Delete pixcell_controlnet.py
Browse files- pixcell_controlnet.py +0 -176
pixcell_controlnet.py
DELETED
@@ -1,176 +0,0 @@
|
|
1 |
-
from dataclasses import dataclass
|
2 |
-
from diffusers.configuration_utils import ConfigMixin
|
3 |
-
from diffusers.models.modeling_utils import ModelMixin
|
4 |
-
import torch
|
5 |
-
import torch.nn as nn
|
6 |
-
from typing import Any, Dict, Optional, Tuple
|
7 |
-
from pixcell_transformer_2d import PixCellTransformer2DModel
|
8 |
-
|
9 |
-
from diffusers.models.controlnet import zero_module
|
10 |
-
from diffusers.models.embeddings import PatchEmbed
|
11 |
-
from diffusers.utils import BaseOutput, is_torch_version
|
12 |
-
|
13 |
-
@dataclass
|
14 |
-
class PixCellControlNetOutput(BaseOutput):
|
15 |
-
controlnet_block_samples: Tuple[torch.Tensor]
|
16 |
-
|
17 |
-
class PixCellControlNet(ModelMixin, ConfigMixin):
|
18 |
-
def __init__(
|
19 |
-
self,
|
20 |
-
base_transformer: PixCellTransformer2DModel,
|
21 |
-
n_blocks: int = None,
|
22 |
-
):
|
23 |
-
super().__init__()
|
24 |
-
|
25 |
-
self.n_blocks = n_blocks
|
26 |
-
|
27 |
-
# Base transformer
|
28 |
-
self.transformer = base_transformer
|
29 |
-
|
30 |
-
# Input patch embedding is frozen
|
31 |
-
# self.transformer.pos_embed.requires_grad = False
|
32 |
-
|
33 |
-
# Condition patch embedding
|
34 |
-
interpolation_scale = (
|
35 |
-
self.transformer.config.interpolation_scale
|
36 |
-
if self.transformer.config.interpolation_scale is not None
|
37 |
-
else max(self.transformer.config.sample_size // 64, 1)
|
38 |
-
)
|
39 |
-
self.cond_pos_embed = zero_module(PatchEmbed(
|
40 |
-
height=self.transformer.config.sample_size,
|
41 |
-
width=self.transformer.config.sample_size,
|
42 |
-
patch_size=self.transformer.config.patch_size,
|
43 |
-
in_channels=self.transformer.config.in_channels,
|
44 |
-
embed_dim=self.transformer.inner_dim,
|
45 |
-
interpolation_scale=interpolation_scale,
|
46 |
-
))
|
47 |
-
|
48 |
-
|
49 |
-
# Do not use all transformer blocks for controlnet
|
50 |
-
if self.n_blocks is not None:
|
51 |
-
self.transformer.transformer_blocks = self.transformer.transformer_blocks[:self.n_blocks]
|
52 |
-
|
53 |
-
# ControlNet layers
|
54 |
-
self.controlnet_blocks = nn.ModuleList([])
|
55 |
-
for i in range(len(self.transformer.transformer_blocks)):
|
56 |
-
controlnet_block = nn.Linear(self.transformer.inner_dim, self.transformer.inner_dim)
|
57 |
-
controlnet_block = zero_module(controlnet_block)
|
58 |
-
self.controlnet_blocks.append(controlnet_block)
|
59 |
-
|
60 |
-
if self.n_blocks is not None:
|
61 |
-
if i+1 == self.n_blocks:
|
62 |
-
break
|
63 |
-
|
64 |
-
def forward(
|
65 |
-
self,
|
66 |
-
hidden_states: torch.Tensor,
|
67 |
-
conditioning: torch.Tensor,
|
68 |
-
encoder_hidden_states: Optional[torch.Tensor] = None,
|
69 |
-
timestep: Optional[torch.LongTensor] = None,
|
70 |
-
conditioning_scale: float = 1.0,
|
71 |
-
added_cond_kwargs: Dict[str, torch.Tensor] = None,
|
72 |
-
cross_attention_kwargs: Dict[str, Any] = None,
|
73 |
-
attention_mask: Optional[torch.Tensor] = None,
|
74 |
-
encoder_attention_mask: Optional[torch.Tensor] = None,
|
75 |
-
return_dict: bool = True,
|
76 |
-
):
|
77 |
-
if self.transformer.use_additional_conditions and added_cond_kwargs is None:
|
78 |
-
raise ValueError("`added_cond_kwargs` cannot be None when using additional conditions for `adaln_single`.")
|
79 |
-
|
80 |
-
# ensure attention_mask is a bias, and give it a singleton query_tokens dimension.
|
81 |
-
# we may have done this conversion already, e.g. if we came here via UNet2DConditionModel#forward.
|
82 |
-
# we can tell by counting dims; if ndim == 2: it's a mask rather than a bias.
|
83 |
-
# expects mask of shape:
|
84 |
-
# [batch, key_tokens]
|
85 |
-
# adds singleton query_tokens dimension:
|
86 |
-
# [batch, 1, key_tokens]
|
87 |
-
# this helps to broadcast it as a bias over attention scores, which will be in one of the following shapes:
|
88 |
-
# [batch, heads, query_tokens, key_tokens] (e.g. torch sdp attn)
|
89 |
-
# [batch * heads, query_tokens, key_tokens] (e.g. xformers or classic attn)
|
90 |
-
if attention_mask is not None and attention_mask.ndim == 2:
|
91 |
-
# assume that mask is expressed as:
|
92 |
-
# (1 = keep, 0 = discard)
|
93 |
-
# convert mask into a bias that can be added to attention scores:
|
94 |
-
# (keep = +0, discard = -10000.0)
|
95 |
-
attention_mask = (1 - attention_mask.to(hidden_states.dtype)) * -10000.0
|
96 |
-
attention_mask = attention_mask.unsqueeze(1)
|
97 |
-
|
98 |
-
# convert encoder_attention_mask to a bias the same way we do for attention_mask
|
99 |
-
if encoder_attention_mask is not None and encoder_attention_mask.ndim == 2:
|
100 |
-
encoder_attention_mask = (1 - encoder_attention_mask.to(hidden_states.dtype)) * -10000.0
|
101 |
-
encoder_attention_mask = encoder_attention_mask.unsqueeze(1)
|
102 |
-
|
103 |
-
# 1. Input
|
104 |
-
batch_size = hidden_states.shape[0]
|
105 |
-
height, width = (
|
106 |
-
hidden_states.shape[-2] // self.transformer.config.patch_size,
|
107 |
-
hidden_states.shape[-1] // self.transformer.config.patch_size,
|
108 |
-
)
|
109 |
-
hidden_states = self.transformer.pos_embed(hidden_states)
|
110 |
-
|
111 |
-
# Conditioning
|
112 |
-
hidden_states = hidden_states + self.cond_pos_embed(conditioning)
|
113 |
-
|
114 |
-
timestep, embedded_timestep = self.transformer.adaln_single(
|
115 |
-
timestep, added_cond_kwargs, batch_size=batch_size, hidden_dtype=hidden_states.dtype
|
116 |
-
)
|
117 |
-
|
118 |
-
if self.transformer.caption_projection is not None:
|
119 |
-
# Add positional embeddings to conditions if >1 UNI are given
|
120 |
-
if self.transformer.y_pos_embed is not None:
|
121 |
-
encoder_hidden_states = self.transformer.y_pos_embed(encoder_hidden_states)
|
122 |
-
encoder_hidden_states = self.transformer.caption_projection(encoder_hidden_states)
|
123 |
-
encoder_hidden_states = encoder_hidden_states.view(batch_size, -1, hidden_states.shape[-1])
|
124 |
-
|
125 |
-
# 2. Blocks
|
126 |
-
block_outputs = ()
|
127 |
-
|
128 |
-
for block in self.transformer.transformer_blocks:
|
129 |
-
if torch.is_grad_enabled() and self.transformer.gradient_checkpointing:
|
130 |
-
|
131 |
-
def create_custom_forward(module, return_dict=None):
|
132 |
-
def custom_forward(*inputs):
|
133 |
-
if return_dict is not None:
|
134 |
-
return module(*inputs, return_dict=return_dict)
|
135 |
-
else:
|
136 |
-
return module(*inputs)
|
137 |
-
|
138 |
-
return custom_forward
|
139 |
-
|
140 |
-
ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
|
141 |
-
hidden_states = torch.utils.checkpoint.checkpoint(
|
142 |
-
create_custom_forward(block),
|
143 |
-
hidden_states,
|
144 |
-
attention_mask,
|
145 |
-
encoder_hidden_states,
|
146 |
-
encoder_attention_mask,
|
147 |
-
timestep,
|
148 |
-
cross_attention_kwargs,
|
149 |
-
None,
|
150 |
-
**ckpt_kwargs,
|
151 |
-
)
|
152 |
-
else:
|
153 |
-
hidden_states = block(
|
154 |
-
hidden_states,
|
155 |
-
attention_mask=attention_mask,
|
156 |
-
encoder_hidden_states=encoder_hidden_states,
|
157 |
-
encoder_attention_mask=encoder_attention_mask,
|
158 |
-
timestep=timestep,
|
159 |
-
cross_attention_kwargs=cross_attention_kwargs,
|
160 |
-
class_labels=None,
|
161 |
-
)
|
162 |
-
|
163 |
-
block_outputs = block_outputs + (hidden_states,)
|
164 |
-
|
165 |
-
# 3. controlnet blocks
|
166 |
-
controlnet_outputs = ()
|
167 |
-
for t_output, controlnet_block in zip(block_outputs, self.controlnet_blocks):
|
168 |
-
b_output = controlnet_block(t_output)
|
169 |
-
controlnet_outputs = controlnet_outputs + (b_output,)
|
170 |
-
|
171 |
-
controlnet_outputs = [sample * conditioning_scale for sample in controlnet_outputs]
|
172 |
-
|
173 |
-
if not return_dict:
|
174 |
-
return (controlnet_outputs,)
|
175 |
-
|
176 |
-
return PixCellControlNetOutput(controlnet_block_samples=controlnet_outputs)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|