kevinlu4588's picture
Training gradient ascent models and sanity checking car erasure
e225ed6
raw
history blame
19.6 kB
from PIL import Image
from matplotlib import pyplot as plt
import textwrap
import argparse
import torch
import copy
import os
import re
import numpy as np
from diffusers import AutoencoderKL, UNet2DConditionModel
from PIL import Image
from tqdm.auto import tqdm
from transformers import CLIPTextModel, CLIPTokenizer, CLIPFeatureExtractor
from diffusers.schedulers import EulerAncestralDiscreteScheduler
from diffusers.schedulers.scheduling_ddim import DDIMScheduler
from diffusers.schedulers.scheduling_ddpm import DDPMScheduler
from diffusers.schedulers.scheduling_lms_discrete import LMSDiscreteScheduler
from diffusers.pipelines.stable_diffusion import StableDiffusionSafetyChecker
def to_gif(images, path):
images[0].save(path, save_all=True,
append_images=images[1:], loop=0, duration=len(images) * 20)
def figure_to_image(figure):
figure.set_dpi(300)
figure.canvas.draw()
return Image.frombytes('RGB', figure.canvas.get_width_height(), figure.canvas.tostring_rgb())
def image_grid(images, outpath=None, column_titles=None, row_titles=None):
n_rows = len(images)
n_cols = len(images[0])
fig, axs = plt.subplots(nrows=n_rows, ncols=n_cols,
figsize=(n_cols, n_rows), squeeze=False)
for row, _images in enumerate(images):
for column, image in enumerate(_images):
ax = axs[row][column]
ax.imshow(image)
if column_titles and row == 0:
ax.set_title(textwrap.fill(
column_titles[column], width=12), fontsize='x-small')
if row_titles and column == 0:
ax.set_ylabel(row_titles[row], rotation=0, fontsize='x-small', labelpad=1.6 * len(row_titles[row]))
ax.set_xticks([])
ax.set_yticks([])
plt.subplots_adjust(wspace=0, hspace=0)
if outpath is not None:
plt.savefig(outpath, bbox_inches='tight', dpi=300)
plt.close()
else:
plt.tight_layout(pad=0)
image = figure_to_image(plt.gcf())
plt.close()
return image
def get_module(module, module_name):
if isinstance(module_name, str):
module_name = module_name.split('.')
if len(module_name) == 0:
return module
else:
module = getattr(module, module_name[0])
return get_module(module, module_name[1:])
def set_module(module, module_name, new_module):
if isinstance(module_name, str):
module_name = module_name.split('.')
if len(module_name) == 1:
return setattr(module, module_name[0], new_module)
else:
module = getattr(module, module_name[0])
return set_module(module, module_name[1:], new_module)
def freeze(module):
for parameter in module.parameters():
parameter.requires_grad = False
def unfreeze(module):
for parameter in module.parameters():
parameter.requires_grad = True
def get_concat_h(im1, im2):
dst = Image.new('RGB', (im1.width + im2.width, im1.height))
dst.paste(im1, (0, 0))
dst.paste(im2, (im1.width, 0))
return dst
def get_concat_v(im1, im2):
dst = Image.new('RGB', (im1.width, im1.height + im2.height))
dst.paste(im1, (0, 0))
dst.paste(im2, (0, im1.height))
return dst
class StableDiffuser(torch.nn.Module):
def __init__(self,
scheduler='LMS'
):
super().__init__()
# Load the autoencoder model which will be used to decode the latents into image space.
self.vae = AutoencoderKL.from_pretrained(
"CompVis/stable-diffusion-v1-4", subfolder="vae")
# Load the tokenizer and text encoder to tokenize and encode the text.
self.tokenizer = CLIPTokenizer.from_pretrained(
"openai/clip-vit-large-patch14")
self.text_encoder = CLIPTextModel.from_pretrained(
"openai/clip-vit-large-patch14")
# The UNet model for generating the latents.
self.unet = UNet2DConditionModel.from_pretrained(
"CompVis/stable-diffusion-v1-4", subfolder="unet")
self.feature_extractor = CLIPFeatureExtractor.from_pretrained("CompVis/stable-diffusion-v1-4", subfolder="feature_extractor")
self.safety_checker = StableDiffusionSafetyChecker.from_pretrained("CompVis/stable-diffusion-v1-4", subfolder="safety_checker")
if scheduler == 'LMS':
self.scheduler = LMSDiscreteScheduler(beta_start=0.00085, beta_end=0.012, beta_schedule="scaled_linear", num_train_timesteps=1000)
elif scheduler == 'DDIM':
self.scheduler = DDIMScheduler.from_pretrained("CompVis/stable-diffusion-v1-4", subfolder="scheduler")
elif scheduler == 'DDPM':
self.scheduler = DDPMScheduler.from_pretrained("CompVis/stable-diffusion-v1-4", subfolder="scheduler")
self.eval()
def get_noise(self, batch_size, img_size, generator=None):
param = list(self.parameters())[0]
return torch.randn(
(batch_size, self.unet.in_channels, img_size // 8, img_size // 8),
generator=generator).type(param.dtype).to(param.device)
def add_noise(self, latents, noise, step):
return self.scheduler.add_noise(latents, noise, torch.tensor([self.scheduler.timesteps[step]]))
def text_tokenize(self, prompts):
return self.tokenizer(prompts, padding="max_length", max_length=self.tokenizer.model_max_length, truncation=True, return_tensors="pt")
def text_detokenize(self, tokens):
return [self.tokenizer.decode(token) for token in tokens if token != self.tokenizer.vocab_size - 1]
def text_encode(self, tokens):
return self.text_encoder(tokens.input_ids.to(self.unet.device))[0]
def decode(self, latents):
return self.vae.decode(1 / self.vae.config.scaling_factor * latents).sample
def encode(self, tensors):
return self.vae.encode(tensors).latent_dist.mode() * 0.18215
def to_image(self, image):
image = (image / 2 + 0.5).clamp(0, 1)
image = image.detach().cpu().permute(0, 2, 3, 1).numpy()
images = (image * 255).round().astype("uint8")
pil_images = [Image.fromarray(image) for image in images]
return pil_images
def set_scheduler_timesteps(self, n_steps):
self.scheduler.set_timesteps(n_steps, device=self.unet.device)
def get_initial_latents(self, n_imgs, img_size, n_prompts, generator=None):
noise = self.get_noise(n_imgs, img_size, generator=generator).repeat(n_prompts, 1, 1, 1)
latents = noise * self.scheduler.init_noise_sigma
return latents
def get_text_embeddings(self, prompts, n_imgs):
text_tokens = self.text_tokenize(prompts)
text_embeddings = self.text_encode(text_tokens)
unconditional_tokens = self.text_tokenize([""] * len(prompts))
unconditional_embeddings = self.text_encode(unconditional_tokens)
text_embeddings = torch.cat([unconditional_embeddings, text_embeddings]).repeat_interleave(n_imgs, dim=0)
return text_embeddings
def predict_noise(self,
iteration,
latents,
text_embeddings,
guidance_scale=7.5
):
# expand the latents if we are doing classifier-free guidance to avoid doing two forward passes.
latents = torch.cat([latents] * 2)
latents = self.scheduler.scale_model_input(
latents, self.scheduler.timesteps[iteration])
# predict the noise residual
noise_prediction = self.unet(
latents, self.scheduler.timesteps[iteration], encoder_hidden_states=text_embeddings).sample
# perform guidance
noise_prediction_uncond, noise_prediction_text = noise_prediction.chunk(2)
noise_prediction = noise_prediction_uncond + guidance_scale * \
(noise_prediction_text - noise_prediction_uncond)
return noise_prediction
@torch.no_grad()
def diffusion(self,
latents,
text_embeddings,
end_iteration=1000,
start_iteration=0,
return_steps=False,
pred_x0=False,
trace_args=None,
show_progress=True,
**kwargs):
latents_steps = []
trace_steps = []
trace = None
for iteration in tqdm(range(start_iteration, end_iteration), disable=not show_progress):
if trace_args:
trace = TraceDict(self, **trace_args)
noise_pred = self.predict_noise(
iteration,
latents,
text_embeddings,
**kwargs)
# compute the previous noisy sample x_t -> x_t-1
output = self.scheduler.step(noise_pred, self.scheduler.timesteps[iteration], latents)
if trace_args:
trace.close()
trace_steps.append(trace)
latents = output.prev_sample
if return_steps or iteration == end_iteration - 1:
output = output.pred_original_sample if pred_x0 else latents
if return_steps:
latents_steps.append(output.cpu())
else:
latents_steps.append(output)
return latents_steps, trace_steps
@torch.no_grad()
def __call__(self,
prompts,
img_size=512,
n_steps=50,
n_imgs=1,
end_iteration=None,
generator=None,
**kwargs
):
assert 0 <= n_steps <= 1000
if not isinstance(prompts, list):
prompts = [prompts]
self.set_scheduler_timesteps(n_steps)
latents = self.get_initial_latents(n_imgs, img_size, len(prompts), generator=generator)
text_embeddings = self.get_text_embeddings(prompts,n_imgs=n_imgs)
end_iteration = end_iteration or n_steps
latents_steps, trace_steps = self.diffusion(
latents,
text_embeddings,
end_iteration=end_iteration,
**kwargs
)
latents_steps = [self.decode(latents.to(self.unet.device)) for latents in latents_steps]
images_steps = [self.to_image(latents) for latents in latents_steps]
# for i in range(len(images_steps)):
# self.safety_checker = self.safety_checker.float()
# safety_checker_input = self.feature_extractor(images_steps[i], return_tensors="pt").to(latents_steps[0].device)
# image, has_nsfw_concept = self.safety_checker(
# images=latents_steps[i].float().cpu().numpy(), clip_input=safety_checker_input.pixel_values.float()
# )
# images_steps[i][0] = self.to_image(torch.from_numpy(image))[0]
images_steps = list(zip(*images_steps))
if trace_steps:
return images_steps, trace_steps
return images_steps
class FineTunedModel(torch.nn.Module):
def __init__(self,
model,
train_method,
):
super().__init__()
self.model = model
self.ft_modules = {}
self.orig_modules = {}
freeze(self.model)
for module_name, module in model.named_modules():
if 'unet' not in module_name:
continue
if module.__class__.__name__ in ["Linear", "Conv2d", "LoRACompatibleLinear", "LoRACompatibleConv"]:
if train_method == 'xattn':
if 'attn2' not in module_name:
continue
elif train_method == 'xattn-strict':
if 'attn2' not in module_name or 'to_q' not in module_name or 'to_k' not in module_name:
continue
elif train_method == 'noxattn':
if 'attn2' in module_name:
continue
elif train_method == 'selfattn':
if 'attn1' not in module_name:
continue
else:
raise NotImplementedError(
f"train_method: {train_method} is not implemented."
)
# print(module_name)
ft_module = copy.deepcopy(module)
self.orig_modules[module_name] = module
self.ft_modules[module_name] = ft_module
unfreeze(ft_module)
self.ft_modules_list = torch.nn.ModuleList(self.ft_modules.values())
self.orig_modules_list = torch.nn.ModuleList(self.orig_modules.values())
@classmethod
def from_checkpoint(cls, model, checkpoint, train_method):
if isinstance(checkpoint, str):
checkpoint = torch.load(checkpoint)
modules = [f"{key}$" for key in list(checkpoint.keys())]
ftm = FineTunedModel(model, train_method=train_method)
ftm.load_state_dict(checkpoint)
return ftm
def __enter__(self):
for key, ft_module in self.ft_modules.items():
set_module(self.model, key, ft_module)
def __exit__(self, exc_type, exc_value, tb):
for key, module in self.orig_modules.items():
set_module(self.model, key, module)
def parameters(self):
parameters = []
for ft_module in self.ft_modules.values():
parameters.extend(list(ft_module.parameters()))
return parameters
def state_dict(self):
state_dict = {key: module.state_dict() for key, module in self.ft_modules.items()}
return state_dict
def load_state_dict(self, state_dict):
for key, sd in state_dict.items():
self.ft_modules[key].load_state_dict(sd)
def train(erase_concept, erase_from, train_method, iterations, negative_guidance, lr, save_path):
nsteps = 50
diffuser = StableDiffuser(scheduler='DDIM').to('cuda')
diffuser.train()
finetuner = FineTunedModel(diffuser, train_method=train_method)
optimizer = torch.optim.Adam(finetuner.parameters(), lr=lr)
criteria = torch.nn.MSELoss()
pbar = tqdm(range(iterations))
erase_concept = erase_concept.split(',')
erase_concept = [a.strip() for a in erase_concept]
erase_from = erase_from.split(',')
erase_from = [a.strip() for a in erase_from]
if len(erase_from)!=len(erase_concept):
if len(erase_from) == 1:
c = erase_from[0]
erase_from = [c for _ in erase_concept]
else:
print(erase_from, erase_concept)
raise Exception("Erase from concepts length need to match erase concepts length")
erase_concept_ = []
for e, f in zip(erase_concept, erase_from):
erase_concept_.append([e,f])
erase_concept = erase_concept_
print(erase_concept)
# del diffuser.vae
# del diffuser.text_encoder
# del diffuser.tokenizer
torch.cuda.empty_cache()
for i in pbar:
with torch.no_grad():
index = np.random.choice(len(erase_concept), 1, replace=False)[0]
erase_concept_sampled = erase_concept[index]
neutral_text_embeddings = diffuser.get_text_embeddings([''],n_imgs=1)
positive_text_embeddings = diffuser.get_text_embeddings([erase_concept_sampled[0]],n_imgs=1)
target_text_embeddings = diffuser.get_text_embeddings([erase_concept_sampled[1]],n_imgs=1)
diffuser.set_scheduler_timesteps(nsteps)
optimizer.zero_grad()
iteration = torch.randint(1, nsteps - 1, (1,)).item()
latents = diffuser.get_initial_latents(1, 512, 1)
with finetuner:
latents_steps, _ = diffuser.diffusion(
latents,
positive_text_embeddings,
start_iteration=0,
end_iteration=iteration,
guidance_scale=3,
show_progress=False
)
diffuser.set_scheduler_timesteps(1000)
iteration = int(iteration / nsteps * 1000)
positive_latents = diffuser.predict_noise(iteration, latents_steps[0], positive_text_embeddings, guidance_scale=1)
neutral_latents = diffuser.predict_noise(iteration, latents_steps[0], neutral_text_embeddings, guidance_scale=1)
target_latents = diffuser.predict_noise(iteration, latents_steps[0], target_text_embeddings, guidance_scale=1)
if erase_concept_sampled[0] == erase_concept_sampled[1]:
target_latents = neutral_latents.clone().detach()
with finetuner:
negative_latents = diffuser.predict_noise(iteration, latents_steps[0], target_text_embeddings, guidance_scale=1)
positive_latents.requires_grad = False
neutral_latents.requires_grad = False
loss = criteria(negative_latents, target_latents - (negative_guidance*(positive_latents - neutral_latents))) #loss = criteria(e_n, e_0) works the best try 5000 epochs
loss.backward()
optimizer.step()
try:
torch.save(finetuner.state_dict(), save_path)
print(f"Model saved successfully at {save_path}")
except Exception as e:
print(f"Error saving model: {e}")
del diffuser, loss, optimizer, finetuner, negative_latents, neutral_latents, positive_latents, latents_steps, latents
torch.cuda.empty_cache()
if __name__ == '__main__':
parser = argparse.ArgumentParser(
prog = 'TrainESD',
description = 'Finetuning stable diffusion to erase the concepts')
parser.add_argument('--erase_concept', help='concept to erase', type=str, required=True)
parser.add_argument('--erase_from', help='target concept to erase from', type=str, required=False, default = None)
parser.add_argument('--train_method', help='Type of method (xattn, noxattn, full, xattn-strict', type=str, required=True)
parser.add_argument('--iterations', help='Number of iterations', type=int, default=200)
parser.add_argument('--lr', help='Learning rate', type=float, default=2e-5)
parser.add_argument('--negative_guidance', help='Negative guidance value', type=float, required=False, default=1)
parser.add_argument('--save_path', help='Path to save model', type=str, default='models/')
parser.add_argument('--device', help='cuda device to train on', type=str, required=False, default='cuda:0')
args = parser.parse_args()
prompt = args.erase_concept #'car'
erase_concept = args.erase_concept
erase_from = args.erase_from
if erase_from is None:
erase_from = erase_concept
train_method = args.train_method #'noxattn'
iterations = args.iterations #200
negative_guidance = args.negative_guidance #1
lr = args.lr #1e-5
name = f"esd-{erase_concept.lower().replace(' ','').replace(',','')}_from_{erase_from.lower().replace(' ','').replace(',','')}-{train_method}_{negative_guidance}-epochs_{iterations}"
if not os.path.exists(args.save_path):
os.makedirs(args.save_path, exist_ok = True)
save_path = os.path.join(args.save_path, f"{name}.pt")
print("save path", save_path)
train(erase_concept=erase_concept, erase_from=erase_from, train_method=train_method, iterations=iterations, negative_guidance=negative_guidance, lr=lr, save_path=save_path)