XiangpengYang's picture
update all
a043943
raw
history blame
17 kB
"""
Collect all function in prompt_attention folder.
Provide a API `make_controller' to return an initialized AttentionControlEdit class object in the main validation loop.
"""
from typing import Optional, Union, Tuple, List, Dict
import abc
import numpy as np
import copy
from einops import rearrange
import torch
import torch.nn.functional as F
import video_diffusion.prompt_attention.ptp_utils as ptp_utils
from video_diffusion.prompt_attention.visualization import show_cross_attention,show_cross_attention_plus_org_img,show_self_attention_comp,aggregate_attention
from video_diffusion.prompt_attention.attention_store import AttentionStore, AttentionControl
from video_diffusion.prompt_attention.attention_register import register_attention_control
device = torch.device('cuda') if torch.cuda.is_available() else torch.device('cpu')
from PIL import Image
import os
from video_diffusion.common.image_util import save_gif_mp4_folder_type,make_grid
import cv2
import math
from PIL import Image, ImageDraw
import numpy as np
import math
import os
class EmptyControl:
def step_callback(self, x_t):
return x_t
def between_steps(self):
return
def __call__(self, attn, is_cross: bool, place_in_unet: str):
return attn
def apply_jet_colormap(weight):
# 将权重规范化到0-255
weight = 255*(weight - weight.min()) / (weight.max() - weight.min()+1e-6)
weight = weight.astype(np.uint8)
# 应用Jet颜色映射
color_mapped_weight = cv2.applyColorMap(weight, cv2.COLORMAP_JET)
return color_mapped_weight
def show_self_attention_comp(self_attention_map, video, h_index:int, w_index:int, res: int, frames:int, place_in_unet: List[str], step:int ):
attention_maps = self_attention_map.reshape(frames, res, res, frames, res, res)
weights = attention_maps[0,h_index,w_index,:,:,:]
attention_list = []
video_frames = []
#video f,c,h,w
for i in range(frames):
weight = weights[i].cpu().numpy()
weight_colored = apply_jet_colormap(weight)
weight_colored = weight_colored[:, :, ::-1] # BGR到RGB的转换
weight_colored = np.array(Image.fromarray(weight_colored).resize((256, 256)))
attention_list.append(weight_colored)
frame = video[i].permute(1,2,0).cpu().numpy()
mean = np.array((0.48145466, 0.4578275, 0.40821073)).reshape((1, 1, 3)) # [h, w, c]
varas = np.array((0.26862954, 0.26130258, 0.27577711)).reshape((1, 1, 3))
frame = frame * varas + mean
frame = (frame - frame.min()) / (frame.max() - frame.min() + 1e-6) * 255
frame = frame.astype(np.uint8)
video_frames.append(frame)
alpha = 0.5
overlay_frames = []
for frame, attention in zip(video_frames, attention_list):
attention_resized = cv2.resize(attention, (frame.shape[1], frame.shape[0]))
overlay_frame = cv2.addWeighted(frame, alpha, attention_resized, 1 - alpha, 0)
overlay_frames.append(overlay_frame)
print('vis self attn')
save_path = "with_st_layout_vis_self_attn/vis_self_attn"
os.makedirs(save_path, exist_ok=True)
video_save_path = f'{save_path}/self-attn-{place_in_unet}-{step}-query-frame0-h{h_index}-w{w_index}.gif'
save_gif_mp4_folder_type(overlay_frames, video_save_path,save_gif=False)
def draw_grid_on_image(image, grid_size, line_color="gray"):
draw = ImageDraw.Draw(image)
w, h = image.size
for i in range(0, w, grid_size):
draw.line([(i, 0), (i, h)], fill=line_color)
for i in range(0, h, grid_size):
draw.line([(0, i), (w, i)], fill=line_color)
return image
def identify_self_attention_max_min(sim, video, h_index:int, w_index:int, res: int, frames:int, place_in_unet: str, step:int):
attention_maps = sim.reshape(frames, res, res, frames, res, res)
weights = attention_maps[0, h_index, w_index, :, :, :]
flattened_weights = weights.reshape(-1)
global_max_index = flattened_weights.argmax().cpu().numpy()
global_min_index = flattened_weights.argmin().cpu().numpy()
print('weights.shape',weights.shape)
frame_max, h_max, w_max = np.unravel_index(global_max_index, weights.shape)
frame_min, h_min, w_min = np.unravel_index(global_min_index, weights.shape)
video_frames = []
query_frame_index = 0
query_h = h_index
query_w = w_index
for i in range(frames):
frame = video[i].permute(1, 2, 0).cpu().numpy()
mean = np.array((0.48145466, 0.4578275, 0.40821073)).reshape((1, 1, 3))
varas = np.array((0.26862954, 0.26130258, 0.27577711)).reshape((1, 1, 3))
frame = (frame * varas + mean) * 255
frame = np.clip(frame, 0, 255).astype(np.uint8)
frame_img = Image.fromarray(frame)
grid_size = 512 // res
frame_img = draw_grid_on_image(frame_img, grid_size)
draw = ImageDraw.Draw(frame_img)
if i == frame_max:
max_pixel_pos = (w_max * grid_size, h_max * grid_size)
draw.rectangle([max_pixel_pos, (max_pixel_pos[0] + grid_size, max_pixel_pos[1] + grid_size)], outline="red", width=2)
if i == frame_min:
min_pixel_pos = (w_min * grid_size, h_min * grid_size)
draw.rectangle([min_pixel_pos, (min_pixel_pos[0] + grid_size, min_pixel_pos[1] + grid_size)], outline="blue", width=2)
if i == query_frame_index:
query_pixel_pos = (query_w * grid_size, query_h * grid_size)
draw.rectangle([query_pixel_pos, (query_pixel_pos[0] + grid_size, query_pixel_pos[1] + grid_size)], outline="yellow", width=2)
video_frames.append(frame_img)
save_path = "/visualization/correspondence_with_query"
os.makedirs(save_path, exist_ok=True)
video_save_path = os.path.join(save_path, f'self-attn-{place_in_unet}-{step}-query-frame0-h{h_index}-w{w_index}.gif')
save_gif_mp4_folder_type(video_frames, video_save_path, save_gif=False)
class ST_Layout_Attn_Control(AttentionControl, abc.ABC):
def __init__(self, end_step=15, total_steps=50, step_idx=None, text_cond=None, sreg_maps=None, creg_maps=None, reg_sizes=None,reg_sizes_c=None, time_steps=None,clip_length=None,attention_type=None):
"""
Spatial-Temporal Layout-guided Attention (ST-Layout Attn) for Stable-Diffusion model
note: without vis cross attention weight function.
Args:
end_step: the step to end st-layout attn control
total_steps: the total number of steps
step_idx: list the steps to apply mutual self-attention control
text_cond: discrete text embedding for each region.
sreg_maps: spatial-temporal self-attention qk condition maps.
creg_maps: cross-attention qk condition maps
reg_sizes/reg_sizes_c: size regularzation maps for each instance in self_attn/cross_attention
clip_length: frames len of video
attention_type: FullyFrameAttention_sliced_attn/FullyFrameAttention/SparseCausalAttention
"""
super().__init__()
self.total_steps = total_steps
self.step_idx = list(range(0, end_step))
self.total_infer_steps = 50
self.text_cond = text_cond
self.sreg_maps = sreg_maps
self.creg_maps = creg_maps
self.reg_sizes = reg_sizes
self.reg_sizes_c = reg_sizes_c
self.clip_length = clip_length
self.attention_type = attention_type
self.sreg = .3
self.creg = 1.
self.count = 0
self.reg_part = .3
self.time_steps = time_steps
print("Modulated Ctrl at denoising steps: ", self.step_idx)
def forward(self, sim, is_cross, place_in_unet, **kwargs):
"""
Attention forward function
"""
#print("self.cur_step",self.cur_step)
if self.cur_step not in self.step_idx:
return super().forward(sim, is_cross, place_in_unet, **kwargs)
### sim for "SparseCausalAttention": (frames, heads=8,res, 2*res)
### sim for "FullyFrameAttention" : 1, heads, frame*res,frane*res [1, 8, 12288, 12288])
num_heads = sim.shape[1]
if num_heads == 1:
self.attention_type == "FullyFrameAttention_sliced_attn"
treg = torch.pow((self.time_steps[self.cur_step]-1)/1000, 5)
if not is_cross:
min_value = sim.min(-1)[0].unsqueeze(-1)
max_value = sim.max(-1)[0].unsqueeze(-1)
if self.attention_type == "SparseCausalAttention":
mask = self.sreg_maps[sim.size(2)].repeat(1,num_heads,1,1)
size_reg = self.reg_sizes[sim.size(2)].repeat(1,num_heads,1,1)
elif self.attention_type == "FullyFrameAttention":
mask = self.sreg_maps[sim.size(2)//self.clip_length].repeat(1,num_heads,1,1)
size_reg = self.reg_sizes[sim.size(2)//self.clip_length].repeat(1,num_heads,1,1)
elif self.attention_type == "FullyFrameAttention_sliced_attn":
mask = self.sreg_maps[sim.size(2)//self.clip_length]
size_reg = self.reg_sizes[sim.size(2)//self.clip_length]
else:
print("unknown attention type")
exit()
# if place_in_unet == "up" and res == 32:
# # h_index 11 w_index =15
# show_self_attention_comp(sim,video=self.video,h_index=11,w_index=15,res=32,frames=self.clip_length,place_in_unet="up",step=self.cur_step)
#if place_in_unet == "up" and res == 8:
# identify_self_attention_max_min(sim,video=self.video,h_index=3,w_index=4,res=8,frames=self.clip_length,place_in_unet="up",step=self.cur_step)
sim += (mask>0)*size_reg*self.sreg*treg*(max_value-sim)
sim -= ~(mask>0)*size_reg*self.sreg*treg*(sim-min_value)
else:
min_value = sim.min(-1)[0].unsqueeze(-1)
max_value = sim.max(-1)[0].unsqueeze(-1)
mask = self.creg_maps[sim.size(2)].repeat(1,num_heads,1,1)
size_reg = self.reg_sizes_c[sim.size(2)].repeat(1,num_heads,1,1)
sim += (mask>0)*size_reg*self.creg*treg*(max_value-sim)
sim -= ~(mask>0)*size_reg*self.creg*treg*(sim-min_value)
self.count +=1
return sim
class Attention_Record_Processor(AttentionStore, abc.ABC):
""" record ddim inversion self attention and cross attention """
def __init__(self, additional_attention_store: AttentionStore =None,save_self_attention: bool=True,disk_store=False):
super(Attention_Record_Processor, self).__init__(
save_self_attention=save_self_attention,
disk_store=disk_store)
self.additional_attention_store = additional_attention_store
self.attention_position_counter_dict = {
'down_cross': 0,
'mid_cross': 0,
'up_cross': 0,
'down_self': 0,
'mid_self': 0,
'up_self': 0,
}
#print("Modulated Ctrl at denoising steps: ", self.step_idx)
def update_attention_position_dict(self, current_attention_key):
self.attention_position_counter_dict[current_attention_key] +=1
def forward(self, sim, is_cross: bool, place_in_unet: str,**kwargs):
super(Attention_Record_Processor, self).forward(sim, is_cross, place_in_unet,**kwargs)
key = f"{place_in_unet}_{'cross' if is_cross else 'self'}"
self.update_attention_position_dict(key)
return sim
def between_steps(self):
super().between_steps()
self.step_store = self.get_empty_store()
self.attention_position_counter_dict = {
'down_cross': 0,
'mid_cross': 0,
'up_cross': 0,
'down_self': 0,
'mid_self': 0,
'up_self': 0,
}
return
class ST_Layout_Attn_ControlEdit(AttentionStore, abc.ABC):
def __init__(self, end_step=15, total_steps=50, step_idx=None, text_cond=None, sreg_maps=None, creg_maps=None, reg_sizes=None,reg_sizes_c=None,
time_steps=None,
clip_length=None,attention_type=None,
additional_attention_store: AttentionStore =None,
save_self_attention: bool=True,
disk_store=False,
video = None,
):
"""
Spatial-Temporal Layout-guided Attention (ST-Layout Attn) for Stable-Diffusion model
note: with vis cross attention weight function.
Args:
end_step: the step to end st-layout attn control
total_steps: the total number of steps
step_idx: list the steps to apply mutual self-attention control
text_cond: discrete text embedding for each region.
sreg_maps: spatial-temporal self-attention qk condition maps.
creg_maps: cross-attention qk condition maps
reg_sizes/reg_sizes_c: size regularzation maps for each instance in self_attn/cross_attention
clip_length: frames len of video
attention_type: FullyFrameAttention_sliced_attn/FullyFrameAttention/SparseCausalAttention
"""
super(ST_Layout_Attn_ControlEdit, self).__init__(
save_self_attention=save_self_attention,
disk_store=disk_store)
self.total_steps = total_steps
self.step_idx = list(range(0, end_step))
self.total_infer_steps = 50
self.text_cond = text_cond
self.sreg_maps = sreg_maps
self.creg_maps = creg_maps
self.reg_sizes = reg_sizes
self.reg_sizes_c = reg_sizes_c
self.clip_length = clip_length
self.attention_type = attention_type
self.sreg = .3
self.creg = 1.
self.count = 0
self.reg_part = .3
self.time_steps = time_steps
self.additional_attention_store = additional_attention_store
self.attention_position_counter_dict = {
'down_cross': 0,
'mid_cross': 0,
'up_cross': 0,
'down_self': 0,
'mid_self': 0,
'up_self': 0,
}
self.video = video
def update_attention_position_dict(self, current_attention_key):
self.attention_position_counter_dict[current_attention_key] +=1
def forward(self, sim, is_cross: bool, place_in_unet: str,**kwargs):
super(ST_Layout_Attn_ControlEdit, self).forward(sim, is_cross, place_in_unet,**kwargs)
# print("self.cur_step",self.cur_step)
key = f"{place_in_unet}_{'cross' if is_cross else 'self'}"
self.update_attention_position_dict(key)
if self.cur_step not in self.step_idx:
return sim
num_heads = sim.shape[1]
if num_heads == 1:
self.attention_type == "FullyFrameAttention_sliced_attn"
treg = torch.pow((self.time_steps[self.cur_step]-1)/1000, 5)
if not is_cross:
## Modulate self-attention
min_value = sim.min(-1)[0].unsqueeze(-1)
max_value = sim.max(-1)[0].unsqueeze(-1)
if self.attention_type == "SparseCausalAttention":
mask = self.sreg_maps[sim.size(2)].repeat(1,num_heads,1,1)
size_reg = self.reg_sizes[sim.size(2)].repeat(1,num_heads,1,1)
elif self.attention_type == "FullyFrameAttention":
mask = self.sreg_maps[sim.size(2)//self.clip_length].repeat(1,num_heads,1,1)
size_reg = self.reg_sizes[sim.size(2)//self.clip_length].repeat(1,num_heads,1,1)
elif self.attention_type == "FullyFrameAttention_sliced_attn":
mask = self.sreg_maps[sim.size(2)//self.clip_length]
size_reg = self.reg_sizes[sim.size(2)//self.clip_length]
else:
print("unknown attention type")
exit()
sim += (mask>0)*size_reg*self.sreg*treg*(max_value-sim)
sim -= ~(mask>0)*size_reg*self.sreg*treg*(sim-min_value)
else:
#Modulate cross-attention
min_value = sim.min(-1)[0].unsqueeze(-1)
max_value = sim.max(-1)[0].unsqueeze(-1)
mask = self.creg_maps[sim.size(2)].repeat(1,num_heads,1,1)
size_reg = self.reg_sizes_c[sim.size(2)].repeat(1,num_heads,1,1)
sim += (mask>0)*size_reg*self.creg*treg*(max_value-sim)
sim -= ~(mask>0)*size_reg*self.creg*treg*(sim-min_value)
self.count +=1
return sim
def between_steps(self):
super().between_steps()
self.step_store = self.get_empty_store()
self.attention_position_counter_dict = {
'down_cross': 0,
'mid_cross': 0,
'up_cross': 0,
'down_self': 0,
'mid_self': 0,
'up_self': 0,
}
return