VideoGrain / video_diffusion /pipelines /ddim_spatial_temporal.py
XiangpengYang's picture
update all
a043943
# code mostly taken from https://github.com/huggingface/diffusers
import inspect
from typing import Callable, List, Optional, Union
import PIL
import torch
import numpy as np
from einops import rearrange
from tqdm import tqdm
from diffusers.utils import deprecate, logging
from diffusers.pipelines.stable_diffusion import StableDiffusionPipelineOutput
from .stable_diffusion import SpatioTemporalStableDiffusionPipeline
from diffusers.models import AutoencoderKL
from transformers import CLIPTextModel, CLIPTokenizer
import torch.nn.functional as F
from omegaconf import OmegaConf
from video_diffusion.prompt_attention.attention_register import register_attention_control
from video_diffusion.prompt_attention.attention_util import ST_Layout_Attn_Control,ST_Layout_Attn_ControlEdit,Attention_Record_Processor
from video_diffusion.prompt_attention import attention_util
from video_diffusion.prompt_attention.sd_study_utils import *
from video_diffusion.prompt_attention.attention_store import AttentionStore
from video_diffusion.common.image_util import save_gif_mp4_folder_type
from PIL import Image
from einops import rearrange
from ..models.controlnet3d import ControlNetModel
from ..models.unet_3d_condition import UNetPseudo3DConditionModel
from diffusers.schedulers import (
DDIMScheduler,
DDIMInverseScheduler,
)
import os
import nltk
nltk.download('punkt')
logger = logging.get_logger(__name__) # pylint: disable=invalid-name
class DDIMSpatioTemporalStableDiffusionPipeline(SpatioTemporalStableDiffusionPipeline):
r"""
Pipeline for text-to-video generation using Spatio-Temporal Stable Diffusion.
"""
def __init__(
self,
vae: AutoencoderKL,
text_encoder: CLIPTextModel,
tokenizer: CLIPTokenizer,
unet: UNetPseudo3DConditionModel,
controlnet: ControlNetModel,
scheduler: DDIMScheduler,
inverse_scheduler: DDIMInverseScheduler,
disk_store: bool=False,
logdir=None,
):
super().__init__(vae, text_encoder, tokenizer, unet, controlnet, scheduler,inverse_scheduler)
self.store_controller = attention_util.AttentionStore(disk_store=disk_store)
self.logdir=logdir
r"""
Pipeline for text-to-video generation using Spatio-Temporal Stable Diffusion.
"""
def check_inputs(self, prompt, height, width, callback_steps, strength=None):
if not isinstance(prompt, str) and not isinstance(prompt, list):
raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
if strength is not None:
if strength <= 0 or strength > 1:
raise ValueError(f"The value of strength should in (0.0, 1.0] but is {strength}")
if height % 8 != 0 or width % 8 != 0:
raise ValueError(
f"`height` and `width` have to be divisible by 8 but are {height} and {width}."
)
if (callback_steps is None) or (
callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0)
):
raise ValueError(
f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
f" {type(callback_steps)}."
)
@torch.no_grad()
def prepare_source_latents(self, image, batch_size, num_images_per_prompt,
# dtype, device,
text_embeddings,
generator=None):
# Not sure if image need to change device and type
# image = image.to(device=device, dtype=dtype)
print("generator is list:",isinstance(generator, list))
batch_size = batch_size * num_images_per_prompt
if isinstance(generator, list) and len(generator) != batch_size:
raise ValueError(
f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
f" size of {batch_size}. Make sure the batch size matches the length of the generators."
)
if isinstance(generator, list):
init_latents = [
self.vae.encode(image[i : i + 1]).latent_dist.sample(generator[i]) for i in range(batch_size)
]
init_latents = torch.cat(init_latents, dim=0)
else:
## org is
#init_latents = self.vae.encode(image).latent_dist.sample(generator)
init_latents = self.vae.encode(image).latent_dist.mean
init_latents = 0.18215 * init_latents
if batch_size > init_latents.shape[0] and batch_size % init_latents.shape[0] == 0:
# expand init_latents for batch_size
deprecation_message = (
f"You have passed {batch_size} text prompts (`prompt`), but only {init_latents.shape[0]} initial"
" images (`image`). Initial images are now duplicating to match the number of text prompts. Note"
" that this behavior is deprecated and will be removed in a version 1.0.0. Please make sure to update"
" your script to pass as many initial images as text prompts to suppress this warning."
)
deprecate("len(prompt) != len(image)", "1.0.0", deprecation_message, standard_warn=False)
additional_image_per_prompt = batch_size // init_latents.shape[0]
init_latents = torch.cat([init_latents] * additional_image_per_prompt, dim=0)
elif batch_size > init_latents.shape[0] and batch_size % init_latents.shape[0] != 0:
raise ValueError(
f"Cannot duplicate `image` of batch size {init_latents.shape[0]} to {batch_size} text prompts."
)
else:
init_latents = torch.cat([init_latents], dim=0)
# get latents
init_latents_bcfhw = rearrange(init_latents, "(b f) c h w -> b c f h w", b=batch_size)
return init_latents_bcfhw
def prepare_latents_ddim_inverted(self, image, batch_size,
source_prompt,
do_classifier_free_guidance,
control = None,
controlnet_conditioning_scale=None,
use_pnp=None,
cluster_inversion_feature = None,
**kwargs,
):
weight_dtype = image.dtype
device = self._execution_device
print('device',device)
timesteps = self.scheduler.timesteps
saved_features0 = []
saved_features1 = []
saved_features2 = []
saved_q4 = []
saved_k4 = []
saved_q5 = []
saved_k5 = []
saved_q6 = []
saved_k6 = []
saved_q7 = []
saved_k7 = []
saved_q8 = []
saved_k8 = []
saved_q9 = []
saved_k9 = []
#ddim inverse
num_inverse_steps = 50
self.inverse_scheduler.set_timesteps(num_inverse_steps, device=device)
inverse_timesteps, num_inverse_steps = self.get_inverse_timesteps(num_inverse_steps, 1, device)
num_warmup_steps = len(inverse_timesteps) - num_inverse_steps * self.inverse_scheduler.order
#============ddim inversion==========*
prompt_embeds = self._encode_prompt(
source_prompt,
device=device,
num_images_per_prompt=1,
do_classifier_free_guidance=do_classifier_free_guidance,
negative_prompt=None,
)
latents = self.prepare_video_latents(image, batch_size, self.unet.dtype, device)
bz, c, clip_length, downsample_height, downsample_width = latents.shape
del self.store_controller
self.store_controller = attention_util.AttentionStore()
attention_maps_list = []
self_attention_maps_list = []
cond_embeddings_list = []
editor = Attention_Record_Processor(additional_attention_store=self.store_controller)
attention_util.register_attention_control(self, editor, prompt_embeds, clip_length,downsample_height,downsample_width,ddim_inversion=True)
with self.progress_bar(total=num_inverse_steps-1) as progress_bar:
for i, t in enumerate(inverse_timesteps[1:]):
# expand the latents if we are doing classifier free guidance
latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
latent_model_input = self.inverse_scheduler.scale_model_input(latent_model_input, t)
down_block_res_samples, mid_block_res_sample = self.controlnet(latent_model_input, t, encoder_hidden_states=prompt_embeds,controlnet_cond=control,return_dict=False)
down_block_res_samples = [
down_block_res_sample * controlnet_conditioning_scale
for down_block_res_sample in down_block_res_samples
]
mid_block_res_sample *= controlnet_conditioning_scale
noise_pred = self.unet(
latent_model_input,
t,
encoder_hidden_states=prompt_embeds,
down_block_additional_residuals=down_block_res_samples,
mid_block_additional_residual=mid_block_res_sample,
**kwargs,
).sample
if use_pnp and t.cpu() in timesteps:
saved_features0.append(self.unet.up_blocks[1].resnets[0].out_layers_features.cpu())
saved_features1.append(self.unet.up_blocks[1].resnets[1].out_layers_features.cpu())
saved_features2.append(self.unet.up_blocks[2].resnets[0].out_layers_features.cpu())
saved_q4.append(self.unet.up_blocks[1].attentions[1].transformer_blocks[0].attn1.q.cpu())
saved_k4.append(self.unet.up_blocks[1].attentions[1].transformer_blocks[0].attn1.k.cpu())
saved_q5.append(self.unet.up_blocks[1].attentions[2].transformer_blocks[0].attn1.q.cpu())
saved_k5.append(self.unet.up_blocks[1].attentions[2].transformer_blocks[0].attn1.k.cpu())
saved_q6.append(self.unet.up_blocks[2].attentions[0].transformer_blocks[0].attn1.q.cpu())
saved_k6.append(self.unet.up_blocks[2].attentions[0].transformer_blocks[0].attn1.k.cpu())
saved_q7.append(self.unet.up_blocks[2].attentions[1].transformer_blocks[0].attn1.q.cpu())
saved_k7.append(self.unet.up_blocks[2].attentions[1].transformer_blocks[0].attn1.k.cpu())
saved_q8.append(self.unet.up_blocks[2].attentions[2].transformer_blocks[0].attn1.q.cpu())
saved_k8.append(self.unet.up_blocks[2].attentions[2].transformer_blocks[0].attn1.k.cpu())
saved_q9.append(self.unet.up_blocks[3].attentions[0].transformer_blocks[0].attn1.q.cpu())
saved_k9.append(self.unet.up_blocks[3].attentions[0].transformer_blocks[0].attn1.k.cpu())
if do_classifier_free_guidance:
noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
noise_pred = noise_pred_uncond + 1 * (noise_pred_text - noise_pred_uncond)
# compute the previous noisy sample x_t -> x_t-1
latents = self.inverse_scheduler.step(noise_pred, t, latents).prev_sample.to(dtype=weight_dtype)
if i == len(inverse_timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.inverse_scheduler.order == 0):
progress_bar.update()
if use_pnp:
saved_features0.reverse()
saved_features1.reverse()
saved_features2.reverse()
saved_q4.reverse()
saved_k4.reverse()
saved_q5.reverse()
saved_k5.reverse()
saved_q6.reverse()
saved_k6.reverse()
saved_q7.reverse()
saved_k7.reverse()
saved_q8.reverse()
saved_k8.reverse()
saved_q9.reverse()
saved_k9.reverse()
attn_inversion_dict = {
'features0': saved_features0, 'features1': saved_features1, 'features2': saved_features2,
'q4': saved_q4,'k4': saved_k4,'q5': saved_q5,'k5': saved_k5,'q6': saved_q6,'k6': saved_k6,
'q7': saved_q7,'k7': saved_k7,'q8': saved_q8,'k8': saved_k8,'q9': saved_q9,'k9': saved_k9
}
else:
attn_inversion_dict = None
if cluster_inversion_feature:
logger.info('cluster ddim inversion feature')
inv_self_avg_dict={}
inv_cross_avg_dict={}
element_name = 'attn'
attn_size = 32
for element_name in ['attn']:
inv_self_avg_dict[element_name]={}
inv_cross_avg_dict[element_name]={}
self_attn_avg = editor.aggregate_attention(from_where=("up", "down", "mid"),
res=attn_size,is_cross=False)
cross_attn_avg = editor.aggregate_attention(from_where=("up", "down", "mid"),
res=attn_size,is_cross=True)
print('self_attn_avg',self_attn_avg.shape)
print('cross_attn_avg', cross_attn_avg.shape)
inv_self_avg_dict[element_name][attn_size]=self_attn_avg
inv_cross_avg_dict[element_name][attn_size]=cross_attn_avg
os.makedirs(os.path.join(self.logdir, "attn_inv"), exist_ok=True)
os.makedirs(os.path.join(self.logdir, "sd_study"), exist_ok=True)
with open(os.path.join(self.logdir,
"attn_inv/inv_self_avg_dict.pkl"),
'wb') as f:
pkl.dump(inv_self_avg_dict, f)
with open(os.path.join(self.logdir,
"attn_inv/inv_cross_avg_dict.pkl"),
'wb') as f:
pkl.dump(inv_cross_avg_dict, f)
num_segments=3
draw_pca(inv_self_avg_dict, resolution=32, dict_key='attn',
save_path=os.path.join(self.logdir, 'sd_study'),
special_name='inv_self')
run_clusters(inv_self_avg_dict, resolution=32, dict_key='attn',
save_path=os.path.join(self.logdir, 'sd_study'),
special_name='inv_self',num_segments=num_segments)
cross_attn_visualization = attention_util.show_cross_attention_plus_org_img(self.tokenizer, source_prompt,
image, editor, 32, ["up", "down", "mid"], save_path= os.path.join(self.logdir,'sd_study'),attention_maps=cross_attn_avg)
dict_key='attn'
special_name='inv_self'
resolution = 32
threshold=0.1
tokenized_prompt = nltk.word_tokenize(source_prompt)
nouns = [(i, word) for (i, (word, pos)) in enumerate(nltk.pos_tag(tokenized_prompt)) if pos[:2] == 'NN']
print(nouns)
npy_name=f'cluster_{dict_key}_{resolution}_{special_name}.npy'
save_path=os.path.join(self.logdir, 'sd_study')
abs_filename=os.path.join(self.logdir, "attn_inv", f"inv_cross_avg_dict.pkl")
inv_cross_avg_dict=read_pkl(abs_filename)
video_cross_attention = inv_cross_avg_dict['attn'][32]
video_clusters=np.load(os.path.join(save_path, npy_name))
t = video_clusters.shape[0]
for i in range(t):
clusters = video_clusters[i]
cross_attention = video_cross_attention[i]
c2noun, c2mask = cluster2noun_(clusters, threshold, num_segments, nouns,cross_attention)
print('c2noun',c2noun)
merged_mask={}
for index in range(len(c2noun)):
# mask_ = merged_mask[class_name]
item=c2noun[index]
mask_ = c2mask[index]
mask_ = torch.from_numpy(mask_)
mask_ = F.interpolate(mask_.float().unsqueeze(0).unsqueeze(0), size=512, mode='nearest').round().bool().squeeze(0).squeeze(0)
output_name = os.path.join(f"{save_path}",
f"frame_{i}_{item}_{index}.png")
save_mask(mask_, output_name)
return latents, attn_inversion_dict
def get_timesteps(self, num_inference_steps, strength, device):
# get the original timestep using init_timestep
init_timestep = min(int(num_inference_steps * strength), num_inference_steps)
t_start = max(num_inference_steps - init_timestep, 0)
timesteps = self.scheduler.timesteps[t_start:]
return timesteps, num_inference_steps - t_start
def get_inverse_timesteps(self, num_inference_steps, strength, device):
# get the original timestep using init_timestep
init_timestep = min(int(num_inference_steps * strength), num_inference_steps)
t_start = max(num_inference_steps - init_timestep, 0)
# safety for t_start overflow to prevent empty timsteps slice
if t_start == 0:
return self.inverse_scheduler.timesteps, num_inference_steps
timesteps = self.inverse_scheduler.timesteps[:-t_start]
return timesteps, num_inference_steps - t_start
def prepare_latents(
self,
batch_size,
num_channels_latents,
frames,
height,
width,
dtype,
device,
generator,
latents=None,
):
print("self.vae_scale_factor",self.vae_scale_factor)
shape = (
batch_size,
num_channels_latents,
frames,
height // self.vae_scale_factor,
width // self.vae_scale_factor,
)
if isinstance(generator, list) and len(generator) != batch_size:
raise ValueError(
f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
f" size of {batch_size}. Make sure the batch size matches the length of the generators."
)
if latents is None:
rand_device = "cpu" if device.type == "mps" else device
if isinstance(generator, list):
shape = (1,) + shape[1:]
latents = [
torch.randn(shape, generator=generator[i], device=rand_device, dtype=dtype)
for i in range(batch_size)
]
latents = torch.cat(latents, dim=0).to(device)
else:
latents = torch.randn(shape, generator=generator, device=rand_device, dtype=dtype).to(
device
)
else:
if latents.shape != shape:
raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {shape}")
latents = latents.to(device)
# scale the initial noise by the standard deviation required by the scheduler
latents = latents * self.scheduler.init_noise_sigma
return latents
def prepare_video_latents(self, frames, batch_size, dtype, device, generator=None):
if not isinstance(frames, (torch.Tensor, PIL.Image.Image, list)):
raise ValueError(
f"`image` has to be of type `torch.Tensor`, `PIL.Image.Image` or list but is {type(frames)}"
)
if isinstance(generator, list) and len(generator) != batch_size:
raise ValueError(
f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
f" size of {batch_size}. Make sure the batch size matches the length of the generators."
)
if isinstance(generator, list):
latents = [
self.vae.encode(frames[i : i + 1]).latent_dist.sample(generator[i]) for i in range(batch_size)
]
latents = torch.cat(latents, dim=0)
else:
latents = self.vae.encode(frames).latent_dist.sample(generator)
latents = self.vae.config.scaling_factor * latents
latents = rearrange(latents, "(b f) c h w ->b c f h w", b=batch_size)
return latents
def clean_features(self):
self.unet.up_blocks[1].resnets[0].out_layers_inject_features = None
self.unet.up_blocks[1].resnets[1].out_layers_inject_features = None
self.unet.up_blocks[2].resnets[0].out_layers_inject_features = None
self.unet.up_blocks[1].attentions[1].transformer_blocks[0].attn1.inject_q = None
self.unet.up_blocks[1].attentions[1].transformer_blocks[0].attn1.inject_k = None
self.unet.up_blocks[1].attentions[2].transformer_blocks[0].attn1.inject_q = None
self.unet.up_blocks[1].attentions[2].transformer_blocks[0].attn1.inject_k = None
self.unet.up_blocks[2].attentions[0].transformer_blocks[0].attn1.inject_q = None
self.unet.up_blocks[2].attentions[0].transformer_blocks[0].attn1.inject_k = None
self.unet.up_blocks[2].attentions[1].transformer_blocks[0].attn1.inject_q = None
self.unet.up_blocks[2].attentions[1].transformer_blocks[0].attn1.inject_k = None
self.unet.up_blocks[2].attentions[2].transformer_blocks[0].attn1.inject_q = None
self.unet.up_blocks[2].attentions[2].transformer_blocks[0].attn1.inject_k = None
self.unet.up_blocks[3].attentions[0].transformer_blocks[0].attn1.inject_q = None
self.unet.up_blocks[3].attentions[0].transformer_blocks[0].attn1.inject_k = None
def _get_attention_type(self):
sub_nets = self.unet.named_children()
for net in sub_nets:
if hasattr(net[1], 'children'):
for net in net[1].named_children():
if hasattr(net[1], 'children'):
for net in net[1].named_children():
if net[1].__class__.__name__ == "SpatioTemporalTransformerModel":
for net in net[1].named_children():
if hasattr(net[1], 'children'):
for net in net[1].named_children():
if net[1].__class__.__name__ == "SpatioTemporalTransformerBlock":
for net in net[1].named_children():
if net[1].__class__.__name__ == "SparseCausalAttention":
attention_type = "SparseCausalAttention"
elif net[1].__class__.__name__ == "FullyFrameAttention":
attention_type = "FullyFrameAttention"
#print("attention_type:",attention_type)
return attention_type
def _prepare_attention_layout(self,bsz,height,width,layouts,prompts,clip_length,attention_type,device):
## current layouts f s c h w
## org layouts s c h w
#print("prompt:",prompts)
# sp_sz =self.unet.sample_size
sp_sz = height*width
frames, seg_cls, c, h ,w = layouts.shape
text_input = self.tokenizer(prompts, padding="max_length", return_length=True, return_overflowing_tokens=False,
max_length=self.tokenizer.model_max_length, truncation=True, return_tensors="pt")
cond_embeddings = self.text_encoder(text_input.input_ids.to(device))[0]
uncond_input = self.tokenizer([""]*bsz, padding="max_length", max_length=self.tokenizer.model_max_length,
truncation=True, return_tensors="pt")
uncond_embeddings = self.text_encoder(uncond_input.input_ids.to(device))[0]
for i in range(1,len(prompts)):
wlen = text_input['length'][i] - 2
widx = text_input['input_ids'][i][1:1+wlen]
for j in range(77):
if (text_input['input_ids'][0][j:j+wlen] == widx).sum() == wlen:
break
###########################
###### prep for sreg ######
###########################
sreg_maps = {}
reg_sizes = {}
reg_sizes_c = {}
device = layouts.device
frame_index_pre = torch.arange(frames)+(-1)
frame_index_pre = frame_index_pre.clip(0, frames-1)
for r in range(4):
layouts_s_frames = []
if attention_type == "SparseCausalAttention":
layouts_s_sparse_attn = []
h = int(height/np.power(2,r))
w= int(width/np.power(2,r))
#layouts torch.Size([70, 2, 1, 64, 64])
# layouts_interpolate = F.interpolate(layouts.squeeze(2), (res, res), mode='nearest').unsqueeze(2)
layouts_interpolate = F.interpolate(layouts.squeeze(2), (h, w), mode='nearest').unsqueeze(2)
layouts_interpolate = layouts_interpolate.view(frames,seg_cls,1,-1) ## frames,seg_cls,1,res^2
### implementation of sparse casual attn and fully frame attn
for i in range(frames):
#layouts_f = layouts[i]
layouts_s = layouts_interpolate[i]
if attention_type == "SparseCausalAttention":
### prepare for SparseCausalAttention query, key/value
query= layouts_s
query = query.view(query.size(0),-1,1).to(device) ### segcls,res^2,1 #[cls, 4096, 1]
### key should be segcls,1,2xres^2
key = torch.cat((layouts_interpolate[0],layouts_interpolate[frame_index_pre[i]]),dim=-1).to(device)
#([cls, 1, 8192])
layouts_s_cross_frame_attn= (query * key).sum(0).unsqueeze(0).repeat(bsz,1,1) ## 1,4096,8192
layouts_s_sparse_attn.append(layouts_s_cross_frame_attn)
layouts_s = (layouts_s.view(layouts_s.size(0),1,-1)*layouts_s.view(layouts_s.size(0),-1,1)).sum(0).unsqueeze(0).repeat(bsz,1,1)
layouts_s_frames.append(layouts_s)
layouts_s_frames = torch.stack(layouts_s_frames,dim=0)
if attention_type == "SparseCausalAttention":
layouts_s_sparse_attn = torch.stack(layouts_s_sparse_attn,dim=0)
sreg_maps[h*w] = layouts_s_sparse_attn
reg_sizes[h*w] = 1-1.*layouts_s_frames.sum(-1, keepdim=True)/(np.power(clip_length, 2))
reg_sizes_c[h*w] = 1-1.*layouts_s_frames.sum(-1, keepdim=True)/(np.power(clip_length, 2))
#### code for check error#####
# num_nonzero = torch.count_nonzero(layouts_s_frames)
# print("num_nonzero",num_nonzero)
# print("layouts_s_frames",layouts_s_frames.shape)
# print("layouts_s_frames",layouts_s_frames)
# print("reg_size final shape:", (1-1.*layouts_s_frames.sum(-1, keepdim=True)/(np.power(res, 2))).shape)
# print("reg_size", (1-1.*layouts_s_frames.sum(-1, keepdim=True)/(np.power(res, 2))))
#### code for check error#####
#print("layouts_s",layouts_s.shape)
#print("layouts_s.view(layouts_s.size(0),-1,1)",*layouts_s.view(layouts_s.size(0),-1,1).shape)
if attention_type == "FullyFrameAttention":
layouts_s= rearrange(layouts_interpolate,"f s c res -> s c (f res)")
if r==0:
layout_s = None
reg_sizes[h*w] = None
sreg_maps[h*w] = None
reg_sizes_c[h*w] = None
else:
layouts_s = (layouts_s*layouts_s.view(layouts_s.size(0),-1,1)).sum(0).unsqueeze(0).repeat(bsz,1,1).to(torch.float16)
sreg_maps[h*w] = layouts_s
reg_sizes[h*w] = 1-1.*layouts_s.sum(-1, keepdim=True)/((h*clip_length)*(w*clip_length))
reg_sizes_c[h*w] = 1-1.*layouts_s_frames.sum(-1, keepdim=True)/(h*w)
#print("layouts_s",layouts_s.shape)
# if res == 64:
# reg_sizes[np.power(res, 2)] = None
# else:
# reg_sizes[np.power(res, 2)] = 1-1.*layouts_s.sum(-1, keepdim=True)/(np.power(res*clip_length, 2))
# #sreg_maps[np.power(res, 2)] = layouts_s_frames
# sreg_maps[np.power(res, 2)] = layouts_s
# reg_sizes_c[np.power(res, 2)] = 1-1.*layouts_s_frames.sum(-1, keepdim=True)/(np.power(res, 2))
###########################
###### prep for creg ######
###########################
pww_maps = torch.zeros(frames, 1, 77, height, width).to(device)
for i in range(1,len(prompts)):
wlen = text_input['length'][i] - 2
widx = text_input['input_ids'][i][1:1+wlen]
for j in range(77):
if (text_input['input_ids'][0][j:j+wlen] == widx).sum() == wlen:
for f in range(frames):
pww_maps[f,:,j:j+wlen,:,:] = layouts[f,i-1:i] # frames, seg_cls, c, h ,w = layouts.shape
cond_embeddings[0][j:j+wlen] = cond_embeddings[i][1:1+wlen]
print(prompts[i], i, '-th segment is handled.')
break
# print("cond_embeddings",cond_embeddings)
creg_maps = {}
for r in range(4):
pww_maps_frames = []
h = int(height/np.power(2,r))
w = int(width/np.power(2,r))
for i in range(frames):
pww_map_frame = pww_maps[i]
pww_map_frame.view(1,77,height,width)
pww_map_frame = F.interpolate(pww_map_frame, (h, w), mode='nearest')
pww_map_frame = pww_map_frame.view(1, 77, -1).permute(0, 2, 1).repeat(bsz,1,1) # 重新调整形状
pww_maps_frames.append(pww_map_frame)
# 使用 torch.cat 连接处理后的所有帧
layout_c = torch.stack(pww_maps_frames, dim=0)
# print("layout_c",layout_c)
creg_maps[h*w] = layout_c
###########################
#### prep for text_emb ####
###########################
text_cond = torch.cat([uncond_embeddings, cond_embeddings[:1].repeat(bsz,1,1)])
return text_cond, sreg_maps, creg_maps, reg_sizes, reg_sizes_c
@torch.no_grad()
def __call__(
self,
prompt: Union[str, List[str]],
image: Union[torch.FloatTensor, PIL.Image.Image] = None,
latent_mask: Union[torch.FloatTensor, PIL.Image.Image] = None,
layouts: Union[torch.FloatTensor, PIL.Image.Image] = None,
blending_percentage: float=0.25,
modulated_percentage: float=0.3,
height: Optional[int] = None,
width: Optional[int] = None,
strength: float = None,
num_inference_steps: int = 50,
clip_length: int = 8,
guidance_scale: float = 7.5,
source_prompt: Optional[Union[str, List[str]]] = None,
negative_prompt: Optional[Union[str, List[str]]] = None,
num_images_per_prompt: Optional[int] = 1,
eta: float = 0.0,
generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
control: Optional[torch.FloatTensor] = None,
latents: Optional[torch.FloatTensor] = None,
output_type: Optional[str] = "pil",
return_dict: bool = True,
callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
callback_steps: Optional[int] = 1,
logdir: str=None,
controlnet_conditioning_scale: float = 1.0,
use_pnp: bool = False,
cluster_inversion_feature: bool = False,
vis_cross_attn: bool = False,
attn_inversion_dict: dict=None,
**kwargs,
):
# 0. Default height and width to unet
t , c , height, width = image.shape
prompt = OmegaConf.to_container(prompt, resolve=True)
# 1. Check inputs. Raise error if not correct
self.check_inputs(prompt, height, width, callback_steps, strength)
# 2. Define call parameters
batch_size = 1
weight_dtype = image.dtype
device = self._execution_device
# corresponds to doing no classifier free guidance.
do_classifier_free_guidance = guidance_scale > 1.0
extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
self.scheduler.set_timesteps(num_inference_steps, device=device)
if latents is None:
latents, attn_inversion_dict = self.prepare_latents_ddim_inverted(
image, batch_size, source_prompt,
do_classifier_free_guidance, generator,
control, controlnet_conditioning_scale, use_pnp, cluster_inversion_feature
)
print("use inversion latents")
## prepare text embedding, self attention map, cross attention map
_, _, _, downsample_height, downsample_width = latents.shape
attention_type = self._get_attention_type()
text_cond, sreg_maps, creg_maps, reg_sizes,reg_sizes_c = self._prepare_attention_layout(batch_size,downsample_height,downsample_width,
layouts,prompt,clip_length,attention_type,device)
time_steps = self.scheduler.timesteps
#============do visualization for st-layout attn===============#
self.store_controller = attention_util.AttentionStore()
editor = ST_Layout_Attn_ControlEdit(text_cond=text_cond,sreg_maps=sreg_maps,creg_maps=creg_maps,reg_sizes=reg_sizes,reg_sizes_c=reg_sizes_c,
time_steps=time_steps,clip_length=clip_length,attention_type=attention_type,
additional_attention_store=self.store_controller,
save_self_attention = True,
disk_store = False,
video = image,
)
attention_util.register_attention_control(self, editor, text_cond, clip_length, downsample_height,downsample_width,ddim_inversion=False)
#============do visualization for st-layout attn===============#
# editor = ST_Layout_Attn_Control(text_cond=text_cond,sreg_maps=sreg_maps,creg_maps=creg_maps,reg_sizes=reg_sizes,reg_sizes_c=reg_sizes_c,
# time_steps=time_steps,clip_length=clip_length,attention_type=attention_type)
# register_attention_control(self, editor, text_cond, clip_length,downsample_height,downsample_width,ddim_inversion=False)
# 3. Encode input prompt
prompt = prompt[:1]
text_embeddings = self._encode_prompt(
prompt, device, num_images_per_prompt, do_classifier_free_guidance, negative_prompt
)
source_latents = self.prepare_source_latents(
image, batch_size, num_images_per_prompt,
# text_embeddings.dtype, device,
text_embeddings,
generator,
)
# 7. Denoising loop
num_warmup_steps = len(time_steps) - num_inference_steps * self.scheduler.order
with self.progress_bar(total=num_inference_steps* (1-blending_percentage)) as progress_bar:
for i, t in enumerate(time_steps[int(len(time_steps) * blending_percentage):]):
latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
# inject features
if use_pnp and i < kwargs["inject_step"]:
self.unet.up_blocks[1].resnets[0].out_layers_inject_features = attn_inversion_dict['features0'][i].to(device)
self.unet.up_blocks[1].resnets[1].out_layers_inject_features = attn_inversion_dict['features1'][i].to(device)
self.unet.up_blocks[2].resnets[0].out_layers_inject_features = attn_inversion_dict['features2'][i].to(device)
self.unet.up_blocks[1].attentions[1].transformer_blocks[0].attn1.inject_q = attn_inversion_dict['q4'][i].to(device)
self.unet.up_blocks[1].attentions[1].transformer_blocks[0].attn1.inject_k = attn_inversion_dict['k4'][i].to(device)
self.unet.up_blocks[1].attentions[2].transformer_blocks[0].attn1.inject_q = attn_inversion_dict['q5'][i].to(device)
self.unet.up_blocks[1].attentions[2].transformer_blocks[0].attn1.inject_k = attn_inversion_dict['k5'][i].to(device)
self.unet.up_blocks[2].attentions[0].transformer_blocks[0].attn1.inject_q = attn_inversion_dict['q6'][i].to(device)
self.unet.up_blocks[2].attentions[0].transformer_blocks[0].attn1.inject_k = attn_inversion_dict['k6'][i].to(device)
self.unet.up_blocks[2].attentions[1].transformer_blocks[0].attn1.inject_q = attn_inversion_dict['q7'][i].to(device)
self.unet.up_blocks[2].attentions[1].transformer_blocks[0].attn1.inject_k = attn_inversion_dict['k7'][i].to(device)
self.unet.up_blocks[2].attentions[2].transformer_blocks[0].attn1.inject_q = attn_inversion_dict['q8'][i].to(device)
self.unet.up_blocks[2].attentions[2].transformer_blocks[0].attn1.inject_k = attn_inversion_dict['k8'][i].to(device)
self.unet.up_blocks[3].attentions[0].transformer_blocks[0].attn1.inject_q = attn_inversion_dict['q9'][i].to(device)
self.unet.up_blocks[3].attentions[0].transformer_blocks[0].attn1.inject_k = attn_inversion_dict['k9'][i].to(device)
else:
self.clean_features()
down_block_res_samples, mid_block_res_sample = self.controlnet(
latent_model_input,
t,
encoder_hidden_states=text_embeddings,
controlnet_cond=control,
return_dict=False,
)
down_block_res_samples = [
down_block_res_sample * controlnet_conditioning_scale
for down_block_res_sample in down_block_res_samples
]
mid_block_res_sample *= controlnet_conditioning_scale
noise_pred = self.unet(
latent_model_input,
t,
encoder_hidden_states=text_embeddings,
down_block_additional_residuals=down_block_res_samples,
mid_block_additional_residual=mid_block_res_sample,
**kwargs,
).sample.to(dtype=weight_dtype)
# perform guidance
if do_classifier_free_guidance:
# print("do_classifier_free_guidance")
noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
noise_pred = noise_pred_uncond + guidance_scale * (
noise_pred_text - noise_pred_uncond
)
# compute the previous noisy sample x_t -> x_t-1
latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample
# Blending
noise_source_latents = self.scheduler.add_noise(
source_latents, torch.randn_like(latents), t
)
latents = latents * latent_mask + noise_source_latents * (1 - latent_mask)
# call the callback, if provided
if i == len(time_steps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
progress_bar.update()
if callback is not None and i % callback_steps == 0:
callback(i, t, latents)
### vis cross attn
# image shape fchw
if vis_cross_attn:
save_path = os.path.join(logdir,'visualization_denoise')
os.makedirs(save_path, exist_ok=True)
attention_output = attention_util.show_cross_attention_plus_org_img(self.tokenizer,prompt, image, editor, 32, ["up","down"],save_path=save_path)
# 8. Post-processing
image = self.decode_latents(latents)
# 9. Run safety checker
has_nsfw_concept = None
# 10. Convert to PIL
if output_type == "pil":
image = self.numpy_to_pil(image)
if not return_dict:
return (image, has_nsfw_concept)
torch.cuda.empty_cache()
return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept)