Spaces:
Runtime error
Runtime error
File size: 7,915 Bytes
5fc5efa |
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 |
import os
import cv2
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
from typing import Optional, Union, Tuple, List, Callable, Dict
from torchvision.utils import save_image
from einops import rearrange, repeat
class AttentionBase:
def __init__(self):
self.cur_step = 0
self.num_att_layers = -1
self.cur_att_layer = 0
def after_step(self):
pass
def __call__(self, q, k, v, sim, attn, is_cross, place_in_unet, num_heads, **kwargs):
out = self.forward(q, k, v, sim, attn, is_cross, place_in_unet, num_heads, **kwargs)
self.cur_att_layer += 1
if self.cur_att_layer == self.num_att_layers:
self.cur_att_layer = 0
self.cur_step += 1
# after step
self.after_step()
return out
def forward(self, q, k, v, sim, attn, is_cross, place_in_unet, num_heads, **kwargs):
out = torch.einsum('b i j, b j d -> b i d', attn, v)
out = rearrange(out, '(b h) n d -> b n (h d)', h=num_heads)
return out
def reset(self):
self.cur_step = 0
self.cur_att_layer = 0
class AttentionStore(AttentionBase):
def __init__(self, res=[32], min_step=0, max_step=1000):
super().__init__()
self.res = res
self.min_step = min_step
self.max_step = max_step
self.valid_steps = 0
self.self_attns = [] # store the all attns
self.cross_attns = []
self.self_attns_step = [] # store the attns in each step
self.cross_attns_step = []
def after_step(self):
if self.cur_step > self.min_step and self.cur_step < self.max_step:
self.valid_steps += 1
if len(self.self_attns) == 0:
self.self_attns = self.self_attns_step
self.cross_attns = self.cross_attns_step
else:
for i in range(len(self.self_attns)):
self.self_attns[i] += self.self_attns_step[i]
self.cross_attns[i] += self.cross_attns_step[i]
self.self_attns_step.clear()
self.cross_attns_step.clear()
def forward(self, q, k, v, sim, attn, is_cross, place_in_unet, num_heads, **kwargs):
if attn.shape[1] <= 64 ** 2: # avoid OOM
if is_cross:
self.cross_attns_step.append(attn)
else:
self.self_attns_step.append(attn)
return super().forward(q, k, v, sim, attn, is_cross, place_in_unet, num_heads, **kwargs)
def regiter_attention_editor_diffusers(model, editor: AttentionBase):
"""
Register a attention editor to Diffuser Pipeline, refer from [Prompt-to-Prompt]
"""
def ca_forward(self, place_in_unet):
def forward(x, encoder_hidden_states=None, attention_mask=None, context=None, mask=None):
"""
The attention is similar to the original implementation of LDM CrossAttention class
except adding some modifications on the attention
"""
if encoder_hidden_states is not None:
context = encoder_hidden_states
if attention_mask is not None:
mask = attention_mask
to_out = self.to_out
if isinstance(to_out, nn.modules.container.ModuleList):
to_out = self.to_out[0]
else:
to_out = self.to_out
h = self.heads
q = self.to_q(x)
is_cross = context is not None
context = context if is_cross else x
k = self.to_k(context)
v = self.to_v(context)
q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> (b h) n d', h=h), (q, k, v))
sim = torch.einsum('b i d, b j d -> b i j', q, k) * self.scale
if mask is not None:
mask = rearrange(mask, 'b ... -> b (...)')
max_neg_value = -torch.finfo(sim.dtype).max
mask = repeat(mask, 'b j -> (b h) () j', h=h)
mask = mask[:, None, :].repeat(h, 1, 1)
sim.masked_fill_(~mask, max_neg_value)
attn = sim.softmax(dim=-1)
# the only difference
out = editor(
q, k, v, sim, attn, is_cross, place_in_unet,
self.heads, scale=self.scale)
return to_out(out)
return forward
def register_editor(net, count, place_in_unet):
for name, subnet in net.named_children():
if net.__class__.__name__ == 'Attention': # spatial Transformer layer
net.forward = ca_forward(net, place_in_unet)
return count + 1
elif hasattr(net, 'children'):
count = register_editor(subnet, count, place_in_unet)
return count
cross_att_count = 0
for net_name, net in model.unet.named_children():
if "down" in net_name:
cross_att_count += register_editor(net, 0, "down")
elif "mid" in net_name:
cross_att_count += register_editor(net, 0, "mid")
elif "up" in net_name:
cross_att_count += register_editor(net, 0, "up")
editor.num_att_layers = cross_att_count
def regiter_attention_editor_ldm(model, editor: AttentionBase):
"""
Register a attention editor to Stable Diffusion model, refer from [Prompt-to-Prompt]
"""
def ca_forward(self, place_in_unet):
def forward(x, encoder_hidden_states=None, attention_mask=None, context=None, mask=None):
"""
The attention is similar to the original implementation of LDM CrossAttention class
except adding some modifications on the attention
"""
if encoder_hidden_states is not None:
context = encoder_hidden_states
if attention_mask is not None:
mask = attention_mask
to_out = self.to_out
if isinstance(to_out, nn.modules.container.ModuleList):
to_out = self.to_out[0]
else:
to_out = self.to_out
h = self.heads
q = self.to_q(x)
is_cross = context is not None
context = context if is_cross else x
k = self.to_k(context)
v = self.to_v(context)
q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> (b h) n d', h=h), (q, k, v))
sim = torch.einsum('b i d, b j d -> b i j', q, k) * self.scale
if mask is not None:
mask = rearrange(mask, 'b ... -> b (...)')
max_neg_value = -torch.finfo(sim.dtype).max
mask = repeat(mask, 'b j -> (b h) () j', h=h)
mask = mask[:, None, :].repeat(h, 1, 1)
sim.masked_fill_(~mask, max_neg_value)
attn = sim.softmax(dim=-1)
# the only difference
out = editor(
q, k, v, sim, attn, is_cross, place_in_unet,
self.heads, scale=self.scale)
return to_out(out)
return forward
def register_editor(net, count, place_in_unet):
for name, subnet in net.named_children():
if net.__class__.__name__ == 'CrossAttention': # spatial Transformer layer
net.forward = ca_forward(net, place_in_unet)
return count + 1
elif hasattr(net, 'children'):
count = register_editor(subnet, count, place_in_unet)
return count
cross_att_count = 0
for net_name, net in model.model.diffusion_model.named_children():
if "input" in net_name:
cross_att_count += register_editor(net, 0, "input")
elif "middle" in net_name:
cross_att_count += register_editor(net, 0, "middle")
elif "output" in net_name:
cross_att_count += register_editor(net, 0, "output")
editor.num_att_layers = cross_att_count
|