '')
# use for test
self.feature_list_male = np.fromfile(os.path.join(opt.param, opt.name) + '/man_' + opt.partial + '_feature.bin', dtype=np.float32)
self.feature_list_male.shape = 6247, 512
#girl
self.feature_list_female = np.fromfile(os.path.join(opt.param, opt.name) + '/female_' + opt.partial + '_feature.bin', dtype=np.float32)
self.feature_list_female.shape = 11456, 512
self.feature_list = [self.feature_list_male,self.feature_list_female]
def get_latent(self, input_image):
input_image = (input_image-127.5)/127.5
input_image = np.expand_dims(input_image, axis=2)
input_image = input_image.transpose(2,0,1)
input_image = np.expand_dims(input_image, axis=0)
input_image = input_image.astype('float32')
input_image = transform.to_tensor(jt.array(input_image))
# print(input_image.shape)
mus_mouth = self.net_encoder(input_image)
return mus_mouth
def get_image(self, latent_vec):
# return self.net_decoder(latent_vec)
fakes = self.net_decoder(latent_vec)
fakes = (fakes[0,:,:,:].numpy()+1)/2
fakes = np.transpose(fakes, (1, 2, 0)) * 255.0
fakes = np.clip(fakes, 0, 255)
return fakes.astype(np.uint8)
def get_inter(self, input_image, nearnN=3, sex=1,w_c=1,random_=-1):
generated_f = self.get_latent(input_image)
generated_f = generated_f.numpy()
feature_list = self.feature_list[sex]
list_len = jt.array([feature_list.shape[0]])
# a = jt.random((n,3))
b = jt.code([1, nearnN],
"int32", [jt.array(feature_list),jt.array(generated_f), list_len],
cpu_header="#include <algorithm>",
cpu_src="""
using namespace std;
auto n=out_shape0, k=out_shape1;
int N=@in2(0);
// 使用openmp实现自动并行化
// 存储k近邻的距离和下标
vector<pair<float,int>> id(N);
#pragma omp parallel for
for (int j=0; j<N; j++) {
auto dis = 0.0;
for (int d=0; d<512; d++)
{
auto dx = @in1(0,d)-@in0(j,d);
dis = dis +dx*dx;
}
id[j] = {dis, j};
}
// 使用c++算法库的nth_element排序
nth_element(id.begin(),
id.begin()+k, id.end());
// 将下标输出到计图的变量中
for (int j=0; j<k; j++)
@out(0,j) = id[j].second;
"""
)
idx_sort = b[0].numpy()
if nearnN==1:
vec_mu = feature_list[idx_sort[0]]
vec_mu = vec_mu * w_c + (1 - w_c) * generated_f
return self.get_image(vec_mu), self.get_shadow_image(vec_mu, torch.ones((1,1)).data.cuda(), nearnN), vec_mu
# | vg - sum( wi*vi )| et. sum(wi) = 1
# == | vg - v0 - sum( wi*vi) | et. w = [1,w1,...,wn]
A_0 = [feature_list[idx_sort[0],:]]
A_m = A_0
for i in range(1,nearnN):
A_m = np.concatenate((A_m,[feature_list[idx_sort[i],:]]), axis=0)
A_0 = np.array(A_0)
A_m= np.array(A_m).T
A_m0 = np.concatenate((A_m[:,1:]-A_0.T, np.ones((1,nearnN-1))*10), axis=0)
A = np.dot(A_m0.T, A_m0)
b = np.zeros((1, generated_f.shape[1]+1))
b[0,0:generated_f.shape[1]] = generated_f-A_0
B = np.dot(A_m0.T, b.T)
x = solve(A, B)
xx = np.zeros((nearnN,1))
xx[0,0] = 1 - x.sum()
xx[1:,0] = x[:,0]
# print(time.time()- start_time)
vec_mu = np.dot(A_m, xx).T * w_c + (1-w_c)* generated_f
vec_mu = jt.array(vec_mu.astype('float32'))
return self.get_shadow_image(A_m.T,xx,nearnN), vec_mu
def get_shadow_image(self, mus_mouth, weight, nearnN):
fakes = 0
for i in range(nearnN):
w_i = weight[i]
if w_i<=0:
continue
elif w_i>0.5:
w_i = 0.5
# print(i)
mus_vec = jt.unsqueeze(mus_mouth[[i],:],1)
fake_image = self.net_decoder(jt.array(mus_vec))
# fake_image = fake_image[[0],:,:,:]
if i==0:
fakes = (1-fake_image)/2* w_i
else:
fakes = fakes + (1-fake_image)/2 * w_i
fakes = 1-fakes
fakes = fakes[0,:,:,:].detach().numpy()
fakes = np.transpose(fakes, (1, 2, 0)) * 255.0
fakes = np.clip(fakes, 0, 255)
return fakes.astype(np.uint8)
# helper loading function that can be used by subclasses
def load_network(self, network, network_label, epoch_label, save_dir='', save_path=''):
save_filename = '%s_net_%s.pkl' % (epoch_label, network_label)
if not save_dir:
save_dir = self.save_dir
save_path = os.path.join(save_dir, save_filename)
print("load_path",save_path)
if not os.path.isfile(save_path):
print('%s not exists yet!' % save_path)
else:
network.load(save_path)