File size: 12,716 Bytes
2b534de | 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 | # --------------------------------------------------------- Config -------------------------------------------------
num_workers :int = 1
DDIM_STEPS = 50
BATCH_SIZE = 1
FIXED_CODE = False
# for vis
SAVE_INTERMEDIATES = True
NUM_grid_in_a_column = 5
# ------------------------------------------------------------------------------------------------------------------------
import sys
import os
from pathlib import Path
cur_dir = os.path.dirname(os.path.abspath(__file__))
from confs import *
import torch
import numpy as np
from omegaconf import OmegaConf
from PIL import Image
from tqdm import tqdm
from einops import rearrange
from torchvision.utils import make_grid
from my_py_lib.image_util import imgs_2_grid_A,img_paths_2_grid_A
from pytorch_lightning import seed_everything
from torch import autocast
from contextlib import nullcontext
import torchvision
from ldm.models.diffusion.ddpm import LatentDiffusion
from ldm.util import instantiate_from_config
from ldm.models.diffusion.ddim import DDIMSampler
from Dataset_custom import Dataset_custom
from MoE import offload_unused_tasks__LD
from ldm.models.diffusion.ddpm import LandmarkExtractor
from my_py_lib.torch_util import cleanup_gpu_memory
from gen_lmk_and_mask import gen_lmk_and_mask
# ------------------------------------------------------------------------------------------------------------------------
DDIM_ETA = 0.0
SCALE = 3.0
PRECISION = "full" # "full" or "autocast"
H = 512
W = 512
C = 4
F = 8
# ------------------------------------------------------------------------------------------------------------------------
def load_first_stage_from_sd14(model: LatentDiffusion, sd14_path: Path) -> None:
print(f"Loading first_stage_model from {sd14_path}")
sd14 = torch.load(str(sd14_path), map_location="cpu")
if isinstance(sd14, dict) and "state_dict" in sd14:
sd14_sd = sd14["state_dict"]
else:
sd14_sd = sd14
prefixes = ["first_stage_model.", "model.first_stage_model."]
fs_sd = {}
for prefix in prefixes:
for k, v in sd14_sd.items():
if k.startswith(prefix):
fs_sd[k[len(prefix):]] = v
if fs_sd:
break
if not fs_sd:
raise RuntimeError("Could not find first_stage_model weights in SD v1-4 checkpoint.")
model.first_stage_model.load_state_dict(fs_sd, strict=True)
def save_sample_by_decode(x, model, base_path, segment_id, intermediate_num):
x = model.decode_first_stage(x)
x = torch.clamp((x + 1.0) / 2.0, min=0.0, max=1.0)
x = x.cpu().permute(0, 2, 3, 1).numpy()
for i in range(len(x)):
img = Image.fromarray((x[i] * 255).astype(np.uint8))
save_path = Path(base_path) / segment_id
save_path.mkdir(parents=True, exist_ok=True)
img.save(save_path / f"{intermediate_num}.png")
def get_tensor_clip(normalize=True, toTensor=True):
transform_list = []
if toTensor:
transform_list += [torchvision.transforms.ToTensor()]
if normalize:
transform_list += [
torchvision.transforms.Normalize(
(0.48145466, 0.4578275, 0.40821073),
(0.26862954, 0.26130258, 0.27577711),
)
]
return torchvision.transforms.Compose(transform_list)
def load_model_from_config(ckpt, verbose=1):
if 1:
ckpt = Path(ckpt)
print(f"Loading model from {ckpt}")
pl_sd = torch.load(str(ckpt), map_location="cpu")
if isinstance(pl_sd, dict) and "state_dict" in pl_sd:
sd = pl_sd["state_dict"]
else:
sd = pl_sd
else:
print("DEBUG_skip_load_ckpt")
if 1:
from init_model import get_moe
model: LatentDiffusion = get_moe()
model.ptsM_Generator = LandmarkExtractor(include_visualizer=True, img_256_mode=False)
cleanup_gpu_memory()
if 1:
m, u = model.load_state_dict(sd, strict=False)
if len(m) > 0 and verbose:
print("missing keys:")
pretty_print_torch_module_keys(m)
if len(u) > 0 and verbose:
print("unexpected keys:")
pretty_print_torch_module_keys(u)
load_first_stage_from_sd14(model, SD14_localpath)
offload_unused_tasks__LD(model, TASK, method="del") # for save cuda mem
model.cuda()
model.eval()
return model
def load_pairs(pair_list, tgt, ref):
if tgt and ref:
pairs = [(tgt, ref), ]
elif pair_list:
pairs = []
with open(pair_list, "r") as f:
for line_num, line in enumerate(f, start=1):
line = line.strip()
if not line or line.startswith("#"):
continue
parts = line.split(" ")
if len(parts) != 2:
raise ValueError(f"Invalid pair list line {line_num}: expected white-space-separated tgt/ref. got {parts=}")
pairs.append((parts[0], parts[1]))
else:
raise ValueError("No input pairs provided. Use --tgt/--ref or --pair-list.")
print(f"{pairs=}")
return pairs
def un_norm(x):
return (x + 1.0) / 2.0
def un_norm_clip(x1):
x = x1 * 1.0
reduce = False
if len(x.shape) == 3:
x = x.unsqueeze(0)
reduce = True
x[:, 0, :, :] = x[:, 0, :, :] * 0.26862954 + 0.48145466
x[:, 1, :, :] = x[:, 1, :, :] * 0.26130258 + 0.4578275
x[:, 2, :, :] = x[:, 2, :, :] * 0.27577711 + 0.40821073
if reduce:
x = x.squeeze(0)
return x
if __name__ == "__main__":
pairs = load_pairs(args.pair_list, args.tgt, args.ref)
out_dir = Path(args.out_dir)
result_path = out_dir / "results"
grid_path = out_dir / "grid"
inter_path = out_dir / "intermediates"
inter_pred_path = inter_path / "pred_x0"
inter_noised_path = inter_path / "noised"
out_dir.mkdir(parents=False, exist_ok=True)
result_path.mkdir(parents=False, exist_ok=True)
grid_path.mkdir(parents=False, exist_ok=True)
inter_path.mkdir(parents=False, exist_ok=True)
if SAVE_INTERMEDIATES:
inter_pred_path.mkdir(parents=False, exist_ok=True)
inter_noised_path.mkdir(parents=False, exist_ok=True)
paths_tgt = [p[0] for p in pairs]
paths_ref = [p[1] for p in pairs]
gen_lmk_and_mask(paths_tgt + paths_ref)
seed_everything(42)
model: LatentDiffusion = load_model_from_config(PRETRAIN_CKPT_PATH, )
device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")
model = model.to(device)
sampler = DDIMSampler(model)
dataset = Dataset_custom(
"test",
task=TASK,
paths_tgt=paths_tgt,
paths_ref=paths_ref,
)
dataloader = torch.utils.data.DataLoader(
dataset,
batch_size=BATCH_SIZE,
num_workers=num_workers,
pin_memory=True,
shuffle=False,
drop_last=False,
)
start_code = None
if FIXED_CODE:
start_code = torch.randn([BATCH_SIZE, C, H // F, W // F], device=device)
precision_scope = autocast if PRECISION == "autocast" else nullcontext
grids = []
grid_stems = []
with torch.no_grad():
with precision_scope("cuda"):
with model.ema_scope():
for test_batch, prior, test_model_kwargs, out_stem_batch in tqdm(dataloader):
model.set_task(test_model_kwargs)
bs = test_batch.shape[0]
batch_ = {
**test_model_kwargs,
"GT": torch.zeros_like(test_model_kwargs["inpaint_image"]),
}
batch_, c = model.get_input_and_conditioning(batch_, device=device)
z_inpaint = batch_["z4_inpaint"]
z_inpaint_mask = batch_["tgt_mask_64"]
z_ref = batch_["z_ref"]
z9 = batch_["z9"]
uc = None
if SCALE != 1.0:
uc = model.learnable_vector[TASK].repeat(bs, 1, 1)
shape = [C, H // F, W // F]
local_start_code = start_code
if FIXED_CODE and (local_start_code is None or local_start_code.shape[0] != bs):
local_start_code = torch.randn([bs, C, H // F, W // F], device=device)
samples_ddim, intermediates = sampler.sample(
S=DDIM_STEPS,
conditioning=c,
batch_size=bs,
shape=shape,
verbose=False,
unconditional_guidance_scale=SCALE,
unconditional_conditioning=uc,
eta=DDIM_ETA,
x_T=local_start_code,
log_every_t=100,
z_inpaint=z_inpaint,
z_inpaint_mask=z_inpaint_mask,
z_ref=z_ref,
z9=z9,
)
if SAVE_INTERMEDIATES:
intermediate_pred_x0 = intermediates["pred_x0"]
intermediate_noised = intermediates["x_inter"]
for i in range(len(intermediate_pred_x0)):
for j in range(bs):
stem = f"{out_stem_batch[j]}"
save_sample_by_decode(
intermediate_pred_x0[i][j : j + 1],
model,
inter_pred_path,
stem,
i,
)
save_sample_by_decode(
intermediate_noised[i][j : j + 1],
model,
inter_noised_path,
stem,
i,
)
x_samples_ddim = model.decode_first_stage(samples_ddim)
x_samples_ddim = torch.clamp((x_samples_ddim + 1.0) / 2.0, min=0.0, max=1.0)
x_samples_ddim = x_samples_ddim.cpu().permute(0, 2, 3, 1).numpy()
x_checked_image_torch = torch.from_numpy(x_samples_ddim).permute(0, 3, 1, 2)
for i, x_sample in enumerate(x_checked_image_torch):
stem = f"{out_stem_batch[i]}"
out_path = result_path / f"{stem}.png"
img = Image.fromarray((x_sample.permute(1, 2, 0).numpy() * 255).astype(np.uint8))
img.save(out_path)
print(f"{out_path=}")
for i, x_sample in enumerate(x_checked_image_torch):
all_img = []
all_img.append(un_norm(test_batch[i]).cpu())
if TASK != 2:
ref_img = test_model_kwargs["ref_imgs"].squeeze(1)
ref_img = torchvision.transforms.Resize([512, 512])(ref_img)
ref_img = un_norm_clip(ref_img[i]).cpu()
else:
ref_img = un_norm(test_model_kwargs["ref512"].squeeze(1)[i]).cpu()
all_img.append(ref_img)
all_img.append(x_sample)
grid = torch.stack(all_img, 0)
grid = make_grid(grid)
grid = 255.0 * rearrange(grid, "c h w -> h w c").cpu().numpy()
img = Image.fromarray(grid.astype(np.uint8))
stem = f"{out_stem_batch[i]}"
path_save_img = grid_path / f"grid-{stem}.jpg"
img.save(path_save_img)
print(f"{path_save_img=}")
grids.append(img)
grid_stems.append(stem)
if len(grids) >= NUM_grid_in_a_column:
stem_start = grid_stems[0]
stem_end = grid_stems[-1]
grid_column = imgs_2_grid_A(
grids,
grid_layout='column',
grid_path=os.path.join(grid_path, f"{stem_start}--{stem_end}.jpg"),
)
grids = []
grid_stems = []
model.unset_task()
print(f"Your samples are ready and waiting for you here: {out_dir}")
|