Spaces:
Running
on
A10G
Running
on
A10G
File size: 12,107 Bytes
acc80f0 e1cb97f 2d52659 acc80f0 2822d34 acc80f0 e1cb97f 2822d34 acc80f0 0aa753f acc80f0 2822d34 acc80f0 2822d34 acc80f0 5589998 acc80f0 |
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 |
import torch
import os
from tqdm import tqdm
from PIL import Image, ImageDraw ,ImageFont
from matplotlib import pyplot as plt
import torchvision.transforms as T
import os
import yaml
import numpy as np
import gradio as gr
torch.manual_seed(0)
def load_512(image_path, left=0, right=0, top=0, bottom=0, device=None):
if type(image_path) is str:
image = np.array(Image.open(image_path).convert('RGB'))[:, :, :3]
else:
image = image_path
h, w, c = image.shape
left = min(left, w-1)
right = min(right, w - left - 1)
top = min(top, h - left - 1)
bottom = min(bottom, h - top - 1)
image = image[top:h-bottom, left:w-right]
h, w, c = image.shape
if h < w:
offset = (w - h) // 2
image = image[:, offset:offset + h]
elif w < h:
offset = (h - w) // 2
image = image[offset:offset + w]
image = np.array(Image.fromarray(image).resize((512, 512)))
image = torch.from_numpy(image).float() / 127.5 - 1
image = image.permute(2, 0, 1).unsqueeze(0).to(device)
return image
def load_real_image(folder = "data/", img_name = None, idx = 0, img_size=512, device='cuda'):
from PIL import Image
from glob import glob
if img_name is not None:
path = os.path.join(folder, img_name)
else:
path = glob(folder + "*")[idx]
img = Image.open(path).resize((img_size,
img_size))
img = pil_to_tensor(img).to(device)
if img.shape[1]== 4:
img = img[:,:3,:,:]
return img
def mu_tilde(model, xt,x0, timestep):
"mu_tilde(x_t, x_0) DDPM paper eq. 7"
prev_timestep = timestep - model.scheduler.config.num_train_timesteps // model.scheduler.num_inference_steps
alpha_prod_t_prev = model.scheduler.alphas_cumprod[prev_timestep] if prev_timestep >= 0 else model.scheduler.final_alpha_cumprod
alpha_t = model.scheduler.alphas[timestep]
beta_t = 1 - alpha_t
alpha_bar = model.scheduler.alphas_cumprod[timestep]
return ((alpha_prod_t_prev ** 0.5 * beta_t) / (1-alpha_bar)) * x0 + ((alpha_t**0.5 *(1-alpha_prod_t_prev)) / (1- alpha_bar))*xt
def sample_xts_from_x0(model, x0, num_inference_steps=50):
"""
Samples from P(x_1:T|x_0)
"""
# torch.manual_seed(43256465436)
alpha_bar = model.scheduler.alphas_cumprod
sqrt_one_minus_alpha_bar = (1-alpha_bar) ** 0.5
alphas = model.scheduler.alphas
betas = 1 - alphas
variance_noise_shape = (
num_inference_steps,
model.unet.in_channels,
model.unet.sample_size,
model.unet.sample_size)
timesteps = model.scheduler.timesteps.to(model.device)
t_to_idx = {int(v):k for k,v in enumerate(timesteps)}
xts = torch.zeros(variance_noise_shape).to(x0.device)
for t in reversed(timesteps):
idx = t_to_idx[int(t)]
xts[idx] = x0 * (alpha_bar[t] ** 0.5) + torch.randn_like(x0) * sqrt_one_minus_alpha_bar[t]
xts = torch.cat([xts, x0 ],dim = 0)
return xts
def encode_text(model, prompts):
text_input = model.tokenizer(
prompts,
padding="max_length",
max_length=model.tokenizer.model_max_length,
truncation=True,
return_tensors="pt",
)
with torch.no_grad():
text_encoding = model.text_encoder(text_input.input_ids.to(model.device))[0]
return text_encoding
def forward_step(model, model_output, timestep, sample):
next_timestep = min(model.scheduler.config.num_train_timesteps - 2,
timestep + model.scheduler.config.num_train_timesteps // model.scheduler.num_inference_steps)
# 2. compute alphas, betas
alpha_prod_t = model.scheduler.alphas_cumprod[timestep]
# alpha_prod_t_next = self.scheduler.alphas_cumprod[next_timestep] if next_ltimestep >= 0 else self.scheduler.final_alpha_cumprod
beta_prod_t = 1 - alpha_prod_t
# 3. compute predicted original sample from predicted noise also called
# "predicted x_0" of formula (12) from https://arxiv.org/pdf/2010.02502.pdf
pred_original_sample = (sample - beta_prod_t ** (0.5) * model_output) / alpha_prod_t ** (0.5)
# 5. TODO: simple noising implementatiom
next_sample = model.scheduler.add_noise(pred_original_sample,
model_output,
torch.LongTensor([next_timestep]))
return next_sample
def get_variance(model, timestep): #, prev_timestep):
prev_timestep = timestep - model.scheduler.config.num_train_timesteps // model.scheduler.num_inference_steps
alpha_prod_t = model.scheduler.alphas_cumprod[timestep]
alpha_prod_t_prev = model.scheduler.alphas_cumprod[prev_timestep] if prev_timestep >= 0 else model.scheduler.final_alpha_cumprod
beta_prod_t = 1 - alpha_prod_t
beta_prod_t_prev = 1 - alpha_prod_t_prev
variance = (beta_prod_t_prev / beta_prod_t) * (1 - alpha_prod_t / alpha_prod_t_prev)
return variance
def inversion_forward_process(model, x0,
etas = None,
prog_bar = False,
prompt = "",
cfg_scale = 3.5,
num_inference_steps=50, eps = None
):
if not prompt=="":
text_embeddings = encode_text(model, prompt)
uncond_embedding = encode_text(model, "")
timesteps = model.scheduler.timesteps.to(model.device)
variance_noise_shape = (
num_inference_steps,
model.unet.in_channels,
model.unet.sample_size,
model.unet.sample_size)
if etas is None or (type(etas) in [int, float] and etas == 0):
eta_is_zero = True
zs = None
else:
eta_is_zero = False
if type(etas) in [int, float]: etas = [etas]*model.scheduler.num_inference_steps
xts = sample_xts_from_x0(model, x0, num_inference_steps=num_inference_steps)
alpha_bar = model.scheduler.alphas_cumprod
zs = torch.zeros(size=variance_noise_shape, device=model.device)
t_to_idx = {int(v):k for k,v in enumerate(timesteps)}
xt = x0
op = tqdm(reversed(timesteps)) if prog_bar else reversed(timesteps)
for t in op:
idx = t_to_idx[int(t)]
# 1. predict noise residual
if not eta_is_zero:
xt = xts[idx][None]
with torch.no_grad():
out = model.unet.forward(xt, timestep = t, encoder_hidden_states = uncond_embedding)
if not prompt=="":
cond_out = model.unet.forward(xt, timestep=t, encoder_hidden_states = text_embeddings)
if not prompt=="":
## classifier free guidance
noise_pred = out.sample + cfg_scale * (cond_out.sample - out.sample)
else:
noise_pred = out.sample
if eta_is_zero:
# 2. compute more noisy image and set x_t -> x_t+1
xt = forward_step(model, noise_pred, t, xt)
else:
xtm1 = xts[idx+1][None]
# pred of x0
pred_original_sample = (xt - (1-alpha_bar[t]) ** 0.5 * noise_pred ) / alpha_bar[t] ** 0.5
# direction to xt
prev_timestep = t - model.scheduler.config.num_train_timesteps // model.scheduler.num_inference_steps
alpha_prod_t_prev = model.scheduler.alphas_cumprod[prev_timestep] if prev_timestep >= 0 else model.scheduler.final_alpha_cumprod
variance = get_variance(model, t)
pred_sample_direction = (1 - alpha_prod_t_prev - etas[idx] * variance ) ** (0.5) * noise_pred
mu_xt = alpha_prod_t_prev ** (0.5) * pred_original_sample + pred_sample_direction
z = (xtm1 - mu_xt ) / ( etas[idx] * variance ** 0.5 )
zs[idx] = z
# correction to avoid error accumulation
xtm1 = mu_xt + ( etas[idx] * variance ** 0.5 )*z
xts[idx+1] = xtm1
if not zs is None:
zs[-1] = torch.zeros_like(zs[-1])
return xt, zs, xts
def reverse_step(model, model_output, timestep, sample, eta = 0, variance_noise=None):
# 1. get previous step value (=t-1)
prev_timestep = timestep - model.scheduler.config.num_train_timesteps // model.scheduler.num_inference_steps
# 2. compute alphas, betas
alpha_prod_t = model.scheduler.alphas_cumprod[timestep]
alpha_prod_t_prev = model.scheduler.alphas_cumprod[prev_timestep] if prev_timestep >= 0 else model.scheduler.final_alpha_cumprod
beta_prod_t = 1 - alpha_prod_t
# 3. compute predicted original sample from predicted noise also called
# "predicted x_0" of formula (12) from https://arxiv.org/pdf/2010.02502.pdf
pred_original_sample = (sample - beta_prod_t ** (0.5) * model_output) / alpha_prod_t ** (0.5)
# 5. compute variance: "sigma_t(η)" -> see formula (16)
# σ_t = sqrt((1 − α_t−1)/(1 − α_t)) * sqrt(1 − α_t/α_t−1)
# variance = self.scheduler._get_variance(timestep, prev_timestep)
variance = get_variance(model, timestep) #, prev_timestep)
std_dev_t = eta * variance ** (0.5)
# Take care of asymetric reverse process (asyrp)
model_output_direction = model_output
# 6. compute "direction pointing to x_t" of formula (12) from https://arxiv.org/pdf/2010.02502.pdf
# pred_sample_direction = (1 - alpha_prod_t_prev - std_dev_t**2) ** (0.5) * model_output_direction
pred_sample_direction = (1 - alpha_prod_t_prev - eta * variance) ** (0.5) * model_output_direction
# 7. compute x_t without "random noise" of formula (12) from https://arxiv.org/pdf/2010.02502.pdf
prev_sample = alpha_prod_t_prev ** (0.5) * pred_original_sample + pred_sample_direction
# 8. Add noice if eta > 0
if eta > 0:
if variance_noise is None:
variance_noise = torch.randn(model_output.shape, device=model.device)
sigma_z = eta * variance ** (0.5) * variance_noise
prev_sample = prev_sample + sigma_z
return prev_sample
def inversion_reverse_process(model,
xT,
etas = 0,
prompts = "",
cfg_scales = None,
prog_bar = False,
zs = None,
controller=None,
asyrp = False
):
batch_size = len(prompts)
cfg_scales_tensor = torch.Tensor(cfg_scales).view(-1,1,1,1).to(model.device)
text_embeddings = encode_text(model, prompts)
uncond_embedding = encode_text(model, [""] * batch_size)
if etas is None: etas = 0
if type(etas) in [int, float]: etas = [etas]*model.scheduler.num_inference_steps
assert len(etas) == model.scheduler.num_inference_steps
timesteps = model.scheduler.timesteps.to(model.device)
xt = xT.expand(batch_size, -1, -1, -1)
op = tqdm(timesteps[-zs.shape[0]:]) if prog_bar else timesteps[-zs.shape[0]:]
t_to_idx = {int(v):k for k,v in enumerate(timesteps[-zs.shape[0]:])}
for t in op:
idx = t_to_idx[int(t)]
## Unconditional embedding
with torch.no_grad():
uncond_out = model.unet.forward(xt, timestep = t,
encoder_hidden_states = uncond_embedding)
## Conditional embedding
if prompts:
with torch.no_grad():
cond_out = model.unet.forward(xt, timestep = t,
encoder_hidden_states = text_embeddings)
z = zs[idx] if not zs is None else None
z = z.expand(batch_size, -1, -1, -1)
if prompts:
## classifier free guidance
noise_pred = uncond_out.sample + cfg_scales_tensor * (cond_out.sample - uncond_out.sample)
else:
noise_pred = uncond_out.sample
# 2. compute less noisy image and set x_t -> x_t-1
xt = reverse_step(model, noise_pred, t, xt, eta = etas[idx], variance_noise = z)
if controller is not None:
xt = controller.step_callback(xt)
return xt, zs
|