Spaces:
Sleeping
Sleeping
depthanyvideo
commited on
Commit
•
a1739b9
1
Parent(s):
66993df
init
Browse files- models/__init__.py +7 -0
- models/attention.py +741 -0
- models/attention_processor.py +0 -0
- models/embeddings.py +1539 -0
- models/resnet.py +508 -0
- models/transformers/transformer_temporal_rope.py +230 -0
- models/unets/unet_3d_rope_blocks.py +682 -0
- models/unets/unet_spatio_temporal_rope_condition.py +546 -0
- pipelines/__init__.py +5 -0
- pipelines/dav_pipeline.py +246 -0
- requirements.txt +12 -0
- utils/img_utils.py +112 -0
models/__init__.py
ADDED
@@ -0,0 +1,7 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from .unets.unet_spatio_temporal_rope_condition import (
|
2 |
+
UNetSpatioTemporalRopeConditionModel,
|
3 |
+
)
|
4 |
+
|
5 |
+
__all__ = {
|
6 |
+
"UNetSpatioTemporalRopeConditionModel": UNetSpatioTemporalRopeConditionModel,
|
7 |
+
}
|
models/attention.py
ADDED
@@ -0,0 +1,741 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from typing import Any, Dict, Optional
|
2 |
+
|
3 |
+
import torch
|
4 |
+
import torch.nn.functional as F
|
5 |
+
from torch import nn
|
6 |
+
|
7 |
+
from diffusers.utils import deprecate, logging
|
8 |
+
from diffusers.utils.torch_utils import maybe_allow_in_graph
|
9 |
+
from diffusers.models.activations import GEGLU, GELU, ApproximateGELU
|
10 |
+
from .attention_processor import (
|
11 |
+
Attention,
|
12 |
+
AttnProcessor2_0,
|
13 |
+
JointAttnProcessor2_0,
|
14 |
+
JointAttnROPEProcessor2_0,
|
15 |
+
AttnRopeProcessor2_0,
|
16 |
+
)
|
17 |
+
from .embeddings import SinusoidalPositionalEmbedding
|
18 |
+
from diffusers.models.normalization import (
|
19 |
+
AdaLayerNorm,
|
20 |
+
AdaLayerNormContinuous,
|
21 |
+
AdaLayerNormZero,
|
22 |
+
RMSNorm,
|
23 |
+
)
|
24 |
+
|
25 |
+
|
26 |
+
logger = logging.get_logger(__name__)
|
27 |
+
|
28 |
+
|
29 |
+
def _chunked_feed_forward(
|
30 |
+
ff: nn.Module, hidden_states: torch.Tensor, chunk_dim: int, chunk_size: int
|
31 |
+
):
|
32 |
+
# "feed_forward_chunk_size" can be used to save memory
|
33 |
+
if hidden_states.shape[chunk_dim] % chunk_size != 0:
|
34 |
+
raise ValueError(
|
35 |
+
f"`hidden_states` dimension to be chunked: {hidden_states.shape[chunk_dim]} has to be divisible by chunk size: {chunk_size}. Make sure to set an appropriate `chunk_size` when calling `unet.enable_forward_chunking`."
|
36 |
+
)
|
37 |
+
|
38 |
+
num_chunks = hidden_states.shape[chunk_dim] // chunk_size
|
39 |
+
ff_output = torch.cat(
|
40 |
+
[ff(hid_slice) for hid_slice in hidden_states.chunk(num_chunks, dim=chunk_dim)],
|
41 |
+
dim=chunk_dim,
|
42 |
+
)
|
43 |
+
return ff_output
|
44 |
+
|
45 |
+
|
46 |
+
@maybe_allow_in_graph
|
47 |
+
class GatedSelfAttentionDense(nn.Module):
|
48 |
+
r"""
|
49 |
+
A gated self-attention dense layer that combines visual features and object features.
|
50 |
+
|
51 |
+
Parameters:
|
52 |
+
query_dim (`int`): The number of channels in the query.
|
53 |
+
context_dim (`int`): The number of channels in the context.
|
54 |
+
n_heads (`int`): The number of heads to use for attention.
|
55 |
+
d_head (`int`): The number of channels in each head.
|
56 |
+
"""
|
57 |
+
|
58 |
+
def __init__(self, query_dim: int, context_dim: int, n_heads: int, d_head: int):
|
59 |
+
super().__init__()
|
60 |
+
|
61 |
+
# we need a linear projection since we need cat visual feature and obj feature
|
62 |
+
self.linear = nn.Linear(context_dim, query_dim)
|
63 |
+
|
64 |
+
self.attn = Attention(query_dim=query_dim, heads=n_heads, dim_head=d_head)
|
65 |
+
self.ff = FeedForward(query_dim, activation_fn="geglu")
|
66 |
+
|
67 |
+
self.norm1 = nn.LayerNorm(query_dim)
|
68 |
+
self.norm2 = nn.LayerNorm(query_dim)
|
69 |
+
|
70 |
+
self.register_parameter("alpha_attn", nn.Parameter(torch.tensor(0.0)))
|
71 |
+
self.register_parameter("alpha_dense", nn.Parameter(torch.tensor(0.0)))
|
72 |
+
|
73 |
+
self.enabled = True
|
74 |
+
|
75 |
+
def forward(self, x: torch.Tensor, objs: torch.Tensor) -> torch.Tensor:
|
76 |
+
if not self.enabled:
|
77 |
+
return x
|
78 |
+
|
79 |
+
n_visual = x.shape[1]
|
80 |
+
objs = self.linear(objs)
|
81 |
+
|
82 |
+
x = (
|
83 |
+
x
|
84 |
+
+ self.alpha_attn.tanh()
|
85 |
+
* self.attn(self.norm1(torch.cat([x, objs], dim=1)))[:, :n_visual, :]
|
86 |
+
)
|
87 |
+
x = x + self.alpha_dense.tanh() * self.ff(self.norm2(x))
|
88 |
+
|
89 |
+
return x
|
90 |
+
|
91 |
+
|
92 |
+
@maybe_allow_in_graph
|
93 |
+
class TransformerBlock(nn.Module):
|
94 |
+
r"""
|
95 |
+
A Transformer block following the MMDiT architecture, introduced in Stable Diffusion 3.
|
96 |
+
|
97 |
+
Reference: https://arxiv.org/abs/2403.03206
|
98 |
+
|
99 |
+
Parameters:
|
100 |
+
dim (`int`): The number of channels in the input and output.
|
101 |
+
num_attention_heads (`int`): The number of heads to use for multi-head attention.
|
102 |
+
attention_head_dim (`int`): The number of channels in each head.
|
103 |
+
context_pre_only (`bool`): Boolean to determine if we should add some blocks associated with the
|
104 |
+
processing of `context` conditions.
|
105 |
+
"""
|
106 |
+
|
107 |
+
def __init__(
|
108 |
+
self, dim, num_attention_heads, attention_head_dim, context_pre_only=False
|
109 |
+
):
|
110 |
+
super().__init__()
|
111 |
+
|
112 |
+
self.norm1 = AdaLayerNormZero(dim)
|
113 |
+
|
114 |
+
if hasattr(F, "scaled_dot_product_attention"):
|
115 |
+
processor = AttnProcessor2_0()
|
116 |
+
else:
|
117 |
+
raise ValueError(
|
118 |
+
"The current PyTorch version does not support the `scaled_dot_product_attention` function."
|
119 |
+
)
|
120 |
+
self.attn = Attention(
|
121 |
+
query_dim=dim,
|
122 |
+
cross_attention_dim=None,
|
123 |
+
added_kv_proj_dim=None,
|
124 |
+
dim_head=attention_head_dim // num_attention_heads,
|
125 |
+
heads=num_attention_heads,
|
126 |
+
out_dim=attention_head_dim,
|
127 |
+
context_pre_only=context_pre_only,
|
128 |
+
bias=True,
|
129 |
+
processor=processor,
|
130 |
+
)
|
131 |
+
|
132 |
+
self.norm2 = nn.LayerNorm(dim, elementwise_affine=False, eps=1e-6)
|
133 |
+
self.ff = FeedForward(dim=dim, dim_out=dim, activation_fn="gelu-approximate")
|
134 |
+
|
135 |
+
# let chunk size default to None
|
136 |
+
self._chunk_size = None
|
137 |
+
self._chunk_dim = 0
|
138 |
+
|
139 |
+
# Copied from diffusers.models.attention.BasicTransformerBlock.set_chunk_feed_forward
|
140 |
+
def set_chunk_feed_forward(self, chunk_size: Optional[int], dim: int = 0):
|
141 |
+
# Sets chunk feed-forward
|
142 |
+
self._chunk_size = chunk_size
|
143 |
+
self._chunk_dim = dim
|
144 |
+
|
145 |
+
def forward(
|
146 |
+
self,
|
147 |
+
hidden_states: torch.FloatTensor,
|
148 |
+
temb: torch.FloatTensor,
|
149 |
+
):
|
150 |
+
norm_hidden_states, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.norm1(
|
151 |
+
hidden_states, emb=temb
|
152 |
+
)
|
153 |
+
|
154 |
+
# Attention.
|
155 |
+
attn_output = self.attn(hidden_states=norm_hidden_states)
|
156 |
+
|
157 |
+
# Process attention outputs for the `hidden_states`.
|
158 |
+
attn_output = gate_msa.unsqueeze(1) * attn_output
|
159 |
+
hidden_states = hidden_states + attn_output
|
160 |
+
|
161 |
+
norm_hidden_states = self.norm2(hidden_states)
|
162 |
+
norm_hidden_states = (
|
163 |
+
norm_hidden_states * (1 + scale_mlp[:, None]) + shift_mlp[:, None]
|
164 |
+
)
|
165 |
+
if self._chunk_size is not None:
|
166 |
+
# "feed_forward_chunk_size" can be used to save memory
|
167 |
+
ff_output = _chunked_feed_forward(
|
168 |
+
self.ff, norm_hidden_states, self._chunk_dim, self._chunk_size
|
169 |
+
)
|
170 |
+
else:
|
171 |
+
ff_output = self.ff(norm_hidden_states)
|
172 |
+
ff_output = gate_mlp.unsqueeze(1) * ff_output
|
173 |
+
|
174 |
+
hidden_states = hidden_states + ff_output
|
175 |
+
|
176 |
+
return hidden_states
|
177 |
+
|
178 |
+
|
179 |
+
@maybe_allow_in_graph
|
180 |
+
class BasicTransformerBlock(nn.Module):
|
181 |
+
r"""
|
182 |
+
A basic Transformer block.
|
183 |
+
|
184 |
+
Parameters:
|
185 |
+
dim (`int`): The number of channels in the input and output.
|
186 |
+
num_attention_heads (`int`): The number of heads to use for multi-head attention.
|
187 |
+
attention_head_dim (`int`): The number of channels in each head.
|
188 |
+
dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
|
189 |
+
cross_attention_dim (`int`, *optional*): The size of the encoder_hidden_states vector for cross attention.
|
190 |
+
activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to be used in feed-forward.
|
191 |
+
num_embeds_ada_norm (:
|
192 |
+
obj: `int`, *optional*): The number of diffusion steps used during training. See `Transformer2DModel`.
|
193 |
+
attention_bias (:
|
194 |
+
obj: `bool`, *optional*, defaults to `False`): Configure if the attentions should contain a bias parameter.
|
195 |
+
only_cross_attention (`bool`, *optional*):
|
196 |
+
Whether to use only cross-attention layers. In this case two cross attention layers are used.
|
197 |
+
double_self_attention (`bool`, *optional*):
|
198 |
+
Whether to use two self-attention layers. In this case no cross attention layers are used.
|
199 |
+
upcast_attention (`bool`, *optional*):
|
200 |
+
Whether to upcast the attention computation to float32. This is useful for mixed precision training.
|
201 |
+
norm_elementwise_affine (`bool`, *optional*, defaults to `True`):
|
202 |
+
Whether to use learnable elementwise affine parameters for normalization.
|
203 |
+
norm_type (`str`, *optional*, defaults to `"layer_norm"`):
|
204 |
+
The normalization layer to use. Can be `"layer_norm"`, `"ada_norm"` or `"ada_norm_zero"`.
|
205 |
+
final_dropout (`bool` *optional*, defaults to False):
|
206 |
+
Whether to apply a final dropout after the last feed-forward layer.
|
207 |
+
attention_type (`str`, *optional*, defaults to `"default"`):
|
208 |
+
The type of attention to use. Can be `"default"` or `"gated"` or `"gated-text-image"`.
|
209 |
+
positional_embeddings (`str`, *optional*, defaults to `None`):
|
210 |
+
The type of positional embeddings to apply to.
|
211 |
+
num_positional_embeddings (`int`, *optional*, defaults to `None`):
|
212 |
+
The maximum number of positional embeddings to apply.
|
213 |
+
"""
|
214 |
+
|
215 |
+
def __init__(
|
216 |
+
self,
|
217 |
+
dim: int,
|
218 |
+
num_attention_heads: int,
|
219 |
+
attention_head_dim: int,
|
220 |
+
dropout=0.0,
|
221 |
+
cross_attention_dim: Optional[int] = None,
|
222 |
+
activation_fn: str = "geglu",
|
223 |
+
num_embeds_ada_norm: Optional[int] = None,
|
224 |
+
attention_bias: bool = False,
|
225 |
+
only_cross_attention: bool = False,
|
226 |
+
double_self_attention: bool = False,
|
227 |
+
upcast_attention: bool = False,
|
228 |
+
norm_elementwise_affine: bool = True,
|
229 |
+
norm_type: str = "layer_norm", # 'layer_norm', 'ada_norm', 'ada_norm_zero', 'ada_norm_single', 'ada_norm_continuous', 'layer_norm_i2vgen'
|
230 |
+
norm_eps: float = 1e-5,
|
231 |
+
final_dropout: bool = False,
|
232 |
+
attention_type: str = "default",
|
233 |
+
positional_embeddings: Optional[str] = None,
|
234 |
+
num_positional_embeddings: Optional[int] = None,
|
235 |
+
ada_norm_continous_conditioning_embedding_dim: Optional[int] = None,
|
236 |
+
ada_norm_bias: Optional[int] = None,
|
237 |
+
ff_inner_dim: Optional[int] = None,
|
238 |
+
ff_bias: bool = True,
|
239 |
+
attention_out_bias: bool = True,
|
240 |
+
):
|
241 |
+
super().__init__()
|
242 |
+
self.only_cross_attention = only_cross_attention
|
243 |
+
|
244 |
+
# We keep these boolean flags for backward-compatibility.
|
245 |
+
self.use_ada_layer_norm_zero = (
|
246 |
+
num_embeds_ada_norm is not None
|
247 |
+
) and norm_type == "ada_norm_zero"
|
248 |
+
self.use_ada_layer_norm = (
|
249 |
+
num_embeds_ada_norm is not None
|
250 |
+
) and norm_type == "ada_norm"
|
251 |
+
self.use_ada_layer_norm_single = norm_type == "ada_norm_single"
|
252 |
+
self.use_layer_norm = norm_type == "layer_norm"
|
253 |
+
self.use_ada_layer_norm_continuous = norm_type == "ada_norm_continuous"
|
254 |
+
|
255 |
+
if norm_type in ("ada_norm", "ada_norm_zero") and num_embeds_ada_norm is None:
|
256 |
+
raise ValueError(
|
257 |
+
f"`norm_type` is set to {norm_type}, but `num_embeds_ada_norm` is not defined. Please make sure to"
|
258 |
+
f" define `num_embeds_ada_norm` if setting `norm_type` to {norm_type}."
|
259 |
+
)
|
260 |
+
|
261 |
+
self.norm_type = norm_type
|
262 |
+
self.num_embeds_ada_norm = num_embeds_ada_norm
|
263 |
+
|
264 |
+
if positional_embeddings and (num_positional_embeddings is None):
|
265 |
+
raise ValueError(
|
266 |
+
"If `positional_embedding` type is defined, `num_positition_embeddings` must also be defined."
|
267 |
+
)
|
268 |
+
|
269 |
+
if positional_embeddings == "sinusoidal":
|
270 |
+
self.pos_embed = SinusoidalPositionalEmbedding(
|
271 |
+
dim, max_seq_length=num_positional_embeddings
|
272 |
+
)
|
273 |
+
else:
|
274 |
+
self.pos_embed = None
|
275 |
+
|
276 |
+
# Define 3 blocks. Each block has its own normalization layer.
|
277 |
+
# 1. Self-Attn
|
278 |
+
if norm_type == "ada_norm":
|
279 |
+
self.norm1 = AdaLayerNorm(dim, num_embeds_ada_norm)
|
280 |
+
elif norm_type == "ada_norm_zero":
|
281 |
+
self.norm1 = AdaLayerNormZero(dim, num_embeds_ada_norm)
|
282 |
+
elif norm_type == "ada_norm_continuous":
|
283 |
+
self.norm1 = AdaLayerNormContinuous(
|
284 |
+
dim,
|
285 |
+
ada_norm_continous_conditioning_embedding_dim,
|
286 |
+
norm_elementwise_affine,
|
287 |
+
norm_eps,
|
288 |
+
ada_norm_bias,
|
289 |
+
"rms_norm",
|
290 |
+
)
|
291 |
+
else:
|
292 |
+
self.norm1 = nn.LayerNorm(
|
293 |
+
dim, elementwise_affine=norm_elementwise_affine, eps=norm_eps
|
294 |
+
)
|
295 |
+
|
296 |
+
self.attn1 = Attention(
|
297 |
+
query_dim=dim,
|
298 |
+
heads=num_attention_heads,
|
299 |
+
dim_head=attention_head_dim,
|
300 |
+
dropout=dropout,
|
301 |
+
bias=attention_bias,
|
302 |
+
cross_attention_dim=cross_attention_dim if only_cross_attention else None,
|
303 |
+
upcast_attention=upcast_attention,
|
304 |
+
out_bias=attention_out_bias,
|
305 |
+
)
|
306 |
+
|
307 |
+
# 2. Cross-Attn
|
308 |
+
if cross_attention_dim is not None or double_self_attention:
|
309 |
+
# We currently only use AdaLayerNormZero for self attention where there will only be one attention block.
|
310 |
+
# I.e. the number of returned modulation chunks from AdaLayerZero would not make sense if returned during
|
311 |
+
# the second cross attention block.
|
312 |
+
if norm_type == "ada_norm":
|
313 |
+
self.norm2 = AdaLayerNorm(dim, num_embeds_ada_norm)
|
314 |
+
elif norm_type == "ada_norm_continuous":
|
315 |
+
self.norm2 = AdaLayerNormContinuous(
|
316 |
+
dim,
|
317 |
+
ada_norm_continous_conditioning_embedding_dim,
|
318 |
+
norm_elementwise_affine,
|
319 |
+
norm_eps,
|
320 |
+
ada_norm_bias,
|
321 |
+
"rms_norm",
|
322 |
+
)
|
323 |
+
else:
|
324 |
+
self.norm2 = nn.LayerNorm(dim, norm_eps, norm_elementwise_affine)
|
325 |
+
|
326 |
+
self.attn2 = Attention(
|
327 |
+
query_dim=dim,
|
328 |
+
cross_attention_dim=(
|
329 |
+
cross_attention_dim if not double_self_attention else None
|
330 |
+
),
|
331 |
+
heads=num_attention_heads,
|
332 |
+
dim_head=attention_head_dim,
|
333 |
+
dropout=dropout,
|
334 |
+
bias=attention_bias,
|
335 |
+
upcast_attention=upcast_attention,
|
336 |
+
out_bias=attention_out_bias,
|
337 |
+
) # is self-attn if encoder_hidden_states is none
|
338 |
+
else:
|
339 |
+
self.norm2 = None
|
340 |
+
self.attn2 = None
|
341 |
+
|
342 |
+
# 3. Feed-forward
|
343 |
+
if norm_type == "ada_norm_continuous":
|
344 |
+
self.norm3 = AdaLayerNormContinuous(
|
345 |
+
dim,
|
346 |
+
ada_norm_continous_conditioning_embedding_dim,
|
347 |
+
norm_elementwise_affine,
|
348 |
+
norm_eps,
|
349 |
+
ada_norm_bias,
|
350 |
+
"layer_norm",
|
351 |
+
)
|
352 |
+
|
353 |
+
elif norm_type in [
|
354 |
+
"ada_norm_zero",
|
355 |
+
"ada_norm",
|
356 |
+
"layer_norm",
|
357 |
+
"ada_norm_continuous",
|
358 |
+
]:
|
359 |
+
self.norm3 = nn.LayerNorm(dim, norm_eps, norm_elementwise_affine)
|
360 |
+
elif norm_type == "layer_norm_i2vgen":
|
361 |
+
self.norm3 = None
|
362 |
+
|
363 |
+
self.ff = FeedForward(
|
364 |
+
dim,
|
365 |
+
dropout=dropout,
|
366 |
+
activation_fn=activation_fn,
|
367 |
+
final_dropout=final_dropout,
|
368 |
+
inner_dim=ff_inner_dim,
|
369 |
+
bias=ff_bias,
|
370 |
+
)
|
371 |
+
|
372 |
+
# 4. Fuser
|
373 |
+
if attention_type == "gated" or attention_type == "gated-text-image":
|
374 |
+
self.fuser = GatedSelfAttentionDense(
|
375 |
+
dim, cross_attention_dim, num_attention_heads, attention_head_dim
|
376 |
+
)
|
377 |
+
|
378 |
+
# 5. Scale-shift for PixArt-Alpha.
|
379 |
+
if norm_type == "ada_norm_single":
|
380 |
+
self.scale_shift_table = nn.Parameter(torch.randn(6, dim) / dim**0.5)
|
381 |
+
|
382 |
+
# let chunk size default to None
|
383 |
+
self._chunk_size = None
|
384 |
+
self._chunk_dim = 0
|
385 |
+
|
386 |
+
def set_chunk_feed_forward(self, chunk_size: Optional[int], dim: int = 0):
|
387 |
+
# Sets chunk feed-forward
|
388 |
+
self._chunk_size = chunk_size
|
389 |
+
self._chunk_dim = dim
|
390 |
+
|
391 |
+
def forward(
|
392 |
+
self,
|
393 |
+
hidden_states: torch.Tensor,
|
394 |
+
attention_mask: Optional[torch.Tensor] = None,
|
395 |
+
encoder_hidden_states: Optional[torch.Tensor] = None,
|
396 |
+
encoder_attention_mask: Optional[torch.Tensor] = None,
|
397 |
+
timestep: Optional[torch.LongTensor] = None,
|
398 |
+
cross_attention_kwargs: Dict[str, Any] = None,
|
399 |
+
class_labels: Optional[torch.LongTensor] = None,
|
400 |
+
added_cond_kwargs: Optional[Dict[str, torch.Tensor]] = None,
|
401 |
+
) -> torch.Tensor:
|
402 |
+
if cross_attention_kwargs is not None:
|
403 |
+
if cross_attention_kwargs.get("scale", None) is not None:
|
404 |
+
logger.warning(
|
405 |
+
"Passing `scale` to `cross_attention_kwargs` is deprecated. `scale` will be ignored."
|
406 |
+
)
|
407 |
+
|
408 |
+
# Notice that normalization is always applied before the real computation in the following blocks.
|
409 |
+
# 0. Self-Attention
|
410 |
+
batch_size = hidden_states.shape[0]
|
411 |
+
|
412 |
+
if self.norm_type == "ada_norm":
|
413 |
+
norm_hidden_states = self.norm1(hidden_states, timestep)
|
414 |
+
elif self.norm_type == "ada_norm_zero":
|
415 |
+
norm_hidden_states, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.norm1(
|
416 |
+
hidden_states, timestep, class_labels, hidden_dtype=hidden_states.dtype
|
417 |
+
)
|
418 |
+
elif self.norm_type in ["layer_norm", "layer_norm_i2vgen"]:
|
419 |
+
norm_hidden_states = self.norm1(hidden_states)
|
420 |
+
elif self.norm_type == "ada_norm_continuous":
|
421 |
+
norm_hidden_states = self.norm1(
|
422 |
+
hidden_states, added_cond_kwargs["pooled_text_emb"]
|
423 |
+
)
|
424 |
+
elif self.norm_type == "ada_norm_single":
|
425 |
+
shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = (
|
426 |
+
self.scale_shift_table[None] + timestep.reshape(batch_size, 6, -1)
|
427 |
+
).chunk(6, dim=1)
|
428 |
+
norm_hidden_states = self.norm1(hidden_states)
|
429 |
+
norm_hidden_states = norm_hidden_states * (1 + scale_msa) + shift_msa
|
430 |
+
norm_hidden_states = norm_hidden_states.squeeze(1)
|
431 |
+
else:
|
432 |
+
raise ValueError("Incorrect norm used")
|
433 |
+
|
434 |
+
if self.pos_embed is not None:
|
435 |
+
norm_hidden_states = self.pos_embed(norm_hidden_states)
|
436 |
+
|
437 |
+
# 1. Prepare GLIGEN inputs
|
438 |
+
cross_attention_kwargs = (
|
439 |
+
cross_attention_kwargs.copy() if cross_attention_kwargs is not None else {}
|
440 |
+
)
|
441 |
+
gligen_kwargs = cross_attention_kwargs.pop("gligen", None)
|
442 |
+
|
443 |
+
attn_output = self.attn1(
|
444 |
+
norm_hidden_states,
|
445 |
+
encoder_hidden_states=(
|
446 |
+
encoder_hidden_states if self.only_cross_attention else None
|
447 |
+
),
|
448 |
+
attention_mask=attention_mask,
|
449 |
+
**cross_attention_kwargs,
|
450 |
+
)
|
451 |
+
if self.norm_type == "ada_norm_zero":
|
452 |
+
attn_output = gate_msa.unsqueeze(1) * attn_output
|
453 |
+
elif self.norm_type == "ada_norm_single":
|
454 |
+
attn_output = gate_msa * attn_output
|
455 |
+
|
456 |
+
hidden_states = attn_output + hidden_states
|
457 |
+
if hidden_states.ndim == 4:
|
458 |
+
hidden_states = hidden_states.squeeze(1)
|
459 |
+
|
460 |
+
# 1.2 GLIGEN Control
|
461 |
+
if gligen_kwargs is not None:
|
462 |
+
hidden_states = self.fuser(hidden_states, gligen_kwargs["objs"])
|
463 |
+
|
464 |
+
# 3. Cross-Attention
|
465 |
+
if self.attn2 is not None:
|
466 |
+
if self.norm_type == "ada_norm":
|
467 |
+
norm_hidden_states = self.norm2(hidden_states, timestep)
|
468 |
+
elif self.norm_type in ["ada_norm_zero", "layer_norm", "layer_norm_i2vgen"]:
|
469 |
+
norm_hidden_states = self.norm2(hidden_states)
|
470 |
+
elif self.norm_type == "ada_norm_single":
|
471 |
+
# For PixArt norm2 isn't applied here:
|
472 |
+
# https://github.com/PixArt-alpha/PixArt-alpha/blob/0f55e922376d8b797edd44d25d0e7464b260dcab/diffusion/model/nets/PixArtMS.py#L70C1-L76C103
|
473 |
+
norm_hidden_states = hidden_states
|
474 |
+
elif self.norm_type == "ada_norm_continuous":
|
475 |
+
norm_hidden_states = self.norm2(
|
476 |
+
hidden_states, added_cond_kwargs["pooled_text_emb"]
|
477 |
+
)
|
478 |
+
else:
|
479 |
+
raise ValueError("Incorrect norm")
|
480 |
+
|
481 |
+
if self.pos_embed is not None and self.norm_type != "ada_norm_single":
|
482 |
+
norm_hidden_states = self.pos_embed(norm_hidden_states)
|
483 |
+
|
484 |
+
attn_output = self.attn2(
|
485 |
+
norm_hidden_states,
|
486 |
+
encoder_hidden_states=encoder_hidden_states,
|
487 |
+
attention_mask=encoder_attention_mask,
|
488 |
+
**cross_attention_kwargs,
|
489 |
+
)
|
490 |
+
hidden_states = attn_output + hidden_states
|
491 |
+
|
492 |
+
# 4. Feed-forward
|
493 |
+
# i2vgen doesn't have this norm 🤷♂️
|
494 |
+
if self.norm_type == "ada_norm_continuous":
|
495 |
+
norm_hidden_states = self.norm3(
|
496 |
+
hidden_states, added_cond_kwargs["pooled_text_emb"]
|
497 |
+
)
|
498 |
+
elif not self.norm_type == "ada_norm_single":
|
499 |
+
norm_hidden_states = self.norm3(hidden_states)
|
500 |
+
|
501 |
+
if self.norm_type == "ada_norm_zero":
|
502 |
+
norm_hidden_states = (
|
503 |
+
norm_hidden_states * (1 + scale_mlp[:, None]) + shift_mlp[:, None]
|
504 |
+
)
|
505 |
+
|
506 |
+
if self.norm_type == "ada_norm_single":
|
507 |
+
norm_hidden_states = self.norm2(hidden_states)
|
508 |
+
norm_hidden_states = norm_hidden_states * (1 + scale_mlp) + shift_mlp
|
509 |
+
|
510 |
+
if self._chunk_size is not None:
|
511 |
+
# "feed_forward_chunk_size" can be used to save memory
|
512 |
+
ff_output = _chunked_feed_forward(
|
513 |
+
self.ff, norm_hidden_states, self._chunk_dim, self._chunk_size
|
514 |
+
)
|
515 |
+
else:
|
516 |
+
ff_output = self.ff(norm_hidden_states)
|
517 |
+
|
518 |
+
if self.norm_type == "ada_norm_zero":
|
519 |
+
ff_output = gate_mlp.unsqueeze(1) * ff_output
|
520 |
+
elif self.norm_type == "ada_norm_single":
|
521 |
+
ff_output = gate_mlp * ff_output
|
522 |
+
|
523 |
+
hidden_states = ff_output + hidden_states
|
524 |
+
if hidden_states.ndim == 4:
|
525 |
+
hidden_states = hidden_states.squeeze(1)
|
526 |
+
|
527 |
+
return hidden_states
|
528 |
+
|
529 |
+
|
530 |
+
@maybe_allow_in_graph
|
531 |
+
class TemporalRopeBasicTransformerBlock(nn.Module):
|
532 |
+
r"""
|
533 |
+
A basic Transformer block for video like data.
|
534 |
+
|
535 |
+
Parameters:
|
536 |
+
dim (`int`): The number of channels in the input and output.
|
537 |
+
time_mix_inner_dim (`int`): The number of channels for temporal attention.
|
538 |
+
num_attention_heads (`int`): The number of heads to use for multi-head attention.
|
539 |
+
attention_head_dim (`int`): The number of channels in each head.
|
540 |
+
cross_attention_dim (`int`, *optional*): The size of the encoder_hidden_states vector for cross attention.
|
541 |
+
"""
|
542 |
+
|
543 |
+
def __init__(
|
544 |
+
self,
|
545 |
+
dim: int,
|
546 |
+
time_mix_inner_dim: int,
|
547 |
+
num_attention_heads: int,
|
548 |
+
attention_head_dim: int,
|
549 |
+
cross_attention_dim: Optional[int] = None,
|
550 |
+
):
|
551 |
+
super().__init__()
|
552 |
+
self.is_res = dim == time_mix_inner_dim
|
553 |
+
|
554 |
+
self.norm_in = nn.LayerNorm(dim)
|
555 |
+
|
556 |
+
# Define 3 blocks. Each block has its own normalization layer.
|
557 |
+
# 1. Self-Attn
|
558 |
+
self.ff_in = FeedForward(
|
559 |
+
dim,
|
560 |
+
dim_out=time_mix_inner_dim,
|
561 |
+
activation_fn="geglu",
|
562 |
+
)
|
563 |
+
|
564 |
+
processor = AttnRopeProcessor2_0()
|
565 |
+
|
566 |
+
self.norm1 = nn.LayerNorm(time_mix_inner_dim)
|
567 |
+
self.attn1 = Attention(
|
568 |
+
query_dim=time_mix_inner_dim,
|
569 |
+
heads=num_attention_heads,
|
570 |
+
dim_head=attention_head_dim,
|
571 |
+
cross_attention_dim=None,
|
572 |
+
processor=processor,
|
573 |
+
)
|
574 |
+
|
575 |
+
# 2. Cross-Attn
|
576 |
+
if cross_attention_dim is not None:
|
577 |
+
# We currently only use AdaLayerNormZero for self attention where there will only be one attention block.
|
578 |
+
# I.e. the number of returned modulation chunks from AdaLayerZero would not make sense if returned during
|
579 |
+
# the second cross attention block.
|
580 |
+
self.norm2 = nn.LayerNorm(time_mix_inner_dim)
|
581 |
+
self.attn2 = Attention(
|
582 |
+
query_dim=time_mix_inner_dim,
|
583 |
+
cross_attention_dim=cross_attention_dim,
|
584 |
+
heads=num_attention_heads,
|
585 |
+
dim_head=attention_head_dim,
|
586 |
+
processor=processor,
|
587 |
+
) # is self-attn if encoder_hidden_states is none
|
588 |
+
else:
|
589 |
+
self.norm2 = None
|
590 |
+
self.attn2 = None
|
591 |
+
|
592 |
+
# 3. Feed-forward
|
593 |
+
self.norm3 = nn.LayerNorm(time_mix_inner_dim)
|
594 |
+
self.ff = FeedForward(time_mix_inner_dim, activation_fn="geglu")
|
595 |
+
|
596 |
+
# let chunk size default to None
|
597 |
+
self._chunk_size = None
|
598 |
+
self._chunk_dim = None
|
599 |
+
|
600 |
+
def set_chunk_feed_forward(self, chunk_size: Optional[int], **kwargs):
|
601 |
+
# Sets chunk feed-forward
|
602 |
+
self._chunk_size = chunk_size
|
603 |
+
# chunk dim should be hardcoded to 1 to have better speed vs. memory trade-off
|
604 |
+
self._chunk_dim = 1
|
605 |
+
|
606 |
+
def forward(
|
607 |
+
self,
|
608 |
+
hidden_states: torch.Tensor,
|
609 |
+
num_frames: int,
|
610 |
+
encoder_hidden_states: Optional[torch.Tensor] = None,
|
611 |
+
frame_rotary_emb=None,
|
612 |
+
) -> torch.Tensor:
|
613 |
+
# Notice that normalization is always applied before the real computation in the following blocks.
|
614 |
+
# 0. Self-Attention
|
615 |
+
batch_size = hidden_states.shape[0]
|
616 |
+
|
617 |
+
batch_frames, seq_length, channels = hidden_states.shape
|
618 |
+
batch_size = batch_frames // num_frames
|
619 |
+
|
620 |
+
hidden_states = hidden_states[None, :].reshape(
|
621 |
+
batch_size, num_frames, seq_length, channels
|
622 |
+
)
|
623 |
+
hidden_states = hidden_states.permute(0, 2, 1, 3)
|
624 |
+
hidden_states = hidden_states.reshape(
|
625 |
+
batch_size * seq_length, num_frames, channels
|
626 |
+
)
|
627 |
+
|
628 |
+
residual = hidden_states
|
629 |
+
hidden_states = self.norm_in(hidden_states)
|
630 |
+
|
631 |
+
if self._chunk_size is not None:
|
632 |
+
hidden_states = _chunked_feed_forward(
|
633 |
+
self.ff_in, hidden_states, self._chunk_dim, self._chunk_size
|
634 |
+
)
|
635 |
+
else:
|
636 |
+
hidden_states = self.ff_in(hidden_states)
|
637 |
+
|
638 |
+
if self.is_res:
|
639 |
+
hidden_states = hidden_states + residual
|
640 |
+
|
641 |
+
norm_hidden_states = self.norm1(hidden_states)
|
642 |
+
attn_output = self.attn1(
|
643 |
+
norm_hidden_states,
|
644 |
+
encoder_hidden_states=None,
|
645 |
+
frame_rotary_emb=frame_rotary_emb,
|
646 |
+
)
|
647 |
+
hidden_states = attn_output + hidden_states
|
648 |
+
|
649 |
+
# 3. Cross-Attention
|
650 |
+
if self.attn2 is not None:
|
651 |
+
norm_hidden_states = self.norm2(hidden_states)
|
652 |
+
attn_output = self.attn2(
|
653 |
+
norm_hidden_states,
|
654 |
+
encoder_hidden_states=encoder_hidden_states,
|
655 |
+
frame_rotary_emb=frame_rotary_emb,
|
656 |
+
)
|
657 |
+
hidden_states = attn_output + hidden_states
|
658 |
+
|
659 |
+
# 4. Feed-forward
|
660 |
+
norm_hidden_states = self.norm3(hidden_states)
|
661 |
+
|
662 |
+
if self._chunk_size is not None:
|
663 |
+
ff_output = _chunked_feed_forward(
|
664 |
+
self.ff, norm_hidden_states, self._chunk_dim, self._chunk_size
|
665 |
+
)
|
666 |
+
else:
|
667 |
+
ff_output = self.ff(norm_hidden_states)
|
668 |
+
|
669 |
+
if self.is_res:
|
670 |
+
hidden_states = ff_output + hidden_states
|
671 |
+
else:
|
672 |
+
hidden_states = ff_output
|
673 |
+
|
674 |
+
hidden_states = hidden_states[None, :].reshape(
|
675 |
+
batch_size, seq_length, num_frames, channels
|
676 |
+
)
|
677 |
+
hidden_states = hidden_states.permute(0, 2, 1, 3)
|
678 |
+
hidden_states = hidden_states.reshape(
|
679 |
+
batch_size * num_frames, seq_length, channels
|
680 |
+
)
|
681 |
+
|
682 |
+
return hidden_states
|
683 |
+
|
684 |
+
|
685 |
+
class FeedForward(nn.Module):
|
686 |
+
r"""
|
687 |
+
A feed-forward layer.
|
688 |
+
|
689 |
+
Parameters:
|
690 |
+
dim (`int`): The number of channels in the input.
|
691 |
+
dim_out (`int`, *optional*): The number of channels in the output. If not given, defaults to `dim`.
|
692 |
+
mult (`int`, *optional*, defaults to 4): The multiplier to use for the hidden dimension.
|
693 |
+
dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
|
694 |
+
activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to be used in feed-forward.
|
695 |
+
final_dropout (`bool` *optional*, defaults to False): Apply a final dropout.
|
696 |
+
bias (`bool`, defaults to True): Whether to use a bias in the linear layer.
|
697 |
+
"""
|
698 |
+
|
699 |
+
def __init__(
|
700 |
+
self,
|
701 |
+
dim: int,
|
702 |
+
dim_out: Optional[int] = None,
|
703 |
+
mult: int = 4,
|
704 |
+
dropout: float = 0.0,
|
705 |
+
activation_fn: str = "geglu",
|
706 |
+
final_dropout: bool = False,
|
707 |
+
inner_dim=None,
|
708 |
+
bias: bool = True,
|
709 |
+
):
|
710 |
+
super().__init__()
|
711 |
+
if inner_dim is None:
|
712 |
+
inner_dim = int(dim * mult)
|
713 |
+
dim_out = dim_out if dim_out is not None else dim
|
714 |
+
|
715 |
+
if activation_fn == "gelu":
|
716 |
+
act_fn = GELU(dim, inner_dim, bias=bias)
|
717 |
+
if activation_fn == "gelu-approximate":
|
718 |
+
act_fn = GELU(dim, inner_dim, approximate="tanh", bias=bias)
|
719 |
+
elif activation_fn == "geglu":
|
720 |
+
act_fn = GEGLU(dim, inner_dim, bias=bias)
|
721 |
+
elif activation_fn == "geglu-approximate":
|
722 |
+
act_fn = ApproximateGELU(dim, inner_dim, bias=bias)
|
723 |
+
|
724 |
+
self.net = nn.ModuleList([])
|
725 |
+
# project in
|
726 |
+
self.net.append(act_fn)
|
727 |
+
# project dropout
|
728 |
+
self.net.append(nn.Dropout(dropout))
|
729 |
+
# project out
|
730 |
+
self.net.append(nn.Linear(inner_dim, dim_out, bias=bias))
|
731 |
+
# FF as used in Vision Transformer, MLP-Mixer, etc. have a final dropout
|
732 |
+
if final_dropout:
|
733 |
+
self.net.append(nn.Dropout(dropout))
|
734 |
+
|
735 |
+
def forward(self, hidden_states: torch.Tensor, *args, **kwargs) -> torch.Tensor:
|
736 |
+
if len(args) > 0 or kwargs.get("scale", None) is not None:
|
737 |
+
deprecation_message = "The `scale` argument is deprecated and will be ignored. Please remove it, as passing it will raise an error in the future. `scale` should directly be passed while calling the underlying pipeline component i.e., via `cross_attention_kwargs`."
|
738 |
+
deprecate("scale", "1.0.0", deprecation_message)
|
739 |
+
for module in self.net:
|
740 |
+
hidden_states = module(hidden_states)
|
741 |
+
return hidden_states
|
models/attention_processor.py
ADDED
The diff for this file is too large to render.
See raw diff
|
|
models/embeddings.py
ADDED
@@ -0,0 +1,1539 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Copyright 2024 The HuggingFace Team. All rights reserved.
|
2 |
+
#
|
3 |
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
4 |
+
# you may not use this file except in compliance with the License.
|
5 |
+
# You may obtain a copy of the License at
|
6 |
+
#
|
7 |
+
# http://www.apache.org/licenses/LICENSE-2.0
|
8 |
+
#
|
9 |
+
# Unless required by applicable law or agreed to in writing, software
|
10 |
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
11 |
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
12 |
+
# See the License for the specific language governing permissions and
|
13 |
+
# limitations under the License.
|
14 |
+
import math
|
15 |
+
from typing import List, Optional, Tuple, Union
|
16 |
+
|
17 |
+
import numpy as np
|
18 |
+
import torch
|
19 |
+
import torch.nn.functional as F
|
20 |
+
from torch import nn
|
21 |
+
|
22 |
+
from diffusers.utils import deprecate
|
23 |
+
from diffusers.models.activations import FP32SiLU, get_activation
|
24 |
+
from diffusers.models.attention_processor import Attention
|
25 |
+
|
26 |
+
|
27 |
+
def get_timestep_embedding(
|
28 |
+
timesteps: torch.Tensor,
|
29 |
+
embedding_dim: int,
|
30 |
+
flip_sin_to_cos: bool = False,
|
31 |
+
downscale_freq_shift: float = 1,
|
32 |
+
scale: float = 1,
|
33 |
+
max_period: int = 10000,
|
34 |
+
):
|
35 |
+
"""
|
36 |
+
This matches the implementation in Denoising Diffusion Probabilistic Models: Create sinusoidal timestep embeddings.
|
37 |
+
|
38 |
+
:param timesteps: a 1-D Tensor of N indices, one per batch element.
|
39 |
+
These may be fractional.
|
40 |
+
:param embedding_dim: the dimension of the output. :param max_period: controls the minimum frequency of the
|
41 |
+
embeddings. :return: an [N x dim] Tensor of positional embeddings.
|
42 |
+
"""
|
43 |
+
assert len(timesteps.shape) == 1, "Timesteps should be a 1d-array"
|
44 |
+
|
45 |
+
half_dim = embedding_dim // 2
|
46 |
+
exponent = -math.log(max_period) * torch.arange(
|
47 |
+
start=0, end=half_dim, dtype=torch.float32, device=timesteps.device
|
48 |
+
)
|
49 |
+
exponent = exponent / (half_dim - downscale_freq_shift)
|
50 |
+
|
51 |
+
emb = torch.exp(exponent)
|
52 |
+
emb = timesteps[:, None].float() * emb[None, :]
|
53 |
+
|
54 |
+
# scale embeddings
|
55 |
+
emb = scale * emb
|
56 |
+
|
57 |
+
# concat sine and cosine embeddings
|
58 |
+
emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=-1)
|
59 |
+
|
60 |
+
# flip sine and cosine embeddings
|
61 |
+
if flip_sin_to_cos:
|
62 |
+
emb = torch.cat([emb[:, half_dim:], emb[:, :half_dim]], dim=-1)
|
63 |
+
|
64 |
+
# zero pad
|
65 |
+
if embedding_dim % 2 == 1:
|
66 |
+
emb = torch.nn.functional.pad(emb, (0, 1, 0, 0))
|
67 |
+
return emb
|
68 |
+
|
69 |
+
|
70 |
+
def get_2d_sincos_pos_embed(
|
71 |
+
embed_dim,
|
72 |
+
grid_size,
|
73 |
+
cls_token=False,
|
74 |
+
extra_tokens=0,
|
75 |
+
interpolation_scale=1.0,
|
76 |
+
base_size=16,
|
77 |
+
):
|
78 |
+
"""
|
79 |
+
grid_size: int of the grid height and width return: pos_embed: [grid_size*grid_size, embed_dim] or
|
80 |
+
[1+grid_size*grid_size, embed_dim] (w/ or w/o cls_token)
|
81 |
+
"""
|
82 |
+
if isinstance(grid_size, int):
|
83 |
+
grid_size = (grid_size, grid_size)
|
84 |
+
|
85 |
+
grid_h = (
|
86 |
+
np.arange(grid_size[0], dtype=np.float32)
|
87 |
+
/ (grid_size[0] / base_size)
|
88 |
+
/ interpolation_scale
|
89 |
+
)
|
90 |
+
grid_w = (
|
91 |
+
np.arange(grid_size[1], dtype=np.float32)
|
92 |
+
/ (grid_size[1] / base_size)
|
93 |
+
/ interpolation_scale
|
94 |
+
)
|
95 |
+
grid = np.meshgrid(grid_w, grid_h) # here w goes first
|
96 |
+
grid = np.stack(grid, axis=0)
|
97 |
+
|
98 |
+
grid = grid.reshape([2, 1, grid_size[1], grid_size[0]])
|
99 |
+
pos_embed = get_2d_sincos_pos_embed_from_grid(embed_dim, grid)
|
100 |
+
if cls_token and extra_tokens > 0:
|
101 |
+
pos_embed = np.concatenate(
|
102 |
+
[np.zeros([extra_tokens, embed_dim]), pos_embed], axis=0
|
103 |
+
)
|
104 |
+
return pos_embed
|
105 |
+
|
106 |
+
|
107 |
+
def get_2d_sincos_pos_embed_from_grid(embed_dim, grid):
|
108 |
+
if embed_dim % 2 != 0:
|
109 |
+
raise ValueError("embed_dim must be divisible by 2")
|
110 |
+
|
111 |
+
# use half of dimensions to encode grid_h
|
112 |
+
emb_h = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[0]) # (H*W, D/2)
|
113 |
+
emb_w = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[1]) # (H*W, D/2)
|
114 |
+
|
115 |
+
emb = np.concatenate([emb_h, emb_w], axis=1) # (H*W, D)
|
116 |
+
return emb
|
117 |
+
|
118 |
+
|
119 |
+
def get_1d_sincos_pos_embed_from_grid(embed_dim, pos):
|
120 |
+
"""
|
121 |
+
embed_dim: output dimension for each position pos: a list of positions to be encoded: size (M,) out: (M, D)
|
122 |
+
"""
|
123 |
+
if embed_dim % 2 != 0:
|
124 |
+
raise ValueError("embed_dim must be divisible by 2")
|
125 |
+
|
126 |
+
omega = np.arange(embed_dim // 2, dtype=np.float64)
|
127 |
+
omega /= embed_dim / 2.0
|
128 |
+
omega = 1.0 / 10000**omega # (D/2,)
|
129 |
+
|
130 |
+
pos = pos.reshape(-1) # (M,)
|
131 |
+
out = np.einsum("m,d->md", pos, omega) # (M, D/2), outer product
|
132 |
+
|
133 |
+
emb_sin = np.sin(out) # (M, D/2)
|
134 |
+
emb_cos = np.cos(out) # (M, D/2)
|
135 |
+
|
136 |
+
emb = np.concatenate([emb_sin, emb_cos], axis=1) # (M, D)
|
137 |
+
return emb
|
138 |
+
|
139 |
+
|
140 |
+
class PatchEmbed(nn.Module):
|
141 |
+
"""2D Image to Patch Embedding with support for SD3 cropping."""
|
142 |
+
|
143 |
+
def __init__(
|
144 |
+
self,
|
145 |
+
height=224,
|
146 |
+
width=224,
|
147 |
+
patch_size=16,
|
148 |
+
in_channels=3,
|
149 |
+
embed_dim=768,
|
150 |
+
layer_norm=False,
|
151 |
+
flatten=True,
|
152 |
+
bias=True,
|
153 |
+
interpolation_scale=1,
|
154 |
+
pos_embed_type="sincos",
|
155 |
+
pos_embed_max_size=None, # For SD3 cropping
|
156 |
+
):
|
157 |
+
super().__init__()
|
158 |
+
|
159 |
+
num_patches = (height // patch_size) * (width // patch_size)
|
160 |
+
self.flatten = flatten
|
161 |
+
self.layer_norm = layer_norm
|
162 |
+
self.pos_embed_max_size = pos_embed_max_size
|
163 |
+
|
164 |
+
self.proj = nn.Conv2d(
|
165 |
+
in_channels,
|
166 |
+
embed_dim,
|
167 |
+
kernel_size=(patch_size, patch_size),
|
168 |
+
stride=patch_size,
|
169 |
+
bias=bias,
|
170 |
+
)
|
171 |
+
if layer_norm:
|
172 |
+
self.norm = nn.LayerNorm(embed_dim, elementwise_affine=False, eps=1e-6)
|
173 |
+
else:
|
174 |
+
self.norm = None
|
175 |
+
|
176 |
+
self.patch_size = patch_size
|
177 |
+
self.height, self.width = height // patch_size, width // patch_size
|
178 |
+
self.base_size = height // patch_size
|
179 |
+
self.interpolation_scale = interpolation_scale
|
180 |
+
|
181 |
+
# Calculate positional embeddings based on max size or default
|
182 |
+
if pos_embed_max_size:
|
183 |
+
grid_size = pos_embed_max_size
|
184 |
+
else:
|
185 |
+
grid_size = int(num_patches**0.5)
|
186 |
+
|
187 |
+
if pos_embed_type is None:
|
188 |
+
self.pos_embed = None
|
189 |
+
elif pos_embed_type == "sincos":
|
190 |
+
pos_embed = get_2d_sincos_pos_embed(
|
191 |
+
embed_dim,
|
192 |
+
grid_size,
|
193 |
+
base_size=self.base_size,
|
194 |
+
interpolation_scale=self.interpolation_scale,
|
195 |
+
)
|
196 |
+
persistent = True if pos_embed_max_size else False
|
197 |
+
self.register_buffer(
|
198 |
+
"pos_embed",
|
199 |
+
torch.from_numpy(pos_embed).float().unsqueeze(0),
|
200 |
+
persistent=persistent,
|
201 |
+
)
|
202 |
+
else:
|
203 |
+
raise ValueError(f"Unsupported pos_embed_type: {pos_embed_type}")
|
204 |
+
|
205 |
+
def cropped_pos_embed(self, height, width):
|
206 |
+
"""Crops positional embeddings for SD3 compatibility."""
|
207 |
+
if self.pos_embed_max_size is None:
|
208 |
+
raise ValueError("`pos_embed_max_size` must be set for cropping.")
|
209 |
+
|
210 |
+
height = height // self.patch_size
|
211 |
+
width = width // self.patch_size
|
212 |
+
if height > self.pos_embed_max_size:
|
213 |
+
raise ValueError(
|
214 |
+
f"Height ({height}) cannot be greater than `pos_embed_max_size`: {self.pos_embed_max_size}."
|
215 |
+
)
|
216 |
+
if width > self.pos_embed_max_size:
|
217 |
+
raise ValueError(
|
218 |
+
f"Width ({width}) cannot be greater than `pos_embed_max_size`: {self.pos_embed_max_size}."
|
219 |
+
)
|
220 |
+
|
221 |
+
top = (self.pos_embed_max_size - height) // 2
|
222 |
+
left = (self.pos_embed_max_size - width) // 2
|
223 |
+
spatial_pos_embed = self.pos_embed.reshape(
|
224 |
+
1, self.pos_embed_max_size, self.pos_embed_max_size, -1
|
225 |
+
)
|
226 |
+
spatial_pos_embed = spatial_pos_embed[
|
227 |
+
:, top : top + height, left : left + width, :
|
228 |
+
]
|
229 |
+
spatial_pos_embed = spatial_pos_embed.reshape(
|
230 |
+
1, -1, spatial_pos_embed.shape[-1]
|
231 |
+
)
|
232 |
+
return spatial_pos_embed
|
233 |
+
|
234 |
+
def forward(self, latent):
|
235 |
+
if self.pos_embed_max_size is not None:
|
236 |
+
height, width = latent.shape[-2:]
|
237 |
+
else:
|
238 |
+
height, width = (
|
239 |
+
latent.shape[-2] // self.patch_size,
|
240 |
+
latent.shape[-1] // self.patch_size,
|
241 |
+
)
|
242 |
+
|
243 |
+
latent = self.proj(latent)
|
244 |
+
if self.flatten:
|
245 |
+
latent = latent.flatten(2).transpose(1, 2) # BCHW -> BNC
|
246 |
+
if self.layer_norm:
|
247 |
+
latent = self.norm(latent)
|
248 |
+
if self.pos_embed is None:
|
249 |
+
return latent.to(latent.dtype)
|
250 |
+
# Interpolate or crop positional embeddings as needed
|
251 |
+
if self.pos_embed_max_size:
|
252 |
+
pos_embed = self.cropped_pos_embed(height, width)
|
253 |
+
else:
|
254 |
+
if self.height != height or self.width != width:
|
255 |
+
pos_embed = get_2d_sincos_pos_embed(
|
256 |
+
embed_dim=self.pos_embed.shape[-1],
|
257 |
+
grid_size=(height, width),
|
258 |
+
base_size=self.base_size,
|
259 |
+
interpolation_scale=self.interpolation_scale,
|
260 |
+
)
|
261 |
+
pos_embed = (
|
262 |
+
torch.from_numpy(pos_embed).float().unsqueeze(0).to(latent.device)
|
263 |
+
)
|
264 |
+
else:
|
265 |
+
pos_embed = self.pos_embed
|
266 |
+
|
267 |
+
return (latent + pos_embed).to(latent.dtype)
|
268 |
+
|
269 |
+
|
270 |
+
def get_2d_rotary_pos_embed(embed_dim, crops_coords, grid_size, use_real=True):
|
271 |
+
"""
|
272 |
+
RoPE for image tokens with 2d structure.
|
273 |
+
|
274 |
+
Args:
|
275 |
+
embed_dim: (`int`):
|
276 |
+
The embedding dimension size
|
277 |
+
crops_coords (`Tuple[int]`)
|
278 |
+
The top-left and bottom-right coordinates of the crop.
|
279 |
+
grid_size (`Tuple[int]`):
|
280 |
+
The grid size of the positional embedding.
|
281 |
+
use_real (`bool`):
|
282 |
+
If True, return real part and imaginary part separately. Otherwise, return complex numbers.
|
283 |
+
|
284 |
+
Returns:
|
285 |
+
`torch.Tensor`: positional embdding with shape `( grid_size * grid_size, embed_dim/2)`.
|
286 |
+
"""
|
287 |
+
start, stop = crops_coords
|
288 |
+
grid_h = np.linspace(
|
289 |
+
start[0], stop[0], grid_size[0], endpoint=False, dtype=np.float32
|
290 |
+
)
|
291 |
+
grid_w = np.linspace(
|
292 |
+
start[1], stop[1], grid_size[1], endpoint=False, dtype=np.float32
|
293 |
+
)
|
294 |
+
grid = np.meshgrid(grid_w, grid_h) # here w goes first
|
295 |
+
grid = np.stack(grid, axis=0) # [2, W, H]
|
296 |
+
|
297 |
+
grid = grid.reshape([2, 1, *grid.shape[1:]])
|
298 |
+
pos_embed = get_2d_rotary_pos_embed_from_grid(embed_dim, grid, use_real=use_real)
|
299 |
+
return pos_embed
|
300 |
+
|
301 |
+
|
302 |
+
def get_2d_rotary_pos_embed_from_grid(embed_dim, grid, use_real=False):
|
303 |
+
assert embed_dim % 4 == 0
|
304 |
+
|
305 |
+
# use half of dimensions to encode grid_h
|
306 |
+
emb_h = get_1d_rotary_pos_embed(
|
307 |
+
embed_dim // 2, grid[0].reshape(-1), use_real=use_real
|
308 |
+
) # (H*W, D/4)
|
309 |
+
emb_w = get_1d_rotary_pos_embed(
|
310 |
+
embed_dim // 2, grid[1].reshape(-1), use_real=use_real
|
311 |
+
) # (H*W, D/4)
|
312 |
+
|
313 |
+
if use_real:
|
314 |
+
cos = torch.cat([emb_h[0], emb_w[0]], dim=1) # (H*W, D/2)
|
315 |
+
sin = torch.cat([emb_h[1], emb_w[1]], dim=1) # (H*W, D/2)
|
316 |
+
return cos, sin
|
317 |
+
else:
|
318 |
+
emb = torch.cat([emb_h, emb_w], dim=1) # (H*W, D/2)
|
319 |
+
return emb
|
320 |
+
|
321 |
+
|
322 |
+
def get_1d_rotary_pos_embed(
|
323 |
+
dim: int, pos: Union[np.ndarray, int], theta: float = 10000.0, use_real=False
|
324 |
+
):
|
325 |
+
"""
|
326 |
+
Precompute the frequency tensor for complex exponentials (cis) with given dimensions.
|
327 |
+
|
328 |
+
This function calculates a frequency tensor with complex exponentials using the given dimension 'dim' and the end
|
329 |
+
index 'end'. The 'theta' parameter scales the frequencies. The returned tensor contains complex values in complex64
|
330 |
+
data type.
|
331 |
+
|
332 |
+
Args:
|
333 |
+
dim (`int`): Dimension of the frequency tensor.
|
334 |
+
pos (`np.ndarray` or `int`): Position indices for the frequency tensor. [S] or scalar
|
335 |
+
theta (`float`, *optional*, defaults to 10000.0):
|
336 |
+
Scaling factor for frequency computation. Defaults to 10000.0.
|
337 |
+
use_real (`bool`, *optional*):
|
338 |
+
If True, return real part and imaginary part separately. Otherwise, return complex numbers.
|
339 |
+
|
340 |
+
Returns:
|
341 |
+
`torch.Tensor`: Precomputed frequency tensor with complex exponentials. [S, D/2]
|
342 |
+
"""
|
343 |
+
if isinstance(pos, int):
|
344 |
+
pos = np.arange(pos)
|
345 |
+
freqs = 1.0 / (
|
346 |
+
theta ** (torch.arange(0, dim, 2)[: (dim // 2)].float() / dim)
|
347 |
+
) # [D/2]
|
348 |
+
t = torch.from_numpy(pos).to(freqs.device) # type: ignore # [S]
|
349 |
+
freqs = torch.outer(t, freqs).float() # type: ignore # [S, D/2]
|
350 |
+
if use_real:
|
351 |
+
freqs_cos = freqs.cos().repeat_interleave(2, dim=1) # [S, D]
|
352 |
+
freqs_sin = freqs.sin().repeat_interleave(2, dim=1) # [S, D]
|
353 |
+
return freqs_cos, freqs_sin
|
354 |
+
else:
|
355 |
+
freqs_cis = torch.polar(
|
356 |
+
torch.ones_like(freqs), freqs
|
357 |
+
) # complex64 # [S, D/2]
|
358 |
+
return freqs_cis
|
359 |
+
|
360 |
+
|
361 |
+
def apply_rotary_emb(
|
362 |
+
x: torch.Tensor,
|
363 |
+
freqs_cis: Union[torch.Tensor, Tuple[torch.Tensor]],
|
364 |
+
) -> Tuple[torch.Tensor, torch.Tensor]:
|
365 |
+
"""
|
366 |
+
Apply rotary embeddings to input tensors using the given frequency tensor. This function applies rotary embeddings
|
367 |
+
to the given query or key 'x' tensors using the provided frequency tensor 'freqs_cis'. The input tensors are
|
368 |
+
reshaped as complex numbers, and the frequency tensor is reshaped for broadcasting compatibility. The resulting
|
369 |
+
tensors contain rotary embeddings and are returned as real tensors.
|
370 |
+
|
371 |
+
Args:
|
372 |
+
x (`torch.Tensor`):
|
373 |
+
Query or key tensor to apply rotary embeddings. [B, H, S, D] xk (torch.Tensor): Key tensor to apply
|
374 |
+
freqs_cis (`Tuple[torch.Tensor]`): Precomputed frequency tensor for complex exponentials. ([S, D], [S, D],)
|
375 |
+
|
376 |
+
Returns:
|
377 |
+
Tuple[torch.Tensor, torch.Tensor]: Tuple of modified query tensor and key tensor with rotary embeddings.
|
378 |
+
"""
|
379 |
+
cos, sin = freqs_cis # [S, D]
|
380 |
+
cos = cos[None, None]
|
381 |
+
sin = sin[None, None]
|
382 |
+
cos, sin = cos.to(x.device), sin.to(x.device)
|
383 |
+
|
384 |
+
x_real, x_imag = x.reshape(*x.shape[:-1], -1, 2).unbind(-1) # [B, S, H, D//2]
|
385 |
+
x_rotated = torch.stack([-x_imag, x_real], dim=-1).flatten(3)
|
386 |
+
out = (x.float() * cos + x_rotated.float() * sin).to(x.dtype)
|
387 |
+
|
388 |
+
return out
|
389 |
+
|
390 |
+
|
391 |
+
def rope(pos: torch.Tensor, dim: int, theta=10000.0) -> torch.Tensor:
|
392 |
+
assert dim % 2 == 0, "The dimension must be even."
|
393 |
+
|
394 |
+
scale = torch.arange(0, dim, 2, dtype=torch.float64, device=pos.device) / dim
|
395 |
+
omega = 1.0 / (theta**scale)
|
396 |
+
|
397 |
+
batch_size, seq_length = pos.shape
|
398 |
+
# (B, N, d/2)
|
399 |
+
out = torch.einsum("...n,d->...nd", pos, omega)
|
400 |
+
cos_out = torch.cos(out)
|
401 |
+
sin_out = torch.sin(out)
|
402 |
+
|
403 |
+
stacked_out = torch.stack([cos_out, -sin_out, sin_out, cos_out], dim=-1)
|
404 |
+
# (B, 1, N, d/2, 2, 2)
|
405 |
+
out = stacked_out.view(batch_size, 1, -1, dim // 2, 2, 2)
|
406 |
+
return out.float()
|
407 |
+
|
408 |
+
|
409 |
+
def apply_rope(x, freqs_cis):
|
410 |
+
# (B, num_heads, N, d/2, 1, 2)
|
411 |
+
x_ = x.float().reshape(*x.shape[:-1], -1, 1, 2)
|
412 |
+
# cos * q0 - sin * q1, sin * q0 + cos * q1
|
413 |
+
x_out = freqs_cis[..., 0] * x_[..., 0] + freqs_cis[..., 1] * x_[..., 1]
|
414 |
+
return x_out.reshape(*x.shape).type_as(x)
|
415 |
+
|
416 |
+
|
417 |
+
class TimestepEmbedding(nn.Module):
|
418 |
+
def __init__(
|
419 |
+
self,
|
420 |
+
in_channels: int,
|
421 |
+
time_embed_dim: int,
|
422 |
+
act_fn: str = "silu",
|
423 |
+
out_dim: int = None,
|
424 |
+
post_act_fn: Optional[str] = None,
|
425 |
+
cond_proj_dim=None,
|
426 |
+
sample_proj_bias=True,
|
427 |
+
):
|
428 |
+
super().__init__()
|
429 |
+
|
430 |
+
self.linear_1 = nn.Linear(in_channels, time_embed_dim, sample_proj_bias)
|
431 |
+
|
432 |
+
if cond_proj_dim is not None:
|
433 |
+
self.cond_proj = nn.Linear(cond_proj_dim, in_channels, bias=False)
|
434 |
+
else:
|
435 |
+
self.cond_proj = None
|
436 |
+
|
437 |
+
self.act = get_activation(act_fn)
|
438 |
+
|
439 |
+
if out_dim is not None:
|
440 |
+
time_embed_dim_out = out_dim
|
441 |
+
else:
|
442 |
+
time_embed_dim_out = time_embed_dim
|
443 |
+
self.linear_2 = nn.Linear(time_embed_dim, time_embed_dim_out, sample_proj_bias)
|
444 |
+
|
445 |
+
if post_act_fn is None:
|
446 |
+
self.post_act = None
|
447 |
+
else:
|
448 |
+
self.post_act = get_activation(post_act_fn)
|
449 |
+
|
450 |
+
def forward(self, sample, condition=None):
|
451 |
+
if condition is not None:
|
452 |
+
sample = sample + self.cond_proj(condition)
|
453 |
+
sample = self.linear_1(sample)
|
454 |
+
|
455 |
+
if self.act is not None:
|
456 |
+
sample = self.act(sample)
|
457 |
+
|
458 |
+
sample = self.linear_2(sample)
|
459 |
+
|
460 |
+
if self.post_act is not None:
|
461 |
+
sample = self.post_act(sample)
|
462 |
+
return sample
|
463 |
+
|
464 |
+
|
465 |
+
class Timesteps(nn.Module):
|
466 |
+
def __init__(
|
467 |
+
self, num_channels: int, flip_sin_to_cos: bool, downscale_freq_shift: float
|
468 |
+
):
|
469 |
+
super().__init__()
|
470 |
+
self.num_channels = num_channels
|
471 |
+
self.flip_sin_to_cos = flip_sin_to_cos
|
472 |
+
self.downscale_freq_shift = downscale_freq_shift
|
473 |
+
|
474 |
+
def forward(self, timesteps):
|
475 |
+
t_emb = get_timestep_embedding(
|
476 |
+
timesteps,
|
477 |
+
self.num_channels,
|
478 |
+
flip_sin_to_cos=self.flip_sin_to_cos,
|
479 |
+
downscale_freq_shift=self.downscale_freq_shift,
|
480 |
+
)
|
481 |
+
return t_emb
|
482 |
+
|
483 |
+
|
484 |
+
class GaussianFourierProjection(nn.Module):
|
485 |
+
"""Gaussian Fourier embeddings for noise levels."""
|
486 |
+
|
487 |
+
def __init__(
|
488 |
+
self,
|
489 |
+
embedding_size: int = 256,
|
490 |
+
scale: float = 1.0,
|
491 |
+
set_W_to_weight=True,
|
492 |
+
log=True,
|
493 |
+
flip_sin_to_cos=False,
|
494 |
+
):
|
495 |
+
super().__init__()
|
496 |
+
self.weight = nn.Parameter(
|
497 |
+
torch.randn(embedding_size) * scale, requires_grad=False
|
498 |
+
)
|
499 |
+
self.log = log
|
500 |
+
self.flip_sin_to_cos = flip_sin_to_cos
|
501 |
+
|
502 |
+
if set_W_to_weight:
|
503 |
+
# to delete later
|
504 |
+
self.W = nn.Parameter(
|
505 |
+
torch.randn(embedding_size) * scale, requires_grad=False
|
506 |
+
)
|
507 |
+
|
508 |
+
self.weight = self.W
|
509 |
+
|
510 |
+
def forward(self, x):
|
511 |
+
if self.log:
|
512 |
+
x = torch.log(x)
|
513 |
+
|
514 |
+
x_proj = x[:, None] * self.weight[None, :] * 2 * np.pi
|
515 |
+
|
516 |
+
if self.flip_sin_to_cos:
|
517 |
+
out = torch.cat([torch.cos(x_proj), torch.sin(x_proj)], dim=-1)
|
518 |
+
else:
|
519 |
+
out = torch.cat([torch.sin(x_proj), torch.cos(x_proj)], dim=-1)
|
520 |
+
return out
|
521 |
+
|
522 |
+
|
523 |
+
class SinusoidalPositionalEmbedding(nn.Module):
|
524 |
+
"""Apply positional information to a sequence of embeddings.
|
525 |
+
|
526 |
+
Takes in a sequence of embeddings with shape (batch_size, seq_length, embed_dim) and adds positional embeddings to
|
527 |
+
them
|
528 |
+
|
529 |
+
Args:
|
530 |
+
embed_dim: (int): Dimension of the positional embedding.
|
531 |
+
max_seq_length: Maximum sequence length to apply positional embeddings
|
532 |
+
|
533 |
+
"""
|
534 |
+
|
535 |
+
def __init__(self, embed_dim: int, max_seq_length: int = 32):
|
536 |
+
super().__init__()
|
537 |
+
position = torch.arange(max_seq_length).unsqueeze(1)
|
538 |
+
div_term = torch.exp(
|
539 |
+
torch.arange(0, embed_dim, 2) * (-math.log(10000.0) / embed_dim)
|
540 |
+
)
|
541 |
+
pe = torch.zeros(1, max_seq_length, embed_dim)
|
542 |
+
pe[0, :, 0::2] = torch.sin(position * div_term)
|
543 |
+
pe[0, :, 1::2] = torch.cos(position * div_term)
|
544 |
+
self.register_buffer("pe", pe)
|
545 |
+
|
546 |
+
def forward(self, x):
|
547 |
+
_, seq_length, _ = x.shape
|
548 |
+
x = x + self.pe[:, :seq_length]
|
549 |
+
return x
|
550 |
+
|
551 |
+
|
552 |
+
class ImagePositionalEmbeddings(nn.Module):
|
553 |
+
"""
|
554 |
+
Converts latent image classes into vector embeddings. Sums the vector embeddings with positional embeddings for the
|
555 |
+
height and width of the latent space.
|
556 |
+
|
557 |
+
For more details, see figure 10 of the dall-e paper: https://arxiv.org/abs/2102.12092
|
558 |
+
|
559 |
+
For VQ-diffusion:
|
560 |
+
|
561 |
+
Output vector embeddings are used as input for the transformer.
|
562 |
+
|
563 |
+
Note that the vector embeddings for the transformer are different than the vector embeddings from the VQVAE.
|
564 |
+
|
565 |
+
Args:
|
566 |
+
num_embed (`int`):
|
567 |
+
Number of embeddings for the latent pixels embeddings.
|
568 |
+
height (`int`):
|
569 |
+
Height of the latent image i.e. the number of height embeddings.
|
570 |
+
width (`int`):
|
571 |
+
Width of the latent image i.e. the number of width embeddings.
|
572 |
+
embed_dim (`int`):
|
573 |
+
Dimension of the produced vector embeddings. Used for the latent pixel, height, and width embeddings.
|
574 |
+
"""
|
575 |
+
|
576 |
+
def __init__(
|
577 |
+
self,
|
578 |
+
num_embed: int,
|
579 |
+
height: int,
|
580 |
+
width: int,
|
581 |
+
embed_dim: int,
|
582 |
+
):
|
583 |
+
super().__init__()
|
584 |
+
|
585 |
+
self.height = height
|
586 |
+
self.width = width
|
587 |
+
self.num_embed = num_embed
|
588 |
+
self.embed_dim = embed_dim
|
589 |
+
|
590 |
+
self.emb = nn.Embedding(self.num_embed, embed_dim)
|
591 |
+
self.height_emb = nn.Embedding(self.height, embed_dim)
|
592 |
+
self.width_emb = nn.Embedding(self.width, embed_dim)
|
593 |
+
|
594 |
+
def forward(self, index):
|
595 |
+
emb = self.emb(index)
|
596 |
+
|
597 |
+
height_emb = self.height_emb(
|
598 |
+
torch.arange(self.height, device=index.device).view(1, self.height)
|
599 |
+
)
|
600 |
+
|
601 |
+
# 1 x H x D -> 1 x H x 1 x D
|
602 |
+
height_emb = height_emb.unsqueeze(2)
|
603 |
+
|
604 |
+
width_emb = self.width_emb(
|
605 |
+
torch.arange(self.width, device=index.device).view(1, self.width)
|
606 |
+
)
|
607 |
+
|
608 |
+
# 1 x W x D -> 1 x 1 x W x D
|
609 |
+
width_emb = width_emb.unsqueeze(1)
|
610 |
+
|
611 |
+
pos_emb = height_emb + width_emb
|
612 |
+
|
613 |
+
# 1 x H x W x D -> 1 x L xD
|
614 |
+
pos_emb = pos_emb.view(1, self.height * self.width, -1)
|
615 |
+
|
616 |
+
emb = emb + pos_emb[:, : emb.shape[1], :]
|
617 |
+
|
618 |
+
return emb
|
619 |
+
|
620 |
+
|
621 |
+
class LabelEmbedding(nn.Module):
|
622 |
+
"""
|
623 |
+
Embeds class labels into vector representations. Also handles label dropout for classifier-free guidance.
|
624 |
+
|
625 |
+
Args:
|
626 |
+
num_classes (`int`): The number of classes.
|
627 |
+
hidden_size (`int`): The size of the vector embeddings.
|
628 |
+
dropout_prob (`float`): The probability of dropping a label.
|
629 |
+
"""
|
630 |
+
|
631 |
+
def __init__(self, num_classes, hidden_size, dropout_prob):
|
632 |
+
super().__init__()
|
633 |
+
use_cfg_embedding = dropout_prob > 0
|
634 |
+
self.embedding_table = nn.Embedding(
|
635 |
+
num_classes + use_cfg_embedding, hidden_size
|
636 |
+
)
|
637 |
+
self.num_classes = num_classes
|
638 |
+
self.dropout_prob = dropout_prob
|
639 |
+
|
640 |
+
def token_drop(self, labels, force_drop_ids=None):
|
641 |
+
"""
|
642 |
+
Drops labels to enable classifier-free guidance.
|
643 |
+
"""
|
644 |
+
if force_drop_ids is None:
|
645 |
+
drop_ids = (
|
646 |
+
torch.rand(labels.shape[0], device=labels.device) < self.dropout_prob
|
647 |
+
)
|
648 |
+
else:
|
649 |
+
drop_ids = torch.tensor(force_drop_ids == 1)
|
650 |
+
labels = torch.where(drop_ids, self.num_classes, labels)
|
651 |
+
return labels
|
652 |
+
|
653 |
+
def forward(self, labels: torch.LongTensor, force_drop_ids=None):
|
654 |
+
use_dropout = self.dropout_prob > 0
|
655 |
+
if (self.training and use_dropout) or (force_drop_ids is not None):
|
656 |
+
labels = self.token_drop(labels, force_drop_ids)
|
657 |
+
embeddings = self.embedding_table(labels)
|
658 |
+
return embeddings
|
659 |
+
|
660 |
+
|
661 |
+
class TextImageProjection(nn.Module):
|
662 |
+
def __init__(
|
663 |
+
self,
|
664 |
+
text_embed_dim: int = 1024,
|
665 |
+
image_embed_dim: int = 768,
|
666 |
+
cross_attention_dim: int = 768,
|
667 |
+
num_image_text_embeds: int = 10,
|
668 |
+
):
|
669 |
+
super().__init__()
|
670 |
+
|
671 |
+
self.num_image_text_embeds = num_image_text_embeds
|
672 |
+
self.image_embeds = nn.Linear(
|
673 |
+
image_embed_dim, self.num_image_text_embeds * cross_attention_dim
|
674 |
+
)
|
675 |
+
self.text_proj = nn.Linear(text_embed_dim, cross_attention_dim)
|
676 |
+
|
677 |
+
def forward(self, text_embeds: torch.Tensor, image_embeds: torch.Tensor):
|
678 |
+
batch_size = text_embeds.shape[0]
|
679 |
+
|
680 |
+
# image
|
681 |
+
image_text_embeds = self.image_embeds(image_embeds)
|
682 |
+
image_text_embeds = image_text_embeds.reshape(
|
683 |
+
batch_size, self.num_image_text_embeds, -1
|
684 |
+
)
|
685 |
+
|
686 |
+
# text
|
687 |
+
text_embeds = self.text_proj(text_embeds)
|
688 |
+
|
689 |
+
return torch.cat([image_text_embeds, text_embeds], dim=1)
|
690 |
+
|
691 |
+
|
692 |
+
class ImageProjection(nn.Module):
|
693 |
+
def __init__(
|
694 |
+
self,
|
695 |
+
image_embed_dim: int = 768,
|
696 |
+
cross_attention_dim: int = 768,
|
697 |
+
num_image_text_embeds: int = 32,
|
698 |
+
):
|
699 |
+
super().__init__()
|
700 |
+
|
701 |
+
self.num_image_text_embeds = num_image_text_embeds
|
702 |
+
self.image_embeds = nn.Linear(
|
703 |
+
image_embed_dim, self.num_image_text_embeds * cross_attention_dim
|
704 |
+
)
|
705 |
+
self.norm = nn.LayerNorm(cross_attention_dim)
|
706 |
+
|
707 |
+
def forward(self, image_embeds: torch.Tensor):
|
708 |
+
batch_size = image_embeds.shape[0]
|
709 |
+
|
710 |
+
# image
|
711 |
+
image_embeds = self.image_embeds(image_embeds)
|
712 |
+
image_embeds = image_embeds.reshape(batch_size, self.num_image_text_embeds, -1)
|
713 |
+
image_embeds = self.norm(image_embeds)
|
714 |
+
return image_embeds
|
715 |
+
|
716 |
+
|
717 |
+
class IPAdapterFullImageProjection(nn.Module):
|
718 |
+
def __init__(self, image_embed_dim=1024, cross_attention_dim=1024):
|
719 |
+
super().__init__()
|
720 |
+
from .attention import FeedForward
|
721 |
+
|
722 |
+
self.ff = FeedForward(
|
723 |
+
image_embed_dim, cross_attention_dim, mult=1, activation_fn="gelu"
|
724 |
+
)
|
725 |
+
self.norm = nn.LayerNorm(cross_attention_dim)
|
726 |
+
|
727 |
+
def forward(self, image_embeds: torch.Tensor):
|
728 |
+
return self.norm(self.ff(image_embeds))
|
729 |
+
|
730 |
+
|
731 |
+
class IPAdapterFaceIDImageProjection(nn.Module):
|
732 |
+
def __init__(
|
733 |
+
self, image_embed_dim=1024, cross_attention_dim=1024, mult=1, num_tokens=1
|
734 |
+
):
|
735 |
+
super().__init__()
|
736 |
+
from .attention import FeedForward
|
737 |
+
|
738 |
+
self.num_tokens = num_tokens
|
739 |
+
self.cross_attention_dim = cross_attention_dim
|
740 |
+
self.ff = FeedForward(
|
741 |
+
image_embed_dim,
|
742 |
+
cross_attention_dim * num_tokens,
|
743 |
+
mult=mult,
|
744 |
+
activation_fn="gelu",
|
745 |
+
)
|
746 |
+
self.norm = nn.LayerNorm(cross_attention_dim)
|
747 |
+
|
748 |
+
def forward(self, image_embeds: torch.Tensor):
|
749 |
+
x = self.ff(image_embeds)
|
750 |
+
x = x.reshape(-1, self.num_tokens, self.cross_attention_dim)
|
751 |
+
return self.norm(x)
|
752 |
+
|
753 |
+
|
754 |
+
class CombinedTimestepLabelEmbeddings(nn.Module):
|
755 |
+
def __init__(self, num_classes, embedding_dim, class_dropout_prob=0.1):
|
756 |
+
super().__init__()
|
757 |
+
|
758 |
+
self.time_proj = Timesteps(
|
759 |
+
num_channels=256, flip_sin_to_cos=True, downscale_freq_shift=1
|
760 |
+
)
|
761 |
+
self.timestep_embedder = TimestepEmbedding(
|
762 |
+
in_channels=256, time_embed_dim=embedding_dim
|
763 |
+
)
|
764 |
+
self.class_embedder = LabelEmbedding(
|
765 |
+
num_classes, embedding_dim, class_dropout_prob
|
766 |
+
)
|
767 |
+
|
768 |
+
def forward(self, timestep, class_labels, hidden_dtype=None):
|
769 |
+
timesteps_proj = self.time_proj(timestep)
|
770 |
+
timesteps_emb = self.timestep_embedder(
|
771 |
+
timesteps_proj.to(dtype=hidden_dtype)
|
772 |
+
) # (N, D)
|
773 |
+
|
774 |
+
class_labels = self.class_embedder(class_labels) # (N, D)
|
775 |
+
|
776 |
+
conditioning = timesteps_emb + class_labels # (N, D)
|
777 |
+
|
778 |
+
return conditioning
|
779 |
+
|
780 |
+
|
781 |
+
class CombinedTimestepTextProjEmbeddings(nn.Module):
|
782 |
+
def __init__(self, embedding_dim, pooled_projection_dim):
|
783 |
+
super().__init__()
|
784 |
+
|
785 |
+
self.time_proj = Timesteps(
|
786 |
+
num_channels=256, flip_sin_to_cos=True, downscale_freq_shift=0
|
787 |
+
)
|
788 |
+
self.timestep_embedder = TimestepEmbedding(
|
789 |
+
in_channels=256, time_embed_dim=embedding_dim
|
790 |
+
)
|
791 |
+
self.text_embedder = PixArtAlphaTextProjection(
|
792 |
+
pooled_projection_dim, embedding_dim, act_fn="silu"
|
793 |
+
)
|
794 |
+
|
795 |
+
def forward(self, timestep, pooled_projection):
|
796 |
+
timesteps_proj = self.time_proj(timestep)
|
797 |
+
timesteps_emb = self.timestep_embedder(
|
798 |
+
timesteps_proj.to(dtype=pooled_projection.dtype)
|
799 |
+
) # (N, D)
|
800 |
+
|
801 |
+
pooled_projections = self.text_embedder(pooled_projection)
|
802 |
+
|
803 |
+
conditioning = timesteps_emb + pooled_projections
|
804 |
+
|
805 |
+
return conditioning
|
806 |
+
|
807 |
+
|
808 |
+
class TimestepEmbeddings(nn.Module):
|
809 |
+
def __init__(self, embedding_dim):
|
810 |
+
super().__init__()
|
811 |
+
|
812 |
+
self.time_proj = Timesteps(
|
813 |
+
num_channels=256, flip_sin_to_cos=True, downscale_freq_shift=0
|
814 |
+
)
|
815 |
+
self.timestep_embedder = TimestepEmbedding(
|
816 |
+
in_channels=256, time_embed_dim=embedding_dim
|
817 |
+
)
|
818 |
+
|
819 |
+
def forward(self, timestep):
|
820 |
+
timesteps_proj = self.time_proj(timestep)
|
821 |
+
timesteps_emb = self.timestep_embedder(timesteps_proj) # (N, D)
|
822 |
+
|
823 |
+
conditioning = timesteps_emb
|
824 |
+
|
825 |
+
return conditioning
|
826 |
+
|
827 |
+
|
828 |
+
class HunyuanDiTAttentionPool(nn.Module):
|
829 |
+
# Copied from https://github.com/Tencent/HunyuanDiT/blob/cb709308d92e6c7e8d59d0dff41b74d35088db6a/hydit/modules/poolers.py#L6
|
830 |
+
|
831 |
+
def __init__(
|
832 |
+
self, spacial_dim: int, embed_dim: int, num_heads: int, output_dim: int = None
|
833 |
+
):
|
834 |
+
super().__init__()
|
835 |
+
self.positional_embedding = nn.Parameter(
|
836 |
+
torch.randn(spacial_dim + 1, embed_dim) / embed_dim**0.5
|
837 |
+
)
|
838 |
+
self.k_proj = nn.Linear(embed_dim, embed_dim)
|
839 |
+
self.q_proj = nn.Linear(embed_dim, embed_dim)
|
840 |
+
self.v_proj = nn.Linear(embed_dim, embed_dim)
|
841 |
+
self.c_proj = nn.Linear(embed_dim, output_dim or embed_dim)
|
842 |
+
self.num_heads = num_heads
|
843 |
+
|
844 |
+
def forward(self, x):
|
845 |
+
x = x.permute(1, 0, 2) # NLC -> LNC
|
846 |
+
x = torch.cat([x.mean(dim=0, keepdim=True), x], dim=0) # (L+1)NC
|
847 |
+
x = x + self.positional_embedding[:, None, :].to(x.dtype) # (L+1)NC
|
848 |
+
x, _ = F.multi_head_attention_forward(
|
849 |
+
query=x[:1],
|
850 |
+
key=x,
|
851 |
+
value=x,
|
852 |
+
embed_dim_to_check=x.shape[-1],
|
853 |
+
num_heads=self.num_heads,
|
854 |
+
q_proj_weight=self.q_proj.weight,
|
855 |
+
k_proj_weight=self.k_proj.weight,
|
856 |
+
v_proj_weight=self.v_proj.weight,
|
857 |
+
in_proj_weight=None,
|
858 |
+
in_proj_bias=torch.cat(
|
859 |
+
[self.q_proj.bias, self.k_proj.bias, self.v_proj.bias]
|
860 |
+
),
|
861 |
+
bias_k=None,
|
862 |
+
bias_v=None,
|
863 |
+
add_zero_attn=False,
|
864 |
+
dropout_p=0,
|
865 |
+
out_proj_weight=self.c_proj.weight,
|
866 |
+
out_proj_bias=self.c_proj.bias,
|
867 |
+
use_separate_proj_weight=True,
|
868 |
+
training=self.training,
|
869 |
+
need_weights=False,
|
870 |
+
)
|
871 |
+
return x.squeeze(0)
|
872 |
+
|
873 |
+
|
874 |
+
class HunyuanCombinedTimestepTextSizeStyleEmbedding(nn.Module):
|
875 |
+
def __init__(
|
876 |
+
self,
|
877 |
+
embedding_dim,
|
878 |
+
pooled_projection_dim=1024,
|
879 |
+
seq_len=256,
|
880 |
+
cross_attention_dim=2048,
|
881 |
+
):
|
882 |
+
super().__init__()
|
883 |
+
|
884 |
+
self.time_proj = Timesteps(
|
885 |
+
num_channels=256, flip_sin_to_cos=True, downscale_freq_shift=0
|
886 |
+
)
|
887 |
+
self.timestep_embedder = TimestepEmbedding(
|
888 |
+
in_channels=256, time_embed_dim=embedding_dim
|
889 |
+
)
|
890 |
+
|
891 |
+
self.pooler = HunyuanDiTAttentionPool(
|
892 |
+
seq_len, cross_attention_dim, num_heads=8, output_dim=pooled_projection_dim
|
893 |
+
)
|
894 |
+
# Here we use a default learned embedder layer for future extension.
|
895 |
+
self.style_embedder = nn.Embedding(1, embedding_dim)
|
896 |
+
extra_in_dim = 256 * 6 + embedding_dim + pooled_projection_dim
|
897 |
+
self.extra_embedder = PixArtAlphaTextProjection(
|
898 |
+
in_features=extra_in_dim,
|
899 |
+
hidden_size=embedding_dim * 4,
|
900 |
+
out_features=embedding_dim,
|
901 |
+
act_fn="silu_fp32",
|
902 |
+
)
|
903 |
+
|
904 |
+
def forward(
|
905 |
+
self, timestep, encoder_hidden_states, image_meta_size, style, hidden_dtype=None
|
906 |
+
):
|
907 |
+
timesteps_proj = self.time_proj(timestep)
|
908 |
+
timesteps_emb = self.timestep_embedder(
|
909 |
+
timesteps_proj.to(dtype=hidden_dtype)
|
910 |
+
) # (N, 256)
|
911 |
+
|
912 |
+
# extra condition1: text
|
913 |
+
pooled_projections = self.pooler(encoder_hidden_states) # (N, 1024)
|
914 |
+
|
915 |
+
# extra condition2: image meta size embdding
|
916 |
+
image_meta_size = get_timestep_embedding(image_meta_size.view(-1), 256, True, 0)
|
917 |
+
image_meta_size = image_meta_size.to(dtype=hidden_dtype)
|
918 |
+
image_meta_size = image_meta_size.view(-1, 6 * 256) # (N, 1536)
|
919 |
+
|
920 |
+
# extra condition3: style embedding
|
921 |
+
style_embedding = self.style_embedder(style) # (N, embedding_dim)
|
922 |
+
|
923 |
+
# Concatenate all extra vectors
|
924 |
+
extra_cond = torch.cat(
|
925 |
+
[pooled_projections, image_meta_size, style_embedding], dim=1
|
926 |
+
)
|
927 |
+
conditioning = timesteps_emb + self.extra_embedder(extra_cond) # [B, D]
|
928 |
+
|
929 |
+
return conditioning
|
930 |
+
|
931 |
+
|
932 |
+
class TextTimeEmbedding(nn.Module):
|
933 |
+
def __init__(self, encoder_dim: int, time_embed_dim: int, num_heads: int = 64):
|
934 |
+
super().__init__()
|
935 |
+
self.norm1 = nn.LayerNorm(encoder_dim)
|
936 |
+
self.pool = AttentionPooling(num_heads, encoder_dim)
|
937 |
+
self.proj = nn.Linear(encoder_dim, time_embed_dim)
|
938 |
+
self.norm2 = nn.LayerNorm(time_embed_dim)
|
939 |
+
|
940 |
+
def forward(self, hidden_states):
|
941 |
+
hidden_states = self.norm1(hidden_states)
|
942 |
+
hidden_states = self.pool(hidden_states)
|
943 |
+
hidden_states = self.proj(hidden_states)
|
944 |
+
hidden_states = self.norm2(hidden_states)
|
945 |
+
return hidden_states
|
946 |
+
|
947 |
+
|
948 |
+
class TextImageTimeEmbedding(nn.Module):
|
949 |
+
def __init__(
|
950 |
+
self,
|
951 |
+
text_embed_dim: int = 768,
|
952 |
+
image_embed_dim: int = 768,
|
953 |
+
time_embed_dim: int = 1536,
|
954 |
+
):
|
955 |
+
super().__init__()
|
956 |
+
self.text_proj = nn.Linear(text_embed_dim, time_embed_dim)
|
957 |
+
self.text_norm = nn.LayerNorm(time_embed_dim)
|
958 |
+
self.image_proj = nn.Linear(image_embed_dim, time_embed_dim)
|
959 |
+
|
960 |
+
def forward(self, text_embeds: torch.Tensor, image_embeds: torch.Tensor):
|
961 |
+
# text
|
962 |
+
time_text_embeds = self.text_proj(text_embeds)
|
963 |
+
time_text_embeds = self.text_norm(time_text_embeds)
|
964 |
+
|
965 |
+
# image
|
966 |
+
time_image_embeds = self.image_proj(image_embeds)
|
967 |
+
|
968 |
+
return time_image_embeds + time_text_embeds
|
969 |
+
|
970 |
+
|
971 |
+
class ImageTimeEmbedding(nn.Module):
|
972 |
+
def __init__(self, image_embed_dim: int = 768, time_embed_dim: int = 1536):
|
973 |
+
super().__init__()
|
974 |
+
self.image_proj = nn.Linear(image_embed_dim, time_embed_dim)
|
975 |
+
self.image_norm = nn.LayerNorm(time_embed_dim)
|
976 |
+
|
977 |
+
def forward(self, image_embeds: torch.Tensor):
|
978 |
+
# image
|
979 |
+
time_image_embeds = self.image_proj(image_embeds)
|
980 |
+
time_image_embeds = self.image_norm(time_image_embeds)
|
981 |
+
return time_image_embeds
|
982 |
+
|
983 |
+
|
984 |
+
class ImageHintTimeEmbedding(nn.Module):
|
985 |
+
def __init__(self, image_embed_dim: int = 768, time_embed_dim: int = 1536):
|
986 |
+
super().__init__()
|
987 |
+
self.image_proj = nn.Linear(image_embed_dim, time_embed_dim)
|
988 |
+
self.image_norm = nn.LayerNorm(time_embed_dim)
|
989 |
+
self.input_hint_block = nn.Sequential(
|
990 |
+
nn.Conv2d(3, 16, 3, padding=1),
|
991 |
+
nn.SiLU(),
|
992 |
+
nn.Conv2d(16, 16, 3, padding=1),
|
993 |
+
nn.SiLU(),
|
994 |
+
nn.Conv2d(16, 32, 3, padding=1, stride=2),
|
995 |
+
nn.SiLU(),
|
996 |
+
nn.Conv2d(32, 32, 3, padding=1),
|
997 |
+
nn.SiLU(),
|
998 |
+
nn.Conv2d(32, 96, 3, padding=1, stride=2),
|
999 |
+
nn.SiLU(),
|
1000 |
+
nn.Conv2d(96, 96, 3, padding=1),
|
1001 |
+
nn.SiLU(),
|
1002 |
+
nn.Conv2d(96, 256, 3, padding=1, stride=2),
|
1003 |
+
nn.SiLU(),
|
1004 |
+
nn.Conv2d(256, 4, 3, padding=1),
|
1005 |
+
)
|
1006 |
+
|
1007 |
+
def forward(self, image_embeds: torch.Tensor, hint: torch.Tensor):
|
1008 |
+
# image
|
1009 |
+
time_image_embeds = self.image_proj(image_embeds)
|
1010 |
+
time_image_embeds = self.image_norm(time_image_embeds)
|
1011 |
+
hint = self.input_hint_block(hint)
|
1012 |
+
return time_image_embeds, hint
|
1013 |
+
|
1014 |
+
|
1015 |
+
class AttentionPooling(nn.Module):
|
1016 |
+
# Copied from https://github.com/deep-floyd/IF/blob/2f91391f27dd3c468bf174be5805b4cc92980c0b/deepfloyd_if/model/nn.py#L54
|
1017 |
+
|
1018 |
+
def __init__(self, num_heads, embed_dim, dtype=None):
|
1019 |
+
super().__init__()
|
1020 |
+
self.dtype = dtype
|
1021 |
+
self.positional_embedding = nn.Parameter(
|
1022 |
+
torch.randn(1, embed_dim) / embed_dim**0.5
|
1023 |
+
)
|
1024 |
+
self.k_proj = nn.Linear(embed_dim, embed_dim, dtype=self.dtype)
|
1025 |
+
self.q_proj = nn.Linear(embed_dim, embed_dim, dtype=self.dtype)
|
1026 |
+
self.v_proj = nn.Linear(embed_dim, embed_dim, dtype=self.dtype)
|
1027 |
+
self.num_heads = num_heads
|
1028 |
+
self.dim_per_head = embed_dim // self.num_heads
|
1029 |
+
|
1030 |
+
def forward(self, x):
|
1031 |
+
bs, length, width = x.size()
|
1032 |
+
|
1033 |
+
def shape(x):
|
1034 |
+
# (bs, length, width) --> (bs, length, n_heads, dim_per_head)
|
1035 |
+
x = x.view(bs, -1, self.num_heads, self.dim_per_head)
|
1036 |
+
# (bs, length, n_heads, dim_per_head) --> (bs, n_heads, length, dim_per_head)
|
1037 |
+
x = x.transpose(1, 2)
|
1038 |
+
# (bs, n_heads, length, dim_per_head) --> (bs*n_heads, length, dim_per_head)
|
1039 |
+
x = x.reshape(bs * self.num_heads, -1, self.dim_per_head)
|
1040 |
+
# (bs*n_heads, length, dim_per_head) --> (bs*n_heads, dim_per_head, length)
|
1041 |
+
x = x.transpose(1, 2)
|
1042 |
+
return x
|
1043 |
+
|
1044 |
+
class_token = x.mean(dim=1, keepdim=True) + self.positional_embedding.to(
|
1045 |
+
x.dtype
|
1046 |
+
)
|
1047 |
+
x = torch.cat([class_token, x], dim=1) # (bs, length+1, width)
|
1048 |
+
|
1049 |
+
# (bs*n_heads, class_token_length, dim_per_head)
|
1050 |
+
q = shape(self.q_proj(class_token))
|
1051 |
+
# (bs*n_heads, length+class_token_length, dim_per_head)
|
1052 |
+
k = shape(self.k_proj(x))
|
1053 |
+
v = shape(self.v_proj(x))
|
1054 |
+
|
1055 |
+
# (bs*n_heads, class_token_length, length+class_token_length):
|
1056 |
+
scale = 1 / math.sqrt(math.sqrt(self.dim_per_head))
|
1057 |
+
weight = torch.einsum(
|
1058 |
+
"bct,bcs->bts", q * scale, k * scale
|
1059 |
+
) # More stable with f16 than dividing afterwards
|
1060 |
+
weight = torch.softmax(weight.float(), dim=-1).type(weight.dtype)
|
1061 |
+
|
1062 |
+
# (bs*n_heads, dim_per_head, class_token_length)
|
1063 |
+
a = torch.einsum("bts,bcs->bct", weight, v)
|
1064 |
+
|
1065 |
+
# (bs, length+1, width)
|
1066 |
+
a = a.reshape(bs, -1, 1).transpose(1, 2)
|
1067 |
+
|
1068 |
+
return a[:, 0, :] # cls_token
|
1069 |
+
|
1070 |
+
|
1071 |
+
def get_fourier_embeds_from_boundingbox(embed_dim, box):
|
1072 |
+
"""
|
1073 |
+
Args:
|
1074 |
+
embed_dim: int
|
1075 |
+
box: a 3-D tensor [B x N x 4] representing the bounding boxes for GLIGEN pipeline
|
1076 |
+
Returns:
|
1077 |
+
[B x N x embed_dim] tensor of positional embeddings
|
1078 |
+
"""
|
1079 |
+
|
1080 |
+
batch_size, num_boxes = box.shape[:2]
|
1081 |
+
|
1082 |
+
emb = 100 ** (torch.arange(embed_dim) / embed_dim)
|
1083 |
+
emb = emb[None, None, None].to(device=box.device, dtype=box.dtype)
|
1084 |
+
emb = emb * box.unsqueeze(-1)
|
1085 |
+
|
1086 |
+
emb = torch.stack((emb.sin(), emb.cos()), dim=-1)
|
1087 |
+
emb = emb.permute(0, 1, 3, 4, 2).reshape(batch_size, num_boxes, embed_dim * 2 * 4)
|
1088 |
+
|
1089 |
+
return emb
|
1090 |
+
|
1091 |
+
|
1092 |
+
class GLIGENTextBoundingboxProjection(nn.Module):
|
1093 |
+
def __init__(
|
1094 |
+
self, positive_len, out_dim, feature_type="text-only", fourier_freqs=8
|
1095 |
+
):
|
1096 |
+
super().__init__()
|
1097 |
+
self.positive_len = positive_len
|
1098 |
+
self.out_dim = out_dim
|
1099 |
+
|
1100 |
+
self.fourier_embedder_dim = fourier_freqs
|
1101 |
+
self.position_dim = fourier_freqs * 2 * 4 # 2: sin/cos, 4: xyxy
|
1102 |
+
|
1103 |
+
if isinstance(out_dim, tuple):
|
1104 |
+
out_dim = out_dim[0]
|
1105 |
+
|
1106 |
+
if feature_type == "text-only":
|
1107 |
+
self.linears = nn.Sequential(
|
1108 |
+
nn.Linear(self.positive_len + self.position_dim, 512),
|
1109 |
+
nn.SiLU(),
|
1110 |
+
nn.Linear(512, 512),
|
1111 |
+
nn.SiLU(),
|
1112 |
+
nn.Linear(512, out_dim),
|
1113 |
+
)
|
1114 |
+
self.null_positive_feature = torch.nn.Parameter(
|
1115 |
+
torch.zeros([self.positive_len])
|
1116 |
+
)
|
1117 |
+
|
1118 |
+
elif feature_type == "text-image":
|
1119 |
+
self.linears_text = nn.Sequential(
|
1120 |
+
nn.Linear(self.positive_len + self.position_dim, 512),
|
1121 |
+
nn.SiLU(),
|
1122 |
+
nn.Linear(512, 512),
|
1123 |
+
nn.SiLU(),
|
1124 |
+
nn.Linear(512, out_dim),
|
1125 |
+
)
|
1126 |
+
self.linears_image = nn.Sequential(
|
1127 |
+
nn.Linear(self.positive_len + self.position_dim, 512),
|
1128 |
+
nn.SiLU(),
|
1129 |
+
nn.Linear(512, 512),
|
1130 |
+
nn.SiLU(),
|
1131 |
+
nn.Linear(512, out_dim),
|
1132 |
+
)
|
1133 |
+
self.null_text_feature = torch.nn.Parameter(
|
1134 |
+
torch.zeros([self.positive_len])
|
1135 |
+
)
|
1136 |
+
self.null_image_feature = torch.nn.Parameter(
|
1137 |
+
torch.zeros([self.positive_len])
|
1138 |
+
)
|
1139 |
+
|
1140 |
+
self.null_position_feature = torch.nn.Parameter(
|
1141 |
+
torch.zeros([self.position_dim])
|
1142 |
+
)
|
1143 |
+
|
1144 |
+
def forward(
|
1145 |
+
self,
|
1146 |
+
boxes,
|
1147 |
+
masks,
|
1148 |
+
positive_embeddings=None,
|
1149 |
+
phrases_masks=None,
|
1150 |
+
image_masks=None,
|
1151 |
+
phrases_embeddings=None,
|
1152 |
+
image_embeddings=None,
|
1153 |
+
):
|
1154 |
+
masks = masks.unsqueeze(-1)
|
1155 |
+
|
1156 |
+
# embedding position (it may includes padding as placeholder)
|
1157 |
+
xyxy_embedding = get_fourier_embeds_from_boundingbox(
|
1158 |
+
self.fourier_embedder_dim, boxes
|
1159 |
+
) # B*N*4 -> B*N*C
|
1160 |
+
|
1161 |
+
# learnable null embedding
|
1162 |
+
xyxy_null = self.null_position_feature.view(1, 1, -1)
|
1163 |
+
|
1164 |
+
# replace padding with learnable null embedding
|
1165 |
+
xyxy_embedding = xyxy_embedding * masks + (1 - masks) * xyxy_null
|
1166 |
+
|
1167 |
+
# positionet with text only information
|
1168 |
+
if positive_embeddings is not None:
|
1169 |
+
# learnable null embedding
|
1170 |
+
positive_null = self.null_positive_feature.view(1, 1, -1)
|
1171 |
+
|
1172 |
+
# replace padding with learnable null embedding
|
1173 |
+
positive_embeddings = (
|
1174 |
+
positive_embeddings * masks + (1 - masks) * positive_null
|
1175 |
+
)
|
1176 |
+
|
1177 |
+
objs = self.linears(
|
1178 |
+
torch.cat([positive_embeddings, xyxy_embedding], dim=-1)
|
1179 |
+
)
|
1180 |
+
|
1181 |
+
# positionet with text and image infomation
|
1182 |
+
else:
|
1183 |
+
phrases_masks = phrases_masks.unsqueeze(-1)
|
1184 |
+
image_masks = image_masks.unsqueeze(-1)
|
1185 |
+
|
1186 |
+
# learnable null embedding
|
1187 |
+
text_null = self.null_text_feature.view(1, 1, -1)
|
1188 |
+
image_null = self.null_image_feature.view(1, 1, -1)
|
1189 |
+
|
1190 |
+
# replace padding with learnable null embedding
|
1191 |
+
phrases_embeddings = (
|
1192 |
+
phrases_embeddings * phrases_masks + (1 - phrases_masks) * text_null
|
1193 |
+
)
|
1194 |
+
image_embeddings = (
|
1195 |
+
image_embeddings * image_masks + (1 - image_masks) * image_null
|
1196 |
+
)
|
1197 |
+
|
1198 |
+
objs_text = self.linears_text(
|
1199 |
+
torch.cat([phrases_embeddings, xyxy_embedding], dim=-1)
|
1200 |
+
)
|
1201 |
+
objs_image = self.linears_image(
|
1202 |
+
torch.cat([image_embeddings, xyxy_embedding], dim=-1)
|
1203 |
+
)
|
1204 |
+
objs = torch.cat([objs_text, objs_image], dim=1)
|
1205 |
+
|
1206 |
+
return objs
|
1207 |
+
|
1208 |
+
|
1209 |
+
class PixArtAlphaCombinedTimestepSizeEmbeddings(nn.Module):
|
1210 |
+
"""
|
1211 |
+
For PixArt-Alpha.
|
1212 |
+
|
1213 |
+
Reference:
|
1214 |
+
https://github.com/PixArt-alpha/PixArt-alpha/blob/0f55e922376d8b797edd44d25d0e7464b260dcab/diffusion/model/nets/PixArtMS.py#L164C9-L168C29
|
1215 |
+
"""
|
1216 |
+
|
1217 |
+
def __init__(
|
1218 |
+
self, embedding_dim, size_emb_dim, use_additional_conditions: bool = False
|
1219 |
+
):
|
1220 |
+
super().__init__()
|
1221 |
+
|
1222 |
+
self.outdim = size_emb_dim
|
1223 |
+
self.time_proj = Timesteps(
|
1224 |
+
num_channels=256, flip_sin_to_cos=True, downscale_freq_shift=0
|
1225 |
+
)
|
1226 |
+
self.timestep_embedder = TimestepEmbedding(
|
1227 |
+
in_channels=256, time_embed_dim=embedding_dim
|
1228 |
+
)
|
1229 |
+
|
1230 |
+
self.use_additional_conditions = use_additional_conditions
|
1231 |
+
if use_additional_conditions:
|
1232 |
+
self.additional_condition_proj = Timesteps(
|
1233 |
+
num_channels=256, flip_sin_to_cos=True, downscale_freq_shift=0
|
1234 |
+
)
|
1235 |
+
self.resolution_embedder = TimestepEmbedding(
|
1236 |
+
in_channels=256, time_embed_dim=size_emb_dim
|
1237 |
+
)
|
1238 |
+
self.aspect_ratio_embedder = TimestepEmbedding(
|
1239 |
+
in_channels=256, time_embed_dim=size_emb_dim
|
1240 |
+
)
|
1241 |
+
|
1242 |
+
def forward(self, timestep, resolution, aspect_ratio, batch_size, hidden_dtype):
|
1243 |
+
timesteps_proj = self.time_proj(timestep)
|
1244 |
+
timesteps_emb = self.timestep_embedder(
|
1245 |
+
timesteps_proj.to(dtype=hidden_dtype)
|
1246 |
+
) # (N, D)
|
1247 |
+
|
1248 |
+
if self.use_additional_conditions:
|
1249 |
+
resolution_emb = self.additional_condition_proj(resolution.flatten()).to(
|
1250 |
+
hidden_dtype
|
1251 |
+
)
|
1252 |
+
resolution_emb = self.resolution_embedder(resolution_emb).reshape(
|
1253 |
+
batch_size, -1
|
1254 |
+
)
|
1255 |
+
aspect_ratio_emb = self.additional_condition_proj(
|
1256 |
+
aspect_ratio.flatten()
|
1257 |
+
).to(hidden_dtype)
|
1258 |
+
aspect_ratio_emb = self.aspect_ratio_embedder(aspect_ratio_emb).reshape(
|
1259 |
+
batch_size, -1
|
1260 |
+
)
|
1261 |
+
conditioning = timesteps_emb + torch.cat(
|
1262 |
+
[resolution_emb, aspect_ratio_emb], dim=1
|
1263 |
+
)
|
1264 |
+
else:
|
1265 |
+
conditioning = timesteps_emb
|
1266 |
+
|
1267 |
+
return conditioning
|
1268 |
+
|
1269 |
+
|
1270 |
+
class PixArtAlphaTextProjection(nn.Module):
|
1271 |
+
"""
|
1272 |
+
Projects caption embeddings. Also handles dropout for classifier-free guidance.
|
1273 |
+
|
1274 |
+
Adapted from https://github.com/PixArt-alpha/PixArt-alpha/blob/master/diffusion/model/nets/PixArt_blocks.py
|
1275 |
+
"""
|
1276 |
+
|
1277 |
+
def __init__(self, in_features, hidden_size, out_features=None, act_fn="gelu_tanh"):
|
1278 |
+
super().__init__()
|
1279 |
+
if out_features is None:
|
1280 |
+
out_features = hidden_size
|
1281 |
+
self.linear_1 = nn.Linear(
|
1282 |
+
in_features=in_features, out_features=hidden_size, bias=True
|
1283 |
+
)
|
1284 |
+
if act_fn == "gelu_tanh":
|
1285 |
+
self.act_1 = nn.GELU(approximate="tanh")
|
1286 |
+
elif act_fn == "silu":
|
1287 |
+
self.act_1 = nn.SiLU()
|
1288 |
+
elif act_fn == "silu_fp32":
|
1289 |
+
self.act_1 = FP32SiLU()
|
1290 |
+
else:
|
1291 |
+
raise ValueError(f"Unknown activation function: {act_fn}")
|
1292 |
+
self.linear_2 = nn.Linear(
|
1293 |
+
in_features=hidden_size, out_features=out_features, bias=True
|
1294 |
+
)
|
1295 |
+
|
1296 |
+
def forward(self, caption):
|
1297 |
+
hidden_states = self.linear_1(caption)
|
1298 |
+
hidden_states = self.act_1(hidden_states)
|
1299 |
+
hidden_states = self.linear_2(hidden_states)
|
1300 |
+
return hidden_states
|
1301 |
+
|
1302 |
+
|
1303 |
+
class IPAdapterPlusImageProjectionBlock(nn.Module):
|
1304 |
+
def __init__(
|
1305 |
+
self,
|
1306 |
+
embed_dims: int = 768,
|
1307 |
+
dim_head: int = 64,
|
1308 |
+
heads: int = 16,
|
1309 |
+
ffn_ratio: float = 4,
|
1310 |
+
) -> None:
|
1311 |
+
super().__init__()
|
1312 |
+
from .attention import FeedForward
|
1313 |
+
|
1314 |
+
self.ln0 = nn.LayerNorm(embed_dims)
|
1315 |
+
self.ln1 = nn.LayerNorm(embed_dims)
|
1316 |
+
self.attn = Attention(
|
1317 |
+
query_dim=embed_dims,
|
1318 |
+
dim_head=dim_head,
|
1319 |
+
heads=heads,
|
1320 |
+
out_bias=False,
|
1321 |
+
)
|
1322 |
+
self.ff = nn.Sequential(
|
1323 |
+
nn.LayerNorm(embed_dims),
|
1324 |
+
FeedForward(
|
1325 |
+
embed_dims, embed_dims, activation_fn="gelu", mult=ffn_ratio, bias=False
|
1326 |
+
),
|
1327 |
+
)
|
1328 |
+
|
1329 |
+
def forward(self, x, latents, residual):
|
1330 |
+
encoder_hidden_states = self.ln0(x)
|
1331 |
+
latents = self.ln1(latents)
|
1332 |
+
encoder_hidden_states = torch.cat([encoder_hidden_states, latents], dim=-2)
|
1333 |
+
latents = self.attn(latents, encoder_hidden_states) + residual
|
1334 |
+
latents = self.ff(latents) + latents
|
1335 |
+
return latents
|
1336 |
+
|
1337 |
+
|
1338 |
+
class IPAdapterPlusImageProjection(nn.Module):
|
1339 |
+
"""Resampler of IP-Adapter Plus.
|
1340 |
+
|
1341 |
+
Args:
|
1342 |
+
embed_dims (int): The feature dimension. Defaults to 768. output_dims (int): The number of output channels,
|
1343 |
+
that is the same
|
1344 |
+
number of the channels in the `unet.config.cross_attention_dim`. Defaults to 1024.
|
1345 |
+
hidden_dims (int):
|
1346 |
+
The number of hidden channels. Defaults to 1280. depth (int): The number of blocks. Defaults
|
1347 |
+
to 8. dim_head (int): The number of head channels. Defaults to 64. heads (int): Parallel attention heads.
|
1348 |
+
Defaults to 16. num_queries (int):
|
1349 |
+
The number of queries. Defaults to 8. ffn_ratio (float): The expansion ratio
|
1350 |
+
of feedforward network hidden
|
1351 |
+
layer channels. Defaults to 4.
|
1352 |
+
"""
|
1353 |
+
|
1354 |
+
def __init__(
|
1355 |
+
self,
|
1356 |
+
embed_dims: int = 768,
|
1357 |
+
output_dims: int = 1024,
|
1358 |
+
hidden_dims: int = 1280,
|
1359 |
+
depth: int = 4,
|
1360 |
+
dim_head: int = 64,
|
1361 |
+
heads: int = 16,
|
1362 |
+
num_queries: int = 8,
|
1363 |
+
ffn_ratio: float = 4,
|
1364 |
+
) -> None:
|
1365 |
+
super().__init__()
|
1366 |
+
self.latents = nn.Parameter(
|
1367 |
+
torch.randn(1, num_queries, hidden_dims) / hidden_dims**0.5
|
1368 |
+
)
|
1369 |
+
|
1370 |
+
self.proj_in = nn.Linear(embed_dims, hidden_dims)
|
1371 |
+
|
1372 |
+
self.proj_out = nn.Linear(hidden_dims, output_dims)
|
1373 |
+
self.norm_out = nn.LayerNorm(output_dims)
|
1374 |
+
|
1375 |
+
self.layers = nn.ModuleList(
|
1376 |
+
[
|
1377 |
+
IPAdapterPlusImageProjectionBlock(
|
1378 |
+
hidden_dims, dim_head, heads, ffn_ratio
|
1379 |
+
)
|
1380 |
+
for _ in range(depth)
|
1381 |
+
]
|
1382 |
+
)
|
1383 |
+
|
1384 |
+
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
1385 |
+
"""Forward pass.
|
1386 |
+
|
1387 |
+
Args:
|
1388 |
+
x (torch.Tensor): Input Tensor.
|
1389 |
+
Returns:
|
1390 |
+
torch.Tensor: Output Tensor.
|
1391 |
+
"""
|
1392 |
+
latents = self.latents.repeat(x.size(0), 1, 1)
|
1393 |
+
|
1394 |
+
x = self.proj_in(x)
|
1395 |
+
|
1396 |
+
for block in self.layers:
|
1397 |
+
residual = latents
|
1398 |
+
latents = block(x, latents, residual)
|
1399 |
+
|
1400 |
+
latents = self.proj_out(latents)
|
1401 |
+
return self.norm_out(latents)
|
1402 |
+
|
1403 |
+
|
1404 |
+
class IPAdapterFaceIDPlusImageProjection(nn.Module):
|
1405 |
+
"""FacePerceiverResampler of IP-Adapter Plus.
|
1406 |
+
|
1407 |
+
Args:
|
1408 |
+
embed_dims (int): The feature dimension. Defaults to 768. output_dims (int): The number of output channels,
|
1409 |
+
that is the same
|
1410 |
+
number of the channels in the `unet.config.cross_attention_dim`. Defaults to 1024.
|
1411 |
+
hidden_dims (int):
|
1412 |
+
The number of hidden channels. Defaults to 1280. depth (int): The number of blocks. Defaults
|
1413 |
+
to 8. dim_head (int): The number of head channels. Defaults to 64. heads (int): Parallel attention heads.
|
1414 |
+
Defaults to 16. num_tokens (int): Number of tokens num_queries (int): The number of queries. Defaults to 8.
|
1415 |
+
ffn_ratio (float): The expansion ratio of feedforward network hidden
|
1416 |
+
layer channels. Defaults to 4.
|
1417 |
+
ffproj_ratio (float): The expansion ratio of feedforward network hidden
|
1418 |
+
layer channels (for ID embeddings). Defaults to 4.
|
1419 |
+
"""
|
1420 |
+
|
1421 |
+
def __init__(
|
1422 |
+
self,
|
1423 |
+
embed_dims: int = 768,
|
1424 |
+
output_dims: int = 768,
|
1425 |
+
hidden_dims: int = 1280,
|
1426 |
+
id_embeddings_dim: int = 512,
|
1427 |
+
depth: int = 4,
|
1428 |
+
dim_head: int = 64,
|
1429 |
+
heads: int = 16,
|
1430 |
+
num_tokens: int = 4,
|
1431 |
+
num_queries: int = 8,
|
1432 |
+
ffn_ratio: float = 4,
|
1433 |
+
ffproj_ratio: int = 2,
|
1434 |
+
) -> None:
|
1435 |
+
super().__init__()
|
1436 |
+
from .attention import FeedForward
|
1437 |
+
|
1438 |
+
self.num_tokens = num_tokens
|
1439 |
+
self.embed_dim = embed_dims
|
1440 |
+
self.clip_embeds = None
|
1441 |
+
self.shortcut = False
|
1442 |
+
self.shortcut_scale = 1.0
|
1443 |
+
|
1444 |
+
self.proj = FeedForward(
|
1445 |
+
id_embeddings_dim,
|
1446 |
+
embed_dims * num_tokens,
|
1447 |
+
activation_fn="gelu",
|
1448 |
+
mult=ffproj_ratio,
|
1449 |
+
)
|
1450 |
+
self.norm = nn.LayerNorm(embed_dims)
|
1451 |
+
|
1452 |
+
self.proj_in = nn.Linear(hidden_dims, embed_dims)
|
1453 |
+
|
1454 |
+
self.proj_out = nn.Linear(embed_dims, output_dims)
|
1455 |
+
self.norm_out = nn.LayerNorm(output_dims)
|
1456 |
+
|
1457 |
+
self.layers = nn.ModuleList(
|
1458 |
+
[
|
1459 |
+
IPAdapterPlusImageProjectionBlock(
|
1460 |
+
embed_dims, dim_head, heads, ffn_ratio
|
1461 |
+
)
|
1462 |
+
for _ in range(depth)
|
1463 |
+
]
|
1464 |
+
)
|
1465 |
+
|
1466 |
+
def forward(self, id_embeds: torch.Tensor) -> torch.Tensor:
|
1467 |
+
"""Forward pass.
|
1468 |
+
|
1469 |
+
Args:
|
1470 |
+
id_embeds (torch.Tensor): Input Tensor (ID embeds).
|
1471 |
+
Returns:
|
1472 |
+
torch.Tensor: Output Tensor.
|
1473 |
+
"""
|
1474 |
+
id_embeds = id_embeds.to(self.clip_embeds.dtype)
|
1475 |
+
id_embeds = self.proj(id_embeds)
|
1476 |
+
id_embeds = id_embeds.reshape(-1, self.num_tokens, self.embed_dim)
|
1477 |
+
id_embeds = self.norm(id_embeds)
|
1478 |
+
latents = id_embeds
|
1479 |
+
|
1480 |
+
clip_embeds = self.proj_in(self.clip_embeds)
|
1481 |
+
x = clip_embeds.reshape(-1, clip_embeds.shape[2], clip_embeds.shape[3])
|
1482 |
+
|
1483 |
+
for block in self.layers:
|
1484 |
+
residual = latents
|
1485 |
+
latents = block(x, latents, residual)
|
1486 |
+
|
1487 |
+
latents = self.proj_out(latents)
|
1488 |
+
out = self.norm_out(latents)
|
1489 |
+
if self.shortcut:
|
1490 |
+
out = id_embeds + self.shortcut_scale * out
|
1491 |
+
return out
|
1492 |
+
|
1493 |
+
|
1494 |
+
class MultiIPAdapterImageProjection(nn.Module):
|
1495 |
+
def __init__(
|
1496 |
+
self, IPAdapterImageProjectionLayers: Union[List[nn.Module], Tuple[nn.Module]]
|
1497 |
+
):
|
1498 |
+
super().__init__()
|
1499 |
+
self.image_projection_layers = nn.ModuleList(IPAdapterImageProjectionLayers)
|
1500 |
+
|
1501 |
+
def forward(self, image_embeds: List[torch.Tensor]):
|
1502 |
+
projected_image_embeds = []
|
1503 |
+
|
1504 |
+
# currently, we accept `image_embeds` as
|
1505 |
+
# 1. a tensor (deprecated) with shape [batch_size, embed_dim] or [batch_size, sequence_length, embed_dim]
|
1506 |
+
# 2. list of `n` tensors where `n` is number of ip-adapters, each tensor can hae shape [batch_size, num_images, embed_dim] or [batch_size, num_images, sequence_length, embed_dim]
|
1507 |
+
if not isinstance(image_embeds, list):
|
1508 |
+
deprecation_message = (
|
1509 |
+
"You have passed a tensor as `image_embeds`.This is deprecated and will be removed in a future release."
|
1510 |
+
" Please make sure to update your script to pass `image_embeds` as a list of tensors to supress this warning."
|
1511 |
+
)
|
1512 |
+
deprecate(
|
1513 |
+
"image_embeds not a list",
|
1514 |
+
"1.0.0",
|
1515 |
+
deprecation_message,
|
1516 |
+
standard_warn=False,
|
1517 |
+
)
|
1518 |
+
image_embeds = [image_embeds.unsqueeze(1)]
|
1519 |
+
|
1520 |
+
if len(image_embeds) != len(self.image_projection_layers):
|
1521 |
+
raise ValueError(
|
1522 |
+
f"image_embeds must have the same length as image_projection_layers, got {len(image_embeds)} and {len(self.image_projection_layers)}"
|
1523 |
+
)
|
1524 |
+
|
1525 |
+
for image_embed, image_projection_layer in zip(
|
1526 |
+
image_embeds, self.image_projection_layers
|
1527 |
+
):
|
1528 |
+
batch_size, num_images = image_embed.shape[0], image_embed.shape[1]
|
1529 |
+
image_embed = image_embed.reshape(
|
1530 |
+
(batch_size * num_images,) + image_embed.shape[2:]
|
1531 |
+
)
|
1532 |
+
image_embed = image_projection_layer(image_embed)
|
1533 |
+
image_embed = image_embed.reshape(
|
1534 |
+
(batch_size, num_images) + image_embed.shape[1:]
|
1535 |
+
)
|
1536 |
+
|
1537 |
+
projected_image_embeds.append(image_embed)
|
1538 |
+
|
1539 |
+
return projected_image_embeds
|
models/resnet.py
ADDED
@@ -0,0 +1,508 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from functools import partial
|
2 |
+
from typing import Optional, Tuple, Union
|
3 |
+
|
4 |
+
import torch
|
5 |
+
import torch.nn as nn
|
6 |
+
import torch.nn.functional as F
|
7 |
+
|
8 |
+
from diffusers.utils import deprecate
|
9 |
+
from diffusers.models.activations import get_activation
|
10 |
+
from diffusers.models.attention_processor import SpatialNorm
|
11 |
+
from diffusers.models.downsampling import ( # noqa
|
12 |
+
Downsample2D,
|
13 |
+
downsample_2d,
|
14 |
+
)
|
15 |
+
from diffusers.models.normalization import AdaGroupNorm
|
16 |
+
from diffusers.models.upsampling import ( # noqa
|
17 |
+
Upsample2D,
|
18 |
+
upsample_2d,
|
19 |
+
)
|
20 |
+
|
21 |
+
|
22 |
+
class ResnetBlock2D(nn.Module):
|
23 |
+
r"""
|
24 |
+
A Resnet block.
|
25 |
+
|
26 |
+
Parameters:
|
27 |
+
in_channels (`int`): The number of channels in the input.
|
28 |
+
out_channels (`int`, *optional*, default to be `None`):
|
29 |
+
The number of output channels for the first conv2d layer. If None, same as `in_channels`.
|
30 |
+
dropout (`float`, *optional*, defaults to `0.0`): The dropout probability to use.
|
31 |
+
temb_channels (`int`, *optional*, default to `512`): the number of channels in timestep embedding.
|
32 |
+
groups (`int`, *optional*, default to `32`): The number of groups to use for the first normalization layer.
|
33 |
+
groups_out (`int`, *optional*, default to None):
|
34 |
+
The number of groups to use for the second normalization layer. if set to None, same as `groups`.
|
35 |
+
eps (`float`, *optional*, defaults to `1e-6`): The epsilon to use for the normalization.
|
36 |
+
non_linearity (`str`, *optional*, default to `"swish"`): the activation function to use.
|
37 |
+
time_embedding_norm (`str`, *optional*, default to `"default"` ): Time scale shift config.
|
38 |
+
By default, apply timestep embedding conditioning with a simple shift mechanism. Choose "scale_shift"
|
39 |
+
for a stronger conditioning with scale and shift.
|
40 |
+
kernel (`torch.FloatTensor`, optional, default to None): FIR filter, see
|
41 |
+
[`~models.resnet.FirUpsample2D`] and [`~models.resnet.FirDownsample2D`].
|
42 |
+
output_scale_factor (`float`, *optional*, default to be `1.0`): the scale factor to use for the output.
|
43 |
+
use_in_shortcut (`bool`, *optional*, default to `True`):
|
44 |
+
If `True`, add a 1x1 nn.conv2d layer for skip-connection.
|
45 |
+
up (`bool`, *optional*, default to `False`): If `True`, add an upsample layer.
|
46 |
+
down (`bool`, *optional*, default to `False`): If `True`, add a downsample layer.
|
47 |
+
conv_shortcut_bias (`bool`, *optional*, default to `True`): If `True`, adds a learnable bias to the
|
48 |
+
`conv_shortcut` output.
|
49 |
+
conv_2d_out_channels (`int`, *optional*, default to `None`): the number of channels in the output.
|
50 |
+
If None, same as `out_channels`.
|
51 |
+
"""
|
52 |
+
|
53 |
+
def __init__(
|
54 |
+
self,
|
55 |
+
*,
|
56 |
+
in_channels: int,
|
57 |
+
out_channels: Optional[int] = None,
|
58 |
+
conv_shortcut: bool = False,
|
59 |
+
dropout: float = 0.0,
|
60 |
+
temb_channels: int = 512,
|
61 |
+
groups: int = 32,
|
62 |
+
groups_out: Optional[int] = None,
|
63 |
+
pre_norm: bool = True,
|
64 |
+
eps: float = 1e-6,
|
65 |
+
non_linearity: str = "swish",
|
66 |
+
skip_time_act: bool = False,
|
67 |
+
time_embedding_norm: str = "default", # default, scale_shift,
|
68 |
+
kernel: Optional[torch.FloatTensor] = None,
|
69 |
+
output_scale_factor: float = 1.0,
|
70 |
+
use_in_shortcut: Optional[bool] = None,
|
71 |
+
up: bool = False,
|
72 |
+
down: bool = False,
|
73 |
+
conv_shortcut_bias: bool = True,
|
74 |
+
conv_2d_out_channels: Optional[int] = None,
|
75 |
+
):
|
76 |
+
super().__init__()
|
77 |
+
if time_embedding_norm == "ada_group":
|
78 |
+
raise ValueError(
|
79 |
+
"This class cannot be used with `time_embedding_norm==ada_group`, please use `ResnetBlockCondNorm2D` instead",
|
80 |
+
)
|
81 |
+
if time_embedding_norm == "spatial":
|
82 |
+
raise ValueError(
|
83 |
+
"This class cannot be used with `time_embedding_norm==spatial`, please use `ResnetBlockCondNorm2D` instead",
|
84 |
+
)
|
85 |
+
|
86 |
+
self.pre_norm = True
|
87 |
+
self.in_channels = in_channels
|
88 |
+
out_channels = in_channels if out_channels is None else out_channels
|
89 |
+
self.out_channels = out_channels
|
90 |
+
self.use_conv_shortcut = conv_shortcut
|
91 |
+
self.up = up
|
92 |
+
self.down = down
|
93 |
+
self.output_scale_factor = output_scale_factor
|
94 |
+
self.time_embedding_norm = time_embedding_norm
|
95 |
+
self.skip_time_act = skip_time_act
|
96 |
+
|
97 |
+
linear_cls = nn.Linear
|
98 |
+
conv_cls = nn.Conv2d
|
99 |
+
|
100 |
+
if groups_out is None:
|
101 |
+
groups_out = groups
|
102 |
+
|
103 |
+
self.norm1 = torch.nn.GroupNorm(
|
104 |
+
num_groups=groups, num_channels=in_channels, eps=eps, affine=True
|
105 |
+
)
|
106 |
+
|
107 |
+
self.conv1 = conv_cls(
|
108 |
+
in_channels, out_channels, kernel_size=3, stride=1, padding=1
|
109 |
+
)
|
110 |
+
|
111 |
+
if temb_channels is not None:
|
112 |
+
if self.time_embedding_norm == "default":
|
113 |
+
self.time_emb_proj = linear_cls(temb_channels, out_channels)
|
114 |
+
elif self.time_embedding_norm == "scale_shift":
|
115 |
+
self.time_emb_proj = linear_cls(temb_channels, 2 * out_channels)
|
116 |
+
else:
|
117 |
+
raise ValueError(
|
118 |
+
f"unknown time_embedding_norm : {self.time_embedding_norm} "
|
119 |
+
)
|
120 |
+
else:
|
121 |
+
self.time_emb_proj = None
|
122 |
+
|
123 |
+
self.norm2 = torch.nn.GroupNorm(
|
124 |
+
num_groups=groups_out, num_channels=out_channels, eps=eps, affine=True
|
125 |
+
)
|
126 |
+
|
127 |
+
self.dropout = torch.nn.Dropout(dropout)
|
128 |
+
conv_2d_out_channels = conv_2d_out_channels or out_channels
|
129 |
+
self.conv2 = conv_cls(
|
130 |
+
out_channels, conv_2d_out_channels, kernel_size=3, stride=1, padding=1
|
131 |
+
)
|
132 |
+
|
133 |
+
self.nonlinearity = get_activation(non_linearity)
|
134 |
+
|
135 |
+
self.upsample = self.downsample = None
|
136 |
+
if self.up:
|
137 |
+
if kernel == "fir":
|
138 |
+
fir_kernel = (1, 3, 3, 1)
|
139 |
+
self.upsample = lambda x: upsample_2d(x, kernel=fir_kernel)
|
140 |
+
elif kernel == "sde_vp":
|
141 |
+
self.upsample = partial(F.interpolate, scale_factor=2.0, mode="nearest")
|
142 |
+
else:
|
143 |
+
self.upsample = Upsample2D(in_channels, use_conv=False)
|
144 |
+
elif self.down:
|
145 |
+
if kernel == "fir":
|
146 |
+
fir_kernel = (1, 3, 3, 1)
|
147 |
+
self.downsample = lambda x: downsample_2d(x, kernel=fir_kernel)
|
148 |
+
elif kernel == "sde_vp":
|
149 |
+
self.downsample = partial(F.avg_pool2d, kernel_size=2, stride=2)
|
150 |
+
else:
|
151 |
+
self.downsample = Downsample2D(
|
152 |
+
in_channels, use_conv=False, padding=1, name="op"
|
153 |
+
)
|
154 |
+
|
155 |
+
self.use_in_shortcut = (
|
156 |
+
self.in_channels != conv_2d_out_channels
|
157 |
+
if use_in_shortcut is None
|
158 |
+
else use_in_shortcut
|
159 |
+
)
|
160 |
+
|
161 |
+
self.conv_shortcut = None
|
162 |
+
if self.use_in_shortcut:
|
163 |
+
self.conv_shortcut = conv_cls(
|
164 |
+
in_channels,
|
165 |
+
conv_2d_out_channels,
|
166 |
+
kernel_size=1,
|
167 |
+
stride=1,
|
168 |
+
padding=0,
|
169 |
+
bias=conv_shortcut_bias,
|
170 |
+
)
|
171 |
+
|
172 |
+
def forward(
|
173 |
+
self, input_tensor: torch.FloatTensor, temb: torch.FloatTensor, *args, **kwargs
|
174 |
+
) -> torch.FloatTensor:
|
175 |
+
if len(args) > 0 or kwargs.get("scale", None) is not None:
|
176 |
+
deprecation_message = "The `scale` argument is deprecated and will be ignored. Please remove it, as passing it will raise an error in the future. `scale` should directly be passed while calling the underlying pipeline component i.e., via `cross_attention_kwargs`."
|
177 |
+
deprecate("scale", "1.0.0", deprecation_message)
|
178 |
+
|
179 |
+
hidden_states = input_tensor
|
180 |
+
|
181 |
+
hidden_states = self.norm1(hidden_states)
|
182 |
+
hidden_states = self.nonlinearity(hidden_states)
|
183 |
+
|
184 |
+
if self.upsample is not None:
|
185 |
+
# upsample_nearest_nhwc fails with large batch sizes. see https://github.com/huggingface/diffusers/issues/984
|
186 |
+
if hidden_states.shape[0] >= 64:
|
187 |
+
input_tensor = input_tensor.contiguous()
|
188 |
+
hidden_states = hidden_states.contiguous()
|
189 |
+
input_tensor = self.upsample(input_tensor)
|
190 |
+
hidden_states = self.upsample(hidden_states)
|
191 |
+
elif self.downsample is not None:
|
192 |
+
input_tensor = self.downsample(input_tensor)
|
193 |
+
hidden_states = self.downsample(hidden_states)
|
194 |
+
|
195 |
+
hidden_states = self.conv1(hidden_states)
|
196 |
+
|
197 |
+
if self.time_emb_proj is not None:
|
198 |
+
if not self.skip_time_act:
|
199 |
+
temb = self.nonlinearity(temb)
|
200 |
+
temb = self.time_emb_proj(temb)[:, :, None, None]
|
201 |
+
|
202 |
+
if self.time_embedding_norm == "default":
|
203 |
+
if temb is not None:
|
204 |
+
hidden_states = hidden_states + temb
|
205 |
+
hidden_states = self.norm2(hidden_states)
|
206 |
+
elif self.time_embedding_norm == "scale_shift":
|
207 |
+
if temb is None:
|
208 |
+
raise ValueError(
|
209 |
+
f" `temb` should not be None when `time_embedding_norm` is {self.time_embedding_norm}"
|
210 |
+
)
|
211 |
+
time_scale, time_shift = torch.chunk(temb, 2, dim=1)
|
212 |
+
hidden_states = self.norm2(hidden_states)
|
213 |
+
hidden_states = hidden_states * (1 + time_scale) + time_shift
|
214 |
+
else:
|
215 |
+
hidden_states = self.norm2(hidden_states)
|
216 |
+
|
217 |
+
hidden_states = self.nonlinearity(hidden_states)
|
218 |
+
|
219 |
+
hidden_states = self.dropout(hidden_states)
|
220 |
+
hidden_states = self.conv2(hidden_states)
|
221 |
+
|
222 |
+
if self.conv_shortcut is not None:
|
223 |
+
input_tensor = self.conv_shortcut(input_tensor)
|
224 |
+
|
225 |
+
output_tensor = (input_tensor + hidden_states) / self.output_scale_factor
|
226 |
+
|
227 |
+
return output_tensor
|
228 |
+
|
229 |
+
|
230 |
+
class TemporalResnetBlock(nn.Module):
|
231 |
+
r"""
|
232 |
+
A Resnet block.
|
233 |
+
|
234 |
+
Parameters:
|
235 |
+
in_channels (`int`): The number of channels in the input.
|
236 |
+
out_channels (`int`, *optional*, default to be `None`):
|
237 |
+
The number of output channels for the first conv2d layer. If None, same as `in_channels`.
|
238 |
+
temb_channels (`int`, *optional*, default to `512`): the number of channels in timestep embedding.
|
239 |
+
eps (`float`, *optional*, defaults to `1e-6`): The epsilon to use for the normalization.
|
240 |
+
"""
|
241 |
+
|
242 |
+
def __init__(
|
243 |
+
self,
|
244 |
+
in_channels: int,
|
245 |
+
out_channels: Optional[int] = None,
|
246 |
+
temb_channels: int = 512,
|
247 |
+
eps: float = 1e-6,
|
248 |
+
):
|
249 |
+
super().__init__()
|
250 |
+
self.in_channels = in_channels
|
251 |
+
out_channels = in_channels if out_channels is None else out_channels
|
252 |
+
self.out_channels = out_channels
|
253 |
+
|
254 |
+
kernel_size = (3, 1, 1)
|
255 |
+
padding = [k // 2 for k in kernel_size]
|
256 |
+
|
257 |
+
self.norm1 = torch.nn.GroupNorm(
|
258 |
+
num_groups=32, num_channels=in_channels, eps=eps, affine=True
|
259 |
+
)
|
260 |
+
self.conv1 = nn.Conv3d(
|
261 |
+
in_channels,
|
262 |
+
out_channels,
|
263 |
+
kernel_size=kernel_size,
|
264 |
+
stride=1,
|
265 |
+
padding=padding,
|
266 |
+
)
|
267 |
+
|
268 |
+
if temb_channels is not None:
|
269 |
+
self.time_emb_proj = nn.Linear(temb_channels, out_channels)
|
270 |
+
else:
|
271 |
+
self.time_emb_proj = None
|
272 |
+
|
273 |
+
self.norm2 = torch.nn.GroupNorm(
|
274 |
+
num_groups=32, num_channels=out_channels, eps=eps, affine=True
|
275 |
+
)
|
276 |
+
|
277 |
+
self.dropout = torch.nn.Dropout(0.0)
|
278 |
+
self.conv2 = nn.Conv3d(
|
279 |
+
out_channels,
|
280 |
+
out_channels,
|
281 |
+
kernel_size=kernel_size,
|
282 |
+
stride=1,
|
283 |
+
padding=padding,
|
284 |
+
)
|
285 |
+
|
286 |
+
self.nonlinearity = get_activation("silu")
|
287 |
+
|
288 |
+
self.use_in_shortcut = self.in_channels != out_channels
|
289 |
+
|
290 |
+
self.conv_shortcut = None
|
291 |
+
if self.use_in_shortcut:
|
292 |
+
self.conv_shortcut = nn.Conv3d(
|
293 |
+
in_channels,
|
294 |
+
out_channels,
|
295 |
+
kernel_size=1,
|
296 |
+
stride=1,
|
297 |
+
padding=0,
|
298 |
+
)
|
299 |
+
|
300 |
+
def forward(
|
301 |
+
self, input_tensor: torch.FloatTensor, temb: torch.FloatTensor
|
302 |
+
) -> torch.FloatTensor:
|
303 |
+
hidden_states = input_tensor
|
304 |
+
|
305 |
+
hidden_states = self.norm1(hidden_states)
|
306 |
+
hidden_states = self.nonlinearity(hidden_states)
|
307 |
+
hidden_states = self.conv1(hidden_states)
|
308 |
+
|
309 |
+
if self.time_emb_proj is not None:
|
310 |
+
temb = self.nonlinearity(temb)
|
311 |
+
temb = self.time_emb_proj(temb)[:, :, :, None, None]
|
312 |
+
temb = temb.permute(0, 2, 1, 3, 4)
|
313 |
+
hidden_states = hidden_states + temb
|
314 |
+
|
315 |
+
hidden_states = self.norm2(hidden_states)
|
316 |
+
hidden_states = self.nonlinearity(hidden_states)
|
317 |
+
hidden_states = self.dropout(hidden_states)
|
318 |
+
hidden_states = self.conv2(hidden_states)
|
319 |
+
|
320 |
+
if self.conv_shortcut is not None:
|
321 |
+
input_tensor = self.conv_shortcut(input_tensor)
|
322 |
+
|
323 |
+
output_tensor = input_tensor + hidden_states
|
324 |
+
|
325 |
+
return output_tensor
|
326 |
+
|
327 |
+
|
328 |
+
# VideoResBlock
|
329 |
+
class SpatioTemporalResBlock(nn.Module):
|
330 |
+
r"""
|
331 |
+
A SpatioTemporal Resnet block.
|
332 |
+
|
333 |
+
Parameters:
|
334 |
+
in_channels (`int`): The number of channels in the input.
|
335 |
+
out_channels (`int`, *optional*, default to be `None`):
|
336 |
+
The number of output channels for the first conv2d layer. If None, same as `in_channels`.
|
337 |
+
temb_channels (`int`, *optional*, default to `512`): the number of channels in timestep embedding.
|
338 |
+
eps (`float`, *optional*, defaults to `1e-6`): The epsilon to use for the spatial resenet.
|
339 |
+
temporal_eps (`float`, *optional*, defaults to `eps`): The epsilon to use for the temporal resnet.
|
340 |
+
merge_factor (`float`, *optional*, defaults to `0.5`): The merge factor to use for the temporal mixing.
|
341 |
+
merge_strategy (`str`, *optional*, defaults to `learned_with_images`):
|
342 |
+
The merge strategy to use for the temporal mixing.
|
343 |
+
switch_spatial_to_temporal_mix (`bool`, *optional*, defaults to `False`):
|
344 |
+
If `True`, switch the spatial and temporal mixing.
|
345 |
+
"""
|
346 |
+
|
347 |
+
def __init__(
|
348 |
+
self,
|
349 |
+
in_channels: int,
|
350 |
+
out_channels: Optional[int] = None,
|
351 |
+
temb_channels: int = 512,
|
352 |
+
eps: float = 1e-6,
|
353 |
+
temporal_eps: Optional[float] = None,
|
354 |
+
merge_factor: float = 0.5,
|
355 |
+
merge_strategy="learned_with_images",
|
356 |
+
switch_spatial_to_temporal_mix: bool = False,
|
357 |
+
):
|
358 |
+
super().__init__()
|
359 |
+
|
360 |
+
self.spatial_res_block = ResnetBlock2D(
|
361 |
+
in_channels=in_channels,
|
362 |
+
out_channels=out_channels,
|
363 |
+
temb_channels=temb_channels,
|
364 |
+
eps=eps,
|
365 |
+
)
|
366 |
+
|
367 |
+
self.temporal_res_block = TemporalResnetBlock(
|
368 |
+
in_channels=out_channels if out_channels is not None else in_channels,
|
369 |
+
out_channels=out_channels if out_channels is not None else in_channels,
|
370 |
+
temb_channels=temb_channels,
|
371 |
+
eps=temporal_eps if temporal_eps is not None else eps,
|
372 |
+
)
|
373 |
+
|
374 |
+
self.time_mixer = AlphaBlender(
|
375 |
+
alpha=merge_factor,
|
376 |
+
merge_strategy=merge_strategy,
|
377 |
+
switch_spatial_to_temporal_mix=switch_spatial_to_temporal_mix,
|
378 |
+
)
|
379 |
+
|
380 |
+
def forward(
|
381 |
+
self,
|
382 |
+
hidden_states: torch.FloatTensor,
|
383 |
+
temb: Optional[torch.FloatTensor] = None,
|
384 |
+
image_only_indicator: Optional[torch.Tensor] = None,
|
385 |
+
):
|
386 |
+
num_frames = image_only_indicator.shape[-1]
|
387 |
+
hidden_states = self.spatial_res_block(hidden_states, temb)
|
388 |
+
|
389 |
+
batch_frames, channels, height, width = hidden_states.shape
|
390 |
+
batch_size = batch_frames // num_frames
|
391 |
+
|
392 |
+
hidden_states_mix = (
|
393 |
+
hidden_states[None, :]
|
394 |
+
.reshape(batch_size, num_frames, channels, height, width)
|
395 |
+
.permute(0, 2, 1, 3, 4)
|
396 |
+
)
|
397 |
+
hidden_states = (
|
398 |
+
hidden_states[None, :]
|
399 |
+
.reshape(batch_size, num_frames, channels, height, width)
|
400 |
+
.permute(0, 2, 1, 3, 4)
|
401 |
+
)
|
402 |
+
|
403 |
+
if temb is not None:
|
404 |
+
temb = temb.reshape(batch_size, num_frames, -1)
|
405 |
+
|
406 |
+
hidden_states = self.temporal_res_block(hidden_states, temb)
|
407 |
+
hidden_states = self.time_mixer(
|
408 |
+
x_spatial=hidden_states_mix,
|
409 |
+
x_temporal=hidden_states,
|
410 |
+
image_only_indicator=image_only_indicator,
|
411 |
+
)
|
412 |
+
|
413 |
+
hidden_states = hidden_states.permute(0, 2, 1, 3, 4).reshape(
|
414 |
+
batch_frames, channels, height, width
|
415 |
+
)
|
416 |
+
return hidden_states
|
417 |
+
|
418 |
+
|
419 |
+
class AlphaBlender(nn.Module):
|
420 |
+
r"""
|
421 |
+
A module to blend spatial and temporal features.
|
422 |
+
|
423 |
+
Parameters:
|
424 |
+
alpha (`float`): The initial value of the blending factor.
|
425 |
+
merge_strategy (`str`, *optional*, defaults to `learned_with_images`):
|
426 |
+
The merge strategy to use for the temporal mixing.
|
427 |
+
switch_spatial_to_temporal_mix (`bool`, *optional*, defaults to `False`):
|
428 |
+
If `True`, switch the spatial and temporal mixing.
|
429 |
+
"""
|
430 |
+
|
431 |
+
strategies = ["learned", "fixed", "learned_with_images"]
|
432 |
+
|
433 |
+
def __init__(
|
434 |
+
self,
|
435 |
+
alpha: float,
|
436 |
+
merge_strategy: str = "learned_with_images",
|
437 |
+
switch_spatial_to_temporal_mix: bool = False,
|
438 |
+
):
|
439 |
+
super().__init__()
|
440 |
+
self.merge_strategy = merge_strategy
|
441 |
+
self.switch_spatial_to_temporal_mix = (
|
442 |
+
switch_spatial_to_temporal_mix # For TemporalVAE
|
443 |
+
)
|
444 |
+
|
445 |
+
if merge_strategy not in self.strategies:
|
446 |
+
raise ValueError(f"merge_strategy needs to be in {self.strategies}")
|
447 |
+
|
448 |
+
if self.merge_strategy == "fixed":
|
449 |
+
self.register_buffer("mix_factor", torch.Tensor([alpha]))
|
450 |
+
elif (
|
451 |
+
self.merge_strategy == "learned"
|
452 |
+
or self.merge_strategy == "learned_with_images"
|
453 |
+
):
|
454 |
+
self.register_parameter(
|
455 |
+
"mix_factor", torch.nn.Parameter(torch.Tensor([alpha]))
|
456 |
+
)
|
457 |
+
else:
|
458 |
+
raise ValueError(f"Unknown merge strategy {self.merge_strategy}")
|
459 |
+
|
460 |
+
def get_alpha(self, image_only_indicator: torch.Tensor, ndims: int) -> torch.Tensor:
|
461 |
+
if self.merge_strategy == "fixed":
|
462 |
+
alpha = self.mix_factor
|
463 |
+
|
464 |
+
elif self.merge_strategy == "learned":
|
465 |
+
alpha = torch.sigmoid(self.mix_factor)
|
466 |
+
|
467 |
+
elif self.merge_strategy == "learned_with_images":
|
468 |
+
if image_only_indicator is None:
|
469 |
+
raise ValueError(
|
470 |
+
"Please provide image_only_indicator to use learned_with_images merge strategy"
|
471 |
+
)
|
472 |
+
|
473 |
+
alpha = torch.where(
|
474 |
+
image_only_indicator.bool(),
|
475 |
+
torch.ones(1, 1, device=image_only_indicator.device),
|
476 |
+
torch.sigmoid(self.mix_factor)[..., None],
|
477 |
+
)
|
478 |
+
|
479 |
+
# (batch, channel, frames, height, width)
|
480 |
+
if ndims == 5:
|
481 |
+
alpha = alpha[:, None, :, None, None]
|
482 |
+
# (batch*frames, height*width, channels)
|
483 |
+
elif ndims == 3:
|
484 |
+
alpha = alpha.reshape(-1)[:, None, None]
|
485 |
+
else:
|
486 |
+
raise ValueError(
|
487 |
+
f"Unexpected ndims {ndims}. Dimensions should be 3 or 5"
|
488 |
+
)
|
489 |
+
|
490 |
+
else:
|
491 |
+
raise NotImplementedError
|
492 |
+
|
493 |
+
return alpha
|
494 |
+
|
495 |
+
def forward(
|
496 |
+
self,
|
497 |
+
x_spatial: torch.Tensor,
|
498 |
+
x_temporal: torch.Tensor,
|
499 |
+
image_only_indicator: Optional[torch.Tensor] = None,
|
500 |
+
) -> torch.Tensor:
|
501 |
+
alpha = self.get_alpha(image_only_indicator, x_spatial.ndim)
|
502 |
+
alpha = alpha.to(x_spatial.dtype)
|
503 |
+
|
504 |
+
if self.switch_spatial_to_temporal_mix:
|
505 |
+
alpha = 1.0 - alpha
|
506 |
+
|
507 |
+
x = alpha * x_spatial + (1.0 - alpha) * x_temporal
|
508 |
+
return x
|
models/transformers/transformer_temporal_rope.py
ADDED
@@ -0,0 +1,230 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Copyright 2024 The HuggingFace Team. All rights reserved.
|
2 |
+
#
|
3 |
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
4 |
+
# you may not use this file except in compliance with the License.
|
5 |
+
# You may obtain a copy of the License at
|
6 |
+
#
|
7 |
+
# http://www.apache.org/licenses/LICENSE-2.0
|
8 |
+
#
|
9 |
+
# Unless required by applicable law or agreed to in writing, software
|
10 |
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
11 |
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
12 |
+
# See the License for the specific language governing permissions and
|
13 |
+
# limitations under the License.
|
14 |
+
from dataclasses import dataclass
|
15 |
+
from typing import Any, Dict, Optional
|
16 |
+
|
17 |
+
import torch
|
18 |
+
from torch import nn
|
19 |
+
|
20 |
+
from diffusers.configuration_utils import ConfigMixin, register_to_config
|
21 |
+
from diffusers.utils import BaseOutput
|
22 |
+
from diffusers.models.embeddings import TimestepEmbedding, Timesteps
|
23 |
+
from diffusers.models.modeling_utils import ModelMixin
|
24 |
+
from ..resnet import AlphaBlender
|
25 |
+
from ..attention import (
|
26 |
+
BasicTransformerBlock,
|
27 |
+
TemporalRopeBasicTransformerBlock,
|
28 |
+
)
|
29 |
+
from ..embeddings import rope
|
30 |
+
|
31 |
+
|
32 |
+
@dataclass
|
33 |
+
class TransformerTemporalModelOutput(BaseOutput):
|
34 |
+
"""
|
35 |
+
The output of [`TransformerTemporalModel`].
|
36 |
+
|
37 |
+
Args:
|
38 |
+
sample (`torch.FloatTensor` of shape `(batch_size x num_frames, num_channels, height, width)`):
|
39 |
+
The hidden states output conditioned on `encoder_hidden_states` input.
|
40 |
+
"""
|
41 |
+
|
42 |
+
sample: torch.FloatTensor
|
43 |
+
|
44 |
+
|
45 |
+
class TransformerSpatioTemporalModel(nn.Module):
|
46 |
+
"""
|
47 |
+
A Transformer model for video-like data.
|
48 |
+
|
49 |
+
Parameters:
|
50 |
+
num_attention_heads (`int`, *optional*, defaults to 16): The number of heads to use for multi-head attention.
|
51 |
+
attention_head_dim (`int`, *optional*, defaults to 88): The number of channels in each head.
|
52 |
+
in_channels (`int`, *optional*):
|
53 |
+
The number of channels in the input and output (specify if the input is **continuous**).
|
54 |
+
out_channels (`int`, *optional*):
|
55 |
+
The number of channels in the output (specify if the input is **continuous**).
|
56 |
+
num_layers (`int`, *optional*, defaults to 1): The number of layers of Transformer blocks to use.
|
57 |
+
cross_attention_dim (`int`, *optional*): The number of `encoder_hidden_states` dimensions to use.
|
58 |
+
"""
|
59 |
+
|
60 |
+
def __init__(
|
61 |
+
self,
|
62 |
+
num_attention_heads: int = 16,
|
63 |
+
attention_head_dim: int = 88,
|
64 |
+
in_channels: int = 320,
|
65 |
+
out_channels: Optional[int] = None,
|
66 |
+
num_layers: int = 1,
|
67 |
+
cross_attention_dim: Optional[int] = None,
|
68 |
+
):
|
69 |
+
super().__init__()
|
70 |
+
self.num_attention_heads = num_attention_heads
|
71 |
+
self.attention_head_dim = attention_head_dim
|
72 |
+
|
73 |
+
inner_dim = num_attention_heads * attention_head_dim
|
74 |
+
self.inner_dim = inner_dim
|
75 |
+
|
76 |
+
# 2. Define input layers
|
77 |
+
self.in_channels = in_channels
|
78 |
+
self.norm = torch.nn.GroupNorm(
|
79 |
+
num_groups=32, num_channels=in_channels, eps=1e-6
|
80 |
+
)
|
81 |
+
self.proj_in = nn.Linear(in_channels, inner_dim)
|
82 |
+
|
83 |
+
# 3. Define transformers blocks
|
84 |
+
self.transformer_blocks = nn.ModuleList(
|
85 |
+
[
|
86 |
+
BasicTransformerBlock(
|
87 |
+
inner_dim,
|
88 |
+
num_attention_heads,
|
89 |
+
attention_head_dim,
|
90 |
+
cross_attention_dim=cross_attention_dim,
|
91 |
+
)
|
92 |
+
for d in range(num_layers)
|
93 |
+
]
|
94 |
+
)
|
95 |
+
|
96 |
+
time_mix_inner_dim = inner_dim
|
97 |
+
self.temporal_transformer_blocks = nn.ModuleList(
|
98 |
+
[
|
99 |
+
TemporalRopeBasicTransformerBlock(
|
100 |
+
inner_dim,
|
101 |
+
time_mix_inner_dim,
|
102 |
+
num_attention_heads,
|
103 |
+
attention_head_dim,
|
104 |
+
cross_attention_dim=cross_attention_dim,
|
105 |
+
)
|
106 |
+
for _ in range(num_layers)
|
107 |
+
]
|
108 |
+
)
|
109 |
+
|
110 |
+
self.time_mixer = AlphaBlender(alpha=0.5, merge_strategy="learned_with_images")
|
111 |
+
|
112 |
+
# 4. Define output layers
|
113 |
+
self.out_channels = in_channels if out_channels is None else out_channels
|
114 |
+
# TODO: should use out_channels for continuous projections
|
115 |
+
self.proj_out = nn.Linear(inner_dim, in_channels)
|
116 |
+
|
117 |
+
self.gradient_checkpointing = False
|
118 |
+
|
119 |
+
def forward(
|
120 |
+
self,
|
121 |
+
hidden_states: torch.Tensor,
|
122 |
+
encoder_hidden_states: Optional[torch.Tensor] = None,
|
123 |
+
image_only_indicator: Optional[torch.Tensor] = None,
|
124 |
+
return_dict: bool = True,
|
125 |
+
position_ids: Optional[torch.Tensor] = None,
|
126 |
+
):
|
127 |
+
"""
|
128 |
+
Args:
|
129 |
+
hidden_states (`torch.FloatTensor` of shape `(batch size, channel, height, width)`):
|
130 |
+
Input hidden_states.
|
131 |
+
num_frames (`int`):
|
132 |
+
The number of frames to be processed per batch. This is used to reshape the hidden states.
|
133 |
+
encoder_hidden_states ( `torch.LongTensor` of shape `(batch size, encoder_hidden_states dim)`, *optional*):
|
134 |
+
Conditional embeddings for cross attention layer. If not given, cross-attention defaults to
|
135 |
+
self-attention.
|
136 |
+
image_only_indicator (`torch.LongTensor` of shape `(batch size, num_frames)`, *optional*):
|
137 |
+
A tensor indicating whether the input contains only images. 1 indicates that the input contains only
|
138 |
+
images, 0 indicates that the input contains video frames.
|
139 |
+
return_dict (`bool`, *optional*, defaults to `True`):
|
140 |
+
Whether or not to return a [`~models.transformer_temporal.TransformerTemporalModelOutput`] instead of a plain
|
141 |
+
tuple.
|
142 |
+
|
143 |
+
Returns:
|
144 |
+
[`~models.transformer_temporal.TransformerTemporalModelOutput`] or `tuple`:
|
145 |
+
If `return_dict` is True, an [`~models.transformer_temporal.TransformerTemporalModelOutput`] is
|
146 |
+
returned, otherwise a `tuple` where the first element is the sample tensor.
|
147 |
+
"""
|
148 |
+
# 1. Input
|
149 |
+
batch_frames, _, height, width = hidden_states.shape
|
150 |
+
num_frames = image_only_indicator.shape[-1]
|
151 |
+
batch_size = batch_frames // num_frames
|
152 |
+
|
153 |
+
# (B*F, 1, C)
|
154 |
+
time_context = encoder_hidden_states
|
155 |
+
# (B, 1, C)
|
156 |
+
time_context_first_timestep = time_context[None, :].reshape(
|
157 |
+
batch_size, num_frames, -1, time_context.shape[-1]
|
158 |
+
)[:, 0]
|
159 |
+
|
160 |
+
# (B*N, 1, C)
|
161 |
+
time_context = time_context_first_timestep.repeat_interleave(
|
162 |
+
height * width, dim=0
|
163 |
+
)
|
164 |
+
|
165 |
+
residual = hidden_states
|
166 |
+
|
167 |
+
hidden_states = self.norm(hidden_states)
|
168 |
+
inner_dim = hidden_states.shape[1]
|
169 |
+
hidden_states = hidden_states.permute(0, 2, 3, 1).reshape(
|
170 |
+
batch_frames, height * width, inner_dim
|
171 |
+
)
|
172 |
+
hidden_states = self.proj_in(hidden_states)
|
173 |
+
|
174 |
+
if position_ids is None:
|
175 |
+
# (B, F)
|
176 |
+
frame_rotary_emb = torch.arange(num_frames, device=hidden_states.device)
|
177 |
+
frame_rotary_emb = frame_rotary_emb[None, :].repeat(batch_size, 1)
|
178 |
+
else:
|
179 |
+
frame_rotary_emb = position_ids
|
180 |
+
|
181 |
+
# (B, 1, F, d/2, 2, 2)
|
182 |
+
frame_rotary_emb = rope(frame_rotary_emb, self.attention_head_dim)
|
183 |
+
# (B*N, 1, F, d/2, 2, 2)
|
184 |
+
frame_rotary_emb = frame_rotary_emb.repeat_interleave(height * width, dim=0)
|
185 |
+
|
186 |
+
# 2. Blocks
|
187 |
+
for block, temporal_block in zip(
|
188 |
+
self.transformer_blocks, self.temporal_transformer_blocks
|
189 |
+
):
|
190 |
+
if self.training and self.gradient_checkpointing:
|
191 |
+
hidden_states = torch.utils.checkpoint.checkpoint(
|
192 |
+
block,
|
193 |
+
hidden_states,
|
194 |
+
None,
|
195 |
+
encoder_hidden_states,
|
196 |
+
None,
|
197 |
+
use_reentrant=False,
|
198 |
+
)
|
199 |
+
else:
|
200 |
+
hidden_states = block(
|
201 |
+
hidden_states,
|
202 |
+
encoder_hidden_states=encoder_hidden_states,
|
203 |
+
)
|
204 |
+
|
205 |
+
hidden_states_mix = temporal_block(
|
206 |
+
hidden_states,
|
207 |
+
num_frames=num_frames,
|
208 |
+
encoder_hidden_states=time_context,
|
209 |
+
frame_rotary_emb=frame_rotary_emb,
|
210 |
+
)
|
211 |
+
hidden_states = self.time_mixer(
|
212 |
+
x_spatial=hidden_states,
|
213 |
+
x_temporal=hidden_states_mix,
|
214 |
+
image_only_indicator=image_only_indicator,
|
215 |
+
)
|
216 |
+
|
217 |
+
# 3. Output
|
218 |
+
hidden_states = self.proj_out(hidden_states)
|
219 |
+
hidden_states = (
|
220 |
+
hidden_states.reshape(batch_frames, height, width, inner_dim)
|
221 |
+
.permute(0, 3, 1, 2)
|
222 |
+
.contiguous()
|
223 |
+
)
|
224 |
+
|
225 |
+
output = hidden_states + residual
|
226 |
+
|
227 |
+
if not return_dict:
|
228 |
+
return (output,)
|
229 |
+
|
230 |
+
return TransformerTemporalModelOutput(sample=output)
|
models/unets/unet_3d_rope_blocks.py
ADDED
@@ -0,0 +1,682 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from typing import Any, Dict, Optional, Tuple, Union
|
2 |
+
|
3 |
+
import torch
|
4 |
+
from torch import nn
|
5 |
+
|
6 |
+
from diffusers.utils import is_torch_version, logging
|
7 |
+
from ..resnet import (
|
8 |
+
Downsample2D,
|
9 |
+
SpatioTemporalResBlock,
|
10 |
+
Upsample2D,
|
11 |
+
)
|
12 |
+
|
13 |
+
from ..transformers.transformer_temporal_rope import TransformerSpatioTemporalModel
|
14 |
+
|
15 |
+
logger = logging.get_logger(__name__) # pylint: disable=invalid-name
|
16 |
+
|
17 |
+
|
18 |
+
def get_down_block(
|
19 |
+
down_block_type: str,
|
20 |
+
num_layers: int,
|
21 |
+
in_channels: int,
|
22 |
+
out_channels: int,
|
23 |
+
temb_channels: int,
|
24 |
+
add_downsample: bool,
|
25 |
+
resnet_eps: float,
|
26 |
+
resnet_act_fn: str,
|
27 |
+
num_attention_heads: int,
|
28 |
+
resnet_groups: Optional[int] = None,
|
29 |
+
cross_attention_dim: Optional[int] = None,
|
30 |
+
downsample_padding: Optional[int] = None,
|
31 |
+
dual_cross_attention: bool = False,
|
32 |
+
use_linear_projection: bool = True,
|
33 |
+
only_cross_attention: bool = False,
|
34 |
+
upcast_attention: bool = False,
|
35 |
+
resnet_time_scale_shift: str = "default",
|
36 |
+
temporal_num_attention_heads: int = 8,
|
37 |
+
temporal_max_seq_length: int = 32,
|
38 |
+
transformer_layers_per_block: int = 1,
|
39 |
+
) -> Union[
|
40 |
+
"DownBlockSpatioTemporal",
|
41 |
+
"CrossAttnDownBlockSpatioTemporal",
|
42 |
+
]:
|
43 |
+
if down_block_type == "DownBlockSpatioTemporal":
|
44 |
+
# added for SDV
|
45 |
+
return DownBlockSpatioTemporal(
|
46 |
+
num_layers=num_layers,
|
47 |
+
in_channels=in_channels,
|
48 |
+
out_channels=out_channels,
|
49 |
+
temb_channels=temb_channels,
|
50 |
+
add_downsample=add_downsample,
|
51 |
+
)
|
52 |
+
elif down_block_type == "CrossAttnDownBlockSpatioTemporal":
|
53 |
+
# added for SDV
|
54 |
+
if cross_attention_dim is None:
|
55 |
+
raise ValueError(
|
56 |
+
"cross_attention_dim must be specified for CrossAttnDownBlockSpatioTemporal"
|
57 |
+
)
|
58 |
+
return CrossAttnDownBlockSpatioTemporal(
|
59 |
+
in_channels=in_channels,
|
60 |
+
out_channels=out_channels,
|
61 |
+
temb_channels=temb_channels,
|
62 |
+
num_layers=num_layers,
|
63 |
+
transformer_layers_per_block=transformer_layers_per_block,
|
64 |
+
add_downsample=add_downsample,
|
65 |
+
cross_attention_dim=cross_attention_dim,
|
66 |
+
num_attention_heads=num_attention_heads,
|
67 |
+
)
|
68 |
+
|
69 |
+
raise ValueError(f"{down_block_type} does not exist.")
|
70 |
+
|
71 |
+
|
72 |
+
def get_up_block(
|
73 |
+
up_block_type: str,
|
74 |
+
num_layers: int,
|
75 |
+
in_channels: int,
|
76 |
+
out_channels: int,
|
77 |
+
prev_output_channel: int,
|
78 |
+
temb_channels: int,
|
79 |
+
add_upsample: bool,
|
80 |
+
resnet_eps: float,
|
81 |
+
resnet_act_fn: str,
|
82 |
+
num_attention_heads: int,
|
83 |
+
resolution_idx: Optional[int] = None,
|
84 |
+
resnet_groups: Optional[int] = None,
|
85 |
+
cross_attention_dim: Optional[int] = None,
|
86 |
+
dual_cross_attention: bool = False,
|
87 |
+
use_linear_projection: bool = True,
|
88 |
+
only_cross_attention: bool = False,
|
89 |
+
upcast_attention: bool = False,
|
90 |
+
resnet_time_scale_shift: str = "default",
|
91 |
+
temporal_num_attention_heads: int = 8,
|
92 |
+
temporal_cross_attention_dim: Optional[int] = None,
|
93 |
+
temporal_max_seq_length: int = 32,
|
94 |
+
transformer_layers_per_block: int = 1,
|
95 |
+
dropout: float = 0.0,
|
96 |
+
) -> Union[
|
97 |
+
"UpBlockSpatioTemporal",
|
98 |
+
"CrossAttnUpBlockSpatioTemporal",
|
99 |
+
]:
|
100 |
+
if up_block_type == "UpBlockSpatioTemporal":
|
101 |
+
# added for SDV
|
102 |
+
return UpBlockSpatioTemporal(
|
103 |
+
num_layers=num_layers,
|
104 |
+
in_channels=in_channels,
|
105 |
+
out_channels=out_channels,
|
106 |
+
prev_output_channel=prev_output_channel,
|
107 |
+
temb_channels=temb_channels,
|
108 |
+
resolution_idx=resolution_idx,
|
109 |
+
add_upsample=add_upsample,
|
110 |
+
)
|
111 |
+
elif up_block_type == "CrossAttnUpBlockSpatioTemporal":
|
112 |
+
# added for SDV
|
113 |
+
if cross_attention_dim is None:
|
114 |
+
raise ValueError(
|
115 |
+
"cross_attention_dim must be specified for CrossAttnUpBlockSpatioTemporal"
|
116 |
+
)
|
117 |
+
return CrossAttnUpBlockSpatioTemporal(
|
118 |
+
in_channels=in_channels,
|
119 |
+
out_channels=out_channels,
|
120 |
+
prev_output_channel=prev_output_channel,
|
121 |
+
temb_channels=temb_channels,
|
122 |
+
num_layers=num_layers,
|
123 |
+
transformer_layers_per_block=transformer_layers_per_block,
|
124 |
+
add_upsample=add_upsample,
|
125 |
+
cross_attention_dim=cross_attention_dim,
|
126 |
+
num_attention_heads=num_attention_heads,
|
127 |
+
resolution_idx=resolution_idx,
|
128 |
+
)
|
129 |
+
|
130 |
+
raise ValueError(f"{up_block_type} does not exist.")
|
131 |
+
|
132 |
+
|
133 |
+
class UNetMidBlockSpatioTemporal(nn.Module):
|
134 |
+
def __init__(
|
135 |
+
self,
|
136 |
+
in_channels: int,
|
137 |
+
temb_channels: int,
|
138 |
+
num_layers: int = 1,
|
139 |
+
transformer_layers_per_block: Union[int, Tuple[int]] = 1,
|
140 |
+
num_attention_heads: int = 1,
|
141 |
+
cross_attention_dim: int = 1280,
|
142 |
+
):
|
143 |
+
super().__init__()
|
144 |
+
|
145 |
+
self.has_cross_attention = True
|
146 |
+
self.num_attention_heads = num_attention_heads
|
147 |
+
|
148 |
+
# support for variable transformer layers per block
|
149 |
+
if isinstance(transformer_layers_per_block, int):
|
150 |
+
transformer_layers_per_block = [transformer_layers_per_block] * num_layers
|
151 |
+
|
152 |
+
# there is always at least one resnet
|
153 |
+
resnets = [
|
154 |
+
SpatioTemporalResBlock(
|
155 |
+
in_channels=in_channels,
|
156 |
+
out_channels=in_channels,
|
157 |
+
temb_channels=temb_channels,
|
158 |
+
eps=1e-5,
|
159 |
+
)
|
160 |
+
]
|
161 |
+
attentions = []
|
162 |
+
|
163 |
+
for i in range(num_layers):
|
164 |
+
attentions.append(
|
165 |
+
TransformerSpatioTemporalModel(
|
166 |
+
num_attention_heads,
|
167 |
+
in_channels // num_attention_heads,
|
168 |
+
in_channels=in_channels,
|
169 |
+
num_layers=transformer_layers_per_block[i],
|
170 |
+
cross_attention_dim=cross_attention_dim,
|
171 |
+
)
|
172 |
+
)
|
173 |
+
|
174 |
+
resnets.append(
|
175 |
+
SpatioTemporalResBlock(
|
176 |
+
in_channels=in_channels,
|
177 |
+
out_channels=in_channels,
|
178 |
+
temb_channels=temb_channels,
|
179 |
+
eps=1e-5,
|
180 |
+
)
|
181 |
+
)
|
182 |
+
|
183 |
+
self.attentions = nn.ModuleList(attentions)
|
184 |
+
self.resnets = nn.ModuleList(resnets)
|
185 |
+
|
186 |
+
self.gradient_checkpointing = False
|
187 |
+
|
188 |
+
def forward(
|
189 |
+
self,
|
190 |
+
hidden_states: torch.FloatTensor,
|
191 |
+
temb: Optional[torch.FloatTensor] = None,
|
192 |
+
encoder_hidden_states: Optional[torch.FloatTensor] = None,
|
193 |
+
image_only_indicator: Optional[torch.Tensor] = None,
|
194 |
+
position_ids: Optional[torch.Tensor] = None,
|
195 |
+
) -> torch.FloatTensor:
|
196 |
+
hidden_states = self.resnets[0](
|
197 |
+
hidden_states,
|
198 |
+
temb,
|
199 |
+
image_only_indicator=image_only_indicator,
|
200 |
+
)
|
201 |
+
|
202 |
+
for attn, resnet in zip(self.attentions, self.resnets[1:]):
|
203 |
+
if self.training and self.gradient_checkpointing: # TODO
|
204 |
+
|
205 |
+
def create_custom_forward(module, return_dict=None):
|
206 |
+
def custom_forward(*inputs):
|
207 |
+
if return_dict is not None:
|
208 |
+
return module(*inputs, return_dict=return_dict)
|
209 |
+
else:
|
210 |
+
return module(*inputs)
|
211 |
+
|
212 |
+
return custom_forward
|
213 |
+
|
214 |
+
ckpt_kwargs: Dict[str, Any] = (
|
215 |
+
{"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
|
216 |
+
)
|
217 |
+
hidden_states = attn(
|
218 |
+
hidden_states,
|
219 |
+
encoder_hidden_states=encoder_hidden_states,
|
220 |
+
image_only_indicator=image_only_indicator,
|
221 |
+
return_dict=False,
|
222 |
+
position_ids=position_ids,
|
223 |
+
)[0]
|
224 |
+
hidden_states = torch.utils.checkpoint.checkpoint(
|
225 |
+
create_custom_forward(resnet),
|
226 |
+
hidden_states,
|
227 |
+
temb,
|
228 |
+
image_only_indicator,
|
229 |
+
**ckpt_kwargs,
|
230 |
+
)
|
231 |
+
else:
|
232 |
+
hidden_states = attn(
|
233 |
+
hidden_states,
|
234 |
+
encoder_hidden_states=encoder_hidden_states,
|
235 |
+
image_only_indicator=image_only_indicator,
|
236 |
+
return_dict=False,
|
237 |
+
position_ids=position_ids,
|
238 |
+
)[0]
|
239 |
+
hidden_states = resnet(
|
240 |
+
hidden_states,
|
241 |
+
temb,
|
242 |
+
image_only_indicator=image_only_indicator,
|
243 |
+
)
|
244 |
+
|
245 |
+
return hidden_states
|
246 |
+
|
247 |
+
|
248 |
+
class DownBlockSpatioTemporal(nn.Module):
|
249 |
+
def __init__(
|
250 |
+
self,
|
251 |
+
in_channels: int,
|
252 |
+
out_channels: int,
|
253 |
+
temb_channels: int,
|
254 |
+
num_layers: int = 1,
|
255 |
+
add_downsample: bool = True,
|
256 |
+
):
|
257 |
+
super().__init__()
|
258 |
+
resnets = []
|
259 |
+
|
260 |
+
for i in range(num_layers):
|
261 |
+
in_channels = in_channels if i == 0 else out_channels
|
262 |
+
resnets.append(
|
263 |
+
SpatioTemporalResBlock(
|
264 |
+
in_channels=in_channels,
|
265 |
+
out_channels=out_channels,
|
266 |
+
temb_channels=temb_channels,
|
267 |
+
eps=1e-5,
|
268 |
+
)
|
269 |
+
)
|
270 |
+
|
271 |
+
self.resnets = nn.ModuleList(resnets)
|
272 |
+
|
273 |
+
if add_downsample:
|
274 |
+
self.downsamplers = nn.ModuleList(
|
275 |
+
[
|
276 |
+
Downsample2D(
|
277 |
+
out_channels,
|
278 |
+
use_conv=True,
|
279 |
+
out_channels=out_channels,
|
280 |
+
name="op",
|
281 |
+
)
|
282 |
+
]
|
283 |
+
)
|
284 |
+
else:
|
285 |
+
self.downsamplers = None
|
286 |
+
|
287 |
+
self.gradient_checkpointing = False
|
288 |
+
|
289 |
+
def forward(
|
290 |
+
self,
|
291 |
+
hidden_states: torch.FloatTensor,
|
292 |
+
temb: Optional[torch.FloatTensor] = None,
|
293 |
+
image_only_indicator: Optional[torch.Tensor] = None,
|
294 |
+
position_ids=None,
|
295 |
+
) -> Tuple[torch.FloatTensor, Tuple[torch.FloatTensor, ...]]:
|
296 |
+
output_states = ()
|
297 |
+
for resnet in self.resnets:
|
298 |
+
if self.training and self.gradient_checkpointing:
|
299 |
+
|
300 |
+
def create_custom_forward(module):
|
301 |
+
def custom_forward(*inputs):
|
302 |
+
return module(*inputs)
|
303 |
+
|
304 |
+
return custom_forward
|
305 |
+
|
306 |
+
if is_torch_version(">=", "1.11.0"):
|
307 |
+
hidden_states = torch.utils.checkpoint.checkpoint(
|
308 |
+
create_custom_forward(resnet),
|
309 |
+
hidden_states,
|
310 |
+
temb,
|
311 |
+
image_only_indicator,
|
312 |
+
use_reentrant=False,
|
313 |
+
)
|
314 |
+
else:
|
315 |
+
hidden_states = torch.utils.checkpoint.checkpoint(
|
316 |
+
create_custom_forward(resnet),
|
317 |
+
hidden_states,
|
318 |
+
temb,
|
319 |
+
image_only_indicator,
|
320 |
+
)
|
321 |
+
else:
|
322 |
+
hidden_states = resnet(
|
323 |
+
hidden_states,
|
324 |
+
temb,
|
325 |
+
image_only_indicator=image_only_indicator,
|
326 |
+
)
|
327 |
+
|
328 |
+
output_states = output_states + (hidden_states,)
|
329 |
+
|
330 |
+
if self.downsamplers is not None:
|
331 |
+
for downsampler in self.downsamplers:
|
332 |
+
hidden_states = downsampler(hidden_states)
|
333 |
+
|
334 |
+
output_states = output_states + (hidden_states,)
|
335 |
+
|
336 |
+
return hidden_states, output_states
|
337 |
+
|
338 |
+
|
339 |
+
class CrossAttnDownBlockSpatioTemporal(nn.Module):
|
340 |
+
def __init__(
|
341 |
+
self,
|
342 |
+
in_channels: int,
|
343 |
+
out_channels: int,
|
344 |
+
temb_channels: int,
|
345 |
+
num_layers: int = 1,
|
346 |
+
transformer_layers_per_block: Union[int, Tuple[int]] = 1,
|
347 |
+
num_attention_heads: int = 1,
|
348 |
+
cross_attention_dim: int = 1280,
|
349 |
+
add_downsample: bool = True,
|
350 |
+
):
|
351 |
+
super().__init__()
|
352 |
+
resnets = []
|
353 |
+
attentions = []
|
354 |
+
|
355 |
+
self.has_cross_attention = True
|
356 |
+
self.num_attention_heads = num_attention_heads
|
357 |
+
if isinstance(transformer_layers_per_block, int):
|
358 |
+
transformer_layers_per_block = [transformer_layers_per_block] * num_layers
|
359 |
+
|
360 |
+
for i in range(num_layers):
|
361 |
+
in_channels = in_channels if i == 0 else out_channels
|
362 |
+
resnets.append(
|
363 |
+
SpatioTemporalResBlock(
|
364 |
+
in_channels=in_channels,
|
365 |
+
out_channels=out_channels,
|
366 |
+
temb_channels=temb_channels,
|
367 |
+
eps=1e-6,
|
368 |
+
)
|
369 |
+
)
|
370 |
+
attentions.append(
|
371 |
+
TransformerSpatioTemporalModel(
|
372 |
+
num_attention_heads,
|
373 |
+
out_channels // num_attention_heads,
|
374 |
+
in_channels=out_channels,
|
375 |
+
num_layers=transformer_layers_per_block[i],
|
376 |
+
cross_attention_dim=cross_attention_dim,
|
377 |
+
)
|
378 |
+
)
|
379 |
+
|
380 |
+
self.attentions = nn.ModuleList(attentions)
|
381 |
+
self.resnets = nn.ModuleList(resnets)
|
382 |
+
|
383 |
+
if add_downsample:
|
384 |
+
self.downsamplers = nn.ModuleList(
|
385 |
+
[
|
386 |
+
Downsample2D(
|
387 |
+
out_channels,
|
388 |
+
use_conv=True,
|
389 |
+
out_channels=out_channels,
|
390 |
+
padding=1,
|
391 |
+
name="op",
|
392 |
+
)
|
393 |
+
]
|
394 |
+
)
|
395 |
+
else:
|
396 |
+
self.downsamplers = None
|
397 |
+
|
398 |
+
self.gradient_checkpointing = False
|
399 |
+
|
400 |
+
def forward(
|
401 |
+
self,
|
402 |
+
hidden_states: torch.FloatTensor,
|
403 |
+
temb: Optional[torch.FloatTensor] = None,
|
404 |
+
encoder_hidden_states: Optional[torch.FloatTensor] = None,
|
405 |
+
image_only_indicator: Optional[torch.Tensor] = None,
|
406 |
+
position_ids=None,
|
407 |
+
) -> Tuple[torch.FloatTensor, Tuple[torch.FloatTensor, ...]]:
|
408 |
+
output_states = ()
|
409 |
+
|
410 |
+
blocks = list(zip(self.resnets, self.attentions))
|
411 |
+
for resnet, attn in blocks:
|
412 |
+
if self.training and self.gradient_checkpointing: # TODO
|
413 |
+
|
414 |
+
def create_custom_forward(module, return_dict=None):
|
415 |
+
def custom_forward(*inputs):
|
416 |
+
if return_dict is not None:
|
417 |
+
return module(*inputs, return_dict=return_dict)
|
418 |
+
else:
|
419 |
+
return module(*inputs)
|
420 |
+
|
421 |
+
return custom_forward
|
422 |
+
|
423 |
+
ckpt_kwargs: Dict[str, Any] = (
|
424 |
+
{"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
|
425 |
+
)
|
426 |
+
hidden_states = torch.utils.checkpoint.checkpoint(
|
427 |
+
create_custom_forward(resnet),
|
428 |
+
hidden_states,
|
429 |
+
temb,
|
430 |
+
image_only_indicator,
|
431 |
+
**ckpt_kwargs,
|
432 |
+
)
|
433 |
+
|
434 |
+
hidden_states = attn(
|
435 |
+
hidden_states,
|
436 |
+
encoder_hidden_states=encoder_hidden_states,
|
437 |
+
image_only_indicator=image_only_indicator,
|
438 |
+
return_dict=False,
|
439 |
+
position_ids=position_ids,
|
440 |
+
)[0]
|
441 |
+
else:
|
442 |
+
hidden_states = resnet(
|
443 |
+
hidden_states,
|
444 |
+
temb,
|
445 |
+
image_only_indicator=image_only_indicator,
|
446 |
+
)
|
447 |
+
hidden_states = attn(
|
448 |
+
hidden_states,
|
449 |
+
encoder_hidden_states=encoder_hidden_states,
|
450 |
+
image_only_indicator=image_only_indicator,
|
451 |
+
return_dict=False,
|
452 |
+
position_ids=position_ids,
|
453 |
+
)[0]
|
454 |
+
|
455 |
+
output_states = output_states + (hidden_states,)
|
456 |
+
|
457 |
+
if self.downsamplers is not None:
|
458 |
+
for downsampler in self.downsamplers:
|
459 |
+
hidden_states = downsampler(hidden_states)
|
460 |
+
|
461 |
+
output_states = output_states + (hidden_states,)
|
462 |
+
|
463 |
+
return hidden_states, output_states
|
464 |
+
|
465 |
+
|
466 |
+
class UpBlockSpatioTemporal(nn.Module):
|
467 |
+
def __init__(
|
468 |
+
self,
|
469 |
+
in_channels: int,
|
470 |
+
prev_output_channel: int,
|
471 |
+
out_channels: int,
|
472 |
+
temb_channels: int,
|
473 |
+
resolution_idx: Optional[int] = None,
|
474 |
+
num_layers: int = 1,
|
475 |
+
resnet_eps: float = 1e-6,
|
476 |
+
add_upsample: bool = True,
|
477 |
+
):
|
478 |
+
super().__init__()
|
479 |
+
resnets = []
|
480 |
+
|
481 |
+
for i in range(num_layers):
|
482 |
+
res_skip_channels = in_channels if (i == num_layers - 1) else out_channels
|
483 |
+
resnet_in_channels = prev_output_channel if i == 0 else out_channels
|
484 |
+
|
485 |
+
resnets.append(
|
486 |
+
SpatioTemporalResBlock(
|
487 |
+
in_channels=resnet_in_channels + res_skip_channels,
|
488 |
+
out_channels=out_channels,
|
489 |
+
temb_channels=temb_channels,
|
490 |
+
eps=resnet_eps,
|
491 |
+
)
|
492 |
+
)
|
493 |
+
|
494 |
+
self.resnets = nn.ModuleList(resnets)
|
495 |
+
|
496 |
+
if add_upsample:
|
497 |
+
self.upsamplers = nn.ModuleList(
|
498 |
+
[Upsample2D(out_channels, use_conv=True, out_channels=out_channels)]
|
499 |
+
)
|
500 |
+
else:
|
501 |
+
self.upsamplers = None
|
502 |
+
|
503 |
+
self.gradient_checkpointing = False
|
504 |
+
self.resolution_idx = resolution_idx
|
505 |
+
|
506 |
+
def forward(
|
507 |
+
self,
|
508 |
+
hidden_states: torch.FloatTensor,
|
509 |
+
res_hidden_states_tuple: Tuple[torch.FloatTensor, ...],
|
510 |
+
temb: Optional[torch.FloatTensor] = None,
|
511 |
+
upsample_size: Optional[int] = None,
|
512 |
+
image_only_indicator: Optional[torch.Tensor] = None,
|
513 |
+
position_ids: Optional[torch.Tensor] = None,
|
514 |
+
) -> torch.FloatTensor:
|
515 |
+
for resnet in self.resnets:
|
516 |
+
# pop res hidden states
|
517 |
+
res_hidden_states = res_hidden_states_tuple[-1]
|
518 |
+
res_hidden_states_tuple = res_hidden_states_tuple[:-1]
|
519 |
+
|
520 |
+
hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)
|
521 |
+
|
522 |
+
if self.training and self.gradient_checkpointing:
|
523 |
+
|
524 |
+
def create_custom_forward(module):
|
525 |
+
def custom_forward(*inputs):
|
526 |
+
return module(*inputs)
|
527 |
+
|
528 |
+
return custom_forward
|
529 |
+
|
530 |
+
if is_torch_version(">=", "1.11.0"):
|
531 |
+
hidden_states = torch.utils.checkpoint.checkpoint(
|
532 |
+
create_custom_forward(resnet),
|
533 |
+
hidden_states,
|
534 |
+
temb,
|
535 |
+
image_only_indicator,
|
536 |
+
use_reentrant=False,
|
537 |
+
)
|
538 |
+
else:
|
539 |
+
hidden_states = torch.utils.checkpoint.checkpoint(
|
540 |
+
create_custom_forward(resnet),
|
541 |
+
hidden_states,
|
542 |
+
temb,
|
543 |
+
image_only_indicator,
|
544 |
+
)
|
545 |
+
else:
|
546 |
+
hidden_states = resnet(
|
547 |
+
hidden_states,
|
548 |
+
temb,
|
549 |
+
image_only_indicator=image_only_indicator,
|
550 |
+
)
|
551 |
+
|
552 |
+
if self.upsamplers is not None:
|
553 |
+
for upsampler in self.upsamplers:
|
554 |
+
hidden_states = upsampler(hidden_states, upsample_size)
|
555 |
+
|
556 |
+
return hidden_states
|
557 |
+
|
558 |
+
|
559 |
+
class CrossAttnUpBlockSpatioTemporal(nn.Module):
|
560 |
+
def __init__(
|
561 |
+
self,
|
562 |
+
in_channels: int,
|
563 |
+
out_channels: int,
|
564 |
+
prev_output_channel: int,
|
565 |
+
temb_channels: int,
|
566 |
+
resolution_idx: Optional[int] = None,
|
567 |
+
num_layers: int = 1,
|
568 |
+
transformer_layers_per_block: Union[int, Tuple[int]] = 1,
|
569 |
+
resnet_eps: float = 1e-6,
|
570 |
+
num_attention_heads: int = 1,
|
571 |
+
cross_attention_dim: int = 1280,
|
572 |
+
add_upsample: bool = True,
|
573 |
+
):
|
574 |
+
super().__init__()
|
575 |
+
resnets = []
|
576 |
+
attentions = []
|
577 |
+
|
578 |
+
self.has_cross_attention = True
|
579 |
+
self.num_attention_heads = num_attention_heads
|
580 |
+
|
581 |
+
if isinstance(transformer_layers_per_block, int):
|
582 |
+
transformer_layers_per_block = [transformer_layers_per_block] * num_layers
|
583 |
+
|
584 |
+
for i in range(num_layers):
|
585 |
+
res_skip_channels = in_channels if (i == num_layers - 1) else out_channels
|
586 |
+
resnet_in_channels = prev_output_channel if i == 0 else out_channels
|
587 |
+
|
588 |
+
resnets.append(
|
589 |
+
SpatioTemporalResBlock(
|
590 |
+
in_channels=resnet_in_channels + res_skip_channels,
|
591 |
+
out_channels=out_channels,
|
592 |
+
temb_channels=temb_channels,
|
593 |
+
eps=resnet_eps,
|
594 |
+
)
|
595 |
+
)
|
596 |
+
attentions.append(
|
597 |
+
TransformerSpatioTemporalModel(
|
598 |
+
num_attention_heads,
|
599 |
+
out_channels // num_attention_heads,
|
600 |
+
in_channels=out_channels,
|
601 |
+
num_layers=transformer_layers_per_block[i],
|
602 |
+
cross_attention_dim=cross_attention_dim,
|
603 |
+
)
|
604 |
+
)
|
605 |
+
|
606 |
+
self.attentions = nn.ModuleList(attentions)
|
607 |
+
self.resnets = nn.ModuleList(resnets)
|
608 |
+
|
609 |
+
if add_upsample:
|
610 |
+
self.upsamplers = nn.ModuleList(
|
611 |
+
[Upsample2D(out_channels, use_conv=True, out_channels=out_channels)]
|
612 |
+
)
|
613 |
+
else:
|
614 |
+
self.upsamplers = None
|
615 |
+
|
616 |
+
self.gradient_checkpointing = False
|
617 |
+
self.resolution_idx = resolution_idx
|
618 |
+
|
619 |
+
def forward(
|
620 |
+
self,
|
621 |
+
hidden_states: torch.FloatTensor,
|
622 |
+
res_hidden_states_tuple: Tuple[torch.FloatTensor, ...],
|
623 |
+
temb: Optional[torch.FloatTensor] = None,
|
624 |
+
encoder_hidden_states: Optional[torch.FloatTensor] = None,
|
625 |
+
image_only_indicator: Optional[torch.Tensor] = None,
|
626 |
+
upsample_size: Optional[int] = None,
|
627 |
+
position_ids=None,
|
628 |
+
) -> torch.FloatTensor:
|
629 |
+
for resnet, attn in zip(self.resnets, self.attentions):
|
630 |
+
# pop res hidden states
|
631 |
+
res_hidden_states = res_hidden_states_tuple[-1]
|
632 |
+
res_hidden_states_tuple = res_hidden_states_tuple[:-1]
|
633 |
+
|
634 |
+
hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)
|
635 |
+
|
636 |
+
if self.training and self.gradient_checkpointing: # TODO
|
637 |
+
|
638 |
+
def create_custom_forward(module, return_dict=None):
|
639 |
+
def custom_forward(*inputs):
|
640 |
+
if return_dict is not None:
|
641 |
+
return module(*inputs, return_dict=return_dict)
|
642 |
+
else:
|
643 |
+
return module(*inputs)
|
644 |
+
|
645 |
+
return custom_forward
|
646 |
+
|
647 |
+
ckpt_kwargs: Dict[str, Any] = (
|
648 |
+
{"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
|
649 |
+
)
|
650 |
+
hidden_states = torch.utils.checkpoint.checkpoint(
|
651 |
+
create_custom_forward(resnet),
|
652 |
+
hidden_states,
|
653 |
+
temb,
|
654 |
+
image_only_indicator,
|
655 |
+
**ckpt_kwargs,
|
656 |
+
)
|
657 |
+
hidden_states = attn(
|
658 |
+
hidden_states,
|
659 |
+
encoder_hidden_states=encoder_hidden_states,
|
660 |
+
image_only_indicator=image_only_indicator,
|
661 |
+
return_dict=False,
|
662 |
+
position_ids=position_ids,
|
663 |
+
)[0]
|
664 |
+
else:
|
665 |
+
hidden_states = resnet(
|
666 |
+
hidden_states,
|
667 |
+
temb,
|
668 |
+
image_only_indicator=image_only_indicator,
|
669 |
+
)
|
670 |
+
hidden_states = attn(
|
671 |
+
hidden_states,
|
672 |
+
encoder_hidden_states=encoder_hidden_states,
|
673 |
+
image_only_indicator=image_only_indicator,
|
674 |
+
return_dict=False,
|
675 |
+
position_ids=position_ids,
|
676 |
+
)[0]
|
677 |
+
|
678 |
+
if self.upsamplers is not None:
|
679 |
+
for upsampler in self.upsamplers:
|
680 |
+
hidden_states = upsampler(hidden_states, upsample_size)
|
681 |
+
|
682 |
+
return hidden_states
|
models/unets/unet_spatio_temporal_rope_condition.py
ADDED
@@ -0,0 +1,546 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from dataclasses import dataclass
|
2 |
+
from typing import Dict, Optional, Tuple, Union
|
3 |
+
|
4 |
+
import torch
|
5 |
+
import torch.nn as nn
|
6 |
+
|
7 |
+
from diffusers.configuration_utils import ConfigMixin, register_to_config
|
8 |
+
from diffusers.loaders import UNet2DConditionLoadersMixin
|
9 |
+
from diffusers.utils import BaseOutput, logging
|
10 |
+
from diffusers.models.attention_processor import (
|
11 |
+
CROSS_ATTENTION_PROCESSORS,
|
12 |
+
AttentionProcessor,
|
13 |
+
AttnProcessor,
|
14 |
+
)
|
15 |
+
from diffusers.models.embeddings import TimestepEmbedding, Timesteps
|
16 |
+
from diffusers.models.modeling_utils import ModelMixin
|
17 |
+
from .unet_3d_rope_blocks import (
|
18 |
+
UNetMidBlockSpatioTemporal,
|
19 |
+
get_down_block,
|
20 |
+
get_up_block,
|
21 |
+
)
|
22 |
+
|
23 |
+
logger = logging.get_logger(__name__) # pylint: disable=invalid-name
|
24 |
+
|
25 |
+
|
26 |
+
@dataclass
|
27 |
+
class UNetSpatioTemporalRopeConditionOutput(BaseOutput):
|
28 |
+
"""
|
29 |
+
The output of [`UNetSpatioTemporalConditionModel`].
|
30 |
+
|
31 |
+
Args:
|
32 |
+
sample (`torch.FloatTensor` of shape `(batch_size, num_frames, num_channels, height, width)`):
|
33 |
+
The hidden states output conditioned on `encoder_hidden_states` input. Output of last layer of model.
|
34 |
+
"""
|
35 |
+
|
36 |
+
sample: torch.FloatTensor = None
|
37 |
+
|
38 |
+
|
39 |
+
class UNetSpatioTemporalRopeConditionModel(
|
40 |
+
ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin
|
41 |
+
):
|
42 |
+
r"""
|
43 |
+
A conditional Spatio-Temporal UNet model that takes a noisy video frames, conditional state, and a timestep and returns a sample
|
44 |
+
shaped output.
|
45 |
+
|
46 |
+
This model inherits from [`ModelMixin`]. Check the superclass documentation for it's generic methods implemented
|
47 |
+
for all models (such as downloading or saving).
|
48 |
+
|
49 |
+
Parameters:
|
50 |
+
sample_size (`int` or `Tuple[int, int]`, *optional*, defaults to `None`):
|
51 |
+
Height and width of input/output sample.
|
52 |
+
in_channels (`int`, *optional*, defaults to 8): Number of channels in the input sample.
|
53 |
+
out_channels (`int`, *optional*, defaults to 4): Number of channels in the output.
|
54 |
+
down_block_types (`Tuple[str]`, *optional*, defaults to `("CrossAttnDownBlockSpatioTemporal", "CrossAttnDownBlockSpatioTemporal", "CrossAttnDownBlockSpatioTemporal", "DownBlockSpatioTemporal")`):
|
55 |
+
The tuple of downsample blocks to use.
|
56 |
+
up_block_types (`Tuple[str]`, *optional*, defaults to `("UpBlockSpatioTemporal", "CrossAttnUpBlockSpatioTemporal", "CrossAttnUpBlockSpatioTemporal", "CrossAttnUpBlockSpatioTemporal")`):
|
57 |
+
The tuple of upsample blocks to use.
|
58 |
+
block_out_channels (`Tuple[int]`, *optional*, defaults to `(320, 640, 1280, 1280)`):
|
59 |
+
The tuple of output channels for each block.
|
60 |
+
addition_time_embed_dim: (`int`, defaults to 256):
|
61 |
+
Dimension to to encode the additional time ids.
|
62 |
+
projection_class_embeddings_input_dim (`int`, defaults to 768):
|
63 |
+
The dimension of the projection of encoded `added_time_ids`.
|
64 |
+
layers_per_block (`int`, *optional*, defaults to 2): The number of layers per block.
|
65 |
+
cross_attention_dim (`int` or `Tuple[int]`, *optional*, defaults to 1280):
|
66 |
+
The dimension of the cross attention features.
|
67 |
+
transformer_layers_per_block (`int`, `Tuple[int]`, or `Tuple[Tuple]` , *optional*, defaults to 1):
|
68 |
+
The number of transformer blocks of type [`~models.attention.BasicTransformerBlock`]. Only relevant for
|
69 |
+
[`~models.unet_3d_blocks.CrossAttnDownBlockSpatioTemporal`], [`~models.unet_3d_blocks.CrossAttnUpBlockSpatioTemporal`],
|
70 |
+
[`~models.unet_3d_blocks.UNetMidBlockSpatioTemporal`].
|
71 |
+
num_attention_heads (`int`, `Tuple[int]`, defaults to `(5, 10, 10, 20)`):
|
72 |
+
The number of attention heads.
|
73 |
+
dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
|
74 |
+
"""
|
75 |
+
|
76 |
+
_supports_gradient_checkpointing = True
|
77 |
+
|
78 |
+
@register_to_config
|
79 |
+
def __init__(
|
80 |
+
self,
|
81 |
+
sample_size: Optional[int] = None,
|
82 |
+
in_channels: int = 8,
|
83 |
+
out_channels: int = 4,
|
84 |
+
down_block_types: Tuple[str] = (
|
85 |
+
"CrossAttnDownBlockSpatioTemporal",
|
86 |
+
"CrossAttnDownBlockSpatioTemporal",
|
87 |
+
"CrossAttnDownBlockSpatioTemporal",
|
88 |
+
"DownBlockSpatioTemporal",
|
89 |
+
),
|
90 |
+
up_block_types: Tuple[str] = (
|
91 |
+
"UpBlockSpatioTemporal",
|
92 |
+
"CrossAttnUpBlockSpatioTemporal",
|
93 |
+
"CrossAttnUpBlockSpatioTemporal",
|
94 |
+
"CrossAttnUpBlockSpatioTemporal",
|
95 |
+
),
|
96 |
+
block_out_channels: Tuple[int] = (320, 640, 1280, 1280),
|
97 |
+
addition_time_embed_dim: int = 256,
|
98 |
+
projection_class_embeddings_input_dim: int = 768,
|
99 |
+
layers_per_block: Union[int, Tuple[int]] = 2,
|
100 |
+
cross_attention_dim: Union[int, Tuple[int]] = 1024,
|
101 |
+
transformer_layers_per_block: Union[int, Tuple[int], Tuple[Tuple]] = 1,
|
102 |
+
num_attention_heads: Union[int, Tuple[int]] = (5, 10, 20, 20),
|
103 |
+
num_frames: int = 25,
|
104 |
+
):
|
105 |
+
super().__init__()
|
106 |
+
|
107 |
+
self.sample_size = sample_size
|
108 |
+
|
109 |
+
# Check inputs
|
110 |
+
if len(down_block_types) != len(up_block_types):
|
111 |
+
raise ValueError(
|
112 |
+
f"Must provide the same number of `down_block_types` as `up_block_types`. `down_block_types`: {down_block_types}. `up_block_types`: {up_block_types}."
|
113 |
+
)
|
114 |
+
|
115 |
+
if len(block_out_channels) != len(down_block_types):
|
116 |
+
raise ValueError(
|
117 |
+
f"Must provide the same number of `block_out_channels` as `down_block_types`. `block_out_channels`: {block_out_channels}. `down_block_types`: {down_block_types}."
|
118 |
+
)
|
119 |
+
|
120 |
+
if not isinstance(num_attention_heads, int) and len(num_attention_heads) != len(
|
121 |
+
down_block_types
|
122 |
+
):
|
123 |
+
raise ValueError(
|
124 |
+
f"Must provide the same number of `num_attention_heads` as `down_block_types`. `num_attention_heads`: {num_attention_heads}. `down_block_types`: {down_block_types}."
|
125 |
+
)
|
126 |
+
|
127 |
+
if isinstance(cross_attention_dim, list) and len(cross_attention_dim) != len(
|
128 |
+
down_block_types
|
129 |
+
):
|
130 |
+
raise ValueError(
|
131 |
+
f"Must provide the same number of `cross_attention_dim` as `down_block_types`. `cross_attention_dim`: {cross_attention_dim}. `down_block_types`: {down_block_types}."
|
132 |
+
)
|
133 |
+
|
134 |
+
if not isinstance(layers_per_block, int) and len(layers_per_block) != len(
|
135 |
+
down_block_types
|
136 |
+
):
|
137 |
+
raise ValueError(
|
138 |
+
f"Must provide the same number of `layers_per_block` as `down_block_types`. `layers_per_block`: {layers_per_block}. `down_block_types`: {down_block_types}."
|
139 |
+
)
|
140 |
+
|
141 |
+
# input
|
142 |
+
self.conv_in = nn.Conv2d(
|
143 |
+
in_channels,
|
144 |
+
block_out_channels[0],
|
145 |
+
kernel_size=3,
|
146 |
+
padding=1,
|
147 |
+
)
|
148 |
+
|
149 |
+
# time
|
150 |
+
time_embed_dim = block_out_channels[0] * 4
|
151 |
+
|
152 |
+
self.time_proj = Timesteps(block_out_channels[0], True, downscale_freq_shift=0)
|
153 |
+
timestep_input_dim = block_out_channels[0]
|
154 |
+
|
155 |
+
self.time_embedding = TimestepEmbedding(timestep_input_dim, time_embed_dim)
|
156 |
+
|
157 |
+
self.down_blocks = nn.ModuleList([])
|
158 |
+
self.up_blocks = nn.ModuleList([])
|
159 |
+
|
160 |
+
if isinstance(num_attention_heads, int):
|
161 |
+
num_attention_heads = (num_attention_heads,) * len(down_block_types)
|
162 |
+
|
163 |
+
if isinstance(cross_attention_dim, int):
|
164 |
+
cross_attention_dim = (cross_attention_dim,) * len(down_block_types)
|
165 |
+
|
166 |
+
if isinstance(layers_per_block, int):
|
167 |
+
layers_per_block = [layers_per_block] * len(down_block_types)
|
168 |
+
|
169 |
+
if isinstance(transformer_layers_per_block, int):
|
170 |
+
transformer_layers_per_block = [transformer_layers_per_block] * len(
|
171 |
+
down_block_types
|
172 |
+
)
|
173 |
+
|
174 |
+
blocks_time_embed_dim = time_embed_dim
|
175 |
+
|
176 |
+
# down
|
177 |
+
output_channel = block_out_channels[0]
|
178 |
+
for i, down_block_type in enumerate(down_block_types):
|
179 |
+
input_channel = output_channel
|
180 |
+
output_channel = block_out_channels[i]
|
181 |
+
is_final_block = i == len(block_out_channels) - 1
|
182 |
+
|
183 |
+
down_block = get_down_block(
|
184 |
+
down_block_type,
|
185 |
+
num_layers=layers_per_block[i],
|
186 |
+
transformer_layers_per_block=transformer_layers_per_block[i],
|
187 |
+
in_channels=input_channel,
|
188 |
+
out_channels=output_channel,
|
189 |
+
temb_channels=blocks_time_embed_dim,
|
190 |
+
add_downsample=not is_final_block,
|
191 |
+
resnet_eps=1e-5,
|
192 |
+
cross_attention_dim=cross_attention_dim[i],
|
193 |
+
num_attention_heads=num_attention_heads[i],
|
194 |
+
resnet_act_fn="silu",
|
195 |
+
)
|
196 |
+
self.down_blocks.append(down_block)
|
197 |
+
|
198 |
+
# mid
|
199 |
+
self.mid_block = UNetMidBlockSpatioTemporal(
|
200 |
+
block_out_channels[-1],
|
201 |
+
temb_channels=blocks_time_embed_dim,
|
202 |
+
transformer_layers_per_block=transformer_layers_per_block[-1],
|
203 |
+
cross_attention_dim=cross_attention_dim[-1],
|
204 |
+
num_attention_heads=num_attention_heads[-1],
|
205 |
+
)
|
206 |
+
|
207 |
+
# count how many layers upsample the images
|
208 |
+
self.num_upsamplers = 0
|
209 |
+
|
210 |
+
# up
|
211 |
+
reversed_block_out_channels = list(reversed(block_out_channels))
|
212 |
+
reversed_num_attention_heads = list(reversed(num_attention_heads))
|
213 |
+
reversed_layers_per_block = list(reversed(layers_per_block))
|
214 |
+
reversed_cross_attention_dim = list(reversed(cross_attention_dim))
|
215 |
+
reversed_transformer_layers_per_block = list(
|
216 |
+
reversed(transformer_layers_per_block)
|
217 |
+
)
|
218 |
+
|
219 |
+
output_channel = reversed_block_out_channels[0]
|
220 |
+
for i, up_block_type in enumerate(up_block_types):
|
221 |
+
is_final_block = i == len(block_out_channels) - 1
|
222 |
+
|
223 |
+
prev_output_channel = output_channel
|
224 |
+
output_channel = reversed_block_out_channels[i]
|
225 |
+
input_channel = reversed_block_out_channels[
|
226 |
+
min(i + 1, len(block_out_channels) - 1)
|
227 |
+
]
|
228 |
+
|
229 |
+
# add upsample block for all BUT final layer
|
230 |
+
if not is_final_block:
|
231 |
+
add_upsample = True
|
232 |
+
self.num_upsamplers += 1
|
233 |
+
else:
|
234 |
+
add_upsample = False
|
235 |
+
|
236 |
+
up_block = get_up_block(
|
237 |
+
up_block_type,
|
238 |
+
num_layers=reversed_layers_per_block[i] + 1,
|
239 |
+
transformer_layers_per_block=reversed_transformer_layers_per_block[i],
|
240 |
+
in_channels=input_channel,
|
241 |
+
out_channels=output_channel,
|
242 |
+
prev_output_channel=prev_output_channel,
|
243 |
+
temb_channels=blocks_time_embed_dim,
|
244 |
+
add_upsample=add_upsample,
|
245 |
+
resnet_eps=1e-5,
|
246 |
+
resolution_idx=i,
|
247 |
+
cross_attention_dim=reversed_cross_attention_dim[i],
|
248 |
+
num_attention_heads=reversed_num_attention_heads[i],
|
249 |
+
resnet_act_fn="silu",
|
250 |
+
)
|
251 |
+
self.up_blocks.append(up_block)
|
252 |
+
prev_output_channel = output_channel
|
253 |
+
|
254 |
+
# out
|
255 |
+
self.conv_norm_out = nn.GroupNorm(
|
256 |
+
num_channels=block_out_channels[0], num_groups=32, eps=1e-5
|
257 |
+
)
|
258 |
+
self.conv_act = nn.SiLU()
|
259 |
+
|
260 |
+
self.conv_out = nn.Conv2d(
|
261 |
+
block_out_channels[0],
|
262 |
+
out_channels,
|
263 |
+
kernel_size=3,
|
264 |
+
padding=1,
|
265 |
+
)
|
266 |
+
|
267 |
+
@property
|
268 |
+
def attn_processors(self) -> Dict[str, AttentionProcessor]:
|
269 |
+
r"""
|
270 |
+
Returns:
|
271 |
+
`dict` of attention processors: A dictionary containing all attention processors used in the model with
|
272 |
+
indexed by its weight name.
|
273 |
+
"""
|
274 |
+
# set recursively
|
275 |
+
processors = {}
|
276 |
+
|
277 |
+
def fn_recursive_add_processors(
|
278 |
+
name: str,
|
279 |
+
module: torch.nn.Module,
|
280 |
+
processors: Dict[str, AttentionProcessor],
|
281 |
+
):
|
282 |
+
if hasattr(module, "get_processor"):
|
283 |
+
processors[f"{name}.processor"] = module.get_processor(
|
284 |
+
return_deprecated_lora=True
|
285 |
+
)
|
286 |
+
|
287 |
+
for sub_name, child in module.named_children():
|
288 |
+
fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)
|
289 |
+
|
290 |
+
return processors
|
291 |
+
|
292 |
+
for name, module in self.named_children():
|
293 |
+
fn_recursive_add_processors(name, module, processors)
|
294 |
+
|
295 |
+
return processors
|
296 |
+
|
297 |
+
def set_attn_processor(
|
298 |
+
self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]
|
299 |
+
):
|
300 |
+
r"""
|
301 |
+
Sets the attention processor to use to compute attention.
|
302 |
+
|
303 |
+
Parameters:
|
304 |
+
processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
|
305 |
+
The instantiated processor class or a dictionary of processor classes that will be set as the processor
|
306 |
+
for **all** `Attention` layers.
|
307 |
+
|
308 |
+
If `processor` is a dict, the key needs to define the path to the corresponding cross attention
|
309 |
+
processor. This is strongly recommended when setting trainable attention processors.
|
310 |
+
|
311 |
+
"""
|
312 |
+
count = len(self.attn_processors.keys())
|
313 |
+
|
314 |
+
if isinstance(processor, dict) and len(processor) != count:
|
315 |
+
raise ValueError(
|
316 |
+
f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
|
317 |
+
f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
|
318 |
+
)
|
319 |
+
|
320 |
+
def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor):
|
321 |
+
if hasattr(module, "set_processor"):
|
322 |
+
if not isinstance(processor, dict):
|
323 |
+
module.set_processor(processor)
|
324 |
+
else:
|
325 |
+
module.set_processor(processor.pop(f"{name}.processor"))
|
326 |
+
|
327 |
+
for sub_name, child in module.named_children():
|
328 |
+
fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
|
329 |
+
|
330 |
+
for name, module in self.named_children():
|
331 |
+
fn_recursive_attn_processor(name, module, processor)
|
332 |
+
|
333 |
+
def set_default_attn_processor(self):
|
334 |
+
"""
|
335 |
+
Disables custom attention processors and sets the default attention implementation.
|
336 |
+
"""
|
337 |
+
if all(
|
338 |
+
proc.__class__ in CROSS_ATTENTION_PROCESSORS
|
339 |
+
for proc in self.attn_processors.values()
|
340 |
+
):
|
341 |
+
processor = AttnProcessor()
|
342 |
+
else:
|
343 |
+
raise ValueError(
|
344 |
+
f"Cannot call `set_default_attn_processor` when attention processors are of type {next(iter(self.attn_processors.values()))}"
|
345 |
+
)
|
346 |
+
|
347 |
+
self.set_attn_processor(processor)
|
348 |
+
|
349 |
+
def _set_gradient_checkpointing(self, module, value=False):
|
350 |
+
if hasattr(module, "gradient_checkpointing"):
|
351 |
+
module.gradient_checkpointing = value
|
352 |
+
|
353 |
+
# Copied from diffusers.models.unets.unet_3d_condition.UNet3DConditionModel.enable_forward_chunking
|
354 |
+
def enable_forward_chunking(
|
355 |
+
self, chunk_size: Optional[int] = None, dim: int = 0
|
356 |
+
) -> None:
|
357 |
+
"""
|
358 |
+
Sets the attention processor to use [feed forward
|
359 |
+
chunking](https://huggingface.co/blog/reformer#2-chunked-feed-forward-layers).
|
360 |
+
|
361 |
+
Parameters:
|
362 |
+
chunk_size (`int`, *optional*):
|
363 |
+
The chunk size of the feed-forward layers. If not specified, will run feed-forward layer individually
|
364 |
+
over each tensor of dim=`dim`.
|
365 |
+
dim (`int`, *optional*, defaults to `0`):
|
366 |
+
The dimension over which the feed-forward computation should be chunked. Choose between dim=0 (batch)
|
367 |
+
or dim=1 (sequence length).
|
368 |
+
"""
|
369 |
+
if dim not in [0, 1]:
|
370 |
+
raise ValueError(f"Make sure to set `dim` to either 0 or 1, not {dim}")
|
371 |
+
|
372 |
+
# By default chunk size is 1
|
373 |
+
chunk_size = chunk_size or 1
|
374 |
+
|
375 |
+
def fn_recursive_feed_forward(
|
376 |
+
module: torch.nn.Module, chunk_size: int, dim: int
|
377 |
+
):
|
378 |
+
if hasattr(module, "set_chunk_feed_forward"):
|
379 |
+
module.set_chunk_feed_forward(chunk_size=chunk_size, dim=dim)
|
380 |
+
|
381 |
+
for child in module.children():
|
382 |
+
fn_recursive_feed_forward(child, chunk_size, dim)
|
383 |
+
|
384 |
+
for module in self.children():
|
385 |
+
fn_recursive_feed_forward(module, chunk_size, dim)
|
386 |
+
|
387 |
+
def forward(
|
388 |
+
self,
|
389 |
+
sample: torch.FloatTensor,
|
390 |
+
timestep: Union[torch.Tensor, float, int],
|
391 |
+
encoder_hidden_states: torch.Tensor,
|
392 |
+
return_dict: bool = True,
|
393 |
+
position_ids=None,
|
394 |
+
) -> Union[UNetSpatioTemporalRopeConditionOutput, Tuple]:
|
395 |
+
r"""
|
396 |
+
The [`UNetSpatioTemporalConditionModel`] forward method.
|
397 |
+
|
398 |
+
Args:
|
399 |
+
sample (`torch.FloatTensor`):
|
400 |
+
The noisy input tensor with the following shape `(batch, num_frames, channel, height, width)`.
|
401 |
+
timestep (`torch.FloatTensor` or `float` or `int`): The number of timesteps to denoise an input.
|
402 |
+
encoder_hidden_states (`torch.FloatTensor`):
|
403 |
+
The encoder hidden states with shape `(batch, sequence_length, cross_attention_dim)`.
|
404 |
+
return_dict (`bool`, *optional*, defaults to `True`):
|
405 |
+
Whether or not to return a [`~models.unet_slatio_temporal.UNetSpatioTemporalConditionOutput`] instead of a plain
|
406 |
+
tuple.
|
407 |
+
Returns:
|
408 |
+
[`~models.unet_slatio_temporal.UNetSpatioTemporalConditionOutput`] or `tuple`:
|
409 |
+
If `return_dict` is True, an [`~models.unet_slatio_temporal.UNetSpatioTemporalConditionOutput`] is returned, otherwise
|
410 |
+
a `tuple` is returned where the first element is the sample tensor.
|
411 |
+
"""
|
412 |
+
default_overall_up_factor = 2**self.num_upsamplers
|
413 |
+
|
414 |
+
# upsample size should be forwarded when sample is not a multiple of `default_overall_up_factor`
|
415 |
+
forward_upsample_size = False
|
416 |
+
upsample_size = None
|
417 |
+
|
418 |
+
for dim in sample.shape[-2:]:
|
419 |
+
if dim % default_overall_up_factor != 0:
|
420 |
+
# Forward upsample size to force interpolation output size.
|
421 |
+
forward_upsample_size = True
|
422 |
+
break
|
423 |
+
|
424 |
+
# 1. time
|
425 |
+
timesteps = timestep
|
426 |
+
if not torch.is_tensor(timesteps):
|
427 |
+
# TODO: this requires sync between CPU and GPU. So try to pass timesteps as tensors if you can
|
428 |
+
# This would be a good case for the `match` statement (Python 3.10+)
|
429 |
+
is_mps = sample.device.type == "mps"
|
430 |
+
if isinstance(timestep, float):
|
431 |
+
dtype = torch.float32 if is_mps else torch.float64
|
432 |
+
else:
|
433 |
+
dtype = torch.int32 if is_mps else torch.int64
|
434 |
+
timesteps = torch.tensor([timesteps], dtype=dtype, device=sample.device)
|
435 |
+
elif len(timesteps.shape) == 0:
|
436 |
+
timesteps = timesteps[None].to(sample.device)
|
437 |
+
|
438 |
+
# broadcast to batch dimension in a way that's compatible with ONNX/Core ML
|
439 |
+
batch_size, num_frames = sample.shape[:2]
|
440 |
+
timesteps = timesteps.expand(batch_size)
|
441 |
+
|
442 |
+
t_emb = self.time_proj(timesteps)
|
443 |
+
|
444 |
+
# `Timesteps` does not contain any weights and will always return f32 tensors
|
445 |
+
# but time_embedding might actually be running in fp16. so we need to cast here.
|
446 |
+
# there might be better ways to encapsulate this.
|
447 |
+
t_emb = t_emb.to(dtype=sample.dtype)
|
448 |
+
|
449 |
+
emb = self.time_embedding(t_emb)
|
450 |
+
|
451 |
+
# Flatten the batch and frames dimensions
|
452 |
+
# sample: [batch, frames, channels, height, width] -> [batch * frames, channels, height, width]
|
453 |
+
sample = sample.flatten(0, 1)
|
454 |
+
# Repeat the embeddings num_video_frames times
|
455 |
+
# emb: [batch, channels] -> [batch * frames, channels]
|
456 |
+
emb = emb.repeat_interleave(num_frames, dim=0)
|
457 |
+
# encoder_hidden_states: [batch, 1, channels] -> [batch * frames, 1, channels]
|
458 |
+
encoder_hidden_states = encoder_hidden_states.repeat_interleave(
|
459 |
+
num_frames, dim=0
|
460 |
+
)
|
461 |
+
|
462 |
+
# 2. pre-process
|
463 |
+
sample = self.conv_in(sample)
|
464 |
+
|
465 |
+
image_only_indicator = torch.zeros(
|
466 |
+
batch_size, num_frames, dtype=sample.dtype, device=sample.device
|
467 |
+
)
|
468 |
+
|
469 |
+
down_block_res_samples = (sample,)
|
470 |
+
for downsample_block in self.down_blocks:
|
471 |
+
if (
|
472 |
+
hasattr(downsample_block, "has_cross_attention")
|
473 |
+
and downsample_block.has_cross_attention
|
474 |
+
):
|
475 |
+
sample, res_samples = downsample_block(
|
476 |
+
hidden_states=sample,
|
477 |
+
temb=emb,
|
478 |
+
encoder_hidden_states=encoder_hidden_states,
|
479 |
+
image_only_indicator=image_only_indicator,
|
480 |
+
position_ids=position_ids,
|
481 |
+
)
|
482 |
+
else:
|
483 |
+
sample, res_samples = downsample_block(
|
484 |
+
hidden_states=sample,
|
485 |
+
temb=emb,
|
486 |
+
image_only_indicator=image_only_indicator,
|
487 |
+
position_ids=position_ids,
|
488 |
+
)
|
489 |
+
|
490 |
+
down_block_res_samples += res_samples
|
491 |
+
|
492 |
+
# 4. mid
|
493 |
+
sample = self.mid_block(
|
494 |
+
hidden_states=sample,
|
495 |
+
temb=emb,
|
496 |
+
encoder_hidden_states=encoder_hidden_states,
|
497 |
+
image_only_indicator=image_only_indicator,
|
498 |
+
position_ids=position_ids,
|
499 |
+
)
|
500 |
+
|
501 |
+
# 5. up
|
502 |
+
for i, upsample_block in enumerate(self.up_blocks):
|
503 |
+
is_final_block = i == len(self.up_blocks) - 1
|
504 |
+
|
505 |
+
res_samples = down_block_res_samples[-len(upsample_block.resnets) :]
|
506 |
+
down_block_res_samples = down_block_res_samples[
|
507 |
+
: -len(upsample_block.resnets)
|
508 |
+
]
|
509 |
+
if not is_final_block and forward_upsample_size:
|
510 |
+
upsample_size = down_block_res_samples[-1].shape[2:]
|
511 |
+
|
512 |
+
if (
|
513 |
+
hasattr(upsample_block, "has_cross_attention")
|
514 |
+
and upsample_block.has_cross_attention
|
515 |
+
):
|
516 |
+
sample = upsample_block(
|
517 |
+
hidden_states=sample,
|
518 |
+
temb=emb,
|
519 |
+
res_hidden_states_tuple=res_samples,
|
520 |
+
encoder_hidden_states=encoder_hidden_states,
|
521 |
+
image_only_indicator=image_only_indicator,
|
522 |
+
upsample_size=upsample_size,
|
523 |
+
position_ids=position_ids,
|
524 |
+
)
|
525 |
+
else:
|
526 |
+
sample = upsample_block(
|
527 |
+
hidden_states=sample,
|
528 |
+
temb=emb,
|
529 |
+
res_hidden_states_tuple=res_samples,
|
530 |
+
image_only_indicator=image_only_indicator,
|
531 |
+
upsample_size=upsample_size,
|
532 |
+
position_ids=position_ids,
|
533 |
+
)
|
534 |
+
|
535 |
+
# 6. post-process
|
536 |
+
sample = self.conv_norm_out(sample)
|
537 |
+
sample = self.conv_act(sample)
|
538 |
+
sample = self.conv_out(sample)
|
539 |
+
|
540 |
+
# 7. Reshape back to original shape
|
541 |
+
sample = sample.reshape(batch_size, num_frames, *sample.shape[1:])
|
542 |
+
|
543 |
+
if not return_dict:
|
544 |
+
return (sample,)
|
545 |
+
|
546 |
+
return UNetSpatioTemporalRopeConditionOutput(sample=sample)
|
pipelines/__init__.py
ADDED
@@ -0,0 +1,5 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from .dav_pipeline import DAVPipeline
|
2 |
+
|
3 |
+
__all__ = {
|
4 |
+
"DAVPipeline": DAVPipeline,
|
5 |
+
}
|
pipelines/dav_pipeline.py
ADDED
@@ -0,0 +1,246 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
import tqdm
|
3 |
+
import numpy as np
|
4 |
+
from diffusers import DiffusionPipeline
|
5 |
+
from diffusers.utils import BaseOutput
|
6 |
+
import matplotlib
|
7 |
+
|
8 |
+
|
9 |
+
def colorize_depth(depth, cmap="Spectral"):
|
10 |
+
# colorize
|
11 |
+
cm = matplotlib.colormaps[cmap]
|
12 |
+
# (B, N, H, W, 3)
|
13 |
+
depth_colored = cm(depth, bytes=False)[..., 0:3] # value from 0 to 1
|
14 |
+
return depth_colored
|
15 |
+
|
16 |
+
|
17 |
+
class DAVOutput(BaseOutput):
|
18 |
+
r"""
|
19 |
+
Output class for zero-shot text-to-video pipeline.
|
20 |
+
|
21 |
+
Args:
|
22 |
+
frames (`[List[PIL.Image.Image]`, `np.ndarray`]):
|
23 |
+
List of denoised PIL images of length `batch_size` or NumPy array of shape `(batch_size, height, width,
|
24 |
+
num_channels)`.
|
25 |
+
"""
|
26 |
+
|
27 |
+
disparity: np.ndarray
|
28 |
+
disparity_colored: np.ndarray
|
29 |
+
image: np.ndarray
|
30 |
+
|
31 |
+
|
32 |
+
class DAVPipeline(DiffusionPipeline):
|
33 |
+
def __init__(self, vae, unet, unet_interp, scheduler):
|
34 |
+
super().__init__()
|
35 |
+
self.register_modules(
|
36 |
+
vae=vae, unet=unet, unet_interp=unet_interp, scheduler=scheduler
|
37 |
+
)
|
38 |
+
|
39 |
+
def encode(self, input):
|
40 |
+
num_frames = input.shape[1]
|
41 |
+
input = input.flatten(0, 1)
|
42 |
+
latent = self.vae.encode(input.to(self.vae.dtype)).latent_dist.mode()
|
43 |
+
latent = latent * self.vae.config.scaling_factor
|
44 |
+
latent = latent.reshape(-1, num_frames, *latent.shape[1:])
|
45 |
+
return latent
|
46 |
+
|
47 |
+
def decode(self, latents, decode_chunk_size=16):
|
48 |
+
# [batch, frames, channels, height, width] -> [batch*frames, channels, height, width]
|
49 |
+
num_frames = latents.shape[1]
|
50 |
+
latents = latents.flatten(0, 1)
|
51 |
+
latents = latents / self.vae.config.scaling_factor
|
52 |
+
|
53 |
+
# decode decode_chunk_size frames at a time to avoid OOM
|
54 |
+
frames = []
|
55 |
+
for i in range(0, latents.shape[0], decode_chunk_size):
|
56 |
+
num_frames_in = latents[i : i + decode_chunk_size].shape[0]
|
57 |
+
frame = self.vae.decode(
|
58 |
+
latents[i : i + decode_chunk_size].to(self.vae.dtype),
|
59 |
+
num_frames=num_frames_in,
|
60 |
+
).sample
|
61 |
+
frames.append(frame)
|
62 |
+
frames = torch.cat(frames, dim=0)
|
63 |
+
|
64 |
+
# [batch, frames, channels, height, width]
|
65 |
+
frames = frames.reshape(-1, num_frames, *frames.shape[1:])
|
66 |
+
return frames.to(torch.float32)
|
67 |
+
|
68 |
+
def single_infer(self, rgb, position_ids=None, num_inference_steps=None):
|
69 |
+
rgb_latent = self.encode(rgb)
|
70 |
+
noise_latent = torch.randn_like(rgb_latent)
|
71 |
+
|
72 |
+
self.scheduler.set_timesteps(num_inference_steps, device=rgb.device)
|
73 |
+
timesteps = self.scheduler.timesteps
|
74 |
+
|
75 |
+
image_embeddings = torch.zeros((noise_latent.shape[0], 1, 1024)).to(
|
76 |
+
noise_latent
|
77 |
+
)
|
78 |
+
|
79 |
+
for i, t in enumerate(timesteps):
|
80 |
+
latent_model_input = noise_latent
|
81 |
+
|
82 |
+
latent_model_input = torch.cat([latent_model_input, rgb_latent], dim=2)
|
83 |
+
|
84 |
+
# [batch_size, num_frame, 4, h, w]
|
85 |
+
model_output = self.unet(
|
86 |
+
latent_model_input,
|
87 |
+
t,
|
88 |
+
encoder_hidden_states=image_embeddings,
|
89 |
+
position_ids=position_ids,
|
90 |
+
).sample
|
91 |
+
|
92 |
+
# compute the previous noisy sample x_t -> x_t-1
|
93 |
+
noise_latent = self.scheduler.step(
|
94 |
+
model_output, t, noise_latent
|
95 |
+
).prev_sample
|
96 |
+
|
97 |
+
return noise_latent
|
98 |
+
|
99 |
+
def single_interp_infer(
|
100 |
+
self, rgb, masked_depth_latent, mask, num_inference_steps=None
|
101 |
+
):
|
102 |
+
rgb_latent = self.encode(rgb)
|
103 |
+
noise_latent = torch.randn_like(rgb_latent)
|
104 |
+
|
105 |
+
self.scheduler.set_timesteps(num_inference_steps, device=rgb.device)
|
106 |
+
timesteps = self.scheduler.timesteps
|
107 |
+
|
108 |
+
image_embeddings = torch.zeros((noise_latent.shape[0], 1, 1024)).to(
|
109 |
+
noise_latent
|
110 |
+
)
|
111 |
+
|
112 |
+
for i, t in enumerate(timesteps):
|
113 |
+
latent_model_input = noise_latent
|
114 |
+
|
115 |
+
latent_model_input = torch.cat(
|
116 |
+
[latent_model_input, rgb_latent, masked_depth_latent, mask], dim=2
|
117 |
+
)
|
118 |
+
|
119 |
+
# [batch_size, num_frame, 4, h, w]
|
120 |
+
model_output = self.unet_interp(
|
121 |
+
latent_model_input, t, encoder_hidden_states=image_embeddings
|
122 |
+
).sample
|
123 |
+
|
124 |
+
# compute the previous noisy sample x_t -> x_t-1
|
125 |
+
noise_latent = self.scheduler.step(
|
126 |
+
model_output, t, noise_latent
|
127 |
+
).prev_sample
|
128 |
+
|
129 |
+
return noise_latent
|
130 |
+
|
131 |
+
def __call__(
|
132 |
+
self,
|
133 |
+
image,
|
134 |
+
num_frames,
|
135 |
+
num_overlap_frames,
|
136 |
+
num_interp_frames,
|
137 |
+
decode_chunk_size,
|
138 |
+
num_inference_steps,
|
139 |
+
):
|
140 |
+
self.vae.to(dtype=torch.float16)
|
141 |
+
|
142 |
+
# (1, N, 3, H, W)
|
143 |
+
image = image.unsqueeze(0)
|
144 |
+
B, N = image.shape[:2]
|
145 |
+
rgb = image * 2 - 1 # [-1, 1]
|
146 |
+
|
147 |
+
if N <= num_frames or N <= num_interp_frames + 2 - num_overlap_frames:
|
148 |
+
depth_latent = self.single_infer(
|
149 |
+
rgb, num_inference_steps=num_inference_steps
|
150 |
+
)
|
151 |
+
else:
|
152 |
+
assert 2 <= num_overlap_frames <= (num_interp_frames + 2 + 1) // 2
|
153 |
+
assert num_frames % 2 == 0
|
154 |
+
|
155 |
+
key_frame_indices = []
|
156 |
+
for i in range(0, N, num_interp_frames + 2 - num_overlap_frames):
|
157 |
+
if (
|
158 |
+
i + num_interp_frames + 1 >= N
|
159 |
+
or len(key_frame_indices) >= num_frames
|
160 |
+
):
|
161 |
+
break
|
162 |
+
key_frame_indices.append(i)
|
163 |
+
key_frame_indices.append(i + num_interp_frames + 1)
|
164 |
+
|
165 |
+
key_frame_indices = torch.tensor(key_frame_indices, device=rgb.device)
|
166 |
+
|
167 |
+
sorted_key_frame_indices, origin_indices = torch.sort(key_frame_indices)
|
168 |
+
key_rgb = rgb[:, sorted_key_frame_indices]
|
169 |
+
key_depth_latent = self.single_infer(
|
170 |
+
key_rgb,
|
171 |
+
sorted_key_frame_indices.unsqueeze(0).repeat(B, 1),
|
172 |
+
num_inference_steps=num_inference_steps,
|
173 |
+
)
|
174 |
+
key_depth_latent = key_depth_latent[:, origin_indices]
|
175 |
+
|
176 |
+
torch.cuda.empty_cache()
|
177 |
+
|
178 |
+
depth_latent = []
|
179 |
+
pre_latent = None
|
180 |
+
for i in tqdm.tqdm(range(0, len(key_frame_indices), 2)):
|
181 |
+
frame1 = key_depth_latent[:, i]
|
182 |
+
frame2 = key_depth_latent[:, i + 1]
|
183 |
+
masked_depth_latent = torch.zeros(
|
184 |
+
(B, num_interp_frames + 2, *key_depth_latent.shape[2:])
|
185 |
+
).to(key_depth_latent)
|
186 |
+
masked_depth_latent[:, 0] = frame1
|
187 |
+
masked_depth_latent[:, -1] = frame2
|
188 |
+
|
189 |
+
mask = torch.zeros_like(masked_depth_latent)
|
190 |
+
mask[:, [0, -1]] = 1.0
|
191 |
+
|
192 |
+
latent = self.single_interp_infer(
|
193 |
+
rgb[:, key_frame_indices[i] : key_frame_indices[i + 1] + 1],
|
194 |
+
masked_depth_latent,
|
195 |
+
mask,
|
196 |
+
num_inference_steps=num_inference_steps,
|
197 |
+
)
|
198 |
+
latent = latent[:, 1:-1]
|
199 |
+
|
200 |
+
if pre_latent is not None:
|
201 |
+
overlap_a = pre_latent[
|
202 |
+
:, pre_latent.shape[1] - (num_overlap_frames - 2) :
|
203 |
+
]
|
204 |
+
overlap_b = latent[:, : (num_overlap_frames - 2)]
|
205 |
+
ratio = (
|
206 |
+
torch.linspace(0, 1, num_overlap_frames - 2)
|
207 |
+
.to(overlap_a)
|
208 |
+
.view(1, -1, 1, 1, 1)
|
209 |
+
)
|
210 |
+
overlap = overlap_a * (1 - ratio) + overlap_b * ratio
|
211 |
+
pre_latent[:, pre_latent.shape[1] - (num_overlap_frames - 2) :] = (
|
212 |
+
overlap
|
213 |
+
)
|
214 |
+
depth_latent.append(pre_latent)
|
215 |
+
|
216 |
+
pre_latent = latent[:, (num_overlap_frames - 2) if i > 0 else 0 :]
|
217 |
+
|
218 |
+
torch.cuda.empty_cache()
|
219 |
+
|
220 |
+
depth_latent.append(pre_latent)
|
221 |
+
depth_latent = torch.cat(depth_latent, dim=1)
|
222 |
+
|
223 |
+
# dicard the first and last key frames
|
224 |
+
image = image[:, key_frame_indices[0] + 1 : key_frame_indices[-1]]
|
225 |
+
assert depth_latent.shape[1] == image.shape[1]
|
226 |
+
|
227 |
+
disparity = self.decode(depth_latent, decode_chunk_size=decode_chunk_size)
|
228 |
+
disparity = disparity.mean(dim=2, keepdim=False)
|
229 |
+
disparity = torch.clamp(disparity * 0.5 + 0.5, 0.0, 1.0)
|
230 |
+
|
231 |
+
# (N, H, W)
|
232 |
+
disparity = disparity.squeeze(0)
|
233 |
+
# (N, H, W, 3)
|
234 |
+
mid_d, max_d = disparity.min(), disparity.max()
|
235 |
+
disparity_colored = torch.clamp((max_d - disparity) / (max_d - mid_d), 0.0, 1.0)
|
236 |
+
disparity_colored = colorize_depth(disparity_colored.cpu().numpy())
|
237 |
+
disparity_colored = (disparity_colored * 255).astype(np.uint8)
|
238 |
+
image = image.squeeze(0).permute(0, 2, 3, 1).cpu().numpy()
|
239 |
+
image = (image * 255).astype(np.uint8)
|
240 |
+
disparity = disparity.cpu().numpy()
|
241 |
+
|
242 |
+
return DAVOutput(
|
243 |
+
disparity=disparity,
|
244 |
+
disparity_colored=disparity_colored,
|
245 |
+
image=image,
|
246 |
+
)
|
requirements.txt
ADDED
@@ -0,0 +1,12 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
torch==2.4.0
|
2 |
+
torchvision==0.19.0
|
3 |
+
diffusers==0.30.0
|
4 |
+
accelerate==0.31.0
|
5 |
+
transformers==4.43.2
|
6 |
+
huggingface-hub==0.24.2
|
7 |
+
opencv-python
|
8 |
+
tqdm
|
9 |
+
matplotlib
|
10 |
+
scipy
|
11 |
+
pillow
|
12 |
+
easydict
|
utils/img_utils.py
ADDED
@@ -0,0 +1,112 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from PIL import Image
|
2 |
+
import cv2
|
3 |
+
import numpy as np
|
4 |
+
import os
|
5 |
+
import tempfile
|
6 |
+
|
7 |
+
|
8 |
+
def resize(img, size):
|
9 |
+
assert img.dtype == np.uint8
|
10 |
+
pil_image = Image.fromarray(img)
|
11 |
+
pil_image = pil_image.resize(size, Image.LANCZOS)
|
12 |
+
resized_img = np.array(pil_image)
|
13 |
+
return resized_img
|
14 |
+
|
15 |
+
|
16 |
+
def crop(img, start_h, start_w, crop_h, crop_w):
|
17 |
+
img_src = np.zeros((crop_h, crop_w, *img.shape[2:]), dtype=img.dtype)
|
18 |
+
hsize, wsize = crop_h, crop_w
|
19 |
+
dh, dw, sh, sw = start_h, start_w, 0, 0
|
20 |
+
if dh < 0:
|
21 |
+
sh = -dh
|
22 |
+
hsize += dh
|
23 |
+
dh = 0
|
24 |
+
if dh + hsize > img.shape[0]:
|
25 |
+
hsize = img.shape[0] - dh
|
26 |
+
if dw < 0:
|
27 |
+
sw = -dw
|
28 |
+
wsize += dw
|
29 |
+
dw = 0
|
30 |
+
if dw + wsize > img.shape[1]:
|
31 |
+
wsize = img.shape[1] - dw
|
32 |
+
img_src[sh : sh + hsize, sw : sw + wsize] = img[dh : dh + hsize, dw : dw + wsize]
|
33 |
+
return img_src
|
34 |
+
|
35 |
+
|
36 |
+
def imresize_max(img, size, min_side=False):
|
37 |
+
new_img = []
|
38 |
+
for i, _img in enumerate(img):
|
39 |
+
h, w = _img.shape[:2]
|
40 |
+
ori_size = min(h, w) if min_side else max(h, w)
|
41 |
+
_resize = min(size / ori_size, 1.0)
|
42 |
+
new_size = (int(w * _resize), int(h * _resize))
|
43 |
+
new_img.append(resize(_img, new_size))
|
44 |
+
return new_img
|
45 |
+
|
46 |
+
|
47 |
+
def imcrop_multi(img, multiple=32):
|
48 |
+
new_img = []
|
49 |
+
for i, _img in enumerate(img):
|
50 |
+
crop_size = (
|
51 |
+
_img.shape[0] // multiple * multiple,
|
52 |
+
_img.shape[1] // multiple * multiple,
|
53 |
+
)
|
54 |
+
start_h = int(0.5 * max(0, _img.shape[0] - crop_size[0]))
|
55 |
+
start_w = int(0.5 * max(0, _img.shape[1] - crop_size[1]))
|
56 |
+
_img_src = crop(_img, start_h, start_w, crop_size[0], crop_size[1])
|
57 |
+
new_img.append(_img_src)
|
58 |
+
return new_img
|
59 |
+
|
60 |
+
|
61 |
+
def read_video(video_path, max_frames=None):
|
62 |
+
cap = cv2.VideoCapture(video_path)
|
63 |
+
fps = cap.get(cv2.CAP_PROP_FPS)
|
64 |
+
frames = []
|
65 |
+
count = 0
|
66 |
+
while True:
|
67 |
+
ret, frame = cap.read()
|
68 |
+
if not ret:
|
69 |
+
break
|
70 |
+
frame = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
|
71 |
+
frames.append(frame)
|
72 |
+
count += 1
|
73 |
+
if max_frames is not None and count >= max_frames:
|
74 |
+
break
|
75 |
+
cap.release()
|
76 |
+
# (N, H, W, 3)
|
77 |
+
return frames, fps
|
78 |
+
|
79 |
+
|
80 |
+
def read_image(image_path):
|
81 |
+
frame = cv2.imread(image_path)
|
82 |
+
frame = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
|
83 |
+
# (1, H, W, 3)
|
84 |
+
return [frame]
|
85 |
+
|
86 |
+
|
87 |
+
def write_video(video_path, frames, fps):
|
88 |
+
tmp_dir = os.path.join(os.path.dirname(video_path), "tmp")
|
89 |
+
os.makedirs(tmp_dir, exist_ok=True)
|
90 |
+
for i, frame in enumerate(frames):
|
91 |
+
write_image(os.path.join(tmp_dir, f"{i:06d}.png"), frame)
|
92 |
+
# it will cause visual compression artifacts
|
93 |
+
ffmpeg_command = [
|
94 |
+
"ffmpeg",
|
95 |
+
"-f",
|
96 |
+
"image2",
|
97 |
+
"-framerate",
|
98 |
+
f"{fps}",
|
99 |
+
"-i",
|
100 |
+
os.path.join(tmp_dir, "%06d.png"),
|
101 |
+
"-b:v",
|
102 |
+
"5626k",
|
103 |
+
"-y",
|
104 |
+
video_path,
|
105 |
+
]
|
106 |
+
os.system(" ".join(ffmpeg_command))
|
107 |
+
os.system(f"rm -rf {tmp_dir}")
|
108 |
+
|
109 |
+
|
110 |
+
def write_image(image_path, frame):
|
111 |
+
frame = cv2.cvtColor(frame, cv2.COLOR_RGB2BGR)
|
112 |
+
cv2.imwrite(image_path, frame)
|