File size: 11,295 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 | from confs import *
from pathlib import Path
import numpy as np
import cv2
from PIL import Image
import torch
import torch.utils.data as data
import torchvision.transforms as T
from einops import rearrange
import albumentations
from util_face import *
from util_4dataset import *
from util_cv2 import cv2_resize_auto_interpolation
from Mediapipe_Result_Cache import Mediapipe_Result_Cache
def resize_A(img, dataset_name, size=(512, 512), interpolation=None):
is_pil = isinstance(img, Image.Image)
if is_pil:
img = np.array(img)
if img.shape[:2] != (512, 512):
img = cv2_resize_auto_interpolation(img, size, interpolation=interpolation)
if is_pil:
img = Image.fromarray(img)
return img
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
def un_norm(x):
return (x + 1.0) / 2.0
def _dilate(_mask, kernel_size, iterations):
_mask = _mask.astype(np.uint8)
kernel = np.ones((kernel_size, kernel_size), np.uint8)
_mask = cv2.dilate(_mask, kernel, iterations=iterations)
_mask = _mask.astype(bool)
return _mask
def dilate_4_task0(sm_mask):
sm_mask = np.array(sm_mask)
preserve1 = [2, 3, 10, 5]
mask1 = np.isin(sm_mask, preserve1)
mask1 = _dilate(mask1, 7, 1)
preserve2 = [3, 10]
mask2 = np.isin(sm_mask, preserve2)
mask2 = _dilate(mask2, 10, 3)
preserve3 = [1]
mask3 = np.isin(sm_mask, preserve3)
mask3 = _dilate(mask3, 7, 2)
mask = mask1 | mask2 | mask3
return mask
class Dataset_custom(data.Dataset):
mean = (0.5, 0.5, 0.5)
std = (0.5, 0.5, 0.5)
def get_img4clip(
self,
img,
sm_mask,
preserve,
for_clip=True,
add_semantic_head=False,
mask_after_npisin=None,
for_inpaint512=False,
):
sm_mask = np.array(sm_mask)
if mask_after_npisin is None:
if self.task == 0 and 0:
mask = dilate_4_task0(sm_mask)
else:
mask = np.isin(sm_mask, preserve)
if self.task == 0 and 1 and for_inpaint512:
forehead_mask = get_forehead_mask(sm_mask)
mask = mask & ~forehead_mask
else:
mask = mask_after_npisin
if isinstance(img, np.ndarray):
img = Image.fromarray(img)
if add_semantic_head:
mask_before_colorSM = mask
img, mask = add_colorSM(img, sm_mask, preserve, None)
mask = mask_after_npisin__2__tensor(mask)
if for_clip:
image_tensor = get_tensor_clip()(img)
else:
image_tensor = get_tensor(mean=self.mean, std=self.std)(img)
image_tensor = T.Resize([512, 512])(image_tensor)
image_tensor = image_tensor * mask
if for_clip:
image_tensor = 255.0 * rearrange(un_norm_clip(image_tensor), "c h w -> h w c").cpu().numpy()
_size = 224
else:
image_tensor = 255.0 * rearrange(un_norm(image_tensor), "c h w -> h w c").cpu().numpy()
_size = 512
image_tensor = albumentations.Resize(height=_size, width=_size)(image=image_tensor)
image_tensor = Image.fromarray(image_tensor["image"].astype(np.uint8))
if for_clip:
image_tensor = get_tensor_clip()(image_tensor)
else:
image_tensor = get_tensor(mean=self.mean, std=self.std)(image_tensor)
image_tensor = image_tensor * mask
if add_semantic_head:
mask = mask_after_npisin__2__tensor(mask_before_colorSM)
return image_tensor, mask
def __init__(
self,
state,
task,
paths_tgt,
paths_ref,
name="custom",
):
if task == 0:
USE_filter_mediapipe_fail_swap = 1
USE_pts = 1
READ_mediapipe_result_from_cache = 1
elif task == 1:
USE_filter_mediapipe_fail_swap = 0
USE_pts = 0
READ_mediapipe_result_from_cache = 1
elif task == 2:
USE_filter_mediapipe_fail_swap = 1
USE_pts = 1
READ_mediapipe_result_from_cache = 1
elif task == 3:
USE_filter_mediapipe_fail_swap = 0
USE_pts = 1
READ_mediapipe_result_from_cache = 1
self.READ_mediapipe_result_from_cache = READ_mediapipe_result_from_cache
assert state == "test"
self.state = state
self.image_size = 512
self.kernel = np.ones((1, 1), np.uint8)
self.name = name
assert paths_tgt is not None and paths_ref is not None, "paths_tgt and paths_ref are required"
assert len(paths_tgt) == len(paths_ref), "paths_tgt and paths_ref must be the same length"
self.paths_tgt = list(paths_tgt)
self.paths_ref = list(paths_ref)
if READ_mediapipe_result_from_cache:
self.mediapipe_Result_Cache = Mediapipe_Result_Cache()
self.task = task
def __getitem__(self, index):
task = self.task
path_tgt = self.paths_tgt[index]
path_ref = self.paths_ref[index]
img_tgt = Image.open(path_tgt).convert("RGB")
img_tgt = resize_A(img_tgt, self.name)
mask_path = path_img_2_path_mask(path_tgt)
if self.task == 0:
preserve = [1, 2, 3, 10, 5, 6, 7, 9]
if 0:
preserve = [1, 2, 3, 10, 5]
sm_mask_tgt = Image.open(mask_path).convert("L")
sm_mask_tgt = np.array(sm_mask_tgt)
if 0:
mask_tgt = dilate_4_task0(sm_mask_tgt)
else:
mask_tgt = np.isin(sm_mask_tgt, preserve)
if self.task == 0 and 1:
forehead_mask = get_forehead_mask(sm_mask_tgt)
mask_tgt = mask_tgt & ~forehead_mask
elif self.task == 1:
preserve = [4]
mask_tgt = path_img_2_mask(path_tgt, preserve)
elif self.task == 3:
preserve = [1, 2, 3, 10, 4, 5, 6, 7, 9]
mask_tgt = path_img_2_mask(path_tgt, preserve)
elif self.task == 2:
preserve = [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 20, 21]
sm_mask_tgt = Image.open(mask_path).convert("L")
sm_mask_tgt = np.array(sm_mask_tgt)
mask_tgt = np.isin(sm_mask_tgt, preserve)
converted_mask = np.zeros_like(mask_tgt)
converted_mask[mask_tgt] = 255
mask_tgt = Image.fromarray(converted_mask).convert("L")
mask_tensor = 1 - get_tensor(normalize=False, toTensor=True)(mask_tgt)
image_tensor = get_tensor(mean=self.mean, std=self.std)(img_tgt)
image_tensor_resize = T.Resize([self.image_size, self.image_size])(image_tensor)
mask_tensor_resize = T.Resize([self.image_size, self.image_size])(mask_tensor)
if task == 2:
inpaint_tensor_resize = image_tensor_resize
else:
inpaint_tensor_resize = image_tensor_resize * mask_tensor_resize
if 1:
mask_tensor_resize = 1 - mask_tensor_resize
if 1:
mask_path_ref = path_img_2_path_mask(path_ref)
sm_mask_ref = Image.open(mask_path_ref).convert("L")
sm_mask_ref = np.array(sm_mask_ref)
img_ref = cv2.imread(str(path_ref))
img_ref = cv2.cvtColor(img_ref, cv2.COLOR_BGR2RGB)
img_ref = resize_A(img_ref, self.name)
if task != 2:
ref_image_tensor, ref_mask_tensor = self.get_img4clip(
img_ref, sm_mask_ref, preserve, for_clip=True, add_semantic_head=0
)
if task == 3:
ref_image_faceOnly_tensor, _ = self.get_img4clip(
img_ref,
sm_mask_ref,
[1, 2, 3, 10, 5, 6, 7, 9],
for_clip=False,
add_semantic_head=0,
)
else:
ref_image_tensor = inpaint_tensor_resize
ret = {
"inpaint_image": inpaint_tensor_resize,
"inpaint_mask": mask_tensor_resize,
"ref_imgs": ref_image_tensor,
"task": self.task,
}
if self.task == 0:
ret["enInputs"] = {
"face_ID-in": ref_image_tensor,
"face-clip-in": ref_image_tensor,
}
elif self.task == 1:
ret["enInputs"] = {
"hair-clip-in": ref_image_tensor,
}
elif self.task == 2:
tgt_nonBg_tensor, _ = self.get_img4clip(img_tgt, sm_mask_tgt, preserve)
ret["enInputs"] = {
"face_ID-in": tgt_nonBg_tensor,
"head-clip-in": tgt_nonBg_tensor,
}
elif self.task == 3:
ret["enInputs"] = {
"face_ID-in": ref_image_faceOnly_tensor,
"head-clip-in": ref_image_tensor,
}
if (REFNET.ENABLE and REFNET.task2layerNum[task] > 0) or CH14:
if task != 2:
ref_imgs_4unet, ref_mask_4unet = self.get_img4clip(
img_ref, sm_mask_ref, preserve, for_clip=False, add_semantic_head=0
)
else:
ref_imgs_4unet, ref_mask_4unet = self.get_img4clip(
img_tgt,
sm_mask_tgt,
"any",
for_clip=False,
add_semantic_head=0,
mask_after_npisin=np.ones_like(sm_mask_tgt).astype(bool),
)
ref_imgs_4unet = T.Resize([self.image_size, self.image_size])(ref_imgs_4unet)
ref_mask_512 = T.Resize([self.image_size, self.image_size])(ref_mask_4unet)
ret["ref_imgs_4unet"] = ref_imgs_4unet
ret["ref_mask_512"] = ref_mask_512
if self.READ_mediapipe_result_from_cache:
if self.state == "test":
if task == 2:
_p_lmk = path_ref
else:
_p_lmk = path_tgt
else:
_p_lmk = path_tgt
ret["mediapipe_lmkAll"] = self.mediapipe_Result_Cache.get(_p_lmk)
if ret["mediapipe_lmkAll"] is None:
raise RuntimeError(
f"Missing Mediapipe cache for input image: {_p_lmk}. "
"Precompute landmarks and ensure cache exists before inference."
)
if self.state == "test":
prior_image_tensor = "None"
out_stem = f"{Path(path_tgt).stem}-{Path(path_ref).stem}"
if task == 2:
ref512, _ = self.get_img4clip(
img_ref, sm_mask_ref, preserve, for_clip=False, add_semantic_head=0
)
ref512 = T.Resize([self.image_size, self.image_size])(ref512)
ret["ref512"] = ref512
ret = (image_tensor_resize, prior_image_tensor, ret, out_stem)
return ret
def __len__(self):
return len(self.paths_tgt)
|