Upload 4 files
Browse files- cog.yaml +30 -0
- demo.py +307 -0
- predict.py +308 -0
- requirements.txt +10 -0
cog.yaml
ADDED
@@ -0,0 +1,30 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
build:
|
2 |
+
gpu: true
|
3 |
+
python_version: "3.8"
|
4 |
+
system_packages:
|
5 |
+
- "libgl1-mesa-glx"
|
6 |
+
- "libglib2.0-0"
|
7 |
+
- "libsox-fmt-mp3"
|
8 |
+
python_packages:
|
9 |
+
- "torch==1.7.1"
|
10 |
+
- "torchvision==0.8.2"
|
11 |
+
- "numpy==1.18.1"
|
12 |
+
- "ipython==7.21.0"
|
13 |
+
- "Pillow==8.3.1"
|
14 |
+
- "scikit-image==0.18.3"
|
15 |
+
- "librosa==0.7.2"
|
16 |
+
- "tqdm==4.62.3"
|
17 |
+
- "scipy==1.7.1"
|
18 |
+
- "dominate==2.6.0"
|
19 |
+
- "albumentations==0.5.2"
|
20 |
+
- "beautifulsoup4==4.10.0"
|
21 |
+
- "sox==1.4.1"
|
22 |
+
- "h5py==3.4.0"
|
23 |
+
- "numba==0.48"
|
24 |
+
- "moviepy==1.0.3"
|
25 |
+
run:
|
26 |
+
- apt update -y && apt-get install ffmpeg -y
|
27 |
+
- apt-get install sox libsox-fmt-mp3 -y
|
28 |
+
- pip install opencv-python==4.1.2.30
|
29 |
+
|
30 |
+
predict: "predict.py:Predictor"
|
demo.py
ADDED
@@ -0,0 +1,307 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
import subprocess
|
3 |
+
from os.path import join
|
4 |
+
from tqdm import tqdm
|
5 |
+
import numpy as np
|
6 |
+
import torch
|
7 |
+
from collections import OrderedDict
|
8 |
+
import librosa
|
9 |
+
from skimage.io import imread
|
10 |
+
import cv2
|
11 |
+
import scipy.io as sio
|
12 |
+
import argparse
|
13 |
+
import yaml
|
14 |
+
import albumentations as A
|
15 |
+
import albumentations.pytorch
|
16 |
+
from pathlib import Path
|
17 |
+
|
18 |
+
from options.test_audio2feature_options import TestOptions as FeatureOptions
|
19 |
+
from options.test_audio2headpose_options import TestOptions as HeadposeOptions
|
20 |
+
from options.test_feature2face_options import TestOptions as RenderOptions
|
21 |
+
|
22 |
+
from datasets import create_dataset
|
23 |
+
from models import create_model
|
24 |
+
from models.networks import APC_encoder
|
25 |
+
import util.util as util
|
26 |
+
from util.visualizer import Visualizer
|
27 |
+
from funcs import utils
|
28 |
+
from funcs import audio_funcs
|
29 |
+
import soundfile as sf
|
30 |
+
import warnings
|
31 |
+
warnings.filterwarnings("ignore")
|
32 |
+
|
33 |
+
|
34 |
+
|
35 |
+
def write_video_with_audio(audio_path, output_path, prefix='pred_'):
|
36 |
+
fps, fourcc = 60, cv2.VideoWriter_fourcc(*'DIVX')
|
37 |
+
video_tmp_path = join(save_root, 'tmp.avi')
|
38 |
+
out = cv2.VideoWriter(video_tmp_path, fourcc, fps, (Renderopt.loadSize, Renderopt.loadSize))
|
39 |
+
for j in tqdm(range(nframe), position=0, desc='writing video'):
|
40 |
+
img = cv2.imread(join(save_root, prefix + str(j+1) + '.jpg'))
|
41 |
+
out.write(img)
|
42 |
+
out.release()
|
43 |
+
cmd = 'ffmpeg -i "' + video_tmp_path + '" -i "' + audio_path + '" -codec copy -shortest "' + output_path + '"'
|
44 |
+
subprocess.call(cmd, shell=True)
|
45 |
+
os.remove(video_tmp_path) # remove the template video
|
46 |
+
|
47 |
+
|
48 |
+
|
49 |
+
if __name__ == '__main__':
|
50 |
+
parser = argparse.ArgumentParser()
|
51 |
+
parser.add_argument('--id', default='May', help="person name, e.g. Obama1, Obama2, May, Nadella, McStay")
|
52 |
+
parser.add_argument('--driving_audio', default='./data/input/00083.wav', help="path to driving audio")
|
53 |
+
parser.add_argument('--save_intermediates', default=0, help="whether to save intermediate results")
|
54 |
+
parser.add_argument('--device', type=str, default='cpu', help='use cuda for GPU or use cpu for CPU')
|
55 |
+
|
56 |
+
|
57 |
+
############################### I/O Settings ##############################
|
58 |
+
# load config files
|
59 |
+
opt = parser.parse_args()
|
60 |
+
device = torch.device(opt.device)
|
61 |
+
with open(join('./config/', opt.id + '.yaml')) as f:
|
62 |
+
config = yaml.load(f, Loader=yaml.SafeLoader)
|
63 |
+
data_root = join('./data/', opt.id)
|
64 |
+
# create the results folder
|
65 |
+
audio_name = os.path.split(opt.driving_audio)[1][:-4]
|
66 |
+
save_root = join('./results/', opt.id, audio_name)
|
67 |
+
if not os.path.exists(save_root):
|
68 |
+
os.makedirs(save_root)
|
69 |
+
|
70 |
+
|
71 |
+
|
72 |
+
############################ Hyper Parameters #############################
|
73 |
+
h, w, sr, FPS = 512, 512, 16000, 60
|
74 |
+
mouth_indices = np.concatenate([np.arange(4, 11), np.arange(46, 64)])
|
75 |
+
eye_brow_indices = [27, 65, 28, 68, 29, 67, 30, 66, 31, 72, 32, 69, 33, 70, 34, 71]
|
76 |
+
eye_brow_indices = np.array(eye_brow_indices, np.int32)
|
77 |
+
|
78 |
+
|
79 |
+
|
80 |
+
############################ Pre-defined Data #############################
|
81 |
+
mean_pts3d = np.load(join(data_root, 'mean_pts3d.npy'))
|
82 |
+
fit_data = np.load(config['dataset_params']['fit_data_path'])
|
83 |
+
pts3d = np.load(config['dataset_params']['pts3d_path']) - mean_pts3d
|
84 |
+
trans = fit_data['trans'][:,:,0].astype(np.float32)
|
85 |
+
mean_translation = trans.mean(axis=0)
|
86 |
+
candidate_eye_brow = pts3d[10:, eye_brow_indices]
|
87 |
+
std_mean_pts3d = np.load(config['dataset_params']['pts3d_path']).mean(axis=0)
|
88 |
+
# candidates images
|
89 |
+
img_candidates = []
|
90 |
+
for j in range(4):
|
91 |
+
output = imread(join(data_root, 'candidates', f'normalized_full_{j}.jpg'))
|
92 |
+
output = A.pytorch.transforms.ToTensor(normalize={'mean':(0.5,0.5,0.5),
|
93 |
+
'std':(0.5,0.5,0.5)})(image=output)['image']
|
94 |
+
img_candidates.append(output)
|
95 |
+
img_candidates = torch.cat(img_candidates).unsqueeze(0).to(device)
|
96 |
+
|
97 |
+
# shoulders
|
98 |
+
shoulders = np.load(join(data_root, 'normalized_shoulder_points.npy'))
|
99 |
+
shoulder3D = np.load(join(data_root, 'shoulder_points3D.npy'))[1]
|
100 |
+
ref_trans = trans[1]
|
101 |
+
|
102 |
+
# camera matrix, we always use training set intrinsic parameters.
|
103 |
+
camera = utils.camera()
|
104 |
+
camera_intrinsic = np.load(join(data_root, 'camera_intrinsic.npy')).astype(np.float32)
|
105 |
+
APC_feat_database = np.load(join(data_root, 'APC_feature_base.npy'))
|
106 |
+
|
107 |
+
# load reconstruction data
|
108 |
+
scale = sio.loadmat(join(data_root, 'id_scale.mat'))['scale'][0,0]
|
109 |
+
# Audio2Mel_torch = audio_funcs.Audio2Mel(n_fft=512, hop_length=int(16000/120), win_length=int(16000/60), sampling_rate=16000,
|
110 |
+
# n_mel_channels=80, mel_fmin=90, mel_fmax=7600.0).to(device)
|
111 |
+
|
112 |
+
|
113 |
+
|
114 |
+
########################### Experiment Settings ###########################
|
115 |
+
#### user config
|
116 |
+
use_LLE = config['model_params']['APC']['use_LLE']
|
117 |
+
Knear = config['model_params']['APC']['Knear']
|
118 |
+
LLE_percent = config['model_params']['APC']['LLE_percent']
|
119 |
+
headpose_sigma = config['model_params']['Headpose']['sigma']
|
120 |
+
Feat_smooth_sigma = config['model_params']['Audio2Mouth']['smooth']
|
121 |
+
Head_smooth_sigma = config['model_params']['Headpose']['smooth']
|
122 |
+
Feat_center_smooth_sigma, Head_center_smooth_sigma = 0, 0
|
123 |
+
AMP_method = config['model_params']['Audio2Mouth']['AMP'][0]
|
124 |
+
Feat_AMPs = config['model_params']['Audio2Mouth']['AMP'][1:]
|
125 |
+
rot_AMP, trans_AMP = config['model_params']['Headpose']['AMP']
|
126 |
+
shoulder_AMP = config['model_params']['Headpose']['shoulder_AMP']
|
127 |
+
save_feature_maps = config['model_params']['Image2Image']['save_input']
|
128 |
+
|
129 |
+
#### common settings
|
130 |
+
Featopt = FeatureOptions().parse()
|
131 |
+
Headopt = HeadposeOptions().parse()
|
132 |
+
Renderopt = RenderOptions().parse()
|
133 |
+
Featopt.load_epoch = config['model_params']['Audio2Mouth']['ckp_path']
|
134 |
+
Headopt.load_epoch = config['model_params']['Headpose']['ckp_path']
|
135 |
+
Renderopt.dataroot = config['dataset_params']['root']
|
136 |
+
Renderopt.load_epoch = config['model_params']['Image2Image']['ckp_path']
|
137 |
+
Renderopt.size = config['model_params']['Image2Image']['size']
|
138 |
+
## GPU or CPU
|
139 |
+
if opt.device == 'cpu':
|
140 |
+
Featopt.gpu_ids = Headopt.gpu_ids = Renderopt.gpu_ids = []
|
141 |
+
|
142 |
+
|
143 |
+
|
144 |
+
############################# Load Models #################################
|
145 |
+
print('---------- Loading Model: APC-------------')
|
146 |
+
APC_model = APC_encoder(config['model_params']['APC']['mel_dim'],
|
147 |
+
config['model_params']['APC']['hidden_size'],
|
148 |
+
config['model_params']['APC']['num_layers'],
|
149 |
+
config['model_params']['APC']['residual'])
|
150 |
+
APC_model.load_state_dict(torch.load(config['model_params']['APC']['ckp_path'],map_location=device), strict=False)
|
151 |
+
if opt.device == 'cuda':
|
152 |
+
APC_model.cuda()
|
153 |
+
APC_model.eval()
|
154 |
+
print('---------- Loading Model: {} -------------'.format(Featopt.task))
|
155 |
+
Audio2Feature = create_model(Featopt)
|
156 |
+
Audio2Feature.setup(Featopt)
|
157 |
+
Audio2Feature.eval()
|
158 |
+
print('---------- Loading Model: {} -------------'.format(Headopt.task))
|
159 |
+
Audio2Headpose = create_model(Headopt)
|
160 |
+
Audio2Headpose.setup(Headopt)
|
161 |
+
Audio2Headpose.eval()
|
162 |
+
if Headopt.feature_decoder == 'WaveNet':
|
163 |
+
if opt.device == 'cuda':
|
164 |
+
Headopt.A2H_receptive_field = Audio2Headpose.Audio2Headpose.module.WaveNet.receptive_field
|
165 |
+
else:
|
166 |
+
Headopt.A2H_receptive_field = Audio2Headpose.Audio2Headpose.WaveNet.receptive_field
|
167 |
+
print('---------- Loading Model: {} -------------'.format(Renderopt.task))
|
168 |
+
facedataset = create_dataset(Renderopt)
|
169 |
+
Feature2Face = create_model(Renderopt)
|
170 |
+
Feature2Face.setup(Renderopt)
|
171 |
+
Feature2Face.eval()
|
172 |
+
visualizer = Visualizer(Renderopt)
|
173 |
+
|
174 |
+
|
175 |
+
|
176 |
+
############################## Inference ##################################
|
177 |
+
print('Processing audio: {} ...'.format(audio_name))
|
178 |
+
# read audio
|
179 |
+
audio, _ = librosa.load(opt.driving_audio, sr=sr)
|
180 |
+
total_frames = np.int32(audio.shape[0] / sr * FPS)
|
181 |
+
|
182 |
+
|
183 |
+
#### 1. compute APC features
|
184 |
+
print('1. Computing APC features...')
|
185 |
+
mel80 = utils.compute_mel_one_sequence(audio, device=opt.device)
|
186 |
+
mel_nframe = mel80.shape[0]
|
187 |
+
with torch.no_grad():
|
188 |
+
length = torch.Tensor([mel_nframe])
|
189 |
+
mel80_torch = torch.from_numpy(mel80.astype(np.float32)).to(device).unsqueeze(0)
|
190 |
+
hidden_reps = APC_model.forward(mel80_torch, length)[0] # [mel_nframe, 512]
|
191 |
+
hidden_reps = hidden_reps.cpu().numpy()
|
192 |
+
audio_feats = hidden_reps
|
193 |
+
|
194 |
+
|
195 |
+
#### 2. manifold projection
|
196 |
+
if use_LLE:
|
197 |
+
print('2. Manifold projection...')
|
198 |
+
ind = utils.KNN_with_torch(audio_feats, APC_feat_database, K=Knear)
|
199 |
+
weights, feat_fuse = utils.compute_LLE_projection_all_frame(audio_feats, APC_feat_database, ind, audio_feats.shape[0])
|
200 |
+
audio_feats = audio_feats * (1-LLE_percent) + feat_fuse * LLE_percent
|
201 |
+
|
202 |
+
|
203 |
+
#### 3. Audio2Mouth
|
204 |
+
print('3. Audio2Mouth inference...')
|
205 |
+
pred_Feat = Audio2Feature.generate_sequences(audio_feats, sr, FPS, fill_zero=True, opt=Featopt)
|
206 |
+
|
207 |
+
|
208 |
+
#### 4. Audio2Headpose
|
209 |
+
print('4. Headpose inference...')
|
210 |
+
# set history headposes as zero
|
211 |
+
pre_headpose = np.zeros(Headopt.A2H_wavenet_input_channels, np.float32)
|
212 |
+
pred_Head = Audio2Headpose.generate_sequences(audio_feats, pre_headpose, fill_zero=True, sigma_scale=0.3, opt=Headopt)
|
213 |
+
|
214 |
+
|
215 |
+
#### 5. Post-Processing
|
216 |
+
print('5. Post-processing...')
|
217 |
+
nframe = min(pred_Feat.shape[0], pred_Head.shape[0])
|
218 |
+
pred_pts3d = np.zeros([nframe, 73, 3])
|
219 |
+
pred_pts3d[:, mouth_indices] = pred_Feat.reshape(-1, 25, 3)[:nframe]
|
220 |
+
|
221 |
+
## mouth
|
222 |
+
pred_pts3d = utils.landmark_smooth_3d(pred_pts3d, Feat_smooth_sigma, area='only_mouth')
|
223 |
+
pred_pts3d = utils.mouth_pts_AMP(pred_pts3d, True, AMP_method, Feat_AMPs)
|
224 |
+
pred_pts3d = pred_pts3d + mean_pts3d
|
225 |
+
pred_pts3d = utils.solve_intersect_mouth(pred_pts3d) # solve intersect lips if exist
|
226 |
+
|
227 |
+
## headpose
|
228 |
+
pred_Head[:, 0:3] *= rot_AMP
|
229 |
+
pred_Head[:, 3:6] *= trans_AMP
|
230 |
+
pred_headpose = utils.headpose_smooth(pred_Head[:,:6], Head_smooth_sigma).astype(np.float32)
|
231 |
+
pred_headpose[:, 3:] += mean_translation
|
232 |
+
pred_headpose[:, 0] += 180
|
233 |
+
|
234 |
+
## compute projected landmarks
|
235 |
+
pred_landmarks = np.zeros([nframe, 73, 2], dtype=np.float32)
|
236 |
+
final_pts3d = np.zeros([nframe, 73, 3], dtype=np.float32)
|
237 |
+
final_pts3d[:] = std_mean_pts3d.copy()
|
238 |
+
final_pts3d[:, 46:64] = pred_pts3d[:nframe, 46:64]
|
239 |
+
for k in tqdm(range(nframe)):
|
240 |
+
ind = k % candidate_eye_brow.shape[0]
|
241 |
+
final_pts3d[k, eye_brow_indices] = candidate_eye_brow[ind] + mean_pts3d[eye_brow_indices]
|
242 |
+
pred_landmarks[k], _, _ = utils.project_landmarks(camera_intrinsic, camera.relative_rotation,
|
243 |
+
camera.relative_translation, scale,
|
244 |
+
pred_headpose[k], final_pts3d[k])
|
245 |
+
|
246 |
+
## Upper Body Motion
|
247 |
+
pred_shoulders = np.zeros([nframe, 18, 2], dtype=np.float32)
|
248 |
+
pred_shoulders3D = np.zeros([nframe, 18, 3], dtype=np.float32)
|
249 |
+
for k in range(nframe):
|
250 |
+
diff_trans = pred_headpose[k][3:] - ref_trans
|
251 |
+
pred_shoulders3D[k] = shoulder3D + diff_trans * shoulder_AMP
|
252 |
+
# project
|
253 |
+
project = camera_intrinsic.dot(pred_shoulders3D[k].T)
|
254 |
+
project[:2, :] /= project[2, :] # divide z
|
255 |
+
pred_shoulders[k] = project[:2, :].T
|
256 |
+
|
257 |
+
|
258 |
+
#### 6. Image2Image translation & Save resuls
|
259 |
+
print('6. Image2Image translation & Saving results...')
|
260 |
+
for ind in tqdm(range(0, nframe), desc='Image2Image translation inference'):
|
261 |
+
# feature_map: [input_nc, h, w]
|
262 |
+
current_pred_feature_map = facedataset.dataset.get_data_test_mode(pred_landmarks[ind],
|
263 |
+
pred_shoulders[ind],
|
264 |
+
facedataset.dataset.image_pad)
|
265 |
+
input_feature_maps = current_pred_feature_map.unsqueeze(0).to(device)
|
266 |
+
pred_fake = Feature2Face.inference(input_feature_maps, img_candidates)
|
267 |
+
# save results
|
268 |
+
visual_list = [('pred', util.tensor2im(pred_fake[0]))]
|
269 |
+
if save_feature_maps:
|
270 |
+
visual_list += [('input', np.uint8(current_pred_feature_map[0].cpu().numpy() * 255))]
|
271 |
+
visuals = OrderedDict(visual_list)
|
272 |
+
visualizer.save_images(save_root, visuals, str(ind+1))
|
273 |
+
|
274 |
+
|
275 |
+
## make videos
|
276 |
+
# generate corresponding audio, reused for all results
|
277 |
+
tmp_audio_path = join(save_root, 'tmp.wav')
|
278 |
+
tmp_audio_clip = audio[ : np.int32(nframe * sr / FPS)]
|
279 |
+
sf.write(tmp_audio_path, tmp_audio_clip, sr)
|
280 |
+
# librosa.output.write_wav(tmp_audio_path, tmp_audio_clip, sr)
|
281 |
+
|
282 |
+
|
283 |
+
final_path = join(save_root, audio_name + '.avi')
|
284 |
+
write_video_with_audio(tmp_audio_path, final_path, 'pred_')
|
285 |
+
feature_maps_path = join(save_root, audio_name + '_feature_maps.avi')
|
286 |
+
write_video_with_audio(tmp_audio_path, feature_maps_path, 'input_')
|
287 |
+
|
288 |
+
if os.path.exists(tmp_audio_path):
|
289 |
+
os.remove(tmp_audio_path)
|
290 |
+
if not opt.save_intermediates:
|
291 |
+
_img_paths = list(map(lambda x:str(x), list(Path(save_root).glob('*.jpg'))))
|
292 |
+
for i in tqdm(range(len(_img_paths)), desc='deleting intermediate images'):
|
293 |
+
os.remove(_img_paths[i])
|
294 |
+
|
295 |
+
|
296 |
+
print('Finish!')
|
297 |
+
|
298 |
+
|
299 |
+
|
300 |
+
|
301 |
+
|
302 |
+
|
303 |
+
|
304 |
+
|
305 |
+
|
306 |
+
|
307 |
+
|
predict.py
ADDED
@@ -0,0 +1,308 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
import subprocess
|
3 |
+
from os.path import join
|
4 |
+
import yaml
|
5 |
+
import tempfile
|
6 |
+
import argparse
|
7 |
+
from skimage.io import imread
|
8 |
+
import numpy as np
|
9 |
+
import librosa
|
10 |
+
from util import util
|
11 |
+
from tqdm import tqdm
|
12 |
+
import torch
|
13 |
+
from collections import OrderedDict
|
14 |
+
import cv2
|
15 |
+
from moviepy.video.io.ffmpeg_tools import ffmpeg_extract_subclip
|
16 |
+
from cog import BasePredictor, Input, Path
|
17 |
+
import scipy.io as sio
|
18 |
+
import albumentations as A
|
19 |
+
from options.test_audio2feature_options import TestOptions as FeatureOptions
|
20 |
+
from options.test_audio2headpose_options import TestOptions as HeadposeOptions
|
21 |
+
from options.test_feature2face_options import TestOptions as RenderOptions
|
22 |
+
from datasets import create_dataset
|
23 |
+
from models import create_model
|
24 |
+
from models.networks import APC_encoder
|
25 |
+
from util.visualizer import Visualizer
|
26 |
+
from funcs import utils, audio_funcs
|
27 |
+
from demo import write_video_with_audio
|
28 |
+
import warnings
|
29 |
+
|
30 |
+
warnings.filterwarnings("ignore")
|
31 |
+
|
32 |
+
|
33 |
+
class Predictor(BasePredictor):
|
34 |
+
def setup(self):
|
35 |
+
self.parser = argparse.ArgumentParser()
|
36 |
+
self.parser.add_argument('--id', default='May', help="person name, e.g. Obama1, Obama2, May, Nadella, McStay")
|
37 |
+
self.parser.add_argument('--driving_audio', default='data/Input/00083.wav', help="path to driving audio")
|
38 |
+
self.parser.add_argument('--save_intermediates', default=0, help="whether to save intermediate results")
|
39 |
+
|
40 |
+
def predict(self,
|
41 |
+
driving_audio: Path = Input(description='driving audio, if the file is more than 20 seconds, only the first 20 seconds will be processed for video generation'),
|
42 |
+
talking_head: str = Input(description="choose a talking head", choices=['May', 'Obama1', 'Obama2', 'Nadella', 'McStay'], default='May')
|
43 |
+
) -> Path:
|
44 |
+
|
45 |
+
############################### I/O Settings ##############################
|
46 |
+
# load config files
|
47 |
+
opt = self.parser.parse_args('')
|
48 |
+
opt.driving_audio = str(driving_audio)
|
49 |
+
opt.id = talking_head
|
50 |
+
with open(join('config', opt.id + '.yaml')) as f:
|
51 |
+
config = yaml.safe_load(f)
|
52 |
+
data_root = join('data', opt.id)
|
53 |
+
|
54 |
+
############################ Hyper Parameters #############################
|
55 |
+
h, w, sr, FPS = 512, 512, 16000, 60
|
56 |
+
mouth_indices = np.concatenate([np.arange(4, 11), np.arange(46, 64)])
|
57 |
+
eye_brow_indices = [27, 65, 28, 68, 29, 67, 30, 66, 31, 72, 32, 69, 33, 70, 34, 71]
|
58 |
+
eye_brow_indices = np.array(eye_brow_indices, np.int32)
|
59 |
+
|
60 |
+
############################ Pre-defined Data #############################
|
61 |
+
mean_pts3d = np.load(join(data_root, 'mean_pts3d.npy'))
|
62 |
+
fit_data = np.load(config['dataset_params']['fit_data_path'])
|
63 |
+
pts3d = np.load(config['dataset_params']['pts3d_path']) - mean_pts3d
|
64 |
+
trans = fit_data['trans'][:, :, 0].astype(np.float32)
|
65 |
+
mean_translation = trans.mean(axis=0)
|
66 |
+
candidate_eye_brow = pts3d[10:, eye_brow_indices]
|
67 |
+
std_mean_pts3d = np.load(config['dataset_params']['pts3d_path']).mean(axis=0)
|
68 |
+
# candidates images
|
69 |
+
img_candidates = []
|
70 |
+
for j in range(4):
|
71 |
+
output = imread(join(data_root, 'candidates', f'normalized_full_{j}.jpg'))
|
72 |
+
output = A.pytorch.transforms.ToTensor(normalize={'mean': (0.5, 0.5, 0.5),
|
73 |
+
'std': (0.5, 0.5, 0.5)})(image=output)['image']
|
74 |
+
img_candidates.append(output)
|
75 |
+
img_candidates = torch.cat(img_candidates).unsqueeze(0).cuda()
|
76 |
+
|
77 |
+
# shoulders
|
78 |
+
shoulders = np.load(join(data_root, 'normalized_shoulder_points.npy'))
|
79 |
+
shoulder3D = np.load(join(data_root, 'shoulder_points3D.npy'))[1]
|
80 |
+
ref_trans = trans[1]
|
81 |
+
|
82 |
+
# camera matrix, we always use training set intrinsic parameters.
|
83 |
+
camera = utils.camera()
|
84 |
+
camera_intrinsic = np.load(join(data_root, 'camera_intrinsic.npy')).astype(np.float32)
|
85 |
+
APC_feat_database = np.load(join(data_root, 'APC_feature_base.npy'))
|
86 |
+
|
87 |
+
# load reconstruction data
|
88 |
+
scale = sio.loadmat(join(data_root, 'id_scale.mat'))['scale'][0, 0]
|
89 |
+
Audio2Mel_torch = audio_funcs.Audio2Mel(n_fft=512, hop_length=int(16000 / 120), win_length=int(16000 / 60),
|
90 |
+
sampling_rate=16000,
|
91 |
+
n_mel_channels=80, mel_fmin=90, mel_fmax=7600.0).cuda()
|
92 |
+
|
93 |
+
########################### Experiment Settings ###########################
|
94 |
+
#### user config
|
95 |
+
use_LLE = config['model_params']['APC']['use_LLE']
|
96 |
+
Knear = config['model_params']['APC']['Knear']
|
97 |
+
LLE_percent = config['model_params']['APC']['LLE_percent']
|
98 |
+
headpose_sigma = config['model_params']['Headpose']['sigma']
|
99 |
+
Feat_smooth_sigma = config['model_params']['Audio2Mouth']['smooth']
|
100 |
+
Head_smooth_sigma = config['model_params']['Headpose']['smooth']
|
101 |
+
Feat_center_smooth_sigma, Head_center_smooth_sigma = 0, 0
|
102 |
+
AMP_method = config['model_params']['Audio2Mouth']['AMP'][0]
|
103 |
+
Feat_AMPs = config['model_params']['Audio2Mouth']['AMP'][1:]
|
104 |
+
rot_AMP, trans_AMP = config['model_params']['Headpose']['AMP']
|
105 |
+
shoulder_AMP = config['model_params']['Headpose']['shoulder_AMP']
|
106 |
+
save_feature_maps = config['model_params']['Image2Image']['save_input']
|
107 |
+
|
108 |
+
#### common settings
|
109 |
+
Featopt = FeatureOptions().parse()
|
110 |
+
Headopt = HeadposeOptions().parse()
|
111 |
+
Renderopt = RenderOptions().parse()
|
112 |
+
Featopt.load_epoch = config['model_params']['Audio2Mouth']['ckp_path']
|
113 |
+
Headopt.load_epoch = config['model_params']['Headpose']['ckp_path']
|
114 |
+
Renderopt.dataroot = config['dataset_params']['root']
|
115 |
+
Renderopt.load_epoch = config['model_params']['Image2Image']['ckp_path']
|
116 |
+
Renderopt.size = config['model_params']['Image2Image']['size']
|
117 |
+
|
118 |
+
############################# Load Models #################################
|
119 |
+
print('---------- Loading Model: APC-------------')
|
120 |
+
APC_model = APC_encoder(config['model_params']['APC']['mel_dim'],
|
121 |
+
config['model_params']['APC']['hidden_size'],
|
122 |
+
config['model_params']['APC']['num_layers'],
|
123 |
+
config['model_params']['APC']['residual'])
|
124 |
+
# load all 5 here?
|
125 |
+
APC_model.load_state_dict(torch.load(config['model_params']['APC']['ckp_path']), strict=False)
|
126 |
+
APC_model.cuda()
|
127 |
+
APC_model.eval()
|
128 |
+
print('---------- Loading Model: {} -------------'.format(Featopt.task))
|
129 |
+
Audio2Feature = create_model(Featopt)
|
130 |
+
Audio2Feature.setup(Featopt)
|
131 |
+
Audio2Feature.eval()
|
132 |
+
print('---------- Loading Model: {} -------------'.format(Headopt.task))
|
133 |
+
Audio2Headpose = create_model(Headopt)
|
134 |
+
Audio2Headpose.setup(Headopt)
|
135 |
+
Audio2Headpose.eval()
|
136 |
+
if Headopt.feature_decoder == 'WaveNet':
|
137 |
+
Headopt.A2H_receptive_field = Audio2Headpose.Audio2Headpose.module.WaveNet.receptive_field
|
138 |
+
print('---------- Loading Model: {} -------------'.format(Renderopt.task))
|
139 |
+
facedataset = create_dataset(Renderopt)
|
140 |
+
Feature2Face = create_model(Renderopt)
|
141 |
+
Feature2Face.setup(Renderopt)
|
142 |
+
Feature2Face.eval()
|
143 |
+
visualizer = Visualizer(Renderopt)
|
144 |
+
|
145 |
+
# check audio duration and trim audio
|
146 |
+
extension_name = os.path.basename(opt.driving_audio).split('.')[-1]
|
147 |
+
audio_threshold = 10
|
148 |
+
duration = librosa.get_duration(filename=opt.driving_audio)
|
149 |
+
if duration > audio_threshold:
|
150 |
+
print(f'audio file is longer than {audio_threshold} seconds, trimming the first {audio_threshold} seconds '
|
151 |
+
f'for further processing')
|
152 |
+
ffmpeg_extract_subclip(opt.driving_audio, 0, audio_threshold, targetname=f'shorter_input.{extension_name}')
|
153 |
+
opt.driving_audio = f'shorter_input.{extension_name}'
|
154 |
+
|
155 |
+
# create the results folder
|
156 |
+
audio_name = os.path.basename(opt.driving_audio).split('.')[0]
|
157 |
+
save_root = join('results', opt.id, audio_name)
|
158 |
+
os.makedirs(save_root, exist_ok=True)
|
159 |
+
clean_folder(save_root)
|
160 |
+
out_path = Path(tempfile.mkdtemp()) / "out.mp4"
|
161 |
+
|
162 |
+
############################## Inference ##################################
|
163 |
+
print('Processing audio: {} ...'.format(audio_name))
|
164 |
+
# read audio
|
165 |
+
audio, _ = librosa.load(opt.driving_audio, sr=sr)
|
166 |
+
total_frames = np.int32(audio.shape[0] / sr * FPS)
|
167 |
+
|
168 |
+
#### 1. compute APC features
|
169 |
+
print('1. Computing APC features...')
|
170 |
+
mel80 = utils.compute_mel_one_sequence(audio)
|
171 |
+
mel_nframe = mel80.shape[0]
|
172 |
+
with torch.no_grad():
|
173 |
+
length = torch.Tensor([mel_nframe])
|
174 |
+
mel80_torch = torch.from_numpy(mel80.astype(np.float32)).cuda().unsqueeze(0)
|
175 |
+
hidden_reps = APC_model.forward(mel80_torch, length)[0] # [mel_nframe, 512]
|
176 |
+
hidden_reps = hidden_reps.cpu().numpy()
|
177 |
+
audio_feats = hidden_reps
|
178 |
+
|
179 |
+
#### 2. manifold projection
|
180 |
+
if use_LLE:
|
181 |
+
print('2. Manifold projection...')
|
182 |
+
ind = utils.KNN_with_torch(audio_feats, APC_feat_database, K=Knear)
|
183 |
+
weights, feat_fuse = utils.compute_LLE_projection_all_frame(audio_feats, APC_feat_database, ind,
|
184 |
+
audio_feats.shape[0])
|
185 |
+
audio_feats = audio_feats * (1 - LLE_percent) + feat_fuse * LLE_percent
|
186 |
+
|
187 |
+
#### 3. Audio2Mouth
|
188 |
+
print('3. Audio2Mouth inference...')
|
189 |
+
pred_Feat = Audio2Feature.generate_sequences(audio_feats, sr, FPS, fill_zero=True, opt=Featopt)
|
190 |
+
|
191 |
+
#### 4. Audio2Headpose
|
192 |
+
print('4. Headpose inference...')
|
193 |
+
# set history headposes as zero
|
194 |
+
pre_headpose = np.zeros(Headopt.A2H_wavenet_input_channels, np.float32)
|
195 |
+
pred_Head = Audio2Headpose.generate_sequences(audio_feats, pre_headpose, fill_zero=True, sigma_scale=0.3,
|
196 |
+
opt=Headopt)
|
197 |
+
|
198 |
+
#### 5. Post-Processing
|
199 |
+
print('5. Post-processing...')
|
200 |
+
nframe = min(pred_Feat.shape[0], pred_Head.shape[0])
|
201 |
+
pred_pts3d = np.zeros([nframe, 73, 3])
|
202 |
+
pred_pts3d[:, mouth_indices] = pred_Feat.reshape(-1, 25, 3)[:nframe]
|
203 |
+
|
204 |
+
## mouth
|
205 |
+
pred_pts3d = utils.landmark_smooth_3d(pred_pts3d, Feat_smooth_sigma, area='only_mouth')
|
206 |
+
pred_pts3d = utils.mouth_pts_AMP(pred_pts3d, True, AMP_method, Feat_AMPs)
|
207 |
+
pred_pts3d = pred_pts3d + mean_pts3d
|
208 |
+
pred_pts3d = utils.solve_intersect_mouth(pred_pts3d) # solve intersect lips if exist
|
209 |
+
|
210 |
+
## headpose
|
211 |
+
pred_Head[:, 0:3] *= rot_AMP
|
212 |
+
pred_Head[:, 3:6] *= trans_AMP
|
213 |
+
pred_headpose = utils.headpose_smooth(pred_Head[:, :6], Head_smooth_sigma).astype(np.float32)
|
214 |
+
pred_headpose[:, 3:] += mean_translation
|
215 |
+
pred_headpose[:, 0] += 180
|
216 |
+
|
217 |
+
## compute projected landmarks
|
218 |
+
pred_landmarks = np.zeros([nframe, 73, 2], dtype=np.float32)
|
219 |
+
final_pts3d = np.zeros([nframe, 73, 3], dtype=np.float32)
|
220 |
+
final_pts3d[:] = std_mean_pts3d.copy()
|
221 |
+
final_pts3d[:, 46:64] = pred_pts3d[:nframe, 46:64]
|
222 |
+
for k in tqdm(range(nframe)):
|
223 |
+
ind = k % candidate_eye_brow.shape[0]
|
224 |
+
final_pts3d[k, eye_brow_indices] = candidate_eye_brow[ind] + mean_pts3d[eye_brow_indices]
|
225 |
+
pred_landmarks[k], _, _ = utils.project_landmarks(camera_intrinsic, camera.relative_rotation,
|
226 |
+
camera.relative_translation, scale,
|
227 |
+
pred_headpose[k], final_pts3d[k])
|
228 |
+
|
229 |
+
## Upper Body Motion
|
230 |
+
pred_shoulders = np.zeros([nframe, 18, 2], dtype=np.float32)
|
231 |
+
pred_shoulders3D = np.zeros([nframe, 18, 3], dtype=np.float32)
|
232 |
+
for k in range(nframe):
|
233 |
+
diff_trans = pred_headpose[k][3:] - ref_trans
|
234 |
+
pred_shoulders3D[k] = shoulder3D + diff_trans * shoulder_AMP
|
235 |
+
# project
|
236 |
+
project = camera_intrinsic.dot(pred_shoulders3D[k].T)
|
237 |
+
project[:2, :] /= project[2, :] # divide z
|
238 |
+
pred_shoulders[k] = project[:2, :].T
|
239 |
+
|
240 |
+
#### 6. Image2Image translation & Save resuls
|
241 |
+
print('6. Image2Image translation & Saving results...')
|
242 |
+
for ind in tqdm(range(0, nframe), desc='Image2Image translation inference'):
|
243 |
+
# feature_map: [input_nc, h, w]
|
244 |
+
current_pred_feature_map = facedataset.dataset.get_data_test_mode(pred_landmarks[ind],
|
245 |
+
pred_shoulders[ind],
|
246 |
+
facedataset.dataset.image_pad)
|
247 |
+
input_feature_maps = current_pred_feature_map.unsqueeze(0).cuda()
|
248 |
+
pred_fake = Feature2Face.inference(input_feature_maps, img_candidates)
|
249 |
+
# save results
|
250 |
+
visual_list = [('pred', util.tensor2im(pred_fake[0]))]
|
251 |
+
if save_feature_maps:
|
252 |
+
visual_list += [('input', np.uint8(current_pred_feature_map[0].cpu().numpy() * 255))]
|
253 |
+
visuals = OrderedDict(visual_list)
|
254 |
+
visualizer.save_images(save_root, visuals, str(ind + 1))
|
255 |
+
|
256 |
+
## make videos
|
257 |
+
# generate corresponding audio, reused for all results
|
258 |
+
tmp_audio_path = join(save_root, 'tmp.wav')
|
259 |
+
tmp_audio_clip = audio[: np.int32(nframe * sr / FPS)]
|
260 |
+
librosa.output.write_wav(tmp_audio_path, tmp_audio_clip, sr)
|
261 |
+
|
262 |
+
def write_video_with_audio(audio_path, output_path, prefix='pred_'):
|
263 |
+
fps, fourcc = 60, cv2.VideoWriter_fourcc(*'DIVX')
|
264 |
+
video_tmp_path = join(save_root, 'tmp.avi')
|
265 |
+
out = cv2.VideoWriter(video_tmp_path, fourcc, fps, (Renderopt.loadSize, Renderopt.loadSize))
|
266 |
+
for j in tqdm(range(nframe), position=0, desc='writing video'):
|
267 |
+
img = cv2.imread(join(save_root, prefix + str(j + 1) + '.jpg'))
|
268 |
+
out.write(img)
|
269 |
+
out.release()
|
270 |
+
cmd = 'ffmpeg -i "' + video_tmp_path + '" -i "' + audio_path + '" -codec copy -shortest "' + output_path + '"'
|
271 |
+
subprocess.call(cmd, shell=True)
|
272 |
+
os.remove(video_tmp_path) # remove the template video
|
273 |
+
|
274 |
+
temp_out = 'temp_video.avi'
|
275 |
+
write_video_with_audio(tmp_audio_path, temp_out, 'pred_')
|
276 |
+
# convert to mp4
|
277 |
+
cmd = ("ffmpeg -i "
|
278 |
+
+ temp_out + " -strict -2 "
|
279 |
+
+ str(out_path)
|
280 |
+
)
|
281 |
+
subprocess.call(cmd, shell=True)
|
282 |
+
|
283 |
+
if os.path.exists(tmp_audio_path):
|
284 |
+
os.remove(tmp_audio_path)
|
285 |
+
if os.path.exists(temp_out):
|
286 |
+
os.remove(temp_out)
|
287 |
+
if os.path.exists(f'shorter_input.{extension_name}'):
|
288 |
+
os.remove(f'shorter_input.{extension_name}')
|
289 |
+
if not opt.save_intermediates:
|
290 |
+
_img_paths = list(map(lambda x: str(x), list(Path(save_root).glob('*.jpg'))))
|
291 |
+
for i in tqdm(range(len(_img_paths)), desc='deleting intermediate images'):
|
292 |
+
os.remove(_img_paths[i])
|
293 |
+
|
294 |
+
print('Finish!')
|
295 |
+
|
296 |
+
return out_path
|
297 |
+
|
298 |
+
|
299 |
+
def clean_folder(folder):
|
300 |
+
for filename in os.listdir(folder):
|
301 |
+
file_path = os.path.join(folder, filename)
|
302 |
+
try:
|
303 |
+
if os.path.isfile(file_path) or os.path.islink(file_path):
|
304 |
+
os.unlink(file_path)
|
305 |
+
elif os.path.isdir(file_path):
|
306 |
+
shutil.rmtree(file_path)
|
307 |
+
except Exception as e:
|
308 |
+
print('Failed to delete %s. Reason: %s' % (file_path, e))
|
requirements.txt
ADDED
@@ -0,0 +1,10 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
tqdm
|
2 |
+
librosa==0.7.0
|
3 |
+
scikit_image
|
4 |
+
opencv_python==4.4.0.40
|
5 |
+
scipy
|
6 |
+
dominate
|
7 |
+
albumentations==0.5.2
|
8 |
+
numpy
|
9 |
+
beautifulsoup4
|
10 |
+
scikit-image
|