waveydaveygravy
commited on
Commit
•
9531098
1
Parent(s):
094e6e0
Upload 6 files
Browse files- bo_1resized.jpg +0 -0
- bo_1resized.mp4 +0 -0
- bo_1resized_ang_bo_1resized.mp4 +0 -0
- demoworking.py +470 -0
- obama3_hap_M003_neu_1_001.mp4 +0 -0
- scarlett_ang_bo_1resized.mp4 +0 -0
bo_1resized.jpg
ADDED
bo_1resized.mp4
ADDED
Binary file (991 kB). View file
|
|
bo_1resized_ang_bo_1resized.mp4
ADDED
Binary file (156 kB). View file
|
|
demoworking.py
ADDED
@@ -0,0 +1,470 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
#@title demo.py with fixed paths
|
2 |
+
|
3 |
+
import os
|
4 |
+
import numpy as np
|
5 |
+
import torch
|
6 |
+
import yaml
|
7 |
+
from modules.generator import OcclusionAwareSPADEGeneratorEam
|
8 |
+
from modules.keypoint_detector import KPDetector, HEEstimator
|
9 |
+
import argparse
|
10 |
+
import imageio
|
11 |
+
from modules.transformer import Audio2kpTransformerBBoxQDeepPrompt as Audio2kpTransformer
|
12 |
+
from modules.prompt import EmotionDeepPrompt, EmotionalDeformationTransformer
|
13 |
+
from scipy.io import wavfile
|
14 |
+
|
15 |
+
from modules.model_transformer import get_rotation_matrix, keypoint_transformation
|
16 |
+
from skimage import io, img_as_float32
|
17 |
+
from skimage.transform import resize
|
18 |
+
import torchaudio
|
19 |
+
import soundfile as sf
|
20 |
+
from scipy.spatial import ConvexHull
|
21 |
+
|
22 |
+
import torch.nn.functional as F
|
23 |
+
import glob
|
24 |
+
from tqdm import tqdm
|
25 |
+
import gzip
|
26 |
+
|
27 |
+
emo_label = ['ang', 'con', 'dis', 'fea', 'hap', 'neu', 'sad', 'sur']
|
28 |
+
emo_label_full = ['angry', 'contempt', 'disgusted', 'fear', 'happy', 'neutral', 'sad', 'surprised']
|
29 |
+
latent_dim = 16
|
30 |
+
|
31 |
+
MEL_PARAMS_25 = {
|
32 |
+
"n_mels": 80,
|
33 |
+
"n_fft": 2048,
|
34 |
+
"win_length": 640,
|
35 |
+
"hop_length": 640
|
36 |
+
}
|
37 |
+
|
38 |
+
to_melspec = torchaudio.transforms.MelSpectrogram(**MEL_PARAMS_25)
|
39 |
+
mean, std = -4, 4
|
40 |
+
|
41 |
+
expU = torch.from_numpy(np.load('/content/EAT_code/expPCAnorm_fin/U_mead.npy')[:,:32])
|
42 |
+
expmean = torch.from_numpy(np.load('/content/EAT_code/expPCAnorm_fin/mean_mead.npy'))
|
43 |
+
|
44 |
+
root_wav = '/content/EAT_code/demo/video_processed/bo_1resized'
|
45 |
+
def normalize_kp(kp_source, kp_driving, kp_driving_initial,
|
46 |
+
use_relative_movement=True, use_relative_jacobian=True):
|
47 |
+
|
48 |
+
kp_new = {k: v for k, v in kp_driving.items()}
|
49 |
+
if use_relative_movement:
|
50 |
+
kp_value_diff = (kp_driving['value'] - kp_driving_initial['value'])
|
51 |
+
kp_new['value'] = kp_value_diff + kp_source['value']
|
52 |
+
|
53 |
+
if use_relative_jacobian:
|
54 |
+
jacobian_diff = torch.matmul(kp_driving['jacobian'], torch.inverse(kp_driving_initial['jacobian']))
|
55 |
+
kp_new['jacobian'] = torch.matmul(jacobian_diff, kp_source['jacobian'])
|
56 |
+
|
57 |
+
return kp_new
|
58 |
+
|
59 |
+
def _load_tensor(data):
|
60 |
+
wave_path = data
|
61 |
+
wave, sr = sf.read(wave_path)
|
62 |
+
wave_tensor = torch.from_numpy(wave).float()
|
63 |
+
return wave_tensor
|
64 |
+
|
65 |
+
def build_model(config, device_ids=[0]):
|
66 |
+
generator = OcclusionAwareSPADEGeneratorEam(**config['model_params']['generator_params'],
|
67 |
+
**config['model_params']['common_params'])
|
68 |
+
if torch.cuda.is_available():
|
69 |
+
print('cuda is available')
|
70 |
+
generator.to(device_ids[0])
|
71 |
+
|
72 |
+
kp_detector = KPDetector(**config['model_params']['kp_detector_params'],
|
73 |
+
**config['model_params']['common_params'])
|
74 |
+
|
75 |
+
if torch.cuda.is_available():
|
76 |
+
kp_detector.to(device_ids[0])
|
77 |
+
|
78 |
+
|
79 |
+
audio2kptransformer = Audio2kpTransformer(**config['model_params']['audio2kp_params'], face_ea=True)
|
80 |
+
|
81 |
+
if torch.cuda.is_available():
|
82 |
+
audio2kptransformer.to(device_ids[0])
|
83 |
+
|
84 |
+
sidetuning = EmotionalDeformationTransformer(**config['model_params']['audio2kp_params'])
|
85 |
+
|
86 |
+
if torch.cuda.is_available():
|
87 |
+
sidetuning.to(device_ids[0])
|
88 |
+
|
89 |
+
emotionprompt = EmotionDeepPrompt()
|
90 |
+
|
91 |
+
if torch.cuda.is_available():
|
92 |
+
emotionprompt.to(device_ids[0])
|
93 |
+
|
94 |
+
return generator, kp_detector, audio2kptransformer, sidetuning, emotionprompt
|
95 |
+
|
96 |
+
|
97 |
+
def prepare_test_data(img_path, audio_path, opt, emotype, use_otherimg=True):
|
98 |
+
# sr,_ = wavfile.read(audio_path)
|
99 |
+
|
100 |
+
if use_otherimg:
|
101 |
+
source_latent = np.load(img_path.replace('cropped', 'latent')[:-4]+'.npy', allow_pickle=True)
|
102 |
+
else:
|
103 |
+
source_latent = np.load(img_path.replace('images', 'latent')[:-9]+'.npy', allow_pickle=True)
|
104 |
+
he_source = {}
|
105 |
+
for k in source_latent[1].keys():
|
106 |
+
he_source[k] = torch.from_numpy(source_latent[1][k][0]).unsqueeze(0).cuda()
|
107 |
+
|
108 |
+
# source images
|
109 |
+
source_img = img_as_float32(io.imread(img_path)).transpose((2, 0, 1))
|
110 |
+
asp = os.path.basename(audio_path)[:-4]
|
111 |
+
|
112 |
+
# latent code
|
113 |
+
y_trg = emo_label.index(emotype)
|
114 |
+
z_trg = torch.randn(latent_dim)
|
115 |
+
|
116 |
+
# driving latent
|
117 |
+
latent_path_driving = f'{root_wav}/latent_evp_25/{asp}.npy'
|
118 |
+
pose_gz = gzip.GzipFile(f'{root_wav}/poseimg/{asp}.npy.gz', 'r')
|
119 |
+
poseimg = np.load(pose_gz)
|
120 |
+
deepfeature = np.load(f'{root_wav}/deepfeature32/{asp}.npy')
|
121 |
+
driving_latent = np.load(latent_path_driving[:-4]+'.npy', allow_pickle=True)
|
122 |
+
he_driving = driving_latent[1]
|
123 |
+
|
124 |
+
# gt frame number
|
125 |
+
frames = glob.glob(f'{root_wav}/images_evp_25/cropped/*.jpg')
|
126 |
+
num_frames = len(frames)
|
127 |
+
|
128 |
+
wave_tensor = _load_tensor(audio_path)
|
129 |
+
if len(wave_tensor.shape) > 1:
|
130 |
+
wave_tensor = wave_tensor[:, 0]
|
131 |
+
mel_tensor = to_melspec(wave_tensor)
|
132 |
+
mel_tensor = (torch.log(1e-5 + mel_tensor) - mean) / std
|
133 |
+
name_len = min(mel_tensor.shape[1], poseimg.shape[0], deepfeature.shape[0])
|
134 |
+
|
135 |
+
audio_frames = []
|
136 |
+
poseimgs = []
|
137 |
+
deep_feature = []
|
138 |
+
|
139 |
+
pad, deep_pad = np.load('/content/EAT_code/pad.npy', allow_pickle=True)
|
140 |
+
|
141 |
+
if name_len < num_frames:
|
142 |
+
diff = num_frames - name_len
|
143 |
+
if diff > 2:
|
144 |
+
print(f"Attention: the frames are {diff} more than name_len, we will use name_len to replace num_frames")
|
145 |
+
num_frames=name_len
|
146 |
+
for k in he_driving.keys():
|
147 |
+
he_driving[k] = he_driving[k][:name_len, :]
|
148 |
+
for rid in range(0, num_frames):
|
149 |
+
audio = []
|
150 |
+
poses = []
|
151 |
+
deeps = []
|
152 |
+
for i in range(rid - opt['num_w'], rid + opt['num_w'] + 1):
|
153 |
+
if i < 0:
|
154 |
+
audio.append(pad)
|
155 |
+
poses.append(poseimg[0])
|
156 |
+
deeps.append(deep_pad)
|
157 |
+
elif i >= name_len:
|
158 |
+
audio.append(pad)
|
159 |
+
poses.append(poseimg[-1])
|
160 |
+
deeps.append(deep_pad)
|
161 |
+
else:
|
162 |
+
audio.append(mel_tensor[:, i])
|
163 |
+
poses.append(poseimg[i])
|
164 |
+
deeps.append(deepfeature[i])
|
165 |
+
|
166 |
+
audio_frames.append(torch.stack(audio, dim=1))
|
167 |
+
poseimgs.append(poses)
|
168 |
+
deep_feature.append(deeps)
|
169 |
+
audio_frames = torch.stack(audio_frames, dim=0)
|
170 |
+
poseimgs = torch.from_numpy(np.array(poseimgs))
|
171 |
+
deep_feature = torch.from_numpy(np.array(deep_feature)).to(torch.float)
|
172 |
+
return audio_frames, poseimgs, deep_feature, source_img, he_source, he_driving, num_frames, y_trg, z_trg, latent_path_driving
|
173 |
+
|
174 |
+
def load_ckpt(ckpt, kp_detector, generator, audio2kptransformer, sidetuning, emotionprompt):
|
175 |
+
checkpoint = torch.load(ckpt, map_location=torch.device('cpu'))
|
176 |
+
if audio2kptransformer is not None:
|
177 |
+
audio2kptransformer.load_state_dict(checkpoint['audio2kptransformer'])
|
178 |
+
if generator is not None:
|
179 |
+
generator.load_state_dict(checkpoint['generator'])
|
180 |
+
if kp_detector is not None:
|
181 |
+
kp_detector.load_state_dict(checkpoint['kp_detector'])
|
182 |
+
if sidetuning is not None:
|
183 |
+
sidetuning.load_state_dict(checkpoint['sidetuning'])
|
184 |
+
if emotionprompt is not None:
|
185 |
+
emotionprompt.load_state_dict(checkpoint['emotionprompt'])
|
186 |
+
|
187 |
+
import cv2
|
188 |
+
import dlib
|
189 |
+
from tqdm import tqdm
|
190 |
+
from skimage import transform as tf
|
191 |
+
detector = dlib.get_frontal_face_detector()
|
192 |
+
predictor = dlib.shape_predictor('/content/EAT_code/demo/shape_predictor_68_face_landmarks.dat')
|
193 |
+
|
194 |
+
def shape_to_np(shape, dtype="int"):
|
195 |
+
# initialize the list of (x, y)-coordinates
|
196 |
+
coords = np.zeros((shape.num_parts, 2), dtype=dtype)
|
197 |
+
|
198 |
+
# loop over all facial landmarks and convert them
|
199 |
+
# to a 2-tuple of (x, y)-coordinates
|
200 |
+
for i in range(0, shape.num_parts):
|
201 |
+
coords[i] = (shape.part(i).x, shape.part(i).y)
|
202 |
+
|
203 |
+
# return the list of (x, y)-coordinates
|
204 |
+
return coords
|
205 |
+
|
206 |
+
def crop_image(image_path, out_path):
|
207 |
+
template = np.load('/content/EAT_code/demo/bo_1resized_template.npy')
|
208 |
+
image = cv2.imread(image_path)
|
209 |
+
gray = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)
|
210 |
+
rects = detector(gray, 1) #detect human face
|
211 |
+
if len(rects) != 1:
|
212 |
+
return 0
|
213 |
+
for (j, rect) in enumerate(rects):
|
214 |
+
shape = predictor(gray, rect) #detect 68 points
|
215 |
+
shape = shape_to_np(shape)
|
216 |
+
|
217 |
+
pts2 = np.float32(template[:47,:])
|
218 |
+
pts1 = np.float32(shape[:47,:]) #eye and nose
|
219 |
+
tform = tf.SimilarityTransform()
|
220 |
+
tform.estimate( pts2, pts1) #Set the transformation matrix with the explicit parameters.
|
221 |
+
|
222 |
+
dst = tf.warp(image, tform, output_shape=(256, 256))
|
223 |
+
|
224 |
+
dst = np.array(dst * 255, dtype=np.uint8)
|
225 |
+
|
226 |
+
cv2.imwrite(out_path, dst)
|
227 |
+
|
228 |
+
def preprocess_imgs(allimgs, tmp_allimgs_cropped):
|
229 |
+
name_cropped = []
|
230 |
+
for path in tmp_allimgs_cropped:
|
231 |
+
name_cropped.append(os.path.basename(path))
|
232 |
+
for path in allimgs:
|
233 |
+
if os.path.basename(path) in name_cropped:
|
234 |
+
continue
|
235 |
+
else:
|
236 |
+
out_path = path.replace('imgs1/', 'imgs_cropped1/')
|
237 |
+
crop_image(path, out_path)
|
238 |
+
|
239 |
+
from sync_batchnorm import DataParallelWithCallback
|
240 |
+
def load_checkpoints_extractor(config_path, checkpoint_path, cpu=False):
|
241 |
+
|
242 |
+
with open(config_path) as f:
|
243 |
+
config = yaml.load(f, Loader=yaml.FullLoader)
|
244 |
+
|
245 |
+
kp_detector = KPDetector(**config['model_params']['kp_detector_params'],
|
246 |
+
**config['model_params']['common_params'])
|
247 |
+
if not cpu:
|
248 |
+
kp_detector.cuda()
|
249 |
+
|
250 |
+
he_estimator = HEEstimator(**config['model_params']['he_estimator_params'],
|
251 |
+
**config['model_params']['common_params'])
|
252 |
+
if not cpu:
|
253 |
+
he_estimator.cuda()
|
254 |
+
|
255 |
+
if cpu:
|
256 |
+
checkpoint = torch.load(checkpoint_path, map_location=torch.device('cpu'))
|
257 |
+
else:
|
258 |
+
checkpoint = torch.load(checkpoint_path)
|
259 |
+
|
260 |
+
kp_detector.load_state_dict(checkpoint['kp_detector'])
|
261 |
+
he_estimator.load_state_dict(checkpoint['he_estimator'])
|
262 |
+
|
263 |
+
if not cpu:
|
264 |
+
kp_detector = DataParallelWithCallback(kp_detector)
|
265 |
+
he_estimator = DataParallelWithCallback(he_estimator)
|
266 |
+
|
267 |
+
kp_detector.eval()
|
268 |
+
he_estimator.eval()
|
269 |
+
|
270 |
+
return kp_detector, he_estimator
|
271 |
+
|
272 |
+
def estimate_latent(driving_video, kp_detector, he_estimator):
|
273 |
+
with torch.no_grad():
|
274 |
+
predictions = []
|
275 |
+
driving = torch.tensor(np.array(driving_video)[np.newaxis].astype(np.float32)).permute(0, 4, 1, 2, 3).cuda()
|
276 |
+
kp_canonical = kp_detector(driving[:, :, 0])
|
277 |
+
he_drivings = {'yaw': [], 'pitch': [], 'roll': [], 't': [], 'exp': []}
|
278 |
+
|
279 |
+
for frame_idx in range(driving.shape[2]):
|
280 |
+
driving_frame = driving[:, :, frame_idx]
|
281 |
+
he_driving = he_estimator(driving_frame)
|
282 |
+
for k in he_drivings.keys():
|
283 |
+
he_drivings[k].append(he_driving[k])
|
284 |
+
return [kp_canonical, he_drivings]
|
285 |
+
|
286 |
+
def extract_keypoints(extract_list):
|
287 |
+
kp_detector, he_estimator = load_checkpoints_extractor(config_path='/content/EAT_code/config/vox-256-spade.yaml', checkpoint_path='/content/EAT_code/ckpt/pretrain_new_274.pth.tar')
|
288 |
+
if not os.path.exists('./demo/imgs_latent/'):
|
289 |
+
os.makedirs('./demo/imgs_latent/')
|
290 |
+
for imgname in tqdm(extract_list):
|
291 |
+
path_frames = [imgname]
|
292 |
+
filesname=os.path.basename(imgname)[:-4]
|
293 |
+
if os.path.exists(f'./demo/imgs_latent/'+filesname+'.npy'):
|
294 |
+
continue
|
295 |
+
driving_frames = []
|
296 |
+
for im in path_frames:
|
297 |
+
driving_frames.append(imageio.imread(im))
|
298 |
+
driving_video = [resize(frame, (256, 256))[..., :3] for frame in driving_frames]
|
299 |
+
|
300 |
+
kc, he = estimate_latent(driving_video, kp_detector, he_estimator)
|
301 |
+
kc = kc['value'].cpu().numpy()
|
302 |
+
for k in he:
|
303 |
+
he[k] = torch.cat(he[k]).cpu().numpy()
|
304 |
+
np.save('./demo/imgs_latent/'+filesname, [kc, he])
|
305 |
+
|
306 |
+
def preprocess_cropped_imgs(allimgs_cropped):
|
307 |
+
extract_list = []
|
308 |
+
for img_path in allimgs_cropped:
|
309 |
+
if not os.path.exists(img_path.replace('cropped', 'latent')[:-4]+'.npy'):
|
310 |
+
extract_list.append(img_path)
|
311 |
+
if len(extract_list) > 0:
|
312 |
+
print('=========', "Extract latent keypoints from New image", '======')
|
313 |
+
extract_keypoints(extract_list)
|
314 |
+
|
315 |
+
def test(ckpt, emotype, save_dir=" "):
|
316 |
+
# with open("config/vox-transformer2.yaml") as f:
|
317 |
+
with open("/content/EAT_code/config/deepprompt_eam3d_st_tanh_304_3090_all.yaml") as f:
|
318 |
+
config = yaml.load(f, Loader=yaml.FullLoader)
|
319 |
+
cur_path = os.getcwd()
|
320 |
+
generator, kp_detector, audio2kptransformer, sidetuning, emotionprompt = build_model(config)
|
321 |
+
load_ckpt(ckpt, kp_detector=kp_detector, generator=generator, audio2kptransformer=audio2kptransformer, sidetuning=sidetuning, emotionprompt=emotionprompt)
|
322 |
+
|
323 |
+
audio2kptransformer.eval()
|
324 |
+
generator.eval()
|
325 |
+
kp_detector.eval()
|
326 |
+
sidetuning.eval()
|
327 |
+
emotionprompt.eval()
|
328 |
+
|
329 |
+
all_wavs2 = [f'{root_wav}/{os.path.basename(root_wav)}.wav']
|
330 |
+
allimg = glob.glob('/content/EAT_code/demo/imgs1/*.jpg')
|
331 |
+
tmp_allimg_cropped = glob.glob('/content/EAT_code/demo/imgs_cropped1/*.jpg')
|
332 |
+
preprocess_imgs(allimg, tmp_allimg_cropped) # crop and align images
|
333 |
+
|
334 |
+
allimg_cropped = glob.glob('/content/EAT_code/demo/imgs_cropped1/*.jpg')
|
335 |
+
preprocess_cropped_imgs(allimg_cropped) # extract latent keypoints if necessary
|
336 |
+
|
337 |
+
for ind in tqdm(range(len(all_wavs2))):
|
338 |
+
for img_path in tqdm(allimg_cropped):
|
339 |
+
audio_path = all_wavs2[ind]
|
340 |
+
# read in data
|
341 |
+
audio_frames, poseimgs, deep_feature, source_img, he_source, he_driving, num_frames, y_trg, z_trg, latent_path_driving = prepare_test_data(img_path, audio_path, config['model_params']['audio2kp_params'], emotype)
|
342 |
+
|
343 |
+
|
344 |
+
with torch.no_grad():
|
345 |
+
source_img = torch.from_numpy(source_img).unsqueeze(0).cuda()
|
346 |
+
kp_canonical = kp_detector(source_img, with_feature=True) # {'value': value, 'jacobian': jacobian}
|
347 |
+
kp_cano = kp_canonical['value']
|
348 |
+
|
349 |
+
x = {}
|
350 |
+
x['mel'] = audio_frames.unsqueeze(1).unsqueeze(0).cuda()
|
351 |
+
x['z_trg'] = z_trg.unsqueeze(0).cuda()
|
352 |
+
x['y_trg'] = torch.tensor(y_trg, dtype=torch.long).cuda().reshape(1)
|
353 |
+
x['pose'] = poseimgs.cuda()
|
354 |
+
x['deep'] = deep_feature.cuda().unsqueeze(0)
|
355 |
+
x['he_driving'] = {'yaw': torch.from_numpy(he_driving['yaw']).cuda().unsqueeze(0),
|
356 |
+
'pitch': torch.from_numpy(he_driving['pitch']).cuda().unsqueeze(0),
|
357 |
+
'roll': torch.from_numpy(he_driving['roll']).cuda().unsqueeze(0),
|
358 |
+
't': torch.from_numpy(he_driving['t']).cuda().unsqueeze(0),
|
359 |
+
}
|
360 |
+
|
361 |
+
### emotion prompt
|
362 |
+
emoprompt, deepprompt = emotionprompt(x)
|
363 |
+
a2kp_exps = []
|
364 |
+
emo_exps = []
|
365 |
+
T = 5
|
366 |
+
if T == 1:
|
367 |
+
for i in range(x['mel'].shape[1]):
|
368 |
+
xi = {}
|
369 |
+
xi['mel'] = x['mel'][:,i,:,:,:].unsqueeze(1)
|
370 |
+
xi['z_trg'] = x['z_trg']
|
371 |
+
xi['y_trg'] = x['y_trg']
|
372 |
+
xi['pose'] = x['pose'][i,:,:,:,:].unsqueeze(0)
|
373 |
+
xi['deep'] = x['deep'][:,i,:,:,:].unsqueeze(1)
|
374 |
+
xi['he_driving'] = {'yaw': x['he_driving']['yaw'][:,i,:].unsqueeze(0),
|
375 |
+
'pitch': x['he_driving']['pitch'][:,i,:].unsqueeze(0),
|
376 |
+
'roll': x['he_driving']['roll'][:,i,:].unsqueeze(0),
|
377 |
+
't': x['he_driving']['t'][:,i,:].unsqueeze(0),
|
378 |
+
}
|
379 |
+
he_driving_emo_xi, input_st_xi = audio2kptransformer(xi, kp_canonical, emoprompt=emoprompt, deepprompt=deepprompt, side=True) # {'yaw': yaw, 'pitch': pitch, 'roll': roll, 't': t, 'exp': exp}
|
380 |
+
emo_exp = sidetuning(input_st_xi, emoprompt, deepprompt)
|
381 |
+
a2kp_exps.append(he_driving_emo_xi['emo'])
|
382 |
+
emo_exps.append(emo_exp)
|
383 |
+
elif T is not None:
|
384 |
+
for i in range(x['mel'].shape[1]//T+1):
|
385 |
+
if i*T >= x['mel'].shape[1]:
|
386 |
+
break
|
387 |
+
xi = {}
|
388 |
+
xi['mel'] = x['mel'][:,i*T:(i+1)*T,:,:,:]
|
389 |
+
xi['z_trg'] = x['z_trg']
|
390 |
+
xi['y_trg'] = x['y_trg']
|
391 |
+
xi['pose'] = x['pose'][i*T:(i+1)*T,:,:,:,:]
|
392 |
+
xi['deep'] = x['deep'][:,i*T:(i+1)*T,:,:,:]
|
393 |
+
xi['he_driving'] = {'yaw': x['he_driving']['yaw'][:,i*T:(i+1)*T,:],
|
394 |
+
'pitch': x['he_driving']['pitch'][:,i*T:(i+1)*T,:],
|
395 |
+
'roll': x['he_driving']['roll'][:,i*T:(i+1)*T,:],
|
396 |
+
't': x['he_driving']['t'][:,i*T:(i+1)*T,:],
|
397 |
+
}
|
398 |
+
he_driving_emo_xi, input_st_xi = audio2kptransformer(xi, kp_canonical, emoprompt=emoprompt, deepprompt=deepprompt, side=True) # {'yaw': yaw, 'pitch': pitch, 'roll': roll, 't': t, 'exp': exp}
|
399 |
+
emo_exp = sidetuning(input_st_xi, emoprompt, deepprompt)
|
400 |
+
a2kp_exps.append(he_driving_emo_xi['emo'])
|
401 |
+
emo_exps.append(emo_exp)
|
402 |
+
|
403 |
+
if T is None:
|
404 |
+
he_driving_emo, input_st = audio2kptransformer(x, kp_canonical, emoprompt=emoprompt, deepprompt=deepprompt, side=True) # {'yaw': yaw, 'pitch': pitch, 'roll': roll, 't': t, 'exp': exp}
|
405 |
+
emo_exps = sidetuning(input_st, emoprompt, deepprompt).reshape(-1, 45)
|
406 |
+
else:
|
407 |
+
he_driving_emo = {}
|
408 |
+
he_driving_emo['emo'] = torch.cat(a2kp_exps, dim=0)
|
409 |
+
emo_exps = torch.cat(emo_exps, dim=0).reshape(-1, 45)
|
410 |
+
|
411 |
+
exp = he_driving_emo['emo']
|
412 |
+
device = exp.get_device()
|
413 |
+
exp = torch.mm(exp, expU.t().to(device))
|
414 |
+
exp = exp + expmean.expand_as(exp).to(device)
|
415 |
+
exp = exp + emo_exps
|
416 |
+
|
417 |
+
|
418 |
+
source_area = ConvexHull(kp_cano[0].cpu().numpy()).volume
|
419 |
+
exp = exp * source_area
|
420 |
+
|
421 |
+
he_new_driving = {'yaw': torch.from_numpy(he_driving['yaw']).cuda(),
|
422 |
+
'pitch': torch.from_numpy(he_driving['pitch']).cuda(),
|
423 |
+
'roll': torch.from_numpy(he_driving['roll']).cuda(),
|
424 |
+
't': torch.from_numpy(he_driving['t']).cuda(),
|
425 |
+
'exp': exp}
|
426 |
+
he_driving['exp'] = torch.from_numpy(he_driving['exp']).cuda()
|
427 |
+
|
428 |
+
kp_source = keypoint_transformation(kp_canonical, he_source, False)
|
429 |
+
mean_source = torch.mean(kp_source['value'], dim=1)[0]
|
430 |
+
kp_driving = keypoint_transformation(kp_canonical, he_new_driving, False)
|
431 |
+
mean_driving = torch.mean(torch.mean(kp_driving['value'], dim=1), dim=0)
|
432 |
+
kp_driving['value'] = kp_driving['value']+(mean_source-mean_driving).unsqueeze(0).unsqueeze(0)
|
433 |
+
bs = kp_source['value'].shape[0]
|
434 |
+
predictions_gen = []
|
435 |
+
for i in tqdm(range(num_frames)):
|
436 |
+
kp_si = {}
|
437 |
+
kp_si['value'] = kp_source['value'][0].unsqueeze(0)
|
438 |
+
kp_di = {}
|
439 |
+
kp_di['value'] = kp_driving['value'][i].unsqueeze(0)
|
440 |
+
generated = generator(source_img, kp_source=kp_si, kp_driving=kp_di, prompt=emoprompt)
|
441 |
+
predictions_gen.append(
|
442 |
+
(np.transpose(generated['prediction'].data.cpu().numpy(), [0, 2, 3, 1])[0] * 255).astype(np.uint8))
|
443 |
+
|
444 |
+
log_dir = save_dir
|
445 |
+
os.makedirs(os.path.join(log_dir, "temp"), exist_ok=True)
|
446 |
+
|
447 |
+
f_name = os.path.basename(img_path[:-4]) + "_" + emotype + "_" + os.path.basename(latent_path_driving)[:-4] + ".mp4"
|
448 |
+
video_path = os.path.join(log_dir, "temp", f_name)
|
449 |
+
imageio.mimsave(video_path, predictions_gen, fps=25.0)
|
450 |
+
|
451 |
+
save_video = os.path.join(log_dir, f_name)
|
452 |
+
cmd = r'ffmpeg -loglevel error -y -i "%s" -i "%s" -vcodec copy -shortest "%s"' % (video_path, audio_path, save_video)
|
453 |
+
os.system(cmd)
|
454 |
+
os.remove(video_path)
|
455 |
+
|
456 |
+
if __name__ == '__main__':
|
457 |
+
argparser = argparse.ArgumentParser()
|
458 |
+
argparser.add_argument("--save_dir", type=str, default="/content/EAT_code/Results ", help="path of the output video")
|
459 |
+
argparser.add_argument("--name", type=str, default="deepprompt_eam3d_all_final_313", help="path of the output video")
|
460 |
+
argparser.add_argument("--emo", type=str, default="hap", help="emotion type ('ang', 'con', 'dis', 'fea', 'hap', 'neu', 'sad', 'sur')")
|
461 |
+
argparser.add_argument("--root_wav", type=str, default='./demo/video_processed/M003_neu_1_001', help="emotion type ('ang', 'con', 'dis', 'fea', 'hap', 'neu', 'sad', 'sur')")
|
462 |
+
args = argparser.parse_args()
|
463 |
+
|
464 |
+
root_wav=args.root_wav
|
465 |
+
|
466 |
+
if len(args.name) > 1:
|
467 |
+
name = args.name
|
468 |
+
print(name)
|
469 |
+
test(f'/content/EAT_code/ckpt/deepprompt_eam3d_all_final_313.pth.tar', args.emo, save_dir=f'./demo/output/{name}/')
|
470 |
+
|
obama3_hap_M003_neu_1_001.mp4
ADDED
Binary file (80.9 kB). View file
|
|
scarlett_ang_bo_1resized.mp4
ADDED
Binary file (182 kB). View file
|
|