Upload animate.py
Browse files- src/facerender/animate.py +262 -0
src/facerender/animate.py
ADDED
@@ -0,0 +1,262 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
import cv2
|
3 |
+
import yaml
|
4 |
+
import numpy as np
|
5 |
+
import warnings
|
6 |
+
from skimage import img_as_ubyte
|
7 |
+
|
8 |
+
warnings.filterwarnings('ignore')
|
9 |
+
|
10 |
+
|
11 |
+
import imageio
|
12 |
+
import torch
|
13 |
+
import torchvision
|
14 |
+
|
15 |
+
|
16 |
+
from src.facerender.modules.keypoint_detector import HEEstimator, KPDetector
|
17 |
+
from src.facerender.modules.mapping import MappingNet
|
18 |
+
from src.facerender.modules.generator import OcclusionAwareGenerator, OcclusionAwareSPADEGenerator
|
19 |
+
from src.facerender.modules.make_animation import make_animation
|
20 |
+
|
21 |
+
from pydub import AudioSegment
|
22 |
+
from src.utils.face_enhancer import enhancer as face_enhancer
|
23 |
+
from src.utils.paste_pic import paste_pic
|
24 |
+
from src.utils.videoio import save_video_with_watermark
|
25 |
+
|
26 |
+
|
27 |
+
class AnimateFromCoeff():
|
28 |
+
|
29 |
+
def __init__(self, free_view_checkpoint, mapping_checkpoint,
|
30 |
+
config_path, device):
|
31 |
+
|
32 |
+
with open(config_path) as f:
|
33 |
+
config = yaml.safe_load(f)
|
34 |
+
|
35 |
+
generator = OcclusionAwareSPADEGenerator(**config['model_params']['generator_params'],
|
36 |
+
**config['model_params']['common_params'])
|
37 |
+
kp_extractor = KPDetector(**config['model_params']['kp_detector_params'],
|
38 |
+
**config['model_params']['common_params'])
|
39 |
+
he_estimator = HEEstimator(**config['model_params']['he_estimator_params'],
|
40 |
+
**config['model_params']['common_params'])
|
41 |
+
mapping = MappingNet(**config['model_params']['mapping_params'])
|
42 |
+
|
43 |
+
|
44 |
+
generator.to(device)
|
45 |
+
kp_extractor.to(device)
|
46 |
+
he_estimator.to(device)
|
47 |
+
mapping.to(device)
|
48 |
+
|
49 |
+
# Wrap models in DataParallel for multi-GPU support
|
50 |
+
if torch.cuda.device_count() > 1:
|
51 |
+
print(f"Using {torch.cuda.device_count()} GPUs")
|
52 |
+
generator = torch.nn.DataParallel(generator)
|
53 |
+
kp_extractor = torch.nn.DataParallel(kp_extractor)
|
54 |
+
he_estimator = torch.nn.DataParallel(he_estimator)
|
55 |
+
mapping = torch.nn.DataParallel(mapping)
|
56 |
+
|
57 |
+
for param in generator.parameters():
|
58 |
+
param.requires_grad = False
|
59 |
+
for param in kp_extractor.parameters():
|
60 |
+
param.requires_grad = False
|
61 |
+
for param in he_estimator.parameters():
|
62 |
+
param.requires_grad = False
|
63 |
+
for param in mapping.parameters():
|
64 |
+
param.requires_grad = False
|
65 |
+
|
66 |
+
if free_view_checkpoint is not None:
|
67 |
+
self.load_cpk_facevid2vid(free_view_checkpoint, kp_detector=kp_extractor, generator=generator, he_estimator=he_estimator)
|
68 |
+
else:
|
69 |
+
raise AttributeError("Checkpoint should be specified for video head pose estimator.")
|
70 |
+
|
71 |
+
if mapping_checkpoint is not None:
|
72 |
+
self.load_cpk_mapping(mapping_checkpoint, mapping=mapping)
|
73 |
+
else:
|
74 |
+
raise AttributeError("Checkpoint should be specified for video head pose estimator.")
|
75 |
+
|
76 |
+
self.kp_extractor = kp_extractor
|
77 |
+
self.generator = generator
|
78 |
+
self.he_estimator = he_estimator
|
79 |
+
self.mapping = mapping
|
80 |
+
|
81 |
+
self.kp_extractor.eval()
|
82 |
+
self.generator.eval()
|
83 |
+
self.he_estimator.eval()
|
84 |
+
self.mapping.eval()
|
85 |
+
|
86 |
+
self.device = device
|
87 |
+
|
88 |
+
def load_cpk_facevid2vid(self, checkpoint_path, generator=None, discriminator=None,
|
89 |
+
kp_detector=None, he_estimator=None, optimizer_generator=None,
|
90 |
+
optimizer_discriminator=None, optimizer_kp_detector=None,
|
91 |
+
optimizer_he_estimator=None, device="cpu"):
|
92 |
+
checkpoint = torch.load(checkpoint_path, map_location=torch.device(device))
|
93 |
+
|
94 |
+
def adjust_state_dict(state_dict, model):
|
95 |
+
new_state_dict = {}
|
96 |
+
if isinstance(model, torch.nn.DataParallel):
|
97 |
+
for k, v in state_dict.items():
|
98 |
+
new_key = f"module.{k}" # Add 'module.' prefix
|
99 |
+
new_state_dict[new_key] = v
|
100 |
+
else:
|
101 |
+
new_state_dict = state_dict # Keep original state_dict for single-GPU models
|
102 |
+
return new_state_dict
|
103 |
+
|
104 |
+
if generator is not None:
|
105 |
+
generator_state_dict = adjust_state_dict(checkpoint['generator'], generator)
|
106 |
+
generator.load_state_dict(generator_state_dict)
|
107 |
+
if kp_detector is not None:
|
108 |
+
kp_state_dict = adjust_state_dict(checkpoint['kp_detector'],kp_detector)
|
109 |
+
kp_detector.load_state_dict(kp_state_dict)
|
110 |
+
if he_estimator is not None:
|
111 |
+
he_state_dict = adjust_state_dict(checkpoint['he_estimator'],he_estimator)
|
112 |
+
he_estimator.load_state_dict(he_state_dict)
|
113 |
+
if discriminator is not None:
|
114 |
+
try:
|
115 |
+
discriminator_dict =adjust_state_dict(checkpoint['discriminator'],discriminator)
|
116 |
+
discriminator.load_state_dict(discriminator_dict)
|
117 |
+
except:
|
118 |
+
print ('No discriminator in the state-dict. Dicriminator will be randomly initialized')
|
119 |
+
if optimizer_generator is not None:
|
120 |
+
optimizer_generator_dict =adjust_state_dict(checkpoint['optimizer_generator'],optimizer_generator)
|
121 |
+
optimizer_generator.load_state_dict(optimizer_generator_dict)
|
122 |
+
if optimizer_discriminator is not None:
|
123 |
+
try:
|
124 |
+
optimizer_discriminator_dict = adjust_state_dict(checkpoint['optimizer_discriminator'],optimizer_discriminator)
|
125 |
+
optimizer_discriminator.load_state_dict(optimizer_discriminator_dict)
|
126 |
+
except RuntimeError as e:
|
127 |
+
print ('No discriminator optimizer in the state-dict. Optimizer will be not initialized')
|
128 |
+
if optimizer_kp_detector is not None:
|
129 |
+
optimizer_kp_detector_dict = adjust_state_dict(checkpoint['optimizer_kp_detector'],optimizer_kp_detector)
|
130 |
+
optimizer_kp_detector.load_state_dict(optimizer_kp_detector_dict)
|
131 |
+
if optimizer_he_estimator is not None:
|
132 |
+
optimizer_he_estimator_dict = adjust_state_dict(checkpoint['optimizer_he_estimator'],optimizer_he_estimator)
|
133 |
+
optimizer_he_estimator.load_state_dict(optimizer_he_estimator_dict)
|
134 |
+
|
135 |
+
return checkpoint['epoch']
|
136 |
+
|
137 |
+
def load_cpk_mapping(self, checkpoint_path, mapping=None, discriminator=None,
|
138 |
+
optimizer_mapping=None, optimizer_discriminator=None, device='cpu'):
|
139 |
+
checkpoint = torch.load(checkpoint_path, map_location=torch.device(device))
|
140 |
+
|
141 |
+
def adjust_state_dict(state_dict, model):
|
142 |
+
new_state_dict = {}
|
143 |
+
if isinstance(model, torch.nn.DataParallel):
|
144 |
+
for k, v in state_dict.items():
|
145 |
+
new_key = f"module.{k}" # Add 'module.' prefix
|
146 |
+
new_state_dict[new_key] = v
|
147 |
+
else:
|
148 |
+
new_state_dict = state_dict # Keep original state_dict for single-GPU models
|
149 |
+
return new_state_dict
|
150 |
+
|
151 |
+
if mapping is not None:
|
152 |
+
mapping_dict = adjust_state_dict(checkpoint['mapping'],mapping)
|
153 |
+
mapping.load_state_dict(mapping_dict)
|
154 |
+
if discriminator is not None:
|
155 |
+
discriminator_dict = adjust_state_dict(checkpoint['discriminator'],discriminator)
|
156 |
+
discriminator.load_state_dict(discriminator_dict)
|
157 |
+
if optimizer_mapping is not None:
|
158 |
+
optimizer_mapping_dict = adjust_state_dict(checkpoint['optimizer_mapping'],optimizer_mapping)
|
159 |
+
optimizer_mapping.load_state_dict(optimizer_mapping_dict)
|
160 |
+
if optimizer_discriminator is not None:
|
161 |
+
optimizer_discriminator_dict = adjust_state_dict(checkpoint['optimizer_discriminator'],optimizer_discriminator)
|
162 |
+
optimizer_discriminator.load_state_dict(optimizer_discriminator_dict)
|
163 |
+
|
164 |
+
return checkpoint['epoch']
|
165 |
+
|
166 |
+
def generate(self, x, video_save_dir, pic_path, crop_info, enhancer=None, background_enhancer=None, preprocess='crop'):
|
167 |
+
source_image=x['source_image'].type(torch.FloatTensor)
|
168 |
+
source_semantics=x['source_semantics'].type(torch.FloatTensor)
|
169 |
+
target_semantics=x['target_semantics_list'].type(torch.FloatTensor)
|
170 |
+
source_image=source_image.to(self.device)
|
171 |
+
source_semantics=source_semantics.to(self.device)
|
172 |
+
target_semantics=target_semantics.to(self.device)
|
173 |
+
if 'yaw_c_seq' in x:
|
174 |
+
yaw_c_seq = x['yaw_c_seq'].type(torch.FloatTensor)
|
175 |
+
yaw_c_seq = x['yaw_c_seq'].to(self.device)
|
176 |
+
else:
|
177 |
+
yaw_c_seq = None
|
178 |
+
if 'pitch_c_seq' in x:
|
179 |
+
pitch_c_seq = x['pitch_c_seq'].type(torch.FloatTensor)
|
180 |
+
pitch_c_seq = x['pitch_c_seq'].to(self.device)
|
181 |
+
else:
|
182 |
+
pitch_c_seq = None
|
183 |
+
if 'roll_c_seq' in x:
|
184 |
+
roll_c_seq = x['roll_c_seq'].type(torch.FloatTensor)
|
185 |
+
roll_c_seq = x['roll_c_seq'].to(self.device)
|
186 |
+
else:
|
187 |
+
roll_c_seq = None
|
188 |
+
|
189 |
+
frame_num = x['frame_num']
|
190 |
+
|
191 |
+
predictions_video = make_animation(source_image, source_semantics, target_semantics,
|
192 |
+
self.generator, self.kp_extractor, self.he_estimator, self.mapping,
|
193 |
+
yaw_c_seq, pitch_c_seq, roll_c_seq, use_exp = True)
|
194 |
+
|
195 |
+
predictions_video = predictions_video.reshape((-1,)+predictions_video.shape[2:])
|
196 |
+
predictions_video = predictions_video[:frame_num]
|
197 |
+
|
198 |
+
video = []
|
199 |
+
for idx in range(predictions_video.shape[0]):
|
200 |
+
image = predictions_video[idx]
|
201 |
+
image = np.transpose(image.data.cpu().numpy(), [1, 2, 0]).astype(np.float32)
|
202 |
+
video.append(image)
|
203 |
+
result = img_as_ubyte(video)
|
204 |
+
|
205 |
+
### the generated video is 256x256, so we keep the aspect ratio,
|
206 |
+
original_size = crop_info[0]
|
207 |
+
if original_size:
|
208 |
+
result = [ cv2.resize(result_i,(256, int(256.0 * original_size[1]/original_size[0]) )) for result_i in result ]
|
209 |
+
|
210 |
+
video_name = x['video_name'] + '.mp4'
|
211 |
+
path = os.path.join(video_save_dir, 'temp_'+video_name)
|
212 |
+
|
213 |
+
imageio.mimsave(path, result, fps=float(25))
|
214 |
+
|
215 |
+
av_path = os.path.join(video_save_dir, video_name)
|
216 |
+
return_path = av_path
|
217 |
+
|
218 |
+
audio_path = x['audio_path']
|
219 |
+
audio_name = os.path.splitext(os.path.split(audio_path)[-1])[0]
|
220 |
+
new_audio_path = os.path.join(video_save_dir, audio_name+'.wav')
|
221 |
+
print('new_audio_path',new_audio_path)
|
222 |
+
start_time = 0
|
223 |
+
# cog will not keep the .mp3 filename
|
224 |
+
sound = AudioSegment.from_file(audio_path)
|
225 |
+
frames = frame_num
|
226 |
+
end_time = start_time + frames*1/25*1000
|
227 |
+
word1=sound.set_frame_rate(16000)
|
228 |
+
word = word1[start_time:end_time]
|
229 |
+
word.export(new_audio_path, format="wav")
|
230 |
+
|
231 |
+
base64_video,temp_file_path = save_video_with_watermark(path, new_audio_path, av_path, watermark= False)
|
232 |
+
print(f'The generated video is named {video_name} in {video_save_dir}')
|
233 |
+
|
234 |
+
if preprocess.lower() == 'full':
|
235 |
+
# only add watermark to the full image.
|
236 |
+
video_name_full = x['video_name'] + '_full.mp4'
|
237 |
+
full_video_path = os.path.join(video_save_dir, video_name_full)
|
238 |
+
return_path = full_video_path
|
239 |
+
base64_video,temp_file_path = paste_pic(path, pic_path, crop_info, new_audio_path, full_video_path)
|
240 |
+
print(f'The generated video is named {video_save_dir}/{video_name_full}')
|
241 |
+
else:
|
242 |
+
full_video_path = av_path
|
243 |
+
|
244 |
+
#### paste back then enhancers
|
245 |
+
if enhancer:
|
246 |
+
video_name_enhancer = x['video_name'] + '_enhanced.mp4'
|
247 |
+
enhanced_path = os.path.join(video_save_dir, 'temp_'+video_name_enhancer)
|
248 |
+
av_path_enhancer = os.path.join(video_save_dir, video_name_enhancer)
|
249 |
+
return_path = av_path_enhancer
|
250 |
+
enhanced_images = face_enhancer(temp_file_path, method=enhancer, bg_upsampler=background_enhancer)
|
251 |
+
|
252 |
+
imageio.mimsave(enhanced_path, enhanced_images, fps=float(25))
|
253 |
+
|
254 |
+
base64_video,temp_file_path = save_video_with_watermark(enhanced_path, new_audio_path, av_path_enhancer, watermark= False)
|
255 |
+
print(f'The generated video is named {video_save_dir}/{video_name_enhancer}')
|
256 |
+
os.remove(enhanced_path)
|
257 |
+
|
258 |
+
os.remove(path)
|
259 |
+
# os.remove(new_audio_path)
|
260 |
+
|
261 |
+
return return_path,base64_video,temp_file_path, new_audio_path
|
262 |
+
|