Spaces:
Running
on
Zero
Running
on
Zero
SunderAli17
commited on
Commit
•
00fb38e
1
Parent(s):
2560b63
Create min_sdxl.py
Browse files- module/min_sdxl.py +907 -0
module/min_sdxl.py
ADDED
@@ -0,0 +1,907 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Modified from minSDXL by Simo Ryu:
|
2 |
+
# https://github.com/cloneofsimo/minSDXL ,
|
3 |
+
# which is in turn modified from the original code of:
|
4 |
+
# https://github.com/huggingface/diffusers
|
5 |
+
# So has APACHE 2.0 license
|
6 |
+
|
7 |
+
from typing import Optional, Union
|
8 |
+
|
9 |
+
import torch
|
10 |
+
import torch.nn as nn
|
11 |
+
import torch.nn.functional as F
|
12 |
+
import math
|
13 |
+
import inspect
|
14 |
+
|
15 |
+
from collections import namedtuple
|
16 |
+
|
17 |
+
from torch.fft import fftn, fftshift, ifftn, ifftshift
|
18 |
+
|
19 |
+
from diffusers.models.attention_processor import AttnProcessor, AttnProcessor2_0
|
20 |
+
|
21 |
+
# Implementation of FreeU for minsdxl
|
22 |
+
|
23 |
+
def fourier_filter(x_in: "torch.Tensor", threshold: int, scale: int) -> "torch.Tensor":
|
24 |
+
"""Fourier filter as introduced in FreeU (https://arxiv.org/abs/2309.11497).
|
25 |
+
This version of the method comes from here:
|
26 |
+
https://github.com/huggingface/diffusers/pull/5164#issuecomment-1732638706
|
27 |
+
"""
|
28 |
+
x = x_in
|
29 |
+
B, C, H, W = x.shape
|
30 |
+
|
31 |
+
# Non-power of 2 images must be float32
|
32 |
+
if (W & (W - 1)) != 0 or (H & (H - 1)) != 0:
|
33 |
+
x = x.to(dtype=torch.float32)
|
34 |
+
|
35 |
+
# FFT
|
36 |
+
x_freq = fftn(x, dim=(-2, -1))
|
37 |
+
x_freq = fftshift(x_freq, dim=(-2, -1))
|
38 |
+
|
39 |
+
B, C, H, W = x_freq.shape
|
40 |
+
mask = torch.ones((B, C, H, W), device=x.device)
|
41 |
+
|
42 |
+
crow, ccol = H // 2, W // 2
|
43 |
+
mask[..., crow - threshold : crow + threshold, ccol - threshold : ccol + threshold] = scale
|
44 |
+
x_freq = x_freq * mask
|
45 |
+
|
46 |
+
# IFFT
|
47 |
+
x_freq = ifftshift(x_freq, dim=(-2, -1))
|
48 |
+
x_filtered = ifftn(x_freq, dim=(-2, -1)).real
|
49 |
+
|
50 |
+
return x_filtered.to(dtype=x_in.dtype)
|
51 |
+
|
52 |
+
|
53 |
+
def apply_freeu(
|
54 |
+
resolution_idx: int, hidden_states: "torch.Tensor", res_hidden_states: "torch.Tensor", **freeu_kwargs):
|
55 |
+
"""Applies the FreeU mechanism as introduced in https:
|
56 |
+
//arxiv.org/abs/2309.11497. Adapted from the official code repository: https://github.com/ChenyangSi/FreeU.
|
57 |
+
Args:
|
58 |
+
resolution_idx (`int`): Integer denoting the UNet block where FreeU is being applied.
|
59 |
+
hidden_states (`torch.Tensor`): Inputs to the underlying block.
|
60 |
+
res_hidden_states (`torch.Tensor`): Features from the skip block corresponding to the underlying block.
|
61 |
+
s1 (`float`): Scaling factor for stage 1 to attenuate the contributions of the skip features.
|
62 |
+
s2 (`float`): Scaling factor for stage 2 to attenuate the contributions of the skip features.
|
63 |
+
b1 (`float`): Scaling factor for stage 1 to amplify the contributions of backbone features.
|
64 |
+
b2 (`float`): Scaling factor for stage 2 to amplify the contributions of backbone features.
|
65 |
+
"""
|
66 |
+
if resolution_idx == 0:
|
67 |
+
num_half_channels = hidden_states.shape[1] // 2
|
68 |
+
hidden_states[:, :num_half_channels] = hidden_states[:, :num_half_channels] * freeu_kwargs["b1"]
|
69 |
+
res_hidden_states = fourier_filter(res_hidden_states, threshold=1, scale=freeu_kwargs["s1"])
|
70 |
+
if resolution_idx == 1:
|
71 |
+
num_half_channels = hidden_states.shape[1] // 2
|
72 |
+
hidden_states[:, :num_half_channels] = hidden_states[:, :num_half_channels] * freeu_kwargs["b2"]
|
73 |
+
res_hidden_states = fourier_filter(res_hidden_states, threshold=1, scale=freeu_kwargs["s2"])
|
74 |
+
|
75 |
+
return hidden_states, res_hidden_states
|
76 |
+
|
77 |
+
# Diffusers-style LoRA to keep everything in the min_sdxl.py file
|
78 |
+
|
79 |
+
class LoRALinearLayer(nn.Module):
|
80 |
+
r"""
|
81 |
+
A linear layer that is used with LoRA.
|
82 |
+
Parameters:
|
83 |
+
in_features (`int`):
|
84 |
+
Number of input features.
|
85 |
+
out_features (`int`):
|
86 |
+
Number of output features.
|
87 |
+
rank (`int`, `optional`, defaults to 4):
|
88 |
+
The rank of the LoRA layer.
|
89 |
+
network_alpha (`float`, `optional`, defaults to `None`):
|
90 |
+
The value of the network alpha used for stable learning and preventing underflow. This value has the same
|
91 |
+
meaning as the `--network_alpha` option in the kohya-ss trainer script. See
|
92 |
+
https://github.com/darkstorm2150/sd-scripts/blob/main/docs/train_network_README-en.md#execute-learning
|
93 |
+
device (`torch.device`, `optional`, defaults to `None`):
|
94 |
+
The device to use for the layer's weights.
|
95 |
+
dtype (`torch.dtype`, `optional`, defaults to `None`):
|
96 |
+
The dtype to use for the layer's weights.
|
97 |
+
"""
|
98 |
+
|
99 |
+
def __init__(
|
100 |
+
self,
|
101 |
+
in_features: int,
|
102 |
+
out_features: int,
|
103 |
+
rank: int = 4,
|
104 |
+
network_alpha: Optional[float] = None,
|
105 |
+
device: Optional[Union[torch.device, str]] = None,
|
106 |
+
dtype: Optional[torch.dtype] = None,
|
107 |
+
):
|
108 |
+
super().__init__()
|
109 |
+
|
110 |
+
self.down = nn.Linear(in_features, rank, bias=False, device=device, dtype=dtype)
|
111 |
+
self.up = nn.Linear(rank, out_features, bias=False, device=device, dtype=dtype)
|
112 |
+
# This value has the same meaning as the `--network_alpha` option in the kohya-ss trainer script.
|
113 |
+
# See https://github.com/darkstorm2150/sd-scripts/blob/main/docs/train_network_README-en.md#execute-learning
|
114 |
+
self.network_alpha = network_alpha
|
115 |
+
self.rank = rank
|
116 |
+
self.out_features = out_features
|
117 |
+
self.in_features = in_features
|
118 |
+
|
119 |
+
nn.init.normal_(self.down.weight, std=1 / rank)
|
120 |
+
nn.init.zeros_(self.up.weight)
|
121 |
+
|
122 |
+
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
|
123 |
+
orig_dtype = hidden_states.dtype
|
124 |
+
dtype = self.down.weight.dtype
|
125 |
+
|
126 |
+
down_hidden_states = self.down(hidden_states.to(dtype))
|
127 |
+
up_hidden_states = self.up(down_hidden_states)
|
128 |
+
|
129 |
+
if self.network_alpha is not None:
|
130 |
+
up_hidden_states *= self.network_alpha / self.rank
|
131 |
+
|
132 |
+
return up_hidden_states.to(orig_dtype)
|
133 |
+
|
134 |
+
class LoRACompatibleLinear(nn.Linear):
|
135 |
+
"""
|
136 |
+
A Linear layer that can be used with LoRA.
|
137 |
+
"""
|
138 |
+
|
139 |
+
def __init__(self, *args, lora_layer: Optional[LoRALinearLayer] = None, **kwargs):
|
140 |
+
super().__init__(*args, **kwargs)
|
141 |
+
self.lora_layer = lora_layer
|
142 |
+
|
143 |
+
def set_lora_layer(self, lora_layer: Optional[LoRALinearLayer]):
|
144 |
+
self.lora_layer = lora_layer
|
145 |
+
|
146 |
+
def _fuse_lora(self, lora_scale: float = 1.0, safe_fusing: bool = False):
|
147 |
+
if self.lora_layer is None:
|
148 |
+
return
|
149 |
+
|
150 |
+
dtype, device = self.weight.data.dtype, self.weight.data.device
|
151 |
+
|
152 |
+
w_orig = self.weight.data.float()
|
153 |
+
w_up = self.lora_layer.up.weight.data.float()
|
154 |
+
w_down = self.lora_layer.down.weight.data.float()
|
155 |
+
|
156 |
+
if self.lora_layer.network_alpha is not None:
|
157 |
+
w_up = w_up * self.lora_layer.network_alpha / self.lora_layer.rank
|
158 |
+
|
159 |
+
fused_weight = w_orig + (lora_scale * torch.bmm(w_up[None, :], w_down[None, :])[0])
|
160 |
+
|
161 |
+
if safe_fusing and torch.isnan(fused_weight).any().item():
|
162 |
+
raise ValueError(
|
163 |
+
"This LoRA weight seems to be broken. "
|
164 |
+
f"Encountered NaN values when trying to fuse LoRA weights for {self}."
|
165 |
+
"LoRA weights will not be fused."
|
166 |
+
)
|
167 |
+
|
168 |
+
self.weight.data = fused_weight.to(device=device, dtype=dtype)
|
169 |
+
|
170 |
+
# we can drop the lora layer now
|
171 |
+
self.lora_layer = None
|
172 |
+
|
173 |
+
# offload the up and down matrices to CPU to not blow the memory
|
174 |
+
self.w_up = w_up.cpu()
|
175 |
+
self.w_down = w_down.cpu()
|
176 |
+
self._lora_scale = lora_scale
|
177 |
+
|
178 |
+
def _unfuse_lora(self):
|
179 |
+
if not (getattr(self, "w_up", None) is not None and getattr(self, "w_down", None) is not None):
|
180 |
+
return
|
181 |
+
|
182 |
+
fused_weight = self.weight.data
|
183 |
+
dtype, device = fused_weight.dtype, fused_weight.device
|
184 |
+
|
185 |
+
w_up = self.w_up.to(device=device).float()
|
186 |
+
w_down = self.w_down.to(device).float()
|
187 |
+
|
188 |
+
unfused_weight = fused_weight.float() - (self._lora_scale * torch.bmm(w_up[None, :], w_down[None, :])[0])
|
189 |
+
self.weight.data = unfused_weight.to(device=device, dtype=dtype)
|
190 |
+
|
191 |
+
self.w_up = None
|
192 |
+
self.w_down = None
|
193 |
+
|
194 |
+
def forward(self, hidden_states: torch.Tensor, scale: float = 1.0) -> torch.Tensor:
|
195 |
+
if self.lora_layer is None:
|
196 |
+
out = super().forward(hidden_states)
|
197 |
+
return out
|
198 |
+
else:
|
199 |
+
out = super().forward(hidden_states) + (scale * self.lora_layer(hidden_states))
|
200 |
+
return out
|
201 |
+
|
202 |
+
class Timesteps(nn.Module):
|
203 |
+
def __init__(self, num_channels: int = 320):
|
204 |
+
super().__init__()
|
205 |
+
self.num_channels = num_channels
|
206 |
+
|
207 |
+
def forward(self, timesteps):
|
208 |
+
half_dim = self.num_channels // 2
|
209 |
+
exponent = -math.log(10000) * torch.arange(
|
210 |
+
half_dim, dtype=torch.float32, device=timesteps.device
|
211 |
+
)
|
212 |
+
exponent = exponent / (half_dim - 0.0)
|
213 |
+
|
214 |
+
emb = torch.exp(exponent)
|
215 |
+
emb = timesteps[:, None].float() * emb[None, :]
|
216 |
+
|
217 |
+
sin_emb = torch.sin(emb)
|
218 |
+
cos_emb = torch.cos(emb)
|
219 |
+
emb = torch.cat([cos_emb, sin_emb], dim=-1)
|
220 |
+
|
221 |
+
return emb
|
222 |
+
|
223 |
+
|
224 |
+
class TimestepEmbedding(nn.Module):
|
225 |
+
def __init__(self, in_features, out_features):
|
226 |
+
super(TimestepEmbedding, self).__init__()
|
227 |
+
self.linear_1 = nn.Linear(in_features, out_features, bias=True)
|
228 |
+
self.act = nn.SiLU()
|
229 |
+
self.linear_2 = nn.Linear(out_features, out_features, bias=True)
|
230 |
+
|
231 |
+
def forward(self, sample):
|
232 |
+
sample = self.linear_1(sample)
|
233 |
+
sample = self.act(sample)
|
234 |
+
sample = self.linear_2(sample)
|
235 |
+
|
236 |
+
return sample
|
237 |
+
|
238 |
+
|
239 |
+
class ResnetBlock2D(nn.Module):
|
240 |
+
def __init__(self, in_channels, out_channels, conv_shortcut=True):
|
241 |
+
super(ResnetBlock2D, self).__init__()
|
242 |
+
self.norm1 = nn.GroupNorm(32, in_channels, eps=1e-05, affine=True)
|
243 |
+
self.conv1 = nn.Conv2d(
|
244 |
+
in_channels, out_channels, kernel_size=3, stride=1, padding=1
|
245 |
+
)
|
246 |
+
self.time_emb_proj = nn.Linear(1280, out_channels, bias=True)
|
247 |
+
self.norm2 = nn.GroupNorm(32, out_channels, eps=1e-05, affine=True)
|
248 |
+
self.dropout = nn.Dropout(p=0.0, inplace=False)
|
249 |
+
self.conv2 = nn.Conv2d(
|
250 |
+
out_channels, out_channels, kernel_size=3, stride=1, padding=1
|
251 |
+
)
|
252 |
+
self.nonlinearity = nn.SiLU()
|
253 |
+
self.conv_shortcut = None
|
254 |
+
if conv_shortcut:
|
255 |
+
self.conv_shortcut = nn.Conv2d(
|
256 |
+
in_channels, out_channels, kernel_size=1, stride=1
|
257 |
+
)
|
258 |
+
|
259 |
+
def forward(self, input_tensor, temb):
|
260 |
+
hidden_states = input_tensor
|
261 |
+
hidden_states = self.norm1(hidden_states)
|
262 |
+
hidden_states = self.nonlinearity(hidden_states)
|
263 |
+
|
264 |
+
hidden_states = self.conv1(hidden_states)
|
265 |
+
|
266 |
+
temb = self.nonlinearity(temb)
|
267 |
+
temb = self.time_emb_proj(temb)[:, :, None, None]
|
268 |
+
hidden_states = hidden_states + temb
|
269 |
+
hidden_states = self.norm2(hidden_states)
|
270 |
+
|
271 |
+
hidden_states = self.nonlinearity(hidden_states)
|
272 |
+
hidden_states = self.dropout(hidden_states)
|
273 |
+
hidden_states = self.conv2(hidden_states)
|
274 |
+
|
275 |
+
if self.conv_shortcut is not None:
|
276 |
+
input_tensor = self.conv_shortcut(input_tensor)
|
277 |
+
|
278 |
+
output_tensor = input_tensor + hidden_states
|
279 |
+
|
280 |
+
return output_tensor
|
281 |
+
|
282 |
+
|
283 |
+
class Attention(nn.Module):
|
284 |
+
def __init__(
|
285 |
+
self, inner_dim, cross_attention_dim=None, num_heads=None, dropout=0.0, processor=None, scale_qk=True
|
286 |
+
):
|
287 |
+
super(Attention, self).__init__()
|
288 |
+
if num_heads is None:
|
289 |
+
self.head_dim = 64
|
290 |
+
self.num_heads = inner_dim // self.head_dim
|
291 |
+
else:
|
292 |
+
self.num_heads = num_heads
|
293 |
+
self.head_dim = inner_dim // num_heads
|
294 |
+
|
295 |
+
self.scale = self.head_dim**-0.5
|
296 |
+
if cross_attention_dim is None:
|
297 |
+
cross_attention_dim = inner_dim
|
298 |
+
self.to_q = LoRACompatibleLinear(inner_dim, inner_dim, bias=False)
|
299 |
+
self.to_k = LoRACompatibleLinear(cross_attention_dim, inner_dim, bias=False)
|
300 |
+
self.to_v = LoRACompatibleLinear(cross_attention_dim, inner_dim, bias=False)
|
301 |
+
|
302 |
+
self.to_out = nn.ModuleList(
|
303 |
+
[LoRACompatibleLinear(inner_dim, inner_dim), nn.Dropout(dropout, inplace=False)]
|
304 |
+
)
|
305 |
+
|
306 |
+
self.scale_qk = scale_qk
|
307 |
+
if processor is None:
|
308 |
+
processor = (
|
309 |
+
AttnProcessor2_0() if hasattr(F, "scaled_dot_product_attention") and self.scale_qk else AttnProcessor()
|
310 |
+
)
|
311 |
+
self.set_processor(processor)
|
312 |
+
|
313 |
+
def forward(
|
314 |
+
self,
|
315 |
+
hidden_states: torch.FloatTensor,
|
316 |
+
encoder_hidden_states: Optional[torch.FloatTensor] = None,
|
317 |
+
attention_mask: Optional[torch.FloatTensor] = None,
|
318 |
+
**cross_attention_kwargs,
|
319 |
+
) -> torch.Tensor:
|
320 |
+
r"""
|
321 |
+
The forward method of the `Attention` class.
|
322 |
+
Args:
|
323 |
+
hidden_states (`torch.Tensor`):
|
324 |
+
The hidden states of the query.
|
325 |
+
encoder_hidden_states (`torch.Tensor`, *optional*):
|
326 |
+
The hidden states of the encoder.
|
327 |
+
attention_mask (`torch.Tensor`, *optional*):
|
328 |
+
The attention mask to use. If `None`, no mask is applied.
|
329 |
+
**cross_attention_kwargs:
|
330 |
+
Additional keyword arguments to pass along to the cross attention.
|
331 |
+
Returns:
|
332 |
+
`torch.Tensor`: The output of the attention layer.
|
333 |
+
"""
|
334 |
+
# The `Attention` class can call different attention processors / attention functions
|
335 |
+
# here we simply pass along all tensors to the selected processor class
|
336 |
+
# For standard processors that are defined here, `**cross_attention_kwargs` is empty
|
337 |
+
|
338 |
+
attn_parameters = set(inspect.signature(self.processor.__call__).parameters.keys())
|
339 |
+
unused_kwargs = [k for k, _ in cross_attention_kwargs.items() if k not in attn_parameters]
|
340 |
+
if len(unused_kwargs) > 0:
|
341 |
+
print(
|
342 |
+
f"cross_attention_kwargs {unused_kwargs} are not expected by {self.processor.__class__.__name__} and will be ignored."
|
343 |
+
)
|
344 |
+
cross_attention_kwargs = {k: w for k, w in cross_attention_kwargs.items() if k in attn_parameters}
|
345 |
+
|
346 |
+
return self.processor(
|
347 |
+
self,
|
348 |
+
hidden_states,
|
349 |
+
encoder_hidden_states=encoder_hidden_states,
|
350 |
+
attention_mask=attention_mask,
|
351 |
+
**cross_attention_kwargs,
|
352 |
+
)
|
353 |
+
|
354 |
+
def orig_forward(self, hidden_states, encoder_hidden_states=None):
|
355 |
+
q = self.to_q(hidden_states)
|
356 |
+
k = (
|
357 |
+
self.to_k(encoder_hidden_states)
|
358 |
+
if encoder_hidden_states is not None
|
359 |
+
else self.to_k(hidden_states)
|
360 |
+
)
|
361 |
+
v = (
|
362 |
+
self.to_v(encoder_hidden_states)
|
363 |
+
if encoder_hidden_states is not None
|
364 |
+
else self.to_v(hidden_states)
|
365 |
+
)
|
366 |
+
b, t, c = q.size()
|
367 |
+
|
368 |
+
q = q.view(q.size(0), q.size(1), self.num_heads, self.head_dim).transpose(1, 2)
|
369 |
+
k = k.view(k.size(0), k.size(1), self.num_heads, self.head_dim).transpose(1, 2)
|
370 |
+
v = v.view(v.size(0), v.size(1), self.num_heads, self.head_dim).transpose(1, 2)
|
371 |
+
|
372 |
+
# scores = torch.matmul(q, k.transpose(-2, -1)) * self.scale
|
373 |
+
# attn_weights = torch.softmax(scores, dim=-1)
|
374 |
+
# attn_output = torch.matmul(attn_weights, v)
|
375 |
+
|
376 |
+
attn_output = F.scaled_dot_product_attention(
|
377 |
+
q, k, v, attn_mask=None, dropout_p=0.0, is_causal=False, scale=self.scale,
|
378 |
+
)
|
379 |
+
|
380 |
+
attn_output = attn_output.transpose(1, 2).contiguous().view(b, t, c)
|
381 |
+
|
382 |
+
for layer in self.to_out:
|
383 |
+
attn_output = layer(attn_output)
|
384 |
+
|
385 |
+
return attn_output
|
386 |
+
|
387 |
+
def set_processor(self, processor) -> None:
|
388 |
+
r"""
|
389 |
+
Set the attention processor to use.
|
390 |
+
Args:
|
391 |
+
processor (`AttnProcessor`):
|
392 |
+
The attention processor to use.
|
393 |
+
"""
|
394 |
+
# if current processor is in `self._modules` and if passed `processor` is not, we need to
|
395 |
+
# pop `processor` from `self._modules`
|
396 |
+
if (
|
397 |
+
hasattr(self, "processor")
|
398 |
+
and isinstance(self.processor, torch.nn.Module)
|
399 |
+
and not isinstance(processor, torch.nn.Module)
|
400 |
+
):
|
401 |
+
print(f"You are removing possibly trained weights of {self.processor} with {processor}")
|
402 |
+
self._modules.pop("processor")
|
403 |
+
|
404 |
+
self.processor = processor
|
405 |
+
|
406 |
+
def get_processor(self, return_deprecated_lora: bool = False):
|
407 |
+
r"""
|
408 |
+
Get the attention processor in use.
|
409 |
+
Args:
|
410 |
+
return_deprecated_lora (`bool`, *optional*, defaults to `False`):
|
411 |
+
Set to `True` to return the deprecated LoRA attention processor.
|
412 |
+
Returns:
|
413 |
+
"AttentionProcessor": The attention processor in use.
|
414 |
+
"""
|
415 |
+
if not return_deprecated_lora:
|
416 |
+
return self.processor
|
417 |
+
|
418 |
+
# TODO(Sayak, Patrick). The rest of the function is needed to ensure backwards compatible
|
419 |
+
# serialization format for LoRA Attention Processors. It should be deleted once the integration
|
420 |
+
# with PEFT is completed.
|
421 |
+
is_lora_activated = {
|
422 |
+
name: module.lora_layer is not None
|
423 |
+
for name, module in self.named_modules()
|
424 |
+
if hasattr(module, "lora_layer")
|
425 |
+
}
|
426 |
+
|
427 |
+
# 1. if no layer has a LoRA activated we can return the processor as usual
|
428 |
+
if not any(is_lora_activated.values()):
|
429 |
+
return self.processor
|
430 |
+
|
431 |
+
# If doesn't apply LoRA do `add_k_proj` or `add_v_proj`
|
432 |
+
is_lora_activated.pop("add_k_proj", None)
|
433 |
+
is_lora_activated.pop("add_v_proj", None)
|
434 |
+
# 2. else it is not possible that only some layers have LoRA activated
|
435 |
+
if not all(is_lora_activated.values()):
|
436 |
+
raise ValueError(
|
437 |
+
f"Make sure that either all layers or no layers have LoRA activated, but have {is_lora_activated}"
|
438 |
+
)
|
439 |
+
|
440 |
+
# 3. And we need to merge the current LoRA layers into the corresponding LoRA attention processor
|
441 |
+
non_lora_processor_cls_name = self.processor.__class__.__name__
|
442 |
+
lora_processor_cls = getattr(import_module(__name__), "LoRA" + non_lora_processor_cls_name)
|
443 |
+
|
444 |
+
hidden_size = self.inner_dim
|
445 |
+
|
446 |
+
# now create a LoRA attention processor from the LoRA layers
|
447 |
+
if lora_processor_cls in [LoRAAttnProcessor, LoRAAttnProcessor2_0, LoRAXFormersAttnProcessor]:
|
448 |
+
kwargs = {
|
449 |
+
"cross_attention_dim": self.cross_attention_dim,
|
450 |
+
"rank": self.to_q.lora_layer.rank,
|
451 |
+
"network_alpha": self.to_q.lora_layer.network_alpha,
|
452 |
+
"q_rank": self.to_q.lora_layer.rank,
|
453 |
+
"q_hidden_size": self.to_q.lora_layer.out_features,
|
454 |
+
"k_rank": self.to_k.lora_layer.rank,
|
455 |
+
"k_hidden_size": self.to_k.lora_layer.out_features,
|
456 |
+
"v_rank": self.to_v.lora_layer.rank,
|
457 |
+
"v_hidden_size": self.to_v.lora_layer.out_features,
|
458 |
+
"out_rank": self.to_out[0].lora_layer.rank,
|
459 |
+
"out_hidden_size": self.to_out[0].lora_layer.out_features,
|
460 |
+
}
|
461 |
+
|
462 |
+
if hasattr(self.processor, "attention_op"):
|
463 |
+
kwargs["attention_op"] = self.processor.attention_op
|
464 |
+
|
465 |
+
lora_processor = lora_processor_cls(hidden_size, **kwargs)
|
466 |
+
lora_processor.to_q_lora.load_state_dict(self.to_q.lora_layer.state_dict())
|
467 |
+
lora_processor.to_k_lora.load_state_dict(self.to_k.lora_layer.state_dict())
|
468 |
+
lora_processor.to_v_lora.load_state_dict(self.to_v.lora_layer.state_dict())
|
469 |
+
lora_processor.to_out_lora.load_state_dict(self.to_out[0].lora_layer.state_dict())
|
470 |
+
elif lora_processor_cls == LoRAAttnAddedKVProcessor:
|
471 |
+
lora_processor = lora_processor_cls(
|
472 |
+
hidden_size,
|
473 |
+
cross_attention_dim=self.add_k_proj.weight.shape[0],
|
474 |
+
rank=self.to_q.lora_layer.rank,
|
475 |
+
network_alpha=self.to_q.lora_layer.network_alpha,
|
476 |
+
)
|
477 |
+
lora_processor.to_q_lora.load_state_dict(self.to_q.lora_layer.state_dict())
|
478 |
+
lora_processor.to_k_lora.load_state_dict(self.to_k.lora_layer.state_dict())
|
479 |
+
lora_processor.to_v_lora.load_state_dict(self.to_v.lora_layer.state_dict())
|
480 |
+
lora_processor.to_out_lora.load_state_dict(self.to_out[0].lora_layer.state_dict())
|
481 |
+
|
482 |
+
# only save if used
|
483 |
+
if self.add_k_proj.lora_layer is not None:
|
484 |
+
lora_processor.add_k_proj_lora.load_state_dict(self.add_k_proj.lora_layer.state_dict())
|
485 |
+
lora_processor.add_v_proj_lora.load_state_dict(self.add_v_proj.lora_layer.state_dict())
|
486 |
+
else:
|
487 |
+
lora_processor.add_k_proj_lora = None
|
488 |
+
lora_processor.add_v_proj_lora = None
|
489 |
+
else:
|
490 |
+
raise ValueError(f"{lora_processor_cls} does not exist.")
|
491 |
+
|
492 |
+
return lora_processor
|
493 |
+
|
494 |
+
class GEGLU(nn.Module):
|
495 |
+
def __init__(self, in_features, out_features):
|
496 |
+
super(GEGLU, self).__init__()
|
497 |
+
self.proj = nn.Linear(in_features, out_features * 2, bias=True)
|
498 |
+
|
499 |
+
def forward(self, x):
|
500 |
+
x_proj = self.proj(x)
|
501 |
+
x1, x2 = x_proj.chunk(2, dim=-1)
|
502 |
+
return x1 * torch.nn.functional.gelu(x2)
|
503 |
+
|
504 |
+
|
505 |
+
class FeedForward(nn.Module):
|
506 |
+
def __init__(self, in_features, out_features):
|
507 |
+
super(FeedForward, self).__init__()
|
508 |
+
|
509 |
+
self.net = nn.ModuleList(
|
510 |
+
[
|
511 |
+
GEGLU(in_features, out_features * 4),
|
512 |
+
nn.Dropout(p=0.0, inplace=False),
|
513 |
+
nn.Linear(out_features * 4, out_features, bias=True),
|
514 |
+
]
|
515 |
+
)
|
516 |
+
|
517 |
+
def forward(self, x):
|
518 |
+
for layer in self.net:
|
519 |
+
x = layer(x)
|
520 |
+
return x
|
521 |
+
|
522 |
+
|
523 |
+
class BasicTransformerBlock(nn.Module):
|
524 |
+
def __init__(self, hidden_size):
|
525 |
+
super(BasicTransformerBlock, self).__init__()
|
526 |
+
self.norm1 = nn.LayerNorm(hidden_size, eps=1e-05, elementwise_affine=True)
|
527 |
+
self.attn1 = Attention(hidden_size)
|
528 |
+
self.norm2 = nn.LayerNorm(hidden_size, eps=1e-05, elementwise_affine=True)
|
529 |
+
self.attn2 = Attention(hidden_size, 2048)
|
530 |
+
self.norm3 = nn.LayerNorm(hidden_size, eps=1e-05, elementwise_affine=True)
|
531 |
+
self.ff = FeedForward(hidden_size, hidden_size)
|
532 |
+
|
533 |
+
def forward(self, x, encoder_hidden_states=None):
|
534 |
+
residual = x
|
535 |
+
|
536 |
+
x = self.norm1(x)
|
537 |
+
x = self.attn1(x)
|
538 |
+
x = x + residual
|
539 |
+
|
540 |
+
residual = x
|
541 |
+
|
542 |
+
x = self.norm2(x)
|
543 |
+
if encoder_hidden_states is not None:
|
544 |
+
x = self.attn2(x, encoder_hidden_states)
|
545 |
+
else:
|
546 |
+
x = self.attn2(x)
|
547 |
+
x = x + residual
|
548 |
+
|
549 |
+
residual = x
|
550 |
+
|
551 |
+
x = self.norm3(x)
|
552 |
+
x = self.ff(x)
|
553 |
+
x = x + residual
|
554 |
+
return x
|
555 |
+
|
556 |
+
|
557 |
+
class Transformer2DModel(nn.Module):
|
558 |
+
def __init__(self, in_channels, out_channels, n_layers):
|
559 |
+
super(Transformer2DModel, self).__init__()
|
560 |
+
self.norm = nn.GroupNorm(32, in_channels, eps=1e-06, affine=True)
|
561 |
+
self.proj_in = nn.Linear(in_channels, out_channels, bias=True)
|
562 |
+
self.transformer_blocks = nn.ModuleList(
|
563 |
+
[BasicTransformerBlock(out_channels) for _ in range(n_layers)]
|
564 |
+
)
|
565 |
+
self.proj_out = nn.Linear(out_channels, out_channels, bias=True)
|
566 |
+
|
567 |
+
def forward(self, hidden_states, encoder_hidden_states=None):
|
568 |
+
batch, _, height, width = hidden_states.shape
|
569 |
+
res = hidden_states
|
570 |
+
hidden_states = self.norm(hidden_states)
|
571 |
+
inner_dim = hidden_states.shape[1]
|
572 |
+
hidden_states = hidden_states.permute(0, 2, 3, 1).reshape(
|
573 |
+
batch, height * width, inner_dim
|
574 |
+
)
|
575 |
+
hidden_states = self.proj_in(hidden_states)
|
576 |
+
|
577 |
+
for block in self.transformer_blocks:
|
578 |
+
hidden_states = block(hidden_states, encoder_hidden_states)
|
579 |
+
|
580 |
+
hidden_states = self.proj_out(hidden_states)
|
581 |
+
hidden_states = (
|
582 |
+
hidden_states.reshape(batch, height, width, inner_dim)
|
583 |
+
.permute(0, 3, 1, 2)
|
584 |
+
.contiguous()
|
585 |
+
)
|
586 |
+
|
587 |
+
return hidden_states + res
|
588 |
+
|
589 |
+
|
590 |
+
class Downsample2D(nn.Module):
|
591 |
+
def __init__(self, in_channels, out_channels):
|
592 |
+
super(Downsample2D, self).__init__()
|
593 |
+
self.conv = nn.Conv2d(
|
594 |
+
in_channels, out_channels, kernel_size=3, stride=2, padding=1
|
595 |
+
)
|
596 |
+
|
597 |
+
def forward(self, x):
|
598 |
+
return self.conv(x)
|
599 |
+
|
600 |
+
|
601 |
+
class Upsample2D(nn.Module):
|
602 |
+
def __init__(self, in_channels, out_channels):
|
603 |
+
super(Upsample2D, self).__init__()
|
604 |
+
self.conv = nn.Conv2d(
|
605 |
+
in_channels, out_channels, kernel_size=3, stride=1, padding=1
|
606 |
+
)
|
607 |
+
|
608 |
+
def forward(self, x):
|
609 |
+
x = F.interpolate(x, scale_factor=2.0, mode="nearest")
|
610 |
+
return self.conv(x)
|
611 |
+
|
612 |
+
|
613 |
+
class DownBlock2D(nn.Module):
|
614 |
+
def __init__(self, in_channels, out_channels):
|
615 |
+
super(DownBlock2D, self).__init__()
|
616 |
+
self.resnets = nn.ModuleList(
|
617 |
+
[
|
618 |
+
ResnetBlock2D(in_channels, out_channels, conv_shortcut=False),
|
619 |
+
ResnetBlock2D(out_channels, out_channels, conv_shortcut=False),
|
620 |
+
]
|
621 |
+
)
|
622 |
+
self.downsamplers = nn.ModuleList([Downsample2D(out_channels, out_channels)])
|
623 |
+
|
624 |
+
def forward(self, hidden_states, temb):
|
625 |
+
output_states = []
|
626 |
+
for module in self.resnets:
|
627 |
+
hidden_states = module(hidden_states, temb)
|
628 |
+
output_states.append(hidden_states)
|
629 |
+
|
630 |
+
hidden_states = self.downsamplers[0](hidden_states)
|
631 |
+
output_states.append(hidden_states)
|
632 |
+
|
633 |
+
return hidden_states, output_states
|
634 |
+
|
635 |
+
|
636 |
+
class CrossAttnDownBlock2D(nn.Module):
|
637 |
+
def __init__(self, in_channels, out_channels, n_layers, has_downsamplers=True):
|
638 |
+
super(CrossAttnDownBlock2D, self).__init__()
|
639 |
+
self.attentions = nn.ModuleList(
|
640 |
+
[
|
641 |
+
Transformer2DModel(out_channels, out_channels, n_layers),
|
642 |
+
Transformer2DModel(out_channels, out_channels, n_layers),
|
643 |
+
]
|
644 |
+
)
|
645 |
+
self.resnets = nn.ModuleList(
|
646 |
+
[
|
647 |
+
ResnetBlock2D(in_channels, out_channels),
|
648 |
+
ResnetBlock2D(out_channels, out_channels, conv_shortcut=False),
|
649 |
+
]
|
650 |
+
)
|
651 |
+
self.downsamplers = None
|
652 |
+
if has_downsamplers:
|
653 |
+
self.downsamplers = nn.ModuleList(
|
654 |
+
[Downsample2D(out_channels, out_channels)]
|
655 |
+
)
|
656 |
+
|
657 |
+
def forward(self, hidden_states, temb, encoder_hidden_states):
|
658 |
+
output_states = []
|
659 |
+
for resnet, attn in zip(self.resnets, self.attentions):
|
660 |
+
hidden_states = resnet(hidden_states, temb)
|
661 |
+
hidden_states = attn(
|
662 |
+
hidden_states,
|
663 |
+
encoder_hidden_states=encoder_hidden_states,
|
664 |
+
)
|
665 |
+
output_states.append(hidden_states)
|
666 |
+
|
667 |
+
if self.downsamplers is not None:
|
668 |
+
hidden_states = self.downsamplers[0](hidden_states)
|
669 |
+
output_states.append(hidden_states)
|
670 |
+
|
671 |
+
return hidden_states, output_states
|
672 |
+
|
673 |
+
|
674 |
+
class CrossAttnUpBlock2D(nn.Module):
|
675 |
+
def __init__(self, in_channels, out_channels, prev_output_channel, n_layers):
|
676 |
+
super(CrossAttnUpBlock2D, self).__init__()
|
677 |
+
self.attentions = nn.ModuleList(
|
678 |
+
[
|
679 |
+
Transformer2DModel(out_channels, out_channels, n_layers),
|
680 |
+
Transformer2DModel(out_channels, out_channels, n_layers),
|
681 |
+
Transformer2DModel(out_channels, out_channels, n_layers),
|
682 |
+
]
|
683 |
+
)
|
684 |
+
self.resnets = nn.ModuleList(
|
685 |
+
[
|
686 |
+
ResnetBlock2D(prev_output_channel + out_channels, out_channels),
|
687 |
+
ResnetBlock2D(2 * out_channels, out_channels),
|
688 |
+
ResnetBlock2D(out_channels + in_channels, out_channels),
|
689 |
+
]
|
690 |
+
)
|
691 |
+
self.upsamplers = nn.ModuleList([Upsample2D(out_channels, out_channels)])
|
692 |
+
|
693 |
+
def forward(
|
694 |
+
self, hidden_states, res_hidden_states_tuple, temb, encoder_hidden_states
|
695 |
+
):
|
696 |
+
for resnet, attn in zip(self.resnets, self.attentions):
|
697 |
+
# pop res hidden states
|
698 |
+
res_hidden_states = res_hidden_states_tuple[-1]
|
699 |
+
res_hidden_states_tuple = res_hidden_states_tuple[:-1]
|
700 |
+
hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)
|
701 |
+
hidden_states = resnet(hidden_states, temb)
|
702 |
+
hidden_states = attn(
|
703 |
+
hidden_states,
|
704 |
+
encoder_hidden_states=encoder_hidden_states,
|
705 |
+
)
|
706 |
+
|
707 |
+
if self.upsamplers is not None:
|
708 |
+
for upsampler in self.upsamplers:
|
709 |
+
hidden_states = upsampler(hidden_states)
|
710 |
+
|
711 |
+
return hidden_states
|
712 |
+
|
713 |
+
|
714 |
+
class UpBlock2D(nn.Module):
|
715 |
+
def __init__(self, in_channels, out_channels, prev_output_channel):
|
716 |
+
super(UpBlock2D, self).__init__()
|
717 |
+
self.resnets = nn.ModuleList(
|
718 |
+
[
|
719 |
+
ResnetBlock2D(out_channels + prev_output_channel, out_channels),
|
720 |
+
ResnetBlock2D(out_channels * 2, out_channels),
|
721 |
+
ResnetBlock2D(out_channels + in_channels, out_channels),
|
722 |
+
]
|
723 |
+
)
|
724 |
+
|
725 |
+
def forward(self, hidden_states, res_hidden_states_tuple, temb=None):
|
726 |
+
|
727 |
+
is_freeu_enabled = (
|
728 |
+
getattr(self, "s1", None)
|
729 |
+
and getattr(self, "s2", None)
|
730 |
+
and getattr(self, "b1", None)
|
731 |
+
and getattr(self, "b2", None)
|
732 |
+
and getattr(self, "resolution_idx", None)
|
733 |
+
)
|
734 |
+
|
735 |
+
for resnet in self.resnets:
|
736 |
+
res_hidden_states = res_hidden_states_tuple[-1]
|
737 |
+
res_hidden_states_tuple = res_hidden_states_tuple[:-1]
|
738 |
+
|
739 |
+
|
740 |
+
if is_freeu_enabled:
|
741 |
+
hidden_states, res_hidden_states = apply_freeu(
|
742 |
+
self.resolution_idx,
|
743 |
+
hidden_states,
|
744 |
+
res_hidden_states,
|
745 |
+
s1=self.s1,
|
746 |
+
s2=self.s2,
|
747 |
+
b1=self.b1,
|
748 |
+
b2=self.b2,
|
749 |
+
)
|
750 |
+
|
751 |
+
hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)
|
752 |
+
hidden_states = resnet(hidden_states, temb)
|
753 |
+
|
754 |
+
return hidden_states
|
755 |
+
|
756 |
+
class UNetMidBlock2DCrossAttn(nn.Module):
|
757 |
+
def __init__(self, in_features):
|
758 |
+
super(UNetMidBlock2DCrossAttn, self).__init__()
|
759 |
+
self.attentions = nn.ModuleList(
|
760 |
+
[Transformer2DModel(in_features, in_features, n_layers=10)]
|
761 |
+
)
|
762 |
+
self.resnets = nn.ModuleList(
|
763 |
+
[
|
764 |
+
ResnetBlock2D(in_features, in_features, conv_shortcut=False),
|
765 |
+
ResnetBlock2D(in_features, in_features, conv_shortcut=False),
|
766 |
+
]
|
767 |
+
)
|
768 |
+
|
769 |
+
def forward(self, hidden_states, temb=None, encoder_hidden_states=None):
|
770 |
+
hidden_states = self.resnets[0](hidden_states, temb)
|
771 |
+
for attn, resnet in zip(self.attentions, self.resnets[1:]):
|
772 |
+
hidden_states = attn(
|
773 |
+
hidden_states,
|
774 |
+
encoder_hidden_states=encoder_hidden_states,
|
775 |
+
)
|
776 |
+
hidden_states = resnet(hidden_states, temb)
|
777 |
+
|
778 |
+
return hidden_states
|
779 |
+
|
780 |
+
|
781 |
+
class UNet2DConditionModel(nn.Module):
|
782 |
+
def __init__(self):
|
783 |
+
super(UNet2DConditionModel, self).__init__()
|
784 |
+
|
785 |
+
# This is needed to imitate huggingface config behavior
|
786 |
+
# has nothing to do with the model itself
|
787 |
+
# remove this if you don't use diffuser's pipeline
|
788 |
+
self.config = namedtuple(
|
789 |
+
"config", "in_channels addition_time_embed_dim sample_size"
|
790 |
+
)
|
791 |
+
self.config.in_channels = 4
|
792 |
+
self.config.addition_time_embed_dim = 256
|
793 |
+
self.config.sample_size = 128
|
794 |
+
|
795 |
+
self.conv_in = nn.Conv2d(4, 320, kernel_size=3, stride=1, padding=1)
|
796 |
+
self.time_proj = Timesteps()
|
797 |
+
self.time_embedding = TimestepEmbedding(in_features=320, out_features=1280)
|
798 |
+
self.add_time_proj = Timesteps(256)
|
799 |
+
self.add_embedding = TimestepEmbedding(in_features=2816, out_features=1280)
|
800 |
+
self.down_blocks = nn.ModuleList(
|
801 |
+
[
|
802 |
+
DownBlock2D(in_channels=320, out_channels=320),
|
803 |
+
CrossAttnDownBlock2D(in_channels=320, out_channels=640, n_layers=2),
|
804 |
+
CrossAttnDownBlock2D(
|
805 |
+
in_channels=640,
|
806 |
+
out_channels=1280,
|
807 |
+
n_layers=10,
|
808 |
+
has_downsamplers=False,
|
809 |
+
),
|
810 |
+
]
|
811 |
+
)
|
812 |
+
self.up_blocks = nn.ModuleList(
|
813 |
+
[
|
814 |
+
CrossAttnUpBlock2D(
|
815 |
+
in_channels=640,
|
816 |
+
out_channels=1280,
|
817 |
+
prev_output_channel=1280,
|
818 |
+
n_layers=10,
|
819 |
+
),
|
820 |
+
CrossAttnUpBlock2D(
|
821 |
+
in_channels=320,
|
822 |
+
out_channels=640,
|
823 |
+
prev_output_channel=1280,
|
824 |
+
n_layers=2,
|
825 |
+
),
|
826 |
+
UpBlock2D(in_channels=320, out_channels=320, prev_output_channel=640),
|
827 |
+
]
|
828 |
+
)
|
829 |
+
self.mid_block = UNetMidBlock2DCrossAttn(1280)
|
830 |
+
self.conv_norm_out = nn.GroupNorm(32, 320, eps=1e-05, affine=True)
|
831 |
+
self.conv_act = nn.SiLU()
|
832 |
+
self.conv_out = nn.Conv2d(320, 4, kernel_size=3, stride=1, padding=1)
|
833 |
+
|
834 |
+
def forward(
|
835 |
+
self, sample, timesteps, encoder_hidden_states, added_cond_kwargs, **kwargs
|
836 |
+
):
|
837 |
+
# Implement the forward pass through the model
|
838 |
+
timesteps = timesteps.expand(sample.shape[0])
|
839 |
+
t_emb = self.time_proj(timesteps).to(dtype=sample.dtype)
|
840 |
+
|
841 |
+
emb = self.time_embedding(t_emb)
|
842 |
+
|
843 |
+
text_embeds = added_cond_kwargs.get("text_embeds")
|
844 |
+
time_ids = added_cond_kwargs.get("time_ids")
|
845 |
+
|
846 |
+
time_embeds = self.add_time_proj(time_ids.flatten())
|
847 |
+
time_embeds = time_embeds.reshape((text_embeds.shape[0], -1))
|
848 |
+
|
849 |
+
add_embeds = torch.concat([text_embeds, time_embeds], dim=-1)
|
850 |
+
add_embeds = add_embeds.to(emb.dtype)
|
851 |
+
aug_emb = self.add_embedding(add_embeds)
|
852 |
+
|
853 |
+
emb = emb + aug_emb
|
854 |
+
|
855 |
+
sample = self.conv_in(sample)
|
856 |
+
|
857 |
+
# 3. down
|
858 |
+
s0 = sample
|
859 |
+
sample, [s1, s2, s3] = self.down_blocks[0](
|
860 |
+
sample,
|
861 |
+
temb=emb,
|
862 |
+
)
|
863 |
+
|
864 |
+
sample, [s4, s5, s6] = self.down_blocks[1](
|
865 |
+
sample,
|
866 |
+
temb=emb,
|
867 |
+
encoder_hidden_states=encoder_hidden_states,
|
868 |
+
)
|
869 |
+
|
870 |
+
sample, [s7, s8] = self.down_blocks[2](
|
871 |
+
sample,
|
872 |
+
temb=emb,
|
873 |
+
encoder_hidden_states=encoder_hidden_states,
|
874 |
+
)
|
875 |
+
|
876 |
+
# 4. mid
|
877 |
+
sample = self.mid_block(
|
878 |
+
sample, emb, encoder_hidden_states=encoder_hidden_states
|
879 |
+
)
|
880 |
+
|
881 |
+
# 5. up
|
882 |
+
sample = self.up_blocks[0](
|
883 |
+
hidden_states=sample,
|
884 |
+
temb=emb,
|
885 |
+
res_hidden_states_tuple=[s6, s7, s8],
|
886 |
+
encoder_hidden_states=encoder_hidden_states,
|
887 |
+
)
|
888 |
+
|
889 |
+
sample = self.up_blocks[1](
|
890 |
+
hidden_states=sample,
|
891 |
+
temb=emb,
|
892 |
+
res_hidden_states_tuple=[s3, s4, s5],
|
893 |
+
encoder_hidden_states=encoder_hidden_states,
|
894 |
+
)
|
895 |
+
|
896 |
+
sample = self.up_blocks[2](
|
897 |
+
hidden_states=sample,
|
898 |
+
temb=emb,
|
899 |
+
res_hidden_states_tuple=[s0, s1, s2],
|
900 |
+
)
|
901 |
+
|
902 |
+
# 6. post-process
|
903 |
+
sample = self.conv_norm_out(sample)
|
904 |
+
sample = self.conv_act(sample)
|
905 |
+
sample = self.conv_out(sample)
|
906 |
+
|
907 |
+
return [sample]
|