forked from yzhou359/MakeItTalk
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathmain_end2end_cartoon.py
203 lines (169 loc) · 8.49 KB
/
main_end2end_cartoon.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
"""
# Copyright 2020 Adobe
# All Rights Reserved.
# NOTICE: Adobe permits you to use, modify, and distribute this file in
# accordance with the terms of the Adobe license agreement accompanying
# it.
"""
import sys
sys.path.append('thirdparty/AdaptiveWingLoss')
import os, glob
import numpy as np
import argparse
from src.approaches.train_image_translation import Image_translation_block
import pickle
import face_alignment
from thirdparty.autovc.AutoVC_mel_Convertor_retrain_version import AutoVC_mel_Convertor
import shutil
from util.icp import icp
ADD_NAIVE_EYE = False
GEN_AUDIO = True
GEN_FLS = True
DEMO_CH = 'danbooru1'
parser = argparse.ArgumentParser()
parser.add_argument('--jpg', type=str, default=DEMO_CH)
parser.add_argument('--load_AUTOVC_name', type=str, default='examples/ckpt/ckpt_autovc.pth')
parser.add_argument('--load_a2l_G_name', type=str, default='examples/ckpt/ckpt_a2l_db_e_875.pth') #ckpt_audio2landmark_g.pth') #
parser.add_argument('--load_a2l_C_name', type=str, default='examples/ckpt/ckpt_audio2landmark_c.pth')
parser.add_argument('--load_G_name', type=str, default='examples/ckpt/ckpt_116_i2i_comb.pth') #ckpt_i2i_finetune_150.pth') #ckpt_image2image.pth') #
parser.add_argument('--amp_lip_x', type=float, default=2.0)
parser.add_argument('--amp_lip_y', type=float, default=2.0)
parser.add_argument('--amp_pos', type=float, default=0.8)
parser.add_argument('--reuse_train_emb_list', default=['45hn7-LXDX8']) # ['E_kmpT-EfOg']) # ['E_kmpT-EfOg']) # ['45hn7-LXDX8'])
parser.add_argument('--add_audio_in', default=False, action='store_true')
parser.add_argument('--comb_fan_awing', default=False, action='store_true')
parser.add_argument('--output_folder', type=str, default='examples_cartoon')
#### NEW POSE MODEL
parser.add_argument('--test_end2end', default=True, action='store_true')
parser.add_argument('--dump_dir', type=str, default='', help='')
parser.add_argument('--pos_dim', default=7, type=int)
parser.add_argument('--use_prior_net', default=True, action='store_true')
parser.add_argument('--transformer_d_model', default=32, type=int)
parser.add_argument('--transformer_N', default=2, type=int)
parser.add_argument('--transformer_heads', default=2, type=int)
parser.add_argument('--spk_emb_enc_size', default=16, type=int)
parser.add_argument('--init_content_encoder', type=str, default='')
parser.add_argument('--lr', type=float, default=1e-3, help='learning rate')
parser.add_argument('--reg_lr', type=float, default=1e-6, help='weight decay')
parser.add_argument('--write', default=False, action='store_true')
parser.add_argument('--segment_batch_size', type=int, default=512, help='batch size')
parser.add_argument('--emb_coef', default=3.0, type=float)
parser.add_argument('--lambda_laplacian_smooth_loss', default=1.0, type=float)
parser.add_argument('--use_11spk_only', default=False, action='store_true')
opt_parser = parser.parse_args()
shape_3d = np.loadtxt('examples_cartoon/{}_face_close_mouth.txt'.format(opt_parser.jpg))
''' STEP 3: Generate audio data as input to audio branch '''
au_data = []
ains = glob.glob1('examples', '*.wav')
ains.sort()
for ain in ains:
os.system('ffmpeg -y -loglevel error -i examples/{} -ar 16000 examples/tmp.wav'.format(ain))
shutil.copyfile('examples/tmp.wav', 'examples/{}'.format(ain))
print('Processing audio file', ain)
c = AutoVC_mel_Convertor('examples')
au_data_i = c.convert_single_wav_to_autovc_input(audio_filename=os.path.join('examples', ain),
autovc_model_path=opt_parser.load_AUTOVC_name)
au_data += au_data_i
# os.remove(os.path.join('examples', 'tmp.wav'))
if(os.path.isfile('examples/tmp.wav')):
os.remove('examples/tmp.wav')
fl_data = []
rot_tran, rot_quat, anchor_t_shape = [], [], []
for au, info in au_data:
au_length = au.shape[0]
fl = np.zeros(shape=(au_length, 68 * 3))
fl_data.append((fl, info))
rot_tran.append(np.zeros(shape=(au_length, 3, 4)))
rot_quat.append(np.zeros(shape=(au_length, 4)))
anchor_t_shape.append(np.zeros(shape=(au_length, 68 * 3)))
if(os.path.exists(os.path.join('examples', 'dump', 'random_val_fl.pickle'))):
os.remove(os.path.join('examples', 'dump', 'random_val_fl.pickle'))
if(os.path.exists(os.path.join('examples', 'dump', 'random_val_fl_interp.pickle'))):
os.remove(os.path.join('examples', 'dump', 'random_val_fl_interp.pickle'))
if(os.path.exists(os.path.join('examples', 'dump', 'random_val_au.pickle'))):
os.remove(os.path.join('examples', 'dump', 'random_val_au.pickle'))
if (os.path.exists(os.path.join('examples', 'dump', 'random_val_gaze.pickle'))):
os.remove(os.path.join('examples', 'dump', 'random_val_gaze.pickle'))
with open(os.path.join('examples', 'dump', 'random_val_fl.pickle'), 'wb') as fp:
pickle.dump(fl_data, fp)
with open(os.path.join('examples', 'dump', 'random_val_au.pickle'), 'wb') as fp:
pickle.dump(au_data, fp)
with open(os.path.join('examples', 'dump', 'random_val_gaze.pickle'), 'wb') as fp:
gaze = {'rot_trans':rot_tran, 'rot_quat':rot_quat, 'anchor_t_shape':anchor_t_shape}
pickle.dump(gaze, fp)
''' STEP 4: RUN audio->landmark network'''
from src.approaches.train_audio2landmark import Audio2landmark_model
model = Audio2landmark_model(opt_parser, jpg_shape=shape_3d)
model.test()
print('finish gen fls')
''' STEP 5: de-normalize the output to the original image scale '''
fls_names = glob.glob1('examples_cartoon', 'pred_fls_*.txt')
fls_names.sort()
for i in range(0,len(fls_names)):
ains = glob.glob1('examples', '*.wav')
ains.sort()
ain = ains[i]
fl = np.loadtxt(os.path.join('examples_cartoon', fls_names[i])).reshape((-1, 68,3))
output_dir = os.path.join('examples_cartoon', fls_names[i][:-4])
try:
os.makedirs(output_dir)
except:
pass
from util.utils import get_puppet_info
bound, scale, shift = get_puppet_info(DEMO_CH, ROOT_DIR='examples_cartoon')
fls = fl.reshape((-1, 68, 3))
fls[:, :, 0:2] = -fls[:, :, 0:2]
fls[:, :, 0:2] = (fls[:, :, 0:2] / scale)
fls[:, :, 0:2] -= shift.reshape(1, 2)
fls = fls.reshape(-1, 204)
# additional smooth
from scipy.signal import savgol_filter
fls[:, 0:48*3] = savgol_filter(fls[:, 0:48*3], 17, 3, axis=0)
fls[:, 48*3:] = savgol_filter(fls[:, 48*3:], 11, 3, axis=0)
fls = fls.reshape((-1, 68, 3))
if (DEMO_CH in ['paint', 'mulaney', 'cartoonM', 'beer', 'color', 'JohnMulaney', 'vangogh', 'jm', 'roy', 'lineface']):
r = list(range(0, 68))
fls = fls[:, r, :]
fls = fls[:, :, 0:2].reshape(-1, 68 * 2)
fls = np.concatenate((fls, np.tile(bound, (fls.shape[0], 1))), axis=1)
fls = fls.reshape(-1, 160)
else:
r = list(range(0, 48)) + list(range(60, 68))
fls = fls[:, r, :]
fls = fls[:, :, 0:2].reshape(-1, 56 * 2)
fls = np.concatenate((fls, np.tile(bound, (fls.shape[0], 1))), axis=1)
fls = fls.reshape(-1, 112 + bound.shape[1])
np.savetxt(os.path.join(output_dir, 'warped_points.txt'), fls, fmt='%.2f')
# static_points.txt
static_frame = np.loadtxt(os.path.join('examples_cartoon', '{}_face_open_mouth.txt'.format(DEMO_CH)))
static_frame = static_frame[r, 0:2]
static_frame = np.concatenate((static_frame, bound.reshape(-1, 2)), axis=0)
np.savetxt(os.path.join(output_dir, 'reference_points.txt'), static_frame, fmt='%.2f')
# triangle_vtx_index.txt
shutil.copy(os.path.join('examples_cartoon', DEMO_CH + '_delauney_tri.txt'),
os.path.join(output_dir, 'triangulation.txt'))
os.remove(os.path.join('examples_cartoon', fls_names[i]))
# # ==============================================
# # Step 4 : Vector art morphing (only work in WINDOWS)
# # ==============================================
# warp_exe = os.path.join(os.getcwd(), 'facewarp', 'facewarp.exe')
# import os
#
# if (os.path.exists(os.path.join(output_dir, 'output'))):
# shutil.rmtree(os.path.join(output_dir, 'output'))
# os.mkdir(os.path.join(output_dir, 'output'))
# os.chdir('{}'.format(os.path.join(output_dir, 'output')))
# print(os.getcwd())
#
# os.system('{} {} {} {} {} {}'.format(
# warp_exe,
# os.path.join('examples_cartoon', DEMO_CH+'.png'),
# os.path.join(output_dir, 'triangulation.txt'),
# os.path.join(output_dir, 'reference_points.txt'),
# os.path.join(output_dir, 'warped_points.txt'),
# # os.path.join(ROOT_DIR, 'puppets', sys.argv[6]),
# '-novsync -dump'))
# os.system('ffmpeg -y -r 62.5 -f image2 -i "%06d.tga" -i {} -shortest {}'.format(
# ain,
# os.path.join(output_dir, sys.argv[8])
# ))