Create my_controlnet.py
Browse files- controlnet/my_controlnet.py +238 -0
controlnet/my_controlnet.py
ADDED
@@ -0,0 +1,238 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from typing import Any, Dict, Optional, Tuple, Union
|
2 |
+
|
3 |
+
import torch
|
4 |
+
import torch.nn as nn
|
5 |
+
from diffusers import ControlNetModel, ModelMixin
|
6 |
+
from diffusers.configuration_utils import register_to_config
|
7 |
+
from diffusers.models.controlnet import ControlNetOutput
|
8 |
+
|
9 |
+
|
10 |
+
def zero_module(module):
|
11 |
+
for p in module.parameters():
|
12 |
+
nn.init.zeros_(p)
|
13 |
+
return module
|
14 |
+
|
15 |
+
|
16 |
+
class MyControlNetModel(ControlNetModel, ModelMixin):
|
17 |
+
@register_to_config
|
18 |
+
def __init__(
|
19 |
+
self,
|
20 |
+
in_channels: int = 4,
|
21 |
+
conditioning_channels: int = 3,
|
22 |
+
flip_sin_to_cos: bool = True,
|
23 |
+
freq_shift: int = 0,
|
24 |
+
down_block_types: Tuple[str, ...] = (
|
25 |
+
"CrossAttnDownBlock2D",
|
26 |
+
"CrossAttnDownBlock2D",
|
27 |
+
"CrossAttnDownBlock2D",
|
28 |
+
"DownBlock2D",
|
29 |
+
),
|
30 |
+
mid_block_type: Optional[str] = "UNetMidBlock2DCrossAttn",
|
31 |
+
only_cross_attention: Union[bool, Tuple[bool]] = False,
|
32 |
+
block_out_channels: Tuple[int, ...] = (320, 640, 1280, 1280),
|
33 |
+
layers_per_block: int = 2,
|
34 |
+
downsample_padding: int = 1,
|
35 |
+
mid_block_scale_factor: float = 1,
|
36 |
+
act_fn: str = "silu",
|
37 |
+
norm_num_groups: Optional[int] = 32,
|
38 |
+
norm_eps: float = 1e-5,
|
39 |
+
cross_attention_dim: int = 1280,
|
40 |
+
transformer_layers_per_block: Union[int, Tuple[int, ...]] = 1,
|
41 |
+
encoder_hid_dim: Optional[int] = None,
|
42 |
+
encoder_hid_dim_type: Optional[str] = None,
|
43 |
+
attention_head_dim: Union[int, Tuple[int, ...]] = 8,
|
44 |
+
num_attention_heads: Optional[Union[int, Tuple[int, ...]]] = None,
|
45 |
+
use_linear_projection: bool = False,
|
46 |
+
class_embed_type: Optional[str] = None,
|
47 |
+
addition_embed_type: Optional[str] = None,
|
48 |
+
addition_time_embed_dim: Optional[int] = None,
|
49 |
+
num_class_embeds: Optional[int] = None,
|
50 |
+
upcast_attention: bool = False,
|
51 |
+
resnet_time_scale_shift: str = "default",
|
52 |
+
projection_class_embeddings_input_dim: Optional[int] = None,
|
53 |
+
controlnet_conditioning_channel_order: str = "rgb",
|
54 |
+
conditioning_embedding_out_channels: Optional[Tuple[int, ...]] = (
|
55 |
+
16, 32, 96, 256),
|
56 |
+
global_pool_conditions: bool = False,
|
57 |
+
addition_embed_type_num_heads: int = 64):
|
58 |
+
super().__init__(in_channels, conditioning_channels, flip_sin_to_cos, freq_shift, down_block_types, mid_block_type, only_cross_attention, block_out_channels, layers_per_block, downsample_padding, mid_block_scale_factor, act_fn, norm_num_groups, norm_eps, cross_attention_dim, transformer_layers_per_block, encoder_hid_dim, encoder_hid_dim_type,
|
59 |
+
attention_head_dim, num_attention_heads, use_linear_projection, class_embed_type, addition_embed_type, addition_time_embed_dim, num_class_embeds, upcast_attention, resnet_time_scale_shift, projection_class_embeddings_input_dim, controlnet_conditioning_channel_order, conditioning_embedding_out_channels, global_pool_conditions, addition_embed_type_num_heads)
|
60 |
+
self.controlnet_cond_embedding = nn.Identity()
|
61 |
+
conv_in_kernel = 3
|
62 |
+
conv_in_padding = (conv_in_kernel - 1) // 2
|
63 |
+
self.conv_in2 = nn.Conv2d(
|
64 |
+
in_channels, block_out_channels[0], kernel_size=conv_in_kernel, padding=conv_in_padding
|
65 |
+
)
|
66 |
+
zero_module(self.conv_in2)
|
67 |
+
|
68 |
+
def forward(
|
69 |
+
self,
|
70 |
+
sample: torch.Tensor,
|
71 |
+
timestep: Union[torch.Tensor, float, int],
|
72 |
+
encoder_hidden_states: torch.Tensor,
|
73 |
+
controlnet_cond: torch.Tensor,
|
74 |
+
conditioning_scale: float = 1.0,
|
75 |
+
class_labels: Optional[torch.Tensor] = None,
|
76 |
+
timestep_cond: Optional[torch.Tensor] = None,
|
77 |
+
attention_mask: Optional[torch.Tensor] = None,
|
78 |
+
added_cond_kwargs: Optional[Dict[str, torch.Tensor]] = None,
|
79 |
+
cross_attention_kwargs: Optional[Dict[str, Any]] = None,
|
80 |
+
guess_mode: bool = False,
|
81 |
+
return_dict: bool = True,
|
82 |
+
) -> Union[ControlNetOutput, Tuple[Tuple[torch.Tensor, ...], torch.Tensor]]:
|
83 |
+
# check channel order
|
84 |
+
channel_order = self.config.controlnet_conditioning_channel_order
|
85 |
+
|
86 |
+
if channel_order == "rgb":
|
87 |
+
# in rgb order by default
|
88 |
+
...
|
89 |
+
elif channel_order == "bgr":
|
90 |
+
controlnet_cond = torch.flip(controlnet_cond, dims=[1])
|
91 |
+
else:
|
92 |
+
raise ValueError(
|
93 |
+
f"unknown `controlnet_conditioning_channel_order`: {channel_order}")
|
94 |
+
|
95 |
+
# prepare attention_mask
|
96 |
+
if attention_mask is not None:
|
97 |
+
attention_mask = (1 - attention_mask.to(sample.dtype)) * -10000.0
|
98 |
+
attention_mask = attention_mask.unsqueeze(1)
|
99 |
+
|
100 |
+
# 1. time
|
101 |
+
timesteps = timestep
|
102 |
+
if not torch.is_tensor(timesteps):
|
103 |
+
# TODO: this requires sync between CPU and GPU. So try to pass timesteps as tensors if you can
|
104 |
+
# This would be a good case for the `match` statement (Python 3.10+)
|
105 |
+
is_mps = sample.device.type == "mps"
|
106 |
+
if isinstance(timestep, float):
|
107 |
+
dtype = torch.float32 if is_mps else torch.float64
|
108 |
+
else:
|
109 |
+
dtype = torch.int32 if is_mps else torch.int64
|
110 |
+
timesteps = torch.tensor(
|
111 |
+
[timesteps], dtype=dtype, device=sample.device)
|
112 |
+
elif len(timesteps.shape) == 0:
|
113 |
+
timesteps = timesteps[None].to(sample.device)
|
114 |
+
|
115 |
+
# broadcast to batch dimension in a way that's compatible with ONNX/Core ML
|
116 |
+
timesteps = timesteps.expand(sample.shape[0])
|
117 |
+
|
118 |
+
t_emb = self.time_proj(timesteps)
|
119 |
+
|
120 |
+
# timesteps does not contain any weights and will always return f32 tensors
|
121 |
+
# but time_embedding might actually be running in fp16. so we need to cast here.
|
122 |
+
# there might be better ways to encapsulate this.
|
123 |
+
t_emb = t_emb.to(dtype=sample.dtype)
|
124 |
+
|
125 |
+
emb = self.time_embedding(t_emb, timestep_cond)
|
126 |
+
aug_emb = None
|
127 |
+
|
128 |
+
if self.class_embedding is not None:
|
129 |
+
if class_labels is None:
|
130 |
+
raise ValueError(
|
131 |
+
"class_labels should be provided when num_class_embeds > 0")
|
132 |
+
|
133 |
+
if self.config.class_embed_type == "timestep":
|
134 |
+
class_labels = self.time_proj(class_labels)
|
135 |
+
|
136 |
+
class_emb = self.class_embedding(class_labels).to(dtype=self.dtype)
|
137 |
+
emb = emb + class_emb
|
138 |
+
|
139 |
+
if self.config.addition_embed_type is not None:
|
140 |
+
if self.config.addition_embed_type == "text":
|
141 |
+
aug_emb = self.add_embedding(encoder_hidden_states)
|
142 |
+
|
143 |
+
elif self.config.addition_embed_type == "text_time":
|
144 |
+
if "text_embeds" not in added_cond_kwargs:
|
145 |
+
raise ValueError(
|
146 |
+
f"{self.__class__} has the config param `addition_embed_type` set to 'text_time' which requires the keyword argument `text_embeds` to be passed in `added_cond_kwargs`"
|
147 |
+
)
|
148 |
+
text_embeds = added_cond_kwargs.get("text_embeds")
|
149 |
+
if "time_ids" not in added_cond_kwargs:
|
150 |
+
raise ValueError(
|
151 |
+
f"{self.__class__} has the config param `addition_embed_type` set to 'text_time' which requires the keyword argument `time_ids` to be passed in `added_cond_kwargs`"
|
152 |
+
)
|
153 |
+
time_ids = added_cond_kwargs.get("time_ids")
|
154 |
+
time_embeds = self.add_time_proj(time_ids.flatten())
|
155 |
+
time_embeds = time_embeds.reshape((text_embeds.shape[0], -1))
|
156 |
+
|
157 |
+
add_embeds = torch.concat([text_embeds, time_embeds], dim=-1)
|
158 |
+
add_embeds = add_embeds.to(emb.dtype)
|
159 |
+
aug_emb = self.add_embedding(add_embeds)
|
160 |
+
|
161 |
+
emb = emb + aug_emb if aug_emb is not None else emb
|
162 |
+
|
163 |
+
# 2. pre-process
|
164 |
+
sample = self.conv_in(sample)
|
165 |
+
controlnet_cond = self.conv_in2(controlnet_cond)
|
166 |
+
|
167 |
+
sample = sample + controlnet_cond
|
168 |
+
|
169 |
+
# 3. down
|
170 |
+
down_block_res_samples = (sample,)
|
171 |
+
for downsample_block in self.down_blocks:
|
172 |
+
if hasattr(downsample_block, "has_cross_attention") and downsample_block.has_cross_attention:
|
173 |
+
sample, res_samples = downsample_block(
|
174 |
+
hidden_states=sample,
|
175 |
+
temb=emb,
|
176 |
+
encoder_hidden_states=encoder_hidden_states,
|
177 |
+
attention_mask=attention_mask,
|
178 |
+
cross_attention_kwargs=cross_attention_kwargs,
|
179 |
+
)
|
180 |
+
else:
|
181 |
+
sample, res_samples = downsample_block(
|
182 |
+
hidden_states=sample, temb=emb)
|
183 |
+
|
184 |
+
down_block_res_samples += res_samples
|
185 |
+
|
186 |
+
# 4. mid
|
187 |
+
if self.mid_block is not None:
|
188 |
+
if hasattr(self.mid_block, "has_cross_attention") and self.mid_block.has_cross_attention:
|
189 |
+
sample = self.mid_block(
|
190 |
+
sample,
|
191 |
+
emb,
|
192 |
+
encoder_hidden_states=encoder_hidden_states,
|
193 |
+
attention_mask=attention_mask,
|
194 |
+
cross_attention_kwargs=cross_attention_kwargs,
|
195 |
+
)
|
196 |
+
else:
|
197 |
+
sample = self.mid_block(sample, emb)
|
198 |
+
|
199 |
+
# 5. Control net blocks
|
200 |
+
controlnet_down_block_res_samples = ()
|
201 |
+
|
202 |
+
for down_block_res_sample, controlnet_block in zip(down_block_res_samples, self.controlnet_down_blocks):
|
203 |
+
down_block_res_sample = controlnet_block(down_block_res_sample)
|
204 |
+
controlnet_down_block_res_samples = controlnet_down_block_res_samples + \
|
205 |
+
(down_block_res_sample,)
|
206 |
+
|
207 |
+
down_block_res_samples = controlnet_down_block_res_samples
|
208 |
+
|
209 |
+
mid_block_res_sample = self.controlnet_mid_block(sample)
|
210 |
+
|
211 |
+
# 6. scaling
|
212 |
+
if guess_mode and not self.config.global_pool_conditions:
|
213 |
+
# 0.1 to 1.0
|
214 |
+
scales = torch.logspace(-1, 0, len(down_block_res_samples) +
|
215 |
+
1, device=sample.device)
|
216 |
+
scales = scales * conditioning_scale
|
217 |
+
down_block_res_samples = [
|
218 |
+
sample * scale for sample, scale in zip(down_block_res_samples, scales)]
|
219 |
+
mid_block_res_sample = mid_block_res_sample * \
|
220 |
+
scales[-1] # last one
|
221 |
+
else:
|
222 |
+
down_block_res_samples = [
|
223 |
+
sample * conditioning_scale for sample in down_block_res_samples]
|
224 |
+
mid_block_res_sample = mid_block_res_sample * conditioning_scale
|
225 |
+
|
226 |
+
if self.config.global_pool_conditions:
|
227 |
+
down_block_res_samples = [
|
228 |
+
torch.mean(sample, dim=(2, 3), keepdim=True) for sample in down_block_res_samples
|
229 |
+
]
|
230 |
+
mid_block_res_sample = torch.mean(
|
231 |
+
mid_block_res_sample, dim=(2, 3), keepdim=True)
|
232 |
+
|
233 |
+
if not return_dict:
|
234 |
+
return (down_block_res_samples, mid_block_res_sample)
|
235 |
+
|
236 |
+
return ControlNetOutput(
|
237 |
+
down_block_res_samples=down_block_res_samples, mid_block_res_sample=mid_block_res_sample
|
238 |
+
)
|