452 lines
20 KiB
Python
452 lines
20 KiB
Python
|
|
import torch
|
|
|
|
import os
|
|
import json
|
|
|
|
from dataclasses import dataclass
|
|
from einops import rearrange, repeat
|
|
from typing import Any, Dict, Optional, Tuple
|
|
from diffusers.models import Transformer2DModel
|
|
from diffusers.utils import USE_PEFT_BACKEND, BaseOutput, deprecate
|
|
from diffusers.models.embeddings import get_1d_sincos_pos_embed_from_grid, ImagePositionalEmbeddings, CaptionProjection, \
|
|
PatchEmbed, CombinedTimestepSizeEmbeddings
|
|
from diffusers.configuration_utils import ConfigMixin, register_to_config
|
|
from diffusers.models.modeling_utils import ModelMixin
|
|
from diffusers.models.attention import BasicTransformerBlock
|
|
from diffusers.models.lora import LoRACompatibleConv, LoRACompatibleLinear
|
|
|
|
import torch
|
|
import torch.nn.functional as F
|
|
from torch import nn
|
|
from diffusers.utils.torch_utils import maybe_allow_in_graph
|
|
from diffusers.models.embeddings import SinusoidalPositionalEmbedding
|
|
from diffusers.models.normalization import AdaLayerNorm, AdaLayerNormZero
|
|
from diffusers.models.attention_processor import Attention
|
|
from diffusers.models.activations import GEGLU, GELU, ApproximateGELU
|
|
|
|
from dataclasses import dataclass
|
|
|
|
from torch import nn
|
|
|
|
|
|
@maybe_allow_in_graph
|
|
class GatedSelfAttentionDense(nn.Module):
|
|
r"""
|
|
A gated self-attention dense layer that combines visual features and object features.
|
|
|
|
Parameters:
|
|
query_dim (`int`): The number of channels in the query.
|
|
context_dim (`int`): The number of channels in the context.
|
|
n_heads (`int`): The number of heads to use for attention.
|
|
d_head (`int`): The number of channels in each head.
|
|
"""
|
|
|
|
def __init__(self, query_dim: int, context_dim: int, n_heads: int, d_head: int):
|
|
super().__init__()
|
|
|
|
# we need a linear projection since we need cat visual feature and obj feature
|
|
self.linear = nn.Linear(context_dim, query_dim)
|
|
|
|
self.attn = Attention(query_dim=query_dim, heads=n_heads, dim_head=d_head)
|
|
self.ff = FeedForward(query_dim, activation_fn="geglu")
|
|
|
|
self.norm1 = nn.LayerNorm(query_dim)
|
|
self.norm2 = nn.LayerNorm(query_dim)
|
|
|
|
self.register_parameter("alpha_attn", nn.Parameter(torch.tensor(0.0)))
|
|
self.register_parameter("alpha_dense", nn.Parameter(torch.tensor(0.0)))
|
|
|
|
self.enabled = True
|
|
|
|
def forward(self, x: torch.Tensor, objs: torch.Tensor) -> torch.Tensor:
|
|
if not self.enabled:
|
|
return x
|
|
|
|
n_visual = x.shape[1]
|
|
objs = self.linear(objs)
|
|
|
|
x = x + self.alpha_attn.tanh() * self.attn(self.norm1(torch.cat([x, objs], dim=1)))[:, :n_visual, :]
|
|
x = x + self.alpha_dense.tanh() * self.ff(self.norm2(x))
|
|
|
|
return x
|
|
|
|
|
|
class FeedForward(nn.Module):
|
|
r"""
|
|
A feed-forward layer.
|
|
|
|
Parameters:
|
|
dim (`int`): The number of channels in the input.
|
|
dim_out (`int`, *optional*): The number of channels in the output. If not given, defaults to `dim`.
|
|
mult (`int`, *optional*, defaults to 4): The multiplier to use for the hidden dimension.
|
|
dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
|
|
activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to be used in feed-forward.
|
|
final_dropout (`bool` *optional*, defaults to False): Apply a final dropout.
|
|
"""
|
|
|
|
def __init__(
|
|
self,
|
|
dim: int,
|
|
dim_out: Optional[int] = None,
|
|
mult: int = 4,
|
|
dropout: float = 0.0,
|
|
activation_fn: str = "geglu",
|
|
final_dropout: bool = False,
|
|
):
|
|
super().__init__()
|
|
inner_dim = int(dim * mult)
|
|
dim_out = dim_out if dim_out is not None else dim
|
|
linear_cls = LoRACompatibleLinear if not USE_PEFT_BACKEND else nn.Linear
|
|
|
|
if activation_fn == "gelu":
|
|
act_fn = GELU(dim, inner_dim)
|
|
if activation_fn == "gelu-approximate":
|
|
act_fn = GELU(dim, inner_dim, approximate="tanh")
|
|
elif activation_fn == "geglu":
|
|
act_fn = GEGLU(dim, inner_dim)
|
|
elif activation_fn == "geglu-approximate":
|
|
act_fn = ApproximateGELU(dim, inner_dim)
|
|
|
|
self.net = nn.ModuleList([])
|
|
# project in
|
|
self.net.append(act_fn)
|
|
# project dropout
|
|
self.net.append(nn.Dropout(dropout))
|
|
# project out
|
|
self.net.append(linear_cls(inner_dim, dim_out))
|
|
# FF as used in Vision Transformer, MLP-Mixer, etc. have a final dropout
|
|
if final_dropout:
|
|
self.net.append(nn.Dropout(dropout))
|
|
|
|
def forward(self, hidden_states: torch.Tensor, scale: float = 1.0) -> torch.Tensor:
|
|
compatible_cls = (GEGLU,) if USE_PEFT_BACKEND else (GEGLU, LoRACompatibleLinear)
|
|
for module in self.net:
|
|
if isinstance(module, compatible_cls):
|
|
hidden_states = module(hidden_states, scale)
|
|
else:
|
|
hidden_states = module(hidden_states)
|
|
return hidden_states
|
|
|
|
|
|
@maybe_allow_in_graph
|
|
class BasicTransformerBlock_(nn.Module):
|
|
r"""
|
|
A basic Transformer block.
|
|
|
|
Parameters:
|
|
dim (`int`): The number of channels in the input and output.
|
|
num_attention_heads (`int`): The number of heads to use for multi-head attention.
|
|
attention_head_dim (`int`): The number of channels in each head.
|
|
dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
|
|
cross_attention_dim (`int`, *optional*): The size of the encoder_hidden_states vector for cross attention.
|
|
activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to be used in feed-forward.
|
|
num_embeds_ada_norm (:
|
|
obj: `int`, *optional*): The number of diffusion steps used during training. See `Transformer2DModel`.
|
|
attention_bias (:
|
|
obj: `bool`, *optional*, defaults to `False`): Configure if the attentions should contain a bias parameter.
|
|
only_cross_attention (`bool`, *optional*):
|
|
Whether to use only cross-attention layers. In this case two cross attention layers are used.
|
|
double_self_attention (`bool`, *optional*):
|
|
Whether to use two self-attention layers. In this case no cross attention layers are used.
|
|
upcast_attention (`bool`, *optional*):
|
|
Whether to upcast the attention computation to float32. This is useful for mixed precision training.
|
|
norm_elementwise_affine (`bool`, *optional*, defaults to `True`):
|
|
Whether to use learnable elementwise affine parameters for normalization.
|
|
norm_type (`str`, *optional*, defaults to `"layer_norm"`):
|
|
The normalization layer to use. Can be `"layer_norm"`, `"ada_norm"` or `"ada_norm_zero"`.
|
|
final_dropout (`bool` *optional*, defaults to False):
|
|
Whether to apply a final dropout after the last feed-forward layer.
|
|
attention_type (`str`, *optional*, defaults to `"default"`):
|
|
The type of attention to use. Can be `"default"` or `"gated"` or `"gated-text-image"`.
|
|
positional_embeddings (`str`, *optional*, defaults to `None`):
|
|
The type of positional embeddings to apply to.
|
|
num_positional_embeddings (`int`, *optional*, defaults to `None`):
|
|
The maximum number of positional embeddings to apply.
|
|
"""
|
|
|
|
def __init__(
|
|
self,
|
|
dim: int,
|
|
num_attention_heads: int,
|
|
attention_head_dim: int,
|
|
dropout=0.0,
|
|
cross_attention_dim: Optional[int] = None,
|
|
activation_fn: str = "geglu",
|
|
num_embeds_ada_norm: Optional[int] = None,
|
|
attention_bias: bool = False,
|
|
only_cross_attention: bool = False,
|
|
double_self_attention: bool = False,
|
|
upcast_attention: bool = False,
|
|
norm_elementwise_affine: bool = True,
|
|
norm_type: str = "layer_norm", # 'layer_norm', 'ada_norm', 'ada_norm_zero', 'ada_norm_single'
|
|
norm_eps: float = 1e-5,
|
|
final_dropout: bool = False,
|
|
attention_type: str = "default",
|
|
positional_embeddings: Optional[str] = None,
|
|
num_positional_embeddings: Optional[int] = None,
|
|
):
|
|
super().__init__()
|
|
self.only_cross_attention = only_cross_attention
|
|
|
|
self.use_ada_layer_norm_zero = (num_embeds_ada_norm is not None) and norm_type == "ada_norm_zero"
|
|
self.use_ada_layer_norm = (num_embeds_ada_norm is not None) and norm_type == "ada_norm"
|
|
self.use_ada_layer_norm_single = norm_type == "ada_norm_single"
|
|
self.use_layer_norm = norm_type == "layer_norm"
|
|
|
|
if norm_type in ("ada_norm", "ada_norm_zero") and num_embeds_ada_norm is None:
|
|
raise ValueError(
|
|
f"`norm_type` is set to {norm_type}, but `num_embeds_ada_norm` is not defined. Please make sure to"
|
|
f" define `num_embeds_ada_norm` if setting `norm_type` to {norm_type}."
|
|
)
|
|
|
|
if positional_embeddings and (num_positional_embeddings is None):
|
|
raise ValueError(
|
|
"If `positional_embedding` type is defined, `num_positition_embeddings` must also be defined."
|
|
)
|
|
|
|
if positional_embeddings == "sinusoidal":
|
|
self.pos_embed = SinusoidalPositionalEmbedding(dim, max_seq_length=num_positional_embeddings)
|
|
else:
|
|
self.pos_embed = None
|
|
|
|
# Define 3 blocks. Each block has its own normalization layer.
|
|
# 1. Self-Attn
|
|
if self.use_ada_layer_norm:
|
|
self.norm1 = AdaLayerNorm(dim, num_embeds_ada_norm)
|
|
elif self.use_ada_layer_norm_zero:
|
|
self.norm1 = AdaLayerNormZero(dim, num_embeds_ada_norm)
|
|
else:
|
|
self.norm1 = nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine, eps=norm_eps)
|
|
|
|
self.attn1 = Attention(
|
|
query_dim=dim,
|
|
heads=num_attention_heads,
|
|
dim_head=attention_head_dim,
|
|
dropout=dropout,
|
|
bias=attention_bias,
|
|
cross_attention_dim=cross_attention_dim if only_cross_attention else None,
|
|
upcast_attention=upcast_attention,
|
|
)
|
|
|
|
# # 2. Cross-Attn
|
|
# if cross_attention_dim is not None or double_self_attention:
|
|
# # We currently only use AdaLayerNormZero for self attention where there will only be one attention block.
|
|
# # I.e. the number of returned modulation chunks from AdaLayerZero would not make sense if returned during
|
|
# # the second cross attention block.
|
|
# self.norm2 = (
|
|
# AdaLayerNorm(dim, num_embeds_ada_norm)
|
|
# if self.use_ada_layer_norm
|
|
# else nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine, eps=norm_eps)
|
|
# )
|
|
# self.attn2 = Attention(
|
|
# query_dim=dim,
|
|
# cross_attention_dim=cross_attention_dim if not double_self_attention else None,
|
|
# heads=num_attention_heads,
|
|
# dim_head=attention_head_dim,
|
|
# dropout=dropout,
|
|
# bias=attention_bias,
|
|
# upcast_attention=upcast_attention,
|
|
# ) # is self-attn if encoder_hidden_states is none
|
|
# else:
|
|
# self.norm2 = None
|
|
# self.attn2 = None
|
|
|
|
# 3. Feed-forward
|
|
# if not self.use_ada_layer_norm_single:
|
|
# self.norm3 = nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine, eps=norm_eps)
|
|
self.norm3 = nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine, eps=norm_eps)
|
|
|
|
self.ff = FeedForward(dim, dropout=dropout, activation_fn=activation_fn, final_dropout=final_dropout)
|
|
|
|
# 4. Fuser
|
|
if attention_type == "gated" or attention_type == "gated-text-image":
|
|
self.fuser = GatedSelfAttentionDense(dim, cross_attention_dim, num_attention_heads, attention_head_dim)
|
|
|
|
# 5. Scale-shift for PixArt-Alpha.
|
|
if self.use_ada_layer_norm_single:
|
|
self.scale_shift_table = nn.Parameter(torch.randn(6, dim) / dim ** 0.5)
|
|
|
|
# let chunk size default to None
|
|
self._chunk_size = None
|
|
self._chunk_dim = 0
|
|
|
|
def set_chunk_feed_forward(self, chunk_size: Optional[int], dim: int):
|
|
# Sets chunk feed-forward
|
|
self._chunk_size = chunk_size
|
|
self._chunk_dim = dim
|
|
|
|
def forward(
|
|
self,
|
|
hidden_states: torch.FloatTensor,
|
|
attention_mask: Optional[torch.FloatTensor] = None,
|
|
encoder_hidden_states: Optional[torch.FloatTensor] = None,
|
|
encoder_attention_mask: Optional[torch.FloatTensor] = None,
|
|
timestep: Optional[torch.LongTensor] = None,
|
|
cross_attention_kwargs: Dict[str, Any] = None,
|
|
class_labels: Optional[torch.LongTensor] = None,
|
|
) -> torch.FloatTensor:
|
|
# Notice that normalization is always applied before the real computation in the following blocks.
|
|
# 0. Self-Attention
|
|
batch_size = hidden_states.shape[0]
|
|
|
|
if self.use_ada_layer_norm:
|
|
norm_hidden_states = self.norm1(hidden_states, timestep)
|
|
elif self.use_ada_layer_norm_zero:
|
|
norm_hidden_states, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.norm1(
|
|
hidden_states, timestep, class_labels, hidden_dtype=hidden_states.dtype
|
|
)
|
|
elif self.use_layer_norm:
|
|
norm_hidden_states = self.norm1(hidden_states)
|
|
elif self.use_ada_layer_norm_single:
|
|
shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = (
|
|
self.scale_shift_table[None] + timestep.reshape(batch_size, 6, -1)
|
|
).chunk(6, dim=1)
|
|
norm_hidden_states = self.norm1(hidden_states)
|
|
norm_hidden_states = norm_hidden_states * (1 + scale_msa) + shift_msa
|
|
norm_hidden_states = norm_hidden_states.squeeze(1)
|
|
else:
|
|
raise ValueError("Incorrect norm used")
|
|
|
|
if self.pos_embed is not None:
|
|
norm_hidden_states = self.pos_embed(norm_hidden_states)
|
|
|
|
# 1. Retrieve lora scale.
|
|
lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0
|
|
|
|
# 2. Prepare GLIGEN inputs
|
|
cross_attention_kwargs = cross_attention_kwargs.copy() if cross_attention_kwargs is not None else {}
|
|
gligen_kwargs = cross_attention_kwargs.pop("gligen", None)
|
|
|
|
attn_output = self.attn1(
|
|
norm_hidden_states,
|
|
encoder_hidden_states=encoder_hidden_states if self.only_cross_attention else None,
|
|
attention_mask=attention_mask,
|
|
**cross_attention_kwargs,
|
|
)
|
|
if self.use_ada_layer_norm_zero:
|
|
attn_output = gate_msa.unsqueeze(1) * attn_output
|
|
elif self.use_ada_layer_norm_single:
|
|
attn_output = gate_msa * attn_output
|
|
|
|
hidden_states = attn_output + hidden_states
|
|
if hidden_states.ndim == 4:
|
|
hidden_states = hidden_states.squeeze(1)
|
|
|
|
# 2.5 GLIGEN Control
|
|
if gligen_kwargs is not None:
|
|
hidden_states = self.fuser(hidden_states, gligen_kwargs["objs"])
|
|
|
|
# # 3. Cross-Attention
|
|
# if self.attn2 is not None:
|
|
# if self.use_ada_layer_norm:
|
|
# norm_hidden_states = self.norm2(hidden_states, timestep)
|
|
# elif self.use_ada_layer_norm_zero or self.use_layer_norm:
|
|
# norm_hidden_states = self.norm2(hidden_states)
|
|
# elif self.use_ada_layer_norm_single:
|
|
# # For PixArt norm2 isn't applied here:
|
|
# # https://github.com/PixArt-alpha/PixArt-alpha/blob/0f55e922376d8b797edd44d25d0e7464b260dcab/diffusion/model/nets/PixArtMS.py#L70C1-L76C103
|
|
# norm_hidden_states = hidden_states
|
|
# else:
|
|
# raise ValueError("Incorrect norm")
|
|
|
|
# if self.pos_embed is not None and self.use_ada_layer_norm_single is False:
|
|
# norm_hidden_states = self.pos_embed(norm_hidden_states)
|
|
|
|
# attn_output = self.attn2(
|
|
# norm_hidden_states,
|
|
# encoder_hidden_states=encoder_hidden_states,
|
|
# attention_mask=encoder_attention_mask,
|
|
# **cross_attention_kwargs,
|
|
# )
|
|
# hidden_states = attn_output + hidden_states
|
|
|
|
# 4. Feed-forward
|
|
# if not self.use_ada_layer_norm_single:
|
|
# norm_hidden_states = self.norm3(hidden_states)
|
|
|
|
if self.use_ada_layer_norm_zero:
|
|
norm_hidden_states = norm_hidden_states * (1 + scale_mlp[:, None]) + shift_mlp[:, None]
|
|
|
|
if self.use_ada_layer_norm_single:
|
|
# norm_hidden_states = self.norm2(hidden_states)
|
|
norm_hidden_states = self.norm3(hidden_states)
|
|
norm_hidden_states = norm_hidden_states * (1 + scale_mlp) + shift_mlp
|
|
|
|
if self._chunk_size is not None:
|
|
# "feed_forward_chunk_size" can be used to save memory
|
|
if norm_hidden_states.shape[self._chunk_dim] % self._chunk_size != 0:
|
|
raise ValueError(
|
|
f"`hidden_states` dimension to be chunked: {norm_hidden_states.shape[self._chunk_dim]} has to be divisible by chunk size: {self._chunk_size}. Make sure to set an appropriate `chunk_size` when calling `unet.enable_forward_chunking`."
|
|
)
|
|
|
|
num_chunks = norm_hidden_states.shape[self._chunk_dim] // self._chunk_size
|
|
ff_output = torch.cat(
|
|
[
|
|
self.ff(hid_slice, scale=lora_scale)
|
|
for hid_slice in norm_hidden_states.chunk(num_chunks, dim=self._chunk_dim)
|
|
],
|
|
dim=self._chunk_dim,
|
|
)
|
|
else:
|
|
ff_output = self.ff(norm_hidden_states, scale=lora_scale)
|
|
|
|
if self.use_ada_layer_norm_zero:
|
|
ff_output = gate_mlp.unsqueeze(1) * ff_output
|
|
elif self.use_ada_layer_norm_single:
|
|
ff_output = gate_mlp * ff_output
|
|
|
|
hidden_states = ff_output + hidden_states
|
|
if hidden_states.ndim == 4:
|
|
hidden_states = hidden_states.squeeze(1)
|
|
|
|
return hidden_states
|
|
|
|
|
|
class AdaLayerNormSingle(nn.Module):
|
|
r"""
|
|
Norm layer adaptive layer norm single (adaLN-single).
|
|
|
|
As proposed in PixArt-Alpha (see: https://arxiv.org/abs/2310.00426; Section 2.3).
|
|
|
|
Parameters:
|
|
embedding_dim (`int`): The size of each embedding vector.
|
|
use_additional_conditions (`bool`): To use additional conditions for normalization or not.
|
|
"""
|
|
|
|
def __init__(self, embedding_dim: int, use_additional_conditions: bool = False):
|
|
super().__init__()
|
|
|
|
self.emb = CombinedTimestepSizeEmbeddings(
|
|
embedding_dim, size_emb_dim=embedding_dim // 3, use_additional_conditions=use_additional_conditions
|
|
)
|
|
|
|
self.silu = nn.SiLU()
|
|
self.linear = nn.Linear(embedding_dim, 6 * embedding_dim, bias=True)
|
|
|
|
def forward(
|
|
self,
|
|
timestep: torch.Tensor,
|
|
added_cond_kwargs: Dict[str, torch.Tensor] = None,
|
|
batch_size: int = None,
|
|
hidden_dtype: Optional[torch.dtype] = None,
|
|
) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
|
|
# No modulation happening here.
|
|
embedded_timestep = self.emb(timestep, batch_size=batch_size, hidden_dtype=hidden_dtype, resolution=None,
|
|
aspect_ratio=None)
|
|
return self.linear(self.silu(embedded_timestep)), embedded_timestep
|
|
|
|
|
|
@dataclass
|
|
class Transformer3DModelOutput(BaseOutput):
|
|
"""
|
|
The output of [`Transformer2DModel`].
|
|
|
|
Args:
|
|
sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)` or `(batch size, num_vector_embeds - 1, num_latent_pixels)` if [`Transformer2DModel`] is discrete):
|
|
The hidden states output conditioned on the `encoder_hidden_states` input. If discrete, returns probability
|
|
distributions for the unnoised latent pixels.
|
|
"""
|
|
|
|
sample: torch.FloatTensor
|