Spaces:
Sleeping
Sleeping
File size: 10,893 Bytes
95b1715 cfa4803 95b1715 cfa4803 95b1715 cfa4803 95b1715 cfa4803 95b1715 96c63c3 95b1715 cfa4803 95b1715 cfa4803 95b1715 ad8ad02 95b1715 cfa4803 95b1715 cfa4803 95b1715 cfa4803 95b1715 |
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 |
import os
import sys
import torch
import pickle
import argparse
import numpy as np
from editings import ganspace
from editings.styleclip.mapper.styleclip_mapper import StyleCLIPMapper
from editings.styleclip.mapper.gloabl_mapper import StyleCLIPGlobalDirection
from editings.deltaedit.editor import DeltaEditor
STYLESPACE_IDX = [
0,
1,
1,
2,
2,
3,
4,
4,
5,
6,
6,
7,
8,
8,
9,
10,
10,
11,
12,
12,
13,
14,
14,
15,
16,
16,
]
class LatentEditor:
def __init__(self, domain="human_faces", device="cpu"):
self.domain = domain
self.device = torch.device(device)
if self.domain == "human_faces":
self.interfacegan_directions = {
"age": "editings/interfacegan_directions/age.pt",
"smile": "editings/interfacegan_directions/smile.pt",
"rotation": "editings/interfacegan_directions/rotation.pt",
}
self.interfacegan_tensors = {
name: torch.load(path, map_location=self.device)
for name, path in self.interfacegan_directions.items()
}
self.ganspace_pca = torch.load("editings/ganspace_pca/ffhq_pca.pt", map_location=self.device)
self.ganspace_directions = {
"eye_openness": (54, 7, 8, 5),
"trimmed_beard": (58, 7, 9, 7),
"lipstick": (34, 10, 11, 20),
"face_roundness": (37, 0, 5, 20.0),
"nose_length": (51, 4, 5, -30.0),
"eyebrow_thickness": (37, 8, 9, 20.0),
"head_angle_up": (11, 1, 4, -10.5),
"displeased": (36, 4, 7, 10.0),
}
self.styleclip_directions = {
"afro": [False, False, True],
"angry": [False, False, True],
"beyonce": [False, False, False],
"bobcut": [False, False, True],
"bowlcut": [False, False, True],
"curly_hair": [False, False, True],
"hilary_clinton": [False, False, False],
"depp": [False, False, False],
"mohawk": [False, False, True],
"purple_hair": [False, False, False],
"surprised": [False, False, True],
"taylor_swift": [False, False, False],
"trump": [False, False, False],
"zuckerberg": [False, False, False],
}
self.styleclip_global_editor = self.load_styleclip_global()
self.stylespace_directions = {
"black hair": [(12, 479)],
"blond hair": [(12, 479), (12, 266)],
"grey hair": [(11, 286)],
"wavy hair": [(6, 500), (8, 128), (5, 92), (6, 394), (6, 323)],
"bangs": [
(3, 259),
(6, 285),
(5, 414),
(6, 128),
(9, 295),
(6, 322),
(6, 487),
(6, 504),
],
"receding hairline": [(5, 414), (6, 322), (6, 497), (6, 504)],
"smiling": [(6, 501)],
"sslipstick": [(15, 45)],
"sideburns": [(12, 237)],
"goatee": [(9, 421)],
"earrings": [(8, 81)],
"glasses": [(3, 288), (2, 175), (3, 120), (2, 97)],
"wear suit": [(9, 441), (8, 292), (11, 358), (6, 223)],
"gender": [(9, 6)],
}
self.fs_directions = {
"fs_glasses": "editings/bound/Eyeglasses_boundary.npy",
"fs_smiling": "editings/bound/Smiling_boundary.npy",
"fs_makeup": "editings/bound/Heavy_Makeup_boundary.npy"
}
self.deltaedit_editor = DeltaEditor(device=self.device)
elif self.domain == "car":
self.stylespace_directions = {
"front": [(8, 411)],
"headlights": [(8, 441), (9, 355)],
"grill": [(9, 191)],
"trees": [(9, 108)],
"grass_ss": [(12, 107)],
"sky": [(12, 76)],
"hubcap": [(12, 113), (12, 439)],
"car color": [(12, 142), (15, 227)],
"logo": [(9, 185)],
"wheel angle": [(8, 420)],
}
self.ganspace_pca = torch.load("editings/ganspace_pca/cars_pca.pt")
self.ganspace_directions = {
"pose_1": (0, 0, 5, 2),
"pose_2": (0, 0, 5, -2),
"cube": (16, 3, 6, 25),
"color": (22, 9, 11, -8),
"grass": (41, 9, 11, -18)
}
def load_styleclip_global(self):
delta_i_c = torch.from_numpy(np.load("editings/styleclip/global_mapper_data/delta_i_c.npy")).float().to(self.device)
with open("editings/styleclip/global_mapper_data/S_mean_std", "rb") as channels_statistics:
_, s_std = pickle.load(channels_statistics)
s_std = [torch.from_numpy(s_i).float().to(self.device) for s_i in s_std]
with open("editings/styleclip/global_mapper_data/templates.txt", "r") as templates:
text_prompt_templates = templates.readlines()
global_direction_calculator = StyleCLIPGlobalDirection(delta_i_c, s_std, text_prompt_templates, device=self.device)
return global_direction_calculator
def get_styleclip_mapper_edits(self, start_w, factors, direction):
latents_to_display = []
mapper_checkpoint_path = os.path.join(
"pretrained_models/styleclip_mappers",
f"{direction}.pt",
)
ckpt = torch.load(mapper_checkpoint_path, map_location="cpu")
opts = ckpt["opts"]
styleclip_opts = argparse.Namespace(
**{
"mapper_type": "LevelsMapper",
"no_coarse_mapper": self.styleclip_directions[direction][0],
"no_medium_mapper": self.styleclip_directions[direction][1],
"no_fine_mapper": self.styleclip_directions[direction][2],
"stylegan_size": 1024,
"checkpoint_path": mapper_checkpoint_path,
}
)
opts.update(vars(styleclip_opts))
opts = argparse.Namespace(**opts)
style_clip_net = StyleCLIPMapper(opts)
style_clip_net.eval()
style_clip_net.to(self.device)
direction = style_clip_net.mapper(start_w)
for factor in factors:
edited_latent = start_w + factor * direction
latents_to_display.append(edited_latent)
return latents_to_display
def get_styleclip_global_edits(self, start_s, factors, direction):
latents_to_display = []
neutral_text, target_text, disentanglement = direction.split("_")
disentanglement = float(disentanglement)
directions = self.styleclip_global_editor.get_delta_s(neutral_text, target_text, disentanglement)
factors = torch.tensor(factors).to(self.device).view(-1, 1)
srart_ss, start_rgb = start_s
edits_rgb = []
edits_ss = []
for i in range(26):
if i in [1, 4, 7, 10, 13, 16, 19, 22, 25]:
edits_rgb.append(directions[i].view(1, -1).repeat(len(factors), 1))
else:
edits_ss.append(directions[i].view(1, -1).repeat(len(factors), 1))
edited_rgb = []
edited_ss = []
for orig, edit in zip(srart_ss, edits_ss):
edited_ss.append(orig.repeat(len(factors), 1) + edit * factors.repeat(1, orig.size(1)) / 1.5)
for orig, edit in zip(start_rgb, edits_rgb):
edited_rgb.append(orig.repeat(len(factors), 1) + edit * factors.repeat(1, orig.size(1)) / 1.5)
return edited_ss, edited_rgb
def get_deltaedit_edits(self, start_s, factors, direction, original_image):
latents_to_display = []
neutral_text, target_text, disentanglement = direction.split("_")
disentanglement = float(disentanglement)
factors = torch.tensor(factors).to(self.device).view(-1, 1)
srart_ss, edited_rgb = start_s
edits_ss = self.deltaedit_editor.get_delta_s(neutral_text, target_text, disentanglement, original_image, srart_ss)
edited_rgb = [latent.repeat(len(factors), 1) for latent in edited_rgb]
edited_ss = []
for orig, edit in zip(srart_ss, edits_ss):
edited_ss.append(orig.repeat(len(factors), 1) + edit * factors.repeat(1, orig.size(1)))
return edited_ss, edited_rgb
def get_ganspace_edits(self, start_w, factors, direction):
latents_to_display = []
for factor in factors:
ganspace_direction = self.ganspace_directions[direction]
edit_direction = list(ganspace_direction)
edit_direction[-1] = factor
edit_direction = tuple(edit_direction)
new_w = ganspace.edit(start_w, self.ganspace_pca, [edit_direction])
latents_to_display.append(new_w)
return latents_to_display
def get_interface_gan_edits(self, start_w, factors, direction):
latents_to_display = []
for factor in factors:
tensor_direction = self.interfacegan_tensors[direction]
edited_latent = start_w + factor / 2 * tensor_direction
latents_to_display.append(edited_latent)
return latents_to_display
def get_stylespace_edits(self, start_s, factors, direction):
edits = self.stylespace_directions[direction]
start_stylespaces, start_stylespaces_rgb = start_s
device = start_stylespaces[0].device
latents_to_display = []
edited_latent = [
s.clone().repeat(len(factors), 1)
for s in start_stylespaces
]
factors = torch.tensor(factors).to(device)
for layer_num, feat_num in edits:
edited_latent[STYLESPACE_IDX[layer_num]][:, feat_num] += factors * 3
edited_stylespaces_rgb = [
rgb.repeat(len(factors), 1) for rgb in start_stylespaces_rgb
]
return edited_latent, edited_stylespaces_rgb
def get_fs_edits(self, w, factors, direction):
path = self.fs_directions[direction]
boundary = np.load(path)
device = w.device
bs = w.size(0)
w_0 = w.cpu().numpy().reshape(bs, -1)
boundary = boundary.reshape(1, -1).repeat(bs, 0)
edits = [torch.tensor(w_0 + factor * boundary).view(bs, -1, 512).to(device) for factor in factors]
return edits
|