Spaces:
Sleeping
Sleeping
File size: 15,434 Bytes
e9a044b |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 |
import os, sys, glob
import numpy as np
from collections import OrderedDict
from decord import VideoReader, cpu
import cv2
import torch
import torchvision
sys.path.insert(1, os.path.join(sys.path[0], '..', '..'))
from lvdm.models.samplers.ddim import DDIMSampler
from lvdm.models.samplers.ddim_mp import DDIMSampler as DDIMSampler_mp
def get_views(video_length, window_size=16, stride=4):
num_blocks_time = (video_length - window_size) // stride + 1
views = []
for i in range(num_blocks_time):
t_start = int(i * stride)
t_end = t_start + window_size
views.append((t_start,t_end))
return views
def batch_ddim_sampling(model, cond, noise_shape, n_samples=1, ddim_steps=50, ddim_eta=1.0,\
cfg_scale=1.0, temporal_cfg_scale=None, **kwargs):
ddim_sampler = DDIMSampler(model)
uncond_type = model.uncond_type
batch_size = noise_shape[0]
## construct unconditional guidance
if cfg_scale != 1.0:
if uncond_type == "empty_seq":
prompts = batch_size * [""]
#prompts = N * T * [""] ## if is_imgbatch=True
uc_emb = model.get_learned_conditioning(prompts)
elif uncond_type == "zero_embed":
c_emb = cond["c_crossattn"][0] if isinstance(cond, dict) else cond
uc_emb = torch.zeros_like(c_emb)
## process image embedding token
if hasattr(model, 'embedder'):
uc_img = torch.zeros(noise_shape[0],3,224,224).to(model.device)
## img: b c h w >> b l c
uc_img = model.get_image_embeds(uc_img)
uc_emb = torch.cat([uc_emb, uc_img], dim=1)
if isinstance(cond, dict):
uc = {key:cond[key] for key in cond.keys()}
uc.update({'c_crossattn': [uc_emb]})
else:
uc = uc_emb
else:
uc = None
x_T = None
batch_variants = []
#batch_variants1, batch_variants2 = [], []
for _ in range(n_samples):
if ddim_sampler is not None:
kwargs.update({"clean_cond": True})
samples, _ = ddim_sampler.sample(S=ddim_steps,
conditioning=cond,
batch_size=noise_shape[0],
shape=noise_shape[1:],
verbose=False,
unconditional_guidance_scale=cfg_scale,
unconditional_conditioning=uc,
eta=ddim_eta,
temporal_length=noise_shape[2],
conditional_guidance_scale_temporal=temporal_cfg_scale,
x_T=x_T,
**kwargs
)
## reconstruct from latent to pixel space
batch_images = model.decode_first_stage_2DAE(samples)
batch_variants.append(batch_images)
## batch, <samples>, c, t, h, w
batch_variants = torch.stack(batch_variants, dim=1)
return batch_variants
def batch_ddim_sampling_freenoise(model, cond, noise_shape, n_samples=1, ddim_steps=50, ddim_eta=1.0,\
cfg_scale=1.0, temporal_cfg_scale=None, args=None, x_T_total=None, **kwargs):
ddim_sampler = DDIMSampler(model)
uncond_type = model.uncond_type
batch_size = noise_shape[0]
## construct unconditional guidance
if cfg_scale != 1.0:
if uncond_type == "empty_seq":
prompts = batch_size * [""]
#prompts = N * T * [""] ## if is_imgbatch=True
uc_emb = model.get_learned_conditioning(prompts)
elif uncond_type == "zero_embed":
c_emb = cond["c_crossattn"][0] if isinstance(cond, dict) else cond
uc_emb = torch.zeros_like(c_emb)
## process image embedding token
if hasattr(model, 'embedder'):
uc_img = torch.zeros(noise_shape[0],3,224,224).to(model.device)
## img: b c h w >> b l c
uc_img = model.get_image_embeds(uc_img)
uc_emb = torch.cat([uc_emb, uc_img], dim=1)
if isinstance(cond, dict):
uc = {key:cond[key] for key in cond.keys()}
uc.update({'c_crossattn': [uc_emb]})
else:
uc = uc_emb
else:
uc = None
views = get_views(args.frames, args.window_size, args.window_stride)
batch_variants = []
#batch_variants1, batch_variants2 = [], []
for _ in range(n_samples):
x_T = x_T_total[_]
if ddim_sampler is not None:
kwargs.update({"clean_cond": True})
samples, _ = ddim_sampler.sample(S=ddim_steps,
conditioning=cond,
batch_size=noise_shape[0],
shape=noise_shape[1:],
verbose=False,
unconditional_guidance_scale=cfg_scale,
unconditional_conditioning=uc,
eta=ddim_eta,
temporal_length=noise_shape[2],
conditional_guidance_scale_temporal=temporal_cfg_scale,
x_T=x_T,
context_next=views,
**kwargs
)
## reconstruct from latent to pixel space
batch_images = model.decode_first_stage_2DAE(samples)
batch_variants.append(batch_images)
## batch, <samples>, c, t, h, w
batch_variants = torch.stack(batch_variants, dim=1)
return batch_variants
def batch_ddim_sampling_freenoise_mp(model, cond, noise_shape, n_samples=1, ddim_steps=50, ddim_eta=1.0,\
cfg_scale=1.0, temporal_cfg_scale=None, args=None, x_T_total=None, **kwargs):
ddim_sampler = DDIMSampler_mp(model)
uncond_type = model.uncond_type
batch_size = noise_shape[0]
## construct unconditional guidance
if cfg_scale != 1.0:
if uncond_type == "empty_seq":
prompts = batch_size * [""]
#prompts = N * T * [""] ## if is_imgbatch=True
uc_emb = model.get_learned_conditioning(prompts)
elif uncond_type == "zero_embed":
c_emb = cond["c_crossattn"][0] if isinstance(cond, dict) else cond
uc_emb = torch.zeros_like(c_emb)
## process image embedding token
if hasattr(model, 'embedder'):
uc_img = torch.zeros(noise_shape[0],3,224,224).to(model.device)
## img: b c h w >> b l c
uc_img = model.get_image_embeds(uc_img)
uc_emb = torch.cat([uc_emb, uc_img], dim=1)
if isinstance(cond, dict):
uc = {key:cond[key] for key in cond.keys()}
uc.update({'c_crossattn': [uc_emb]})
else:
uc = uc_emb
else:
uc = None
views = get_views(args.frames, args.window_size, args.window_stride)
conditioning = cond['c_crossattn'][0]
len1 = int(args.frames * 3 // 8)
len2 = args.frames - len1 * 2
cond_diff1 = (conditioning[[1]] - conditioning[[0]]) / (len2 - 1)
cond_list1 = []
for i in range(len2):
cond_list1.append((conditioning[[0]] + cond_diff1 * i).unsqueeze(0))
cond1 = torch.cat([conditioning[[0]].unsqueeze(0).repeat(1, len1, 1, 1), torch.cat(cond_list1, 1), conditioning[[1]].unsqueeze(0).repeat(1, len1, 1, 1)], 1)
cond2 = torch.cat([conditioning[[1]].unsqueeze(0).repeat(1, args.frames, 1, 1)], 1)
cond_all = torch.cat([cond1, cond2], 0)
cond['c_crossattn'] = [cond_all]
batch_variants = []
#batch_variants1, batch_variants2 = [], []
for _ in range(n_samples):
x_T = x_T_total[_]
if ddim_sampler is not None:
kwargs.update({"clean_cond": True})
samples, _ = ddim_sampler.sample(S=ddim_steps,
conditioning=cond,
batch_size=noise_shape[0],
shape=noise_shape[1:],
verbose=False,
unconditional_guidance_scale=cfg_scale,
unconditional_conditioning=uc,
eta=ddim_eta,
temporal_length=noise_shape[2],
conditional_guidance_scale_temporal=temporal_cfg_scale,
x_T=x_T,
context_next=views,
**kwargs
)
## reconstruct from latent to pixel space
batch_images = model.decode_first_stage_2DAE(samples)
batch_variants.append(batch_images)
## batch, <samples>, c, t, h, w
batch_variants = torch.stack(batch_variants, dim=1)
return batch_variants
def get_filelist(data_dir, ext='*'):
file_list = glob.glob(os.path.join(data_dir, '*.%s'%ext))
file_list.sort()
return file_list
def get_dirlist(path):
list = []
if (os.path.exists(path)):
files = os.listdir(path)
for file in files:
m = os.path.join(path,file)
if (os.path.isdir(m)):
list.append(m)
list.sort()
return list
def load_model_checkpoint(model, ckpt):
def load_checkpoint(model, ckpt, full_strict):
state_dict = torch.load(ckpt, map_location="cpu")
try:
## deepspeed
new_pl_sd = OrderedDict()
for key in state_dict['module'].keys():
new_pl_sd[key[16:]]=state_dict['module'][key]
model.load_state_dict(new_pl_sd, strict=full_strict)
except:
if "state_dict" in list(state_dict.keys()):
state_dict = state_dict["state_dict"]
model.load_state_dict(state_dict, strict=full_strict)
return model
load_checkpoint(model, ckpt, full_strict=True)
print('>>> model checkpoint loaded.')
return model
def load_prompts(prompt_file):
f = open(prompt_file, 'r')
prompt_list = []
for idx, line in enumerate(f.readlines()):
l = line.strip()
if len(l) != 0:
prompt_list.append(l)
f.close()
return prompt_list
def load_prompts_mp(prompt_file):
f = open(prompt_file, 'r')
prompt_list = []
for idx, line in enumerate(f.readlines()):
l = []
line = line.strip()
prompts = line.split(';')
for prompt in prompts:
prompt = prompt.strip()
if len(prompt) != 0:
l.append(prompt)
if len(l) != 0:
prompt_list.append(l)
f.close()
print(prompt_list)
return prompt_list
def load_video_batch(filepath_list, frame_stride, video_size=(256,256), video_frames=16):
'''
Notice about some special cases:
1. video_frames=-1 means to take all the frames (with fs=1)
2. when the total video frames is less than required, padding strategy will be used (repreated last frame)
'''
fps_list = []
batch_tensor = []
assert frame_stride > 0, "valid frame stride should be a positive integer!"
for filepath in filepath_list:
padding_num = 0
vidreader = VideoReader(filepath, ctx=cpu(0), width=video_size[1], height=video_size[0])
fps = vidreader.get_avg_fps()
total_frames = len(vidreader)
max_valid_frames = (total_frames-1) // frame_stride + 1
if video_frames < 0:
## all frames are collected: fs=1 is a must
required_frames = total_frames
frame_stride = 1
else:
required_frames = video_frames
query_frames = min(required_frames, max_valid_frames)
frame_indices = [frame_stride*i for i in range(query_frames)]
## [t,h,w,c] -> [c,t,h,w]
frames = vidreader.get_batch(frame_indices)
frame_tensor = torch.tensor(frames.asnumpy()).permute(3, 0, 1, 2).float()
frame_tensor = (frame_tensor / 255. - 0.5) * 2
if max_valid_frames < required_frames:
padding_num = required_frames - max_valid_frames
frame_tensor = torch.cat([frame_tensor, *([frame_tensor[:,-1:,:,:]]*padding_num)], dim=1)
print(f'{os.path.split(filepath)[1]} is not long enough: {padding_num} frames padded.')
batch_tensor.append(frame_tensor)
sample_fps = int(fps/frame_stride)
fps_list.append(sample_fps)
return torch.stack(batch_tensor, dim=0)
from PIL import Image
def load_image_batch(filepath_list, image_size=(256,256)):
batch_tensor = []
for filepath in filepath_list:
_, filename = os.path.split(filepath)
_, ext = os.path.splitext(filename)
if ext == '.mp4':
vidreader = VideoReader(filepath, ctx=cpu(0), width=image_size[1], height=image_size[0])
frame = vidreader.get_batch([0])
img_tensor = torch.tensor(frame.asnumpy()).squeeze(0).permute(2, 0, 1).float()
elif ext == '.png' or ext == '.jpg':
img = Image.open(filepath).convert("RGB")
rgb_img = np.array(img, np.float32)
#bgr_img = cv2.imread(filepath, cv2.IMREAD_COLOR)
#bgr_img = cv2.cvtColor(bgr_img, cv2.COLOR_BGR2RGB)
rgb_img = cv2.resize(rgb_img, (image_size[1],image_size[0]), interpolation=cv2.INTER_LINEAR)
img_tensor = torch.from_numpy(rgb_img).permute(2, 0, 1).float()
else:
print(f'ERROR: <{ext}> image loading only support format: [mp4], [png], [jpg]')
raise NotImplementedError
img_tensor = (img_tensor / 255. - 0.5) * 2
batch_tensor.append(img_tensor)
return torch.stack(batch_tensor, dim=0)
def save_videos(batch_tensors, savedir, filenames, fps=10):
# b,samples,c,t,h,w
n_samples = batch_tensors.shape[1]
for idx, vid_tensor in enumerate(batch_tensors):
video = vid_tensor.detach().cpu()
video = torch.clamp(video.float(), -1., 1.)
video = video.permute(2, 0, 1, 3, 4) # t,n,c,h,w
frame_grids = [torchvision.utils.make_grid(framesheet, nrow=int(n_samples)) for framesheet in video] #[3, 1*h, n*w]
grid = torch.stack(frame_grids, dim=0) # stack in temporal dim [t, 3, n*h, w]
grid = (grid + 1.0) / 2.0
grid = (grid * 255).to(torch.uint8).permute(0, 2, 3, 1)
savepath = os.path.join(savedir, f"{filenames[idx]}.mp4")
torchvision.io.write_video(savepath, grid, fps=fps, video_codec='h264', options={'crf': '10'})
|