Spaces:
Configuration error
Configuration error
File size: 10,066 Bytes
a23872f |
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 |
# from functools import cache
import importlib
import gradio as gr
import matplotlib.pyplot as plt
import torch
import torchvision
import wandb
from icecream import ic
from torch import nn
from torchvision.transforms.functional import resize
from tqdm import tqdm
from transformers import CLIPModel, CLIPProcessor
import lpips
from edit import blend_paths
from img_processing import *
from img_processing import custom_to_pil
from loaders import load_default
import glob
# global log
log=False
# ic.disable()
# ic.enable()
def get_resized_tensor(x):
if len(x.shape) == 2:
re = x.unsqueeze(0)
else: re = x
re = resize(re, (10, 10))
return re
class ProcessorGradientFlow():
"""
This wraps the huggingface CLIP processor to allow backprop through the image processing step.
The original processor forces conversion to PIL images, which breaks gradient flow.
"""
def __init__(self, device="cuda") -> None:
self.device = device
self.processor = CLIPProcessor.from_pretrained("openai/clip-vit-large-patch14")
self.image_mean = [0.48145466, 0.4578275, 0.40821073]
self.image_std = [0.26862954, 0.26130258, 0.27577711]
self.normalize = torchvision.transforms.Normalize(
self.image_mean,
self.image_std
)
self.resize = torchvision.transforms.Resize(224)
self.center_crop = torchvision.transforms.CenterCrop(224)
def preprocess_img(self, images):
images = self.center_crop(images)
images = self.resize(images)
images = self.center_crop(images)
images = self.normalize(images)
return images
def __call__(self, images=[], **kwargs):
processed_inputs = self.processor(**kwargs)
processed_inputs["pixel_values"] = self.preprocess_img(images)
processed_inputs = {key:value.to(self.device) for (key, value) in processed_inputs.items()}
return processed_inputs
class ImagePromptOptimizer(nn.Module):
def __init__(self,
vqgan,
clip,
clip_preprocessor,
iterations=100,
lr = 0.01,
save_vector=True,
return_val="vector",
quantize=True,
make_grid=False,
lpips_weight = 6.2) -> None:
super().__init__()
self.latent = None
self.device = vqgan.device
vqgan.eval()
self.vqgan = vqgan
self.clip = clip
self.iterations = iterations
self.lr = lr
self.clip_preprocessor = clip_preprocessor
self.make_grid = make_grid
self.return_val = return_val
self.quantize = quantize
self.disc = load_disc(self.device)
self.lpips_weight = lpips_weight
self.perceptual_loss = lpips.LPIPS(net='vgg').to(self.device)
def disc_loss_fn(self, logits):
return -torch.mean(logits)
def set_latent(self, latent):
self.latent = latent.detach().to(self.device)
def set_params(self, lr, iterations, lpips_weight, reconstruction_steps, attn_mask):
self.attn_mask = attn_mask
self.iterations = iterations
self.lr = lr
self.lpips_weight = lpips_weight
self.reconstruction_steps = reconstruction_steps
def forward(self, vector):
base_latent = self.latent.detach().requires_grad_()
trans_latent = base_latent + vector
if self.quantize:
z_q, *_ = self.vqgan.quantize(trans_latent)
else:
z_q = trans_latent
dec = self.vqgan.decode(z_q)
return dec
def _get_clip_similarity(self, prompts, image, weights=None):
if isinstance(prompts, str):
prompts = [prompts]
elif not isinstance(prompts, list):
raise TypeError("Provide prompts as string or list of strings")
clip_inputs = self.clip_preprocessor(text=prompts,
images=image, return_tensors="pt", padding=True)
clip_outputs = self.clip(**clip_inputs)
similarity_logits = clip_outputs.logits_per_image
if weights:
similarity_logits *= weights
return similarity_logits.sum()
def get_similarity_loss(self, pos_prompts, neg_prompts, image):
pos_logits = self._get_clip_similarity(pos_prompts, image)
if neg_prompts:
neg_logits = self._get_clip_similarity(neg_prompts, image)
else:
neg_logits = torch.tensor([1], device=self.device)
loss = -torch.log(pos_logits) + torch.log(neg_logits)
return loss
def visualize(self, processed_img):
if self.make_grid:
self.index += 1
plt.subplot(1, 13, self.index)
plt.imshow(get_pil(processed_img[0]).detach().cpu())
else:
plt.imshow(get_pil(processed_img[0]).detach().cpu())
plt.show()
def attn_masking(self, grad):
# print("attnmask 1")
# print(f"input grad.shape = {grad.shape}")
# print(f"input grad = {get_resized_tensor(grad)}")
newgrad = grad
if self.attn_mask is not None:
# print("masking mult")
newgrad = grad * (self.attn_mask)
# print("output grad, ", get_resized_tensor(newgrad))
# print("end atn 1")
return newgrad
def attn_masking2(self, grad):
# print("attnmask 2")
# print(f"input grad.shape = {grad.shape}")
# print(f"input grad = {get_resized_tensor(grad)}")
newgrad = grad
if self.attn_mask is not None:
# print("masking mult")
newgrad = grad * ((self.attn_mask - 1) * -1)
# print("output grad, ", get_resized_tensor(newgrad))
# print("end atn 2")
return newgrad
def optimize(self, latent, pos_prompts, neg_prompts):
self.set_latent(latent)
# self.make_grid=True
transformed_img = self(torch.zeros_like(self.latent, requires_grad=True, device=self.device))
original_img = loop_post_process(transformed_img)
vector = torch.randn_like(self.latent, requires_grad=True, device=self.device)
optim = torch.optim.Adam([vector], lr=self.lr)
if self.make_grid:
plt.figure(figsize=(35, 25))
self.index = 1
for i in tqdm(range(self.iterations)):
optim.zero_grad()
transformed_img = self(vector)
processed_img = loop_post_process(transformed_img) #* self.attn_mask
processed_img.retain_grad()
lpips_input = processed_img.clone()
lpips_input.register_hook(self.attn_masking2)
lpips_input.retain_grad()
clip_clone = processed_img.clone()
clip_clone.register_hook(self.attn_masking)
clip_clone.retain_grad()
with torch.autocast("cuda"):
clip_loss = self.get_similarity_loss(pos_prompts, neg_prompts, clip_clone)
print("CLIP loss", clip_loss)
perceptual_loss = self.perceptual_loss(lpips_input, original_img.clone()) * self.lpips_weight
print("LPIPS loss: ", perceptual_loss)
with torch.no_grad():
disc_logits = self.disc(transformed_img)
disc_loss = self.disc_loss_fn(disc_logits)
print(f"disc_loss = {disc_loss}")
disc_loss2 = self.disc(processed_img)
if log:
wandb.log({"Perceptual Loss": perceptual_loss})
wandb.log({"Discriminator Loss": disc_loss})
wandb.log({"CLIP Loss": clip_loss})
clip_loss.backward(retain_graph=True)
perceptual_loss.backward(retain_graph=True)
p2 = processed_img.grad
print("Sum Loss", perceptual_loss + clip_loss)
optim.step()
# if i % self.iterations // 10 == 0:
# self.visualize(transformed_img)
yield vector
if self.make_grid:
plt.savefig(f"plot {pos_prompts[0]}.png")
plt.show()
print("lpips solo op")
for i in range(self.reconstruction_steps):
optim.zero_grad()
transformed_img = self(vector)
processed_img = loop_post_process(transformed_img) #* self.attn_mask
processed_img.retain_grad()
lpips_input = processed_img.clone()
lpips_input.register_hook(self.attn_masking2)
lpips_input.retain_grad()
with torch.autocast("cuda"):
perceptual_loss = self.perceptual_loss(lpips_input, original_img.clone()) * self.lpips_weight
with torch.no_grad():
disc_logits = self.disc(transformed_img)
disc_loss = self.disc_loss_fn(disc_logits)
print(f"disc_loss = {disc_loss}")
disc_loss2 = self.disc(processed_img)
# print(f"disc_loss2 = {disc_loss2}")
if log:
wandb.log({"Perceptual Loss": perceptual_loss})
print("LPIPS loss: ", perceptual_loss)
perceptual_loss.backward(retain_graph=True)
optim.step()
yield vector
# torch.save(vector, "nose_vector.pt")
# print("")
# print("DISC STEPS")
# print("*************")
# for i in range(self.reconstruction_steps):
# optim.zero_grad()
# transformed_img = self(vector)
# processed_img = loop_post_process(transformed_img) #* self.attn_mask
# disc_logits = self.disc(transformed_img)
# disc_loss = self.disc_loss_fn(disc_logits)
# print(f"disc_loss = {disc_loss}")
# if log:
# wandb.log({"Disc Loss": disc_loss})
# print("LPIPS loss: ", perceptual_loss)
# disc_loss.backward(retain_graph=True)
# optim.step()
# yield vector
yield vector if self.return_val == "vector" else self.latent + vector
|