Spaces:
Sleeping
Sleeping
File size: 12,488 Bytes
31bd91f 81003d9 31bd91f 5f77949 31bd91f c2ab61f 31bd91f | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 | import os
import cv2
import yaml
import numpy as np
import warnings
from skimage import img_as_ubyte
import safetensors
import safetensors.torch
warnings.filterwarnings('ignore')
import imageio
import torch
import torchvision
from src.facerender.modules.keypoint_detector import HEEstimator, KPDetector
from src.facerender.modules.mapping import MappingNet
from src.facerender.modules.generator import OcclusionAwareGenerator, OcclusionAwareSPADEGenerator
from src.facerender.modules.make_animation import make_animation
from pydub import AudioSegment
# from src.utils.face_enhancer import enhancer_generator_with_len as face_enhancer
from src.utils.face_enhancer import enhancer_with_len as face_enhancer
from src.utils.paste_pic import paste_pic
from src.utils.videoio import save_video_with_watermark
import time
class AnimateFromCoeff():
def __init__(self, sadtalker_path, device, bf16=False):
with open(sadtalker_path['facerender_yaml']) as f:
config = yaml.safe_load(f)
generator = OcclusionAwareSPADEGenerator(**config['model_params']['generator_params'],
**config['model_params']['common_params'])
kp_extractor = KPDetector(**config['model_params']['kp_detector_params'],
**config['model_params']['common_params'])
he_estimator = HEEstimator(**config['model_params']['he_estimator_params'],
**config['model_params']['common_params'])
mapping = MappingNet(**config['model_params']['mapping_params'])
generator.to(device)
kp_extractor.to(device)
he_estimator.to(device)
mapping.to(device)
for param in generator.parameters():
param.requires_grad = False
for param in kp_extractor.parameters():
param.requires_grad = False
for param in he_estimator.parameters():
param.requires_grad = False
for param in mapping.parameters():
param.requires_grad = False
if sadtalker_path is not None:
if 'checkpoint' in sadtalker_path: # use safe tensor
self.load_cpk_facevid2vid_safetensor(sadtalker_path['checkpoint'], kp_detector=kp_extractor, generator=generator, he_estimator=None)
else:
self.load_cpk_facevid2vid(sadtalker_path['free_view_checkpoint'], kp_detector=kp_extractor, generator=generator, he_estimator=he_estimator)
else:
raise AttributeError("Checkpoint should be specified for video head pose estimator.")
if sadtalker_path['mappingnet_checkpoint'] is not None:
self.load_cpk_mapping(sadtalker_path['mappingnet_checkpoint'], mapping=mapping)
else:
raise AttributeError("Checkpoint should be specified for video head pose estimator.")
def optimize_cpu_model(model):
return model
self.kp_extractor = optimize_cpu_model(kp_extractor)
self.generator = optimize_cpu_model(generator)
self.he_estimator = optimize_cpu_model(he_estimator)
self.mapping = optimize_cpu_model(mapping)
self.kp_extractor.eval()
self.generator.eval()
self.he_estimator.eval()
self.mapping.eval()
self.bf16 = bf16
if self.bf16:
import intel_extension_for_pytorch as ipex
self.generator = ipex.optimize(self.generator, dtype=torch.bfloat16)
self.device = device
def load_cpk_facevid2vid_safetensor(self, checkpoint_path, generator=None,
kp_detector=None, he_estimator=None,
device="cpu"):
checkpoint = safetensors.torch.load_file(checkpoint_path)
if generator is not None:
x_generator = {}
for k,v in checkpoint.items():
if 'generator' in k:
x_generator[k.replace('generator.', '')] = v
generator.load_state_dict(x_generator)
if kp_detector is not None:
x_generator = {}
for k,v in checkpoint.items():
if 'kp_extractor' in k:
x_generator[k.replace('kp_extractor.', '')] = v
kp_detector.load_state_dict(x_generator)
if he_estimator is not None:
x_generator = {}
for k,v in checkpoint.items():
if 'he_estimator' in k:
x_generator[k.replace('he_estimator.', '')] = v
he_estimator.load_state_dict(x_generator)
return None
def load_cpk_facevid2vid(self, checkpoint_path, generator=None, discriminator=None,
kp_detector=None, he_estimator=None, optimizer_generator=None,
optimizer_discriminator=None, optimizer_kp_detector=None,
optimizer_he_estimator=None, device="cpu"):
checkpoint = torch.load(checkpoint_path, map_location=torch.device(device))
if generator is not None:
generator.load_state_dict(checkpoint['generator'])
if kp_detector is not None:
kp_detector.load_state_dict(checkpoint['kp_detector'])
if he_estimator is not None:
he_estimator.load_state_dict(checkpoint['he_estimator'])
if discriminator is not None:
try:
discriminator.load_state_dict(checkpoint['discriminator'])
except:
print ('No discriminator in the state-dict. Dicriminator will be randomly initialized')
if optimizer_generator is not None:
optimizer_generator.load_state_dict(checkpoint['optimizer_generator'])
if optimizer_discriminator is not None:
try:
optimizer_discriminator.load_state_dict(checkpoint['optimizer_discriminator'])
except RuntimeError as e:
print ('No discriminator optimizer in the state-dict. Optimizer will be not initialized')
if optimizer_kp_detector is not None:
optimizer_kp_detector.load_state_dict(checkpoint['optimizer_kp_detector'])
if optimizer_he_estimator is not None:
optimizer_he_estimator.load_state_dict(checkpoint['optimizer_he_estimator'])
return checkpoint['epoch']
def load_cpk_mapping(self, checkpoint_path, mapping=None, discriminator=None,
optimizer_mapping=None, optimizer_discriminator=None, device='cpu'):
checkpoint = torch.load(checkpoint_path, map_location=torch.device(device))
if mapping is not None:
mapping.load_state_dict(checkpoint['mapping'])
if discriminator is not None:
discriminator.load_state_dict(checkpoint['discriminator'])
if optimizer_mapping is not None:
optimizer_mapping.load_state_dict(checkpoint['optimizer_mapping'])
if optimizer_discriminator is not None:
optimizer_discriminator.load_state_dict(checkpoint['optimizer_discriminator'])
return checkpoint['epoch']
@torch.inference_mode()
def generate(self, x, video_save_dir, pic_path, crop_info, enhancer=None, background_enhancer=None, preprocess='crop', img_size=256, rank=0, p_num=1, bf16=False):
source_image=x['source_image'].type(torch.FloatTensor)
source_semantics=x['source_semantics'].type(torch.FloatTensor)
target_semantics=x['target_semantics_list'].type(torch.FloatTensor)
source_image=source_image.to(self.device)
source_semantics=source_semantics.to(self.device)
target_semantics=target_semantics.to(self.device)
if 'yaw_c_seq' in x:
yaw_c_seq = x['yaw_c_seq'].type(torch.FloatTensor)
yaw_c_seq = x['yaw_c_seq'].to(self.device)
else:
yaw_c_seq = None
if 'pitch_c_seq' in x:
pitch_c_seq = x['pitch_c_seq'].type(torch.FloatTensor)
pitch_c_seq = x['pitch_c_seq'].to(self.device)
else:
pitch_c_seq = None
if 'roll_c_seq' in x:
roll_c_seq = x['roll_c_seq'].type(torch.FloatTensor)
roll_c_seq = x['roll_c_seq'].to(self.device)
else:
roll_c_seq = None
frame_num = x['frame_num']
with torch.autocast(device_type="cuda" if "cuda" in str(self.device) else "cpu", dtype=torch.float16):
predictions_video = make_animation(source_image, source_semantics, target_semantics,
self.generator, self.kp_extractor, self.he_estimator, self.mapping,
yaw_c_seq, pitch_c_seq, roll_c_seq, use_exp = True, rank=rank, p_num=p_num, bf16=bf16)
if predictions_video is None and rank != 0:
if enhancer is None:
exit(0)
while not os.path.exists("workspace/rendering_video.mp4"):
time.sleep(0.5)
with open("workspace/rendering_video.mp4", 'r') as f:
full_video_path = f.readline()
else:
predictions_video = predictions_video.reshape((-1,)+predictions_video.shape[2:])
predictions_video = predictions_video[:frame_num]
# Vectorized tensor to numpy conversion
predictions_video_np = predictions_video.data.cpu().numpy()
predictions_video_np = np.transpose(predictions_video_np, [0, 2, 3, 1]).astype(np.float32)
video = list(predictions_video_np)
result = img_as_ubyte(video)
### the generated video is 256x256, so we keep the aspect ratio,
original_size = crop_info[0]
if original_size:
result = [ cv2.resize(result_i,(img_size, int(img_size * original_size[1]/original_size[0]) )) for result_i in result ]
video_name = x['video_name'] + '.mp4'
path = os.path.join(video_save_dir, 'temp_'+video_name)
imageio.mimsave(path, result, fps=float(25))
av_path = os.path.join(video_save_dir, video_name)
return_path = av_path
audio_path = x['audio_path']
audio_name = os.path.splitext(os.path.split(audio_path)[-1])[0]
new_audio_path = os.path.join(video_save_dir, audio_name+'.wav')
start_time = 0
# cog will not keep the .mp3 filename
sound = AudioSegment.from_file(audio_path)
frames = frame_num
end_time = start_time + frames*1/25*1000
word1=sound.set_frame_rate(16000)
word = word1[start_time:end_time]
word.export(new_audio_path, format="wav")
save_video_with_watermark(path, new_audio_path, av_path, watermark= False)
print(f'The generated video is named {video_save_dir}/{video_name}')
if 'full' in preprocess.lower():
# only add watermark to the full image.
video_name_full = x['video_name'] + '_full.mp4'
full_video_path = os.path.join(video_save_dir, video_name_full)
return_path = full_video_path
paste_pic(path, pic_path, crop_info, new_audio_path, full_video_path, extended_crop= True if 'ext' in preprocess.lower() else False)
print(f'The generated video is named {video_save_dir}/{video_name_full}')
else:
full_video_path = av_path
os.makedirs("workspace", exist_ok=True)
with open("workspace/rendering_video.mp4", 'w') as f:
f.write(full_video_path)
#### paste back then enhancers
if enhancer:
video_name_enhancer = x['video_name'] + '_enhanced.mp4'
enhanced_path = os.path.join(video_save_dir, 'temp_'+video_name_enhancer)
av_path_enhancer = os.path.join(video_save_dir, video_name_enhancer)
return_path = av_path_enhancer
enhanced_images = face_enhancer(full_video_path, method=enhancer, bg_upsampler=background_enhancer,
rank=rank, p_num=p_num, bf16=bf16)
imageio.mimsave(enhanced_path, enhanced_images, fps=float(25))
save_video_with_watermark(enhanced_path, new_audio_path, av_path_enhancer, watermark= False)
print(f'The generated video is named {video_save_dir}/{video_name_enhancer}')
os.remove(enhanced_path)
os.remove(path)
os.remove(new_audio_path)
return return_path
|