Ailusion-VTON-DEMO-v1 / ip_adapter /attention_processor.py
Roopansh's picture
Initial Commit
73c83cf
# modified from https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py
import torch
import torch.nn as nn
import torch.nn.functional as F
from diffusers.models.lora import LoRACompatibleLinear
from diffusers.models.lora import LoRALinearLayer,LoRAConv2dLayer
from einops import rearrange
from diffusers.models.transformer_2d import Transformer2DModel
class AttnProcessor(nn.Module):
r"""
Default processor for performing attention-related computations.
"""
def __init__(
self,
hidden_size=None,
cross_attention_dim=None,
):
super().__init__()
def __call__(
self,
attn,
hidden_states,
encoder_hidden_states=None,
attention_mask=None,
temb=None,
):
residual = hidden_states
if attn.spatial_norm is not None:
hidden_states = attn.spatial_norm(hidden_states, temb)
input_ndim = hidden_states.ndim
if input_ndim == 4:
batch_size, channel, height, width = hidden_states.shape
hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
batch_size, sequence_length, _ = (
hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
)
attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
if attn.group_norm is not None:
hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
query = attn.to_q(hidden_states)
if encoder_hidden_states is None:
encoder_hidden_states = hidden_states
elif attn.norm_cross:
encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
key = attn.to_k(encoder_hidden_states)
value = attn.to_v(encoder_hidden_states)
query = attn.head_to_batch_dim(query)
key = attn.head_to_batch_dim(key)
value = attn.head_to_batch_dim(value)
attention_probs = attn.get_attention_scores(query, key, attention_mask)
hidden_states = torch.bmm(attention_probs, value)
hidden_states = attn.batch_to_head_dim(hidden_states)
# linear proj
hidden_states = attn.to_out[0](hidden_states)
# dropout
hidden_states = attn.to_out[1](hidden_states)
if input_ndim == 4:
hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
if attn.residual_connection:
hidden_states = hidden_states + residual
hidden_states = hidden_states / attn.rescale_output_factor
return hidden_states
class IPAttnProcessor(nn.Module):
r"""
Attention processor for IP-Adapater.
Args:
hidden_size (`int`):
The hidden size of the attention layer.
cross_attention_dim (`int`):
The number of channels in the `encoder_hidden_states`.
scale (`float`, defaults to 1.0):
the weight scale of image prompt.
num_tokens (`int`, defaults to 4 when do ip_adapter_plus it should be 16):
The context length of the image features.
"""
def __init__(self, hidden_size, cross_attention_dim=None, scale=1.0, num_tokens=4):
super().__init__()
self.hidden_size = hidden_size
self.cross_attention_dim = cross_attention_dim
self.scale = scale
self.num_tokens = num_tokens
self.to_k_ip = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)
self.to_v_ip = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)
def __call__(
self,
attn,
hidden_states,
encoder_hidden_states=None,
attention_mask=None,
temb=None,
):
residual = hidden_states
if attn.spatial_norm is not None:
hidden_states = attn.spatial_norm(hidden_states, temb)
input_ndim = hidden_states.ndim
if input_ndim == 4:
batch_size, channel, height, width = hidden_states.shape
hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
batch_size, sequence_length, _ = (
hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
)
attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
if attn.group_norm is not None:
hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
query = attn.to_q(hidden_states)
if encoder_hidden_states is None:
encoder_hidden_states = hidden_states
else:
# get encoder_hidden_states, ip_hidden_states
end_pos = encoder_hidden_states.shape[1] - self.num_tokens
encoder_hidden_states, ip_hidden_states = (
encoder_hidden_states[:, :end_pos, :],
encoder_hidden_states[:, end_pos:, :],
)
if attn.norm_cross:
encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
key = attn.to_k(encoder_hidden_states)
value = attn.to_v(encoder_hidden_states)
query = attn.head_to_batch_dim(query)
key = attn.head_to_batch_dim(key)
value = attn.head_to_batch_dim(value)
attention_probs = attn.get_attention_scores(query, key, attention_mask)
hidden_states = torch.bmm(attention_probs, value)
hidden_states = attn.batch_to_head_dim(hidden_states)
# for ip-adapter
ip_key = self.to_k_ip(ip_hidden_states)
ip_value = self.to_v_ip(ip_hidden_states)
ip_key = attn.head_to_batch_dim(ip_key)
ip_value = attn.head_to_batch_dim(ip_value)
ip_attention_probs = attn.get_attention_scores(query, ip_key, None)
ip_hidden_states = torch.bmm(ip_attention_probs, ip_value)
ip_hidden_states = attn.batch_to_head_dim(ip_hidden_states)
hidden_states = hidden_states + self.scale * ip_hidden_states
# linear proj
hidden_states = attn.to_out[0](hidden_states)
# dropout
hidden_states = attn.to_out[1](hidden_states)
if input_ndim == 4:
hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
if attn.residual_connection:
hidden_states = hidden_states + residual
hidden_states = hidden_states / attn.rescale_output_factor
return hidden_states
class AttnProcessor2_0(torch.nn.Module):
r"""
Processor for implementing scaled dot-product attention (enabled by default if you're using PyTorch 2.0).
"""
def __init__(
self,
hidden_size=None,
cross_attention_dim=None,
):
super().__init__()
if not hasattr(F, "scaled_dot_product_attention"):
raise ImportError("AttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.")
def __call__(
self,
attn,
hidden_states,
encoder_hidden_states=None,
attention_mask=None,
temb=None,
scale= 1.0,
):
residual = hidden_states
if attn.spatial_norm is not None:
hidden_states = attn.spatial_norm(hidden_states, temb)
input_ndim = hidden_states.ndim
if input_ndim == 4:
batch_size, channel, height, width = hidden_states.shape
hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
batch_size, sequence_length, _ = (
hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
)
if attention_mask is not None:
attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
# scaled_dot_product_attention expects attention_mask shape to be
# (batch, heads, source_length, target_length)
attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1])
if attn.group_norm is not None:
hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
# args = (scale, )
args = ()
query = attn.to_q(hidden_states, *args)
if encoder_hidden_states is None:
encoder_hidden_states = hidden_states
elif attn.norm_cross:
encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
key = attn.to_k(encoder_hidden_states, *args)
value = attn.to_v(encoder_hidden_states, *args)
inner_dim = key.shape[-1]
head_dim = inner_dim // attn.heads
query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
# the output of sdp = (batch, num_heads, seq_len, head_dim)
# TODO: add support for attn.scale when we move to Torch 2.1
hidden_states = F.scaled_dot_product_attention(
query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
)
hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
hidden_states = hidden_states.to(query.dtype)
# linear proj
hidden_states = attn.to_out[0](hidden_states, *args)
# dropout
hidden_states = attn.to_out[1](hidden_states)
if input_ndim == 4:
hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
if attn.residual_connection:
hidden_states = hidden_states + residual
hidden_states = hidden_states / attn.rescale_output_factor
return hidden_states
class AttnProcessor2_0_attn(torch.nn.Module):
r"""
Processor for implementing scaled dot-product attention (enabled by default if you're using PyTorch 2.0).
"""
def __init__(
self,
hidden_size=None,
cross_attention_dim=None,
):
super().__init__()
if not hasattr(F, "scaled_dot_product_attention"):
raise ImportError("AttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.")
def __call__(
self,
attn,
hidden_states,
encoder_hidden_states=None,
attention_mask=None,
temb=None,
is_cloth_pass=False,
cloth = None,
up_cnt=None,
mid_cnt=None,
down_cnt=None,
inside_up=None,
inside_down=None,
cloth_text=None,
):
residual = hidden_states
if attn.spatial_norm is not None:
hidden_states = attn.spatial_norm(hidden_states, temb)
input_ndim = hidden_states.ndim
if input_ndim == 4:
batch_size, channel, height, width = hidden_states.shape
hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
batch_size, sequence_length, _ = (
hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
)
if attention_mask is not None:
attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
# scaled_dot_product_attention expects attention_mask shape to be
# (batch, heads, source_length, target_length)
attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1])
if attn.group_norm is not None:
hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
query = attn.to_q(hidden_states)
if encoder_hidden_states is None:
encoder_hidden_states = hidden_states
elif attn.norm_cross:
encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
key = attn.to_k(encoder_hidden_states)
value = attn.to_v(encoder_hidden_states)
inner_dim = key.shape[-1]
head_dim = inner_dim // attn.heads
query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
# the output of sdp = (batch, num_heads, seq_len, head_dim)
# TODO: add support for attn.scale when we move to Torch 2.1
hidden_states = F.scaled_dot_product_attention(
query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
)
hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
hidden_states = hidden_states.to(query.dtype)
# linear proj
hidden_states = attn.to_out[0](hidden_states)
# dropout
hidden_states = attn.to_out[1](hidden_states)
if input_ndim == 4:
hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
if attn.residual_connection:
hidden_states = hidden_states + residual
hidden_states = hidden_states / attn.rescale_output_factor
return hidden_states
class AttnProcessor2_0_Lora(torch.nn.Module):
r"""
Processor for implementing scaled dot-product attention (enabled by default if you're using PyTorch 2.0).
"""
def __init__(
self,
scale_lora =1.0,
hidden_size=None,
cross_attention_dim=None,
):
super().__init__()
if not hasattr(F, "scaled_dot_product_attention"):
raise ImportError("AttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.")
self.scale_lora = scale_lora
def __call__(
self,
attn,
hidden_states,
encoder_hidden_states=None,
attention_mask=None,
temb=None,
):
residual = hidden_states
if attn.spatial_norm is not None:
hidden_states = attn.spatial_norm(hidden_states, temb)
input_ndim = hidden_states.ndim
if input_ndim == 4:
batch_size, channel, height, width = hidden_states.shape
hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
batch_size, sequence_length, _ = (
hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
)
if attention_mask is not None:
attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
# scaled_dot_product_attention expects attention_mask shape to be
# (batch, heads, source_length, target_length)
attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1])
if attn.group_norm is not None:
hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
if hasattr(attn,'q_lora'):
query = attn.to_q(hidden_states)
q_lora = attn.q_lora(hidden_states)
query = query + self.scale_lora * q_lora
else:
query = attn.to_q(hidden_states)
if encoder_hidden_states is None:
encoder_hidden_states = hidden_states
elif attn.norm_cross:
encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
if hasattr(attn,'k_lora'):
key = attn.to_k(hidden_states)
k_lora = attn.k_lora(hidden_states)
key = key + self.scale_lora * k_lora
else:
key = attn.to_k(hidden_states)
if hasattr(attn,'v_lora'):
value = attn.to_v(encoder_hidden_states)
v_lora = attn.v_lora(hidden_states)
value = value + self.scale_lora * v_lora
else:
value = attn.to_v(encoder_hidden_states)
inner_dim = key.shape[-1]
head_dim = inner_dim // attn.heads
query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
# the output of sdp = (batch, num_heads, seq_len, head_dim)
# TODO: add support for attn.scale when we move to Torch 2.1
hidden_states = F.scaled_dot_product_attention(
query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
)
hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
hidden_states = hidden_states.to(query.dtype)
# linear proj
if hasattr(attn,'out_lora'):
hidden_states = attn.to_out[0](hidden_states)
out_lora = attn.out_lora(hidden_states)
hidden_states = hidden_states+ self.scale_lora*out_lora
else:
hidden_states = attn.to_out[0](hidden_states)
# dropout
hidden_states = attn.to_out[1](hidden_states)
if input_ndim == 4:
hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
if attn.residual_connection:
hidden_states = hidden_states + residual
hidden_states = hidden_states / attn.rescale_output_factor
return hidden_states
class IPAttnProcessor_clothpass_noip(torch.nn.Module):
r"""
Attention processor for IP-Adapater for PyTorch 2.0.
Args:
hidden_size (`int`):
The hidden size of the attention layer.
cross_attention_dim (`int`):
The number of channels in the `encoder_hidden_states`.
scale (`float`, defaults to 1.0):
the weight scale of image prompt.
num_tokens (`int`, defaults to 4 when do ip_adapter_plus it should be 16):
The context length of the image features.
"""
def __init__(self, hidden_size, cross_attention_dim=None, scale=1.0, num_tokens=4):
super().__init__()
if not hasattr(F, "scaled_dot_product_attention"):
raise ImportError("AttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.")
self.hidden_size = hidden_size
self.cross_attention_dim = cross_attention_dim
self.scale = scale
self.num_tokens = num_tokens
self.to_k_ip = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)
self.to_v_ip = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)
def __call__(
self,
attn,
hidden_states,
encoder_hidden_states=None,
attention_mask=None,
temb=None,
is_cloth_pass=False,
cloth = None,
up_cnt=None,
mid_cnt=None,
down_cnt=None,
inside=None,
):
if is_cloth_pass or up_cnt is None:
residual = hidden_states
if attn.spatial_norm is not None:
hidden_states = attn.spatial_norm(hidden_states, temb)
input_ndim = hidden_states.ndim
if input_ndim == 4:
batch_size, channel, height, width = hidden_states.shape
hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
batch_size, sequence_length, _ = (
hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
)
if attention_mask is not None:
attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
# scaled_dot_product_attention expects attention_mask shape to be
# (batch, heads, source_length, target_length)
attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1])
if attn.group_norm is not None:
hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
query = attn.to_q(hidden_states)
if encoder_hidden_states is None:
encoder_hidden_states = hidden_states
elif attn.norm_cross:
encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
key = attn.to_k(encoder_hidden_states)
value = attn.to_v(encoder_hidden_states)
inner_dim = key.shape[-1]
head_dim = inner_dim // attn.heads
query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
# the output of sdp = (batch, num_heads, seq_len, head_dim)
# TODO: add support for attn.scale when we move to Torch 2.1
hidden_states = F.scaled_dot_product_attention(
query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
)
hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
hidden_states = hidden_states.to(query.dtype)
# linear proj
hidden_states = attn.to_out[0](hidden_states)
# dropout
hidden_states = attn.to_out[1](hidden_states)
if input_ndim == 4:
hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
if attn.residual_connection:
hidden_states = hidden_states + residual
hidden_states = hidden_states / attn.rescale_output_factor
return hidden_states
else:
residual = hidden_states
if attn.spatial_norm is not None:
hidden_states = attn.spatial_norm(hidden_states, temb)
input_ndim = hidden_states.ndim
if input_ndim == 4:
batch_size, channel, height, width = hidden_states.shape
hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
batch_size, sequence_length, _ = (
hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
)
if attention_mask is not None:
print('!!!!attention_mask is not NoNE!!!!')
attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
# scaled_dot_product_attention expects attention_mask shape to be
# (batch, heads, source_length, target_length)
attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1])
if attn.group_norm is not None:
hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
query = attn.to_q(hidden_states)
if encoder_hidden_states is None:
encoder_hidden_states = hidden_states
elif attn.norm_cross:
encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
key = attn.to_k(encoder_hidden_states)
value = attn.to_v(encoder_hidden_states)
inner_dim = key.shape[-1]
head_dim = inner_dim // attn.heads
query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
# the output of sdp = (batch, num_heads, seq_len, head_dim)
# TODO: add support for attn.scale when we move to Torch 2.1
hidden_states = F.scaled_dot_product_attention(
query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
)
hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
hidden_states = hidden_states.to(query.dtype)
# for ip-adapter
# print(up_cnt*3 + inside)
cloth_feature = cloth[up_cnt*3 + inside-1]
cloth_feature = rearrange(cloth_feature, "b c h w -> b (h w) c").contiguous()
# print(cloth_feature.shape)
# print(self.hidden_size)
c_key = self.to_k_c(cloth_feature)
c_value = self.to_v_c(cloth_feature)
c_key = c_key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
c_value = c_value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
# print(ip_value.shape)
#$$ attn_mask?
hidden_states_cloth = F.scaled_dot_product_attention(
query, c_key, c_value, attn_mask=None, dropout_p=0.0, is_causal=False
)
hidden_states_cloth = hidden_states_cloth.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
hidden_states_cloth = hidden_states_cloth.to(query.dtype)
hidden_states = hidden_states + self.scale * hidden_states_cloth
# linear proj
hidden_states = attn.to_out[0](hidden_states)
# dropout
hidden_states = attn.to_out[1](hidden_states)
if input_ndim == 4:
hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
if attn.residual_connection:
hidden_states = hidden_states + residual
hidden_states = hidden_states / attn.rescale_output_factor
return hidden_states
class IPAttnProcessor_clothpass(torch.nn.Module):
r"""
Attention processor for IP-Adapater for PyTorch 2.0.
Args:
hidden_size (`int`):
The hidden size of the attention layer.
cross_attention_dim (`int`):
The number of channels in the `encoder_hidden_states`.
scale (`float`, defaults to 1.0):
the weight scale of image prompt.
num_tokens (`int`, defaults to 4 when do ip_adapter_plus it should be 16):
The context length of the image features.
"""
def __init__(self, hidden_size, cross_attention_dim=None, scale=1.0, num_tokens=4):
super().__init__()
if not hasattr(F, "scaled_dot_product_attention"):
raise ImportError("AttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.")
self.hidden_size = hidden_size
self.cross_attention_dim = cross_attention_dim
self.scale = scale
self.num_tokens = num_tokens
self.to_k_ip = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)
self.to_v_ip = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)
def __call__(
self,
attn,
hidden_states,
encoder_hidden_states=None,
attention_mask=None,
temb=None,
is_cloth_pass=False,
cloth = None,
up_cnt=None,
mid_cnt=None,
down_cnt=None,
inside=None,
):
if is_cloth_pass :
residual = hidden_states
if attn.spatial_norm is not None:
hidden_states = attn.spatial_norm(hidden_states, temb)
input_ndim = hidden_states.ndim
if input_ndim == 4:
batch_size, channel, height, width = hidden_states.shape
hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
batch_size, sequence_length, _ = (
hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
)
if attention_mask is not None:
attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
# scaled_dot_product_attention expects attention_mask shape to be
# (batch, heads, source_length, target_length)
attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1])
if attn.group_norm is not None:
hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
query = attn.to_q(hidden_states)
if encoder_hidden_states is None:
encoder_hidden_states = hidden_states
elif attn.norm_cross:
encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
key = attn.to_k(encoder_hidden_states)
value = attn.to_v(encoder_hidden_states)
inner_dim = key.shape[-1]
head_dim = inner_dim // attn.heads
query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
# the output of sdp = (batch, num_heads, seq_len, head_dim)
# TODO: add support for attn.scale when we move to Torch 2.1
hidden_states = F.scaled_dot_product_attention(
query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
)
hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
hidden_states = hidden_states.to(query.dtype)
# linear proj
hidden_states = attn.to_out[0](hidden_states)
# dropout
hidden_states = attn.to_out[1](hidden_states)
if input_ndim == 4:
hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
if attn.residual_connection:
hidden_states = hidden_states + residual
hidden_states = hidden_states / attn.rescale_output_factor
return hidden_states
elif up_cnt is None:
residual = hidden_states
if attn.spatial_norm is not None:
hidden_states = attn.spatial_norm(hidden_states, temb)
input_ndim = hidden_states.ndim
if input_ndim == 4:
batch_size, channel, height, width = hidden_states.shape
hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
batch_size, sequence_length, _ = (
hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
)
if attention_mask is not None:
attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
# scaled_dot_product_attention expects attention_mask shape to be
# (batch, heads, source_length, target_length)
attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1])
if attn.group_norm is not None:
hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
query = attn.to_q(hidden_states)
if encoder_hidden_states is None:
encoder_hidden_states = hidden_states
else:
# get encoder_hidden_states, ip_hidden_states
end_pos = encoder_hidden_states.shape[1] - self.num_tokens
encoder_hidden_states, ip_hidden_states = (
encoder_hidden_states[:, :end_pos, :],
encoder_hidden_states[:, end_pos:, :],
)
if attn.norm_cross:
encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
key = attn.to_k(encoder_hidden_states)
value = attn.to_v(encoder_hidden_states)
inner_dim = key.shape[-1]
head_dim = inner_dim // attn.heads
query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
# the output of sdp = (batch, num_heads, seq_len, head_dim)
# TODO: add support for attn.scale when we move to Torch 2.1
hidden_states = F.scaled_dot_product_attention(
query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
)
hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
hidden_states = hidden_states.to(query.dtype)
# for ip-adapter
ip_key = self.to_k_ip(ip_hidden_states)
ip_value = self.to_v_ip(ip_hidden_states)
ip_key = ip_key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
ip_value = ip_value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
# the output of sdp = (batch, num_heads, seq_len, head_dim)
# TODO: add support for attn.scale when we move to Torch 2.1
ip_hidden_states = F.scaled_dot_product_attention(
query, ip_key, ip_value, attn_mask=None, dropout_p=0.0, is_causal=False
)
ip_hidden_states = ip_hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
ip_hidden_states = ip_hidden_states.to(query.dtype)
hidden_states = hidden_states + self.scale * ip_hidden_states
# linear proj
hidden_states = attn.to_out[0](hidden_states)
# dropout
hidden_states = attn.to_out[1](hidden_states)
if input_ndim == 4:
hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
if attn.residual_connection:
hidden_states = hidden_states + residual
hidden_states = hidden_states / attn.rescale_output_factor
return hidden_states
else:
residual = hidden_states
if attn.spatial_norm is not None:
hidden_states = attn.spatial_norm(hidden_states, temb)
input_ndim = hidden_states.ndim
if input_ndim == 4:
batch_size, channel, height, width = hidden_states.shape
hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
batch_size, sequence_length, _ = (
hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
)
if attention_mask is not None:
print('!!!!attention_mask is not NoNE!!!!')
attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
# scaled_dot_product_attention expects attention_mask shape to be
# (batch, heads, source_length, target_length)
attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1])
if attn.group_norm is not None:
hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
query = attn.to_q(hidden_states)
if encoder_hidden_states is None:
encoder_hidden_states = hidden_states
else:
# get encoder_hidden_states, ip_hidden_states
end_pos = encoder_hidden_states.shape[1] - self.num_tokens
encoder_hidden_states, ip_hidden_states = (
encoder_hidden_states[:, :end_pos, :],
encoder_hidden_states[:, end_pos:, :],
)
if attn.norm_cross:
encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
key = attn.to_k(encoder_hidden_states)
value = attn.to_v(encoder_hidden_states)
inner_dim = key.shape[-1]
head_dim = inner_dim // attn.heads
query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
# the output of sdp = (batch, num_heads, seq_len, head_dim)
# TODO: add support for attn.scale when we move to Torch 2.1
hidden_states = F.scaled_dot_product_attention(
query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
)
hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
hidden_states = hidden_states.to(query.dtype)
# for ip-adapter
# print(up_cnt*3 + inside)
cloth_feature = cloth[up_cnt*3 + inside-1]
cloth_feature = rearrange(cloth_feature, "b c h w -> b (h w) c").contiguous()
# for ip-adapter
ip_key = self.to_k_ip(ip_hidden_states)
ip_value = self.to_v_ip(ip_hidden_states)
ip_key = ip_key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
ip_value = ip_value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
# the output of sdp = (batch, num_heads, seq_len, head_dim)
# TODO: add support for attn.scale when we move to Torch 2.1
ip_hidden_states = F.scaled_dot_product_attention(
query, ip_key, ip_value, attn_mask=None, dropout_p=0.0, is_causal=False
)
c_key = self.to_k_c(cloth_feature)
c_value = self.to_v_c(cloth_feature)
c_key = c_key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
c_value = c_value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
ip_hidden_states = F.scaled_dot_product_attention(
ip_hidden_states, c_key, c_value, attn_mask=None, dropout_p=0.0, is_causal=False
)
ip_hidden_states = ip_hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
ip_hidden_states = ip_hidden_states.to(query.dtype)
hidden_states = hidden_states + self.scale * ip_hidden_states
# linear proj
hidden_states = attn.to_out[0](hidden_states)
# dropout
hidden_states = attn.to_out[1](hidden_states)
if input_ndim == 4:
hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
if attn.residual_connection:
hidden_states = hidden_states + residual
hidden_states = hidden_states / attn.rescale_output_factor
return hidden_states
class IPAttnProcessor_clothpass_extend(torch.nn.Module):
r"""
Attention processor for IP-Adapater for PyTorch 2.0.
Args:
hidden_size (`int`):
The hidden size of the attention layer.
cross_attention_dim (`int`):
The number of channels in the `encoder_hidden_states`.
scale (`float`, defaults to 1.0):
the weight scale of image prompt.
num_tokens (`int`, defaults to 4 when do ip_adapter_plus it should be 16):
The context length of the image features.
"""
def __init__(self, hidden_size, cross_attention_dim=None, scale=1.0, num_tokens=4):
super().__init__()
if not hasattr(F, "scaled_dot_product_attention"):
raise ImportError("AttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.")
self.hidden_size = hidden_size
self.cross_attention_dim = cross_attention_dim
self.scale = scale
self.num_tokens = num_tokens
self.to_k_ip = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)
self.to_v_ip = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)
def __call__(
self,
attn,
hidden_states,
encoder_hidden_states=None,
attention_mask=None,
temb=None,
is_cloth_pass=False,
cloth = None,
up_cnt=None,
mid_cnt=None,
down_cnt=None,
inside_up=None,
inside_down=None,
):
if is_cloth_pass :
residual = hidden_states
if attn.spatial_norm is not None:
hidden_states = attn.spatial_norm(hidden_states, temb)
input_ndim = hidden_states.ndim
if input_ndim == 4:
batch_size, channel, height, width = hidden_states.shape
hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
batch_size, sequence_length, _ = (
hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
)
if attention_mask is not None:
attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
# scaled_dot_product_attention expects attention_mask shape to be
# (batch, heads, source_length, target_length)
attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1])
if attn.group_norm is not None:
hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
query = attn.to_q(hidden_states)
if encoder_hidden_states is None:
encoder_hidden_states = hidden_states
elif attn.norm_cross:
encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
key = attn.to_k(encoder_hidden_states)
value = attn.to_v(encoder_hidden_states)
inner_dim = key.shape[-1]
head_dim = inner_dim // attn.heads
query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
# the output of sdp = (batch, num_heads, seq_len, head_dim)
# TODO: add support for attn.scale when we move to Torch 2.1
hidden_states = F.scaled_dot_product_attention(
query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
)
hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
hidden_states = hidden_states.to(query.dtype)
# linear proj
hidden_states = attn.to_out[0](hidden_states)
# dropout
hidden_states = attn.to_out[1](hidden_states)
if input_ndim == 4:
hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
if attn.residual_connection:
hidden_states = hidden_states + residual
hidden_states = hidden_states / attn.rescale_output_factor
return hidden_states
# elif up_cnt is None or down_cnt is None:
# residual = hidden_states
# if attn.spatial_norm is not None:
# hidden_states = attn.spatial_norm(hidden_states, temb)
# input_ndim = hidden_states.ndim
# if input_ndim == 4:
# batch_size, channel, height, width = hidden_states.shape
# hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
# batch_size, sequence_length, _ = (
# hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
# )
# if attention_mask is not None:
# attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
# # scaled_dot_product_attention expects attention_mask shape to be
# # (batch, heads, source_length, target_length)
# attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1])
# if attn.group_norm is not None:
# hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
# query = attn.to_q(hidden_states)
# if encoder_hidden_states is None:
# encoder_hidden_states = hidden_states
# else:
# # get encoder_hidden_states, ip_hidden_states
# end_pos = encoder_hidden_states.shape[1] - self.num_tokens
# encoder_hidden_states, ip_hidden_states = (
# encoder_hidden_states[:, :end_pos, :],
# encoder_hidden_states[:, end_pos:, :],
# )
# if attn.norm_cross:
# encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
# key = attn.to_k(encoder_hidden_states)
# value = attn.to_v(encoder_hidden_states)
# inner_dim = key.shape[-1]
# head_dim = inner_dim // attn.heads
# query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
# key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
# value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
# # the output of sdp = (batch, num_heads, seq_len, head_dim)
# # TODO: add support for attn.scale when we move to Torch 2.1
# hidden_states = F.scaled_dot_product_attention(
# query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
# )
# hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
# hidden_states = hidden_states.to(query.dtype)
# # for ip-adapter
# ip_key = self.to_k_ip(ip_hidden_states)
# ip_value = self.to_v_ip(ip_hidden_states)
# ip_key = ip_key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
# ip_value = ip_value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
# # the output of sdp = (batch, num_heads, seq_len, head_dim)
# # TODO: add support for attn.scale when we move to Torch 2.1
# ip_hidden_states = F.scaled_dot_product_attention(
# query, ip_key, ip_value, attn_mask=None, dropout_p=0.0, is_causal=False
# )
# ip_hidden_states = ip_hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
# ip_hidden_states = ip_hidden_states.to(query.dtype)
# hidden_states = hidden_states + self.scale * ip_hidden_states
# # linear proj
# hidden_states = attn.to_out[0](hidden_states)
# # dropout
# hidden_states = attn.to_out[1](hidden_states)
# if input_ndim == 4:
# hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
# if attn.residual_connection:
# hidden_states = hidden_states + residual
# hidden_states = hidden_states / attn.rescale_output_factor
# return hidden_states
elif down_cnt is not None or up_cnt is not None or mid_cnt is not None:
residual = hidden_states
if attn.spatial_norm is not None:
hidden_states = attn.spatial_norm(hidden_states, temb)
input_ndim = hidden_states.ndim
if input_ndim == 4:
batch_size, channel, height, width = hidden_states.shape
hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
batch_size, sequence_length, _ = (
hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
)
if attention_mask is not None:
print('!!!!attention_mask is not NoNE!!!!')
attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
# scaled_dot_product_attention expects attention_mask shape to be
# (batch, heads, source_length, target_length)
attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1])
if attn.group_norm is not None:
hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
query = attn.to_q(hidden_states)
if encoder_hidden_states is None:
encoder_hidden_states = hidden_states
else:
# get encoder_hidden_states, ip_hidden_states
end_pos = encoder_hidden_states.shape[1] - self.num_tokens
encoder_hidden_states, ip_hidden_states = (
encoder_hidden_states[:, :end_pos, :],
encoder_hidden_states[:, end_pos:, :],
)
if attn.norm_cross:
encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
key = attn.to_k(encoder_hidden_states)
value = attn.to_v(encoder_hidden_states)
inner_dim = key.shape[-1]
head_dim = inner_dim // attn.heads
query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
# the output of sdp = (batch, num_heads, seq_len, head_dim)
# TODO: add support for attn.scale when we move to Torch 2.1
hidden_states = F.scaled_dot_product_attention(
query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
)
hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
hidden_states = hidden_states.to(query.dtype)
# # for ip-adapter
# print(self.hidden_size)
# # print(up_cnt*3 + inside)
# print(inside_down)
cloth_feature = cloth[inside_down]
# print(cloth_feature.shape)
# if down_cnt is not None:
# # print("up_index")
# cloth_feature = cloth[down_cnt*3 + inside_down+1]
# # print(up_cnt*3 + inside_up)
# elif mid_cnt is not None:
# cloth_feature = cloth[9]
# else:
# cloth_feature = cloth[11+up_cnt*3 + inside_up]
# print("down_index")
# print(down_cnt*3 + inside_down)
cloth_feature = rearrange(cloth_feature, "b c h w -> b (h w) c").contiguous()
# print(cloth_feature.shape)
# print(self.hidden_size)
# for ip-adapter
ip_key = self.to_k_ip(ip_hidden_states)
ip_value = self.to_v_ip(ip_hidden_states)
ip_key = ip_key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
ip_value = ip_value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
# the output of sdp = (batch, num_heads, seq_len, head_dim)
# TODO: add support for attn.scale when we move to Torch 2.1
ip_hidden_states = F.scaled_dot_product_attention(
query, ip_key, ip_value, attn_mask=None, dropout_p=0.0, is_causal=False
)
c_key = self.to_k_c(cloth_feature)
c_value = self.to_v_c(cloth_feature)
c_key = c_key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
c_value = c_value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
ip_hidden_states = F.scaled_dot_product_attention(
ip_hidden_states, c_key, c_value, attn_mask=None, dropout_p=0.0, is_causal=False
)
ip_hidden_states = ip_hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
ip_hidden_states = ip_hidden_states.to(query.dtype)
hidden_states = hidden_states + self.scale * ip_hidden_states
# linear proj
hidden_states = attn.to_out[0](hidden_states)
# dropout
hidden_states = attn.to_out[1](hidden_states)
if input_ndim == 4:
hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
if attn.residual_connection:
hidden_states = hidden_states + residual
hidden_states = hidden_states / attn.rescale_output_factor
return hidden_states
else:
assert(False)
class IPAttnProcessorMulti2_0_2(torch.nn.Module):
r"""
Attention processor for IP-Adapater for PyTorch 2.0.
Args:
hidden_size (`int`):
The hidden size of the attention layer.
cross_attention_dim (`int`):
The number of channels in the `encoder_hidden_states`.
scale (`float`, defaults to 1.0):
the weight scale of image prompt.
num_tokens (`int`, defaults to 4 when do ip_adapter_plus it should be 16):
The context length of the image features.
"""
def __init__(self, hidden_size, cross_attention_dim=None, scale=1.0, num_tokens=4):
super().__init__()
if not hasattr(F, "scaled_dot_product_attention"):
raise ImportError("AttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.")
self.hidden_size = hidden_size
self.cross_attention_dim = cross_attention_dim
self.scale = scale
self.num_tokens = num_tokens
self.to_k_ip = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)
self.to_v_ip = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)
def __call__(
self,
attn,
hidden_states,
encoder_hidden_states=None,
attention_mask=None,
temb=None,
is_cloth_pass=False,
cloth = None,
up_cnt=None,
mid_cnt=None,
down_cnt=None,
inside=None,
cloth_text=None,
):
if is_cloth_pass or up_cnt is None:
residual = hidden_states
if attn.spatial_norm is not None:
hidden_states = attn.spatial_norm(hidden_states, temb)
input_ndim = hidden_states.ndim
if input_ndim == 4:
batch_size, channel, height, width = hidden_states.shape
hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
batch_size, sequence_length, _ = (
hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
)
if attention_mask is not None:
attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
# scaled_dot_product_attention expects attention_mask shape to be
# (batch, heads, source_length, target_length)
attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1])
if attn.group_norm is not None:
hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
query = attn.to_q(hidden_states)
if encoder_hidden_states is None:
encoder_hidden_states = hidden_states
else:
# get encoder_hidden_states, ip_hidden_states
end_pos = encoder_hidden_states.shape[1] - self.num_tokens
encoder_hidden_states, ip_hidden_states = (
encoder_hidden_states[:, :end_pos, :],
encoder_hidden_states[:, end_pos:, :],
)
if attn.norm_cross:
encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
key = attn.to_k(encoder_hidden_states)
value = attn.to_v(encoder_hidden_states)
inner_dim = key.shape[-1]
head_dim = inner_dim // attn.heads
query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
# the output of sdp = (batch, num_heads, seq_len, head_dim)
# TODO: add support for attn.scale when we move to Torch 2.1
hidden_states = F.scaled_dot_product_attention(
query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
)
hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
hidden_states = hidden_states.to(query.dtype)
# for ip-adapter
ip_key = self.to_k_ip(ip_hidden_states)
ip_value = self.to_v_ip(ip_hidden_states)
ip_key = ip_key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
ip_value = ip_value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
# the output of sdp = (batch, num_heads, seq_len, head_dim)
# TODO: add support for attn.scale when we move to Torch 2.1
ip_hidden_states = F.scaled_dot_product_attention(
query, ip_key, ip_value, attn_mask=None, dropout_p=0.0, is_causal=False
)
ip_hidden_states = ip_hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
ip_hidden_states = ip_hidden_states.to(query.dtype)
hidden_states = hidden_states + self.scale * ip_hidden_states
# linear proj
hidden_states = attn.to_out[0](hidden_states)
# dropout
hidden_states = attn.to_out[1](hidden_states)
if input_ndim == 4:
hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
if attn.residual_connection:
hidden_states = hidden_states + residual
hidden_states = hidden_states / attn.rescale_output_factor
return hidden_states
else:
residual = hidden_states
if attn.spatial_norm is not None:
hidden_states = attn.spatial_norm(hidden_states, temb)
input_ndim = hidden_states.ndim
if input_ndim == 4:
batch_size, channel, height, width = hidden_states.shape
hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
batch_size, sequence_length, _ = (
hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
)
# print(up_cnt)
# print("hidden_states.shape")
# print(hidden_states.shape)
if attention_mask is not None:
attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
# scaled_dot_product_attention expects attention_mask shape to be
# (batch, heads, source_length, target_length)
attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1])
if attn.group_norm is not None:
hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
query = attn.to_q(hidden_states)
if encoder_hidden_states is None:
encoder_hidden_states = hidden_states
else:
# get encoder_hidden_states, ip_hidden_states
end_pos = encoder_hidden_states.shape[1] - self.num_tokens
encoder_hidden_states, ip_hidden_states = (
encoder_hidden_states[:, :end_pos, :],
encoder_hidden_states[:, end_pos:, :],
)
if attn.norm_cross:
encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
key = attn.to_k(encoder_hidden_states)
value = attn.to_v(encoder_hidden_states)
inner_dim = key.shape[-1]
head_dim = inner_dim // attn.heads
query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
# the output of sdp = (batch, num_heads, seq_len, head_dim)
# TODO: add support for attn.scale when we move to Torch 2.1
hidden_states = F.scaled_dot_product_attention(
query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
)
hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
hidden_states = hidden_states.to(query.dtype)
# for ip-adapter
# print(up_cnt*3 + inside)
cloth_feature = cloth[up_cnt*3 + inside-1]
cloth_feature = rearrange(cloth_feature, "b c h w -> b (h w) c").contiguous()
# print(cloth_feature.shape)
# print(self.hidden_size)
# print("cloth_feature.shape")
# print(cloth_feature.shape)
query_cloth = self.q_additional(cloth_feature)
query_cloth = query_cloth.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
ip_key = self.to_k_ip(cloth_text)
ip_value = self.to_v_ip(cloth_text)
ip_key = ip_key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
ip_value = ip_value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
# print(ip_value.shape)
#$$ attn_mask?
hidden_states_cloth = F.scaled_dot_product_attention(
query_cloth, ip_key, ip_value, attn_mask=None, dropout_p=0.0, is_causal=False
)
hidden_states_cloth = hidden_states_cloth.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
hidden_states_cloth = hidden_states_cloth.to(query.dtype)
ip_key = self.k_additional(hidden_states_cloth)
ip_value = self.v_additional(hidden_states_cloth)
ip_key = ip_key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
ip_value = ip_value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
# the output of sdp = (batch, num_heads, seq_len, head_dim)
# TODO: add support for attn.scale when we move to Torch 2.1
ip_hidden_states = F.scaled_dot_product_attention(
query, ip_key, ip_value, attn_mask=None, dropout_p=0.0, is_causal=False
)
ip_hidden_states = ip_hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
ip_hidden_states = ip_hidden_states.to(query.dtype)
hidden_states = hidden_states + self.scale * ip_hidden_states
# linear proj
hidden_states = attn.to_out[0](hidden_states)
# dropout
hidden_states = attn.to_out[1](hidden_states)
if input_ndim == 4:
hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
if attn.residual_connection:
hidden_states = hidden_states + residual
hidden_states = hidden_states / attn.rescale_output_factor
return hidden_states
class IPAttnProcessor2_0_paint(torch.nn.Module):
r"""
Attention processor for IP-Adapater for PyTorch 2.0.
Args:
hidden_size (`int`):
The hidden size of the attention layer.
cross_attention_dim (`int`):
The number of channels in the `encoder_hidden_states`.
scale (`float`, defaults to 1.0):
the weight scale of image prompt.
num_tokens (`int`, defaults to 4 when do ip_adapter_plus it should be 16):
The context length of the image features.
"""
def __init__(self, hidden_size, cross_attention_dim=None, scale=1.0, num_tokens=4):
super().__init__()
if not hasattr(F, "scaled_dot_product_attention"):
raise ImportError("AttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.")
if cross_attention_dim==None:
print("cross_attention_dim is none")
self.hidden_size = hidden_size
self.cross_attention_dim = cross_attention_dim
self.scale = scale
self.num_tokens = num_tokens
self.to_k_ip = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)
self.to_v_ip = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)
def __call__(
self,
attn,
hidden_states,
encoder_hidden_states=None,
attention_mask=None,
temb=None,
):
residual = hidden_states
if attn.spatial_norm is not None:
hidden_states = attn.spatial_norm(hidden_states, temb)
input_ndim = hidden_states.ndim
if input_ndim == 4:
batch_size, channel, height, width = hidden_states.shape
hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
batch_size, sequence_length, _ = (
hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
)
if attention_mask is not None:
attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
# scaled_dot_product_attention expects attention_mask shape to be
# (batch, heads, source_length, target_length)
attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1])
if attn.group_norm is not None:
hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
query = attn.to_q(hidden_states)
if encoder_hidden_states is None:
encoder_hidden_states = hidden_states
elif attn.norm_cross:
encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
key = attn.to_k(encoder_hidden_states)
value = attn.to_v(encoder_hidden_states)
inner_dim = key.shape[-1]
head_dim = inner_dim // attn.heads
query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
# the output of sdp = (batch, num_heads, seq_len, head_dim)
# TODO: add support for attn.scale when we move to Torch 2.1
hidden_states = F.scaled_dot_product_attention(
query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
)
hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
hidden_states = hidden_states.to(query.dtype)
# #######
# # for ip-adapter
ip_key = self.to_k_ip(encoder_hidden_states)
ip_value = self.to_v_ip(encoder_hidden_states)
ip_key = ip_key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
ip_value = ip_value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
# the output of sdp = (batch, num_heads, seq_len, head_dim)
# TODO: add support for attn.scale when we move to Torch 2.1
ip_hidden_states = F.scaled_dot_product_attention(
query, ip_key, ip_value, attn_mask=None, dropout_p=0.0, is_causal=False
)
ip_hidden_states = ip_hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
ip_hidden_states = ip_hidden_states.to(query.dtype)
hidden_states = hidden_states + self.scale * ip_hidden_states
# #######
# linear proj
hidden_states = attn.to_out[0](hidden_states)
# dropout
hidden_states = attn.to_out[1](hidden_states)
if input_ndim == 4:
hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
if attn.residual_connection:
hidden_states = hidden_states + residual
hidden_states = hidden_states / attn.rescale_output_factor
return hidden_states
class IPAttnProcessor2_0_variant(torch.nn.Module):
r"""
Attention processor for IP-Adapater for PyTorch 2.0.
Args:
hidden_size (`int`):
The hidden size of the attention layer.
cross_attention_dim (`int`):
The number of channels in the `encoder_hidden_states`.
scale (`float`, defaults to 1.0):
the weight scale of image prompt.
num_tokens (`int`, defaults to 4 when do ip_adapter_plus it should be 16):
The context length of the image features.
"""
def __init__(self, hidden_size, cross_attention_dim=None, scale=1.0, num_tokens=4):
super().__init__()
if not hasattr(F, "scaled_dot_product_attention"):
raise ImportError("AttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.")
self.hidden_size = hidden_size
self.cross_attention_dim = cross_attention_dim
self.scale = scale
self.num_tokens = num_tokens
self.to_k_ip = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)
self.to_v_ip = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)
def __call__(
self,
attn,
hidden_states,
encoder_hidden_states=None,
attention_mask=None,
temb=None,
):
residual = hidden_states
if attn.spatial_norm is not None:
hidden_states = attn.spatial_norm(hidden_states, temb)
input_ndim = hidden_states.ndim
if input_ndim == 4:
batch_size, channel, height, width = hidden_states.shape
hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
batch_size, sequence_length, _ = (
hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
)
if attention_mask is not None:
attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
# scaled_dot_product_attention expects attention_mask shape to be
# (batch, heads, source_length, target_length)
attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1])
if attn.group_norm is not None:
hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
query = attn.to_q(hidden_states)
if encoder_hidden_states is None:
encoder_hidden_states = hidden_states
else:
# get encoder_hidden_states, ip_hidden_states
end_pos = encoder_hidden_states.shape[1] - self.num_tokens
encoder_hidden_states, ip_hidden_states = (
encoder_hidden_states[:, :end_pos, :],
encoder_hidden_states[:, end_pos:, :],
)
if attn.norm_cross:
encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
key = attn.to_k(encoder_hidden_states)
value = attn.to_v(encoder_hidden_states)
inner_dim = key.shape[-1]
head_dim = inner_dim // attn.heads
query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
# the output of sdp = (batch, num_heads, seq_len, head_dim)
# TODO: add support for attn.scale when we move to Torch 2.1
hidden_states = F.scaled_dot_product_attention(
query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
)
# hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
# hidden_states = hidden_states.to(query.dtype)
# for ip-adapter
ip_key = self.to_k_ip(ip_hidden_states)
ip_value = self.to_v_ip(ip_hidden_states)
ip_key = ip_key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
ip_value = ip_value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
# the output of sdp = (batch, num_heads, seq_len, head_dim)
# TODO: add support for attn.scale when we move to Torch 2.1
ip_hidden_states = F.scaled_dot_product_attention(
hidden_states, ip_key, ip_value, attn_mask=None, dropout_p=0.0, is_causal=False
)
with torch.no_grad():
self.attn_map = query @ ip_key.transpose(-2, -1).softmax(dim=-1)
ip_hidden_states = ip_hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
ip_hidden_states = ip_hidden_states.to(query.dtype)
hidden_states = ip_hidden_states
# linear proj
hidden_states = attn.to_out[0](hidden_states)
# dropout
hidden_states = attn.to_out[1](hidden_states)
if input_ndim == 4:
hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
if attn.residual_connection:
hidden_states = hidden_states + residual
hidden_states = hidden_states / attn.rescale_output_factor
return hidden_states
class IPAttnProcessor2_0(torch.nn.Module):
r"""
Attention processor for IP-Adapater for PyTorch 2.0.
Args:
hidden_size (`int`):
The hidden size of the attention layer.
cross_attention_dim (`int`):
The number of channels in the `encoder_hidden_states`.
scale (`float`, defaults to 1.0):
the weight scale of image prompt.
num_tokens (`int`, defaults to 4 when do ip_adapter_plus it should be 16):
The context length of the image features.
"""
def __init__(self, hidden_size, cross_attention_dim=None, scale=1.0, num_tokens=4):
super().__init__()
if not hasattr(F, "scaled_dot_product_attention"):
raise ImportError("AttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.")
self.hidden_size = hidden_size
self.cross_attention_dim = cross_attention_dim
self.scale = scale
self.num_tokens = num_tokens
self.to_k_ip = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)
self.to_v_ip = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)
def __call__(
self,
attn,
hidden_states,
encoder_hidden_states=None,
attention_mask=None,
temb=None,
scale=1.0
):
residual = hidden_states
if attn.spatial_norm is not None:
hidden_states = attn.spatial_norm(hidden_states, temb)
input_ndim = hidden_states.ndim
if input_ndim == 4:
batch_size, channel, height, width = hidden_states.shape
hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
batch_size, sequence_length, _ = (
hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
)
if attention_mask is not None:
attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
# scaled_dot_product_attention expects attention_mask shape to be
# (batch, heads, source_length, target_length)
attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1])
if attn.group_norm is not None:
hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
# args = (scale, )
args = ()
query = attn.to_q(hidden_states, *args)
if encoder_hidden_states is None:
encoder_hidden_states = hidden_states
else:
# get encoder_hidden_states, ip_hidden_states
end_pos = encoder_hidden_states.shape[1] - self.num_tokens
encoder_hidden_states, ip_hidden_states = (
encoder_hidden_states[:, :end_pos, :],
encoder_hidden_states[:, end_pos:, :],
)
if attn.norm_cross:
encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
key = attn.to_k(encoder_hidden_states, *args)
value = attn.to_v(encoder_hidden_states, *args)
inner_dim = key.shape[-1]
head_dim = inner_dim // attn.heads
query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
# the output of sdp = (batch, num_heads, seq_len, head_dim)
# TODO: add support for attn.scale when we move to Torch 2.1
hidden_states = F.scaled_dot_product_attention(
query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
)
hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
hidden_states = hidden_states.to(query.dtype)
# for ip-adapter
ip_key = self.to_k_ip(ip_hidden_states)
ip_value = self.to_v_ip(ip_hidden_states)
ip_key = ip_key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
ip_value = ip_value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
# the output of sdp = (batch, num_heads, seq_len, head_dim)
# TODO: add support for attn.scale when we move to Torch 2.1
ip_hidden_states = F.scaled_dot_product_attention(
query, ip_key, ip_value, attn_mask=None, dropout_p=0.0, is_causal=False
)
with torch.no_grad():
self.attn_map = query @ ip_key.transpose(-2, -1).softmax(dim=-1)
ip_hidden_states = ip_hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
ip_hidden_states = ip_hidden_states.to(query.dtype)
hidden_states = hidden_states + self.scale * ip_hidden_states
# linear proj
hidden_states = attn.to_out[0](hidden_states, *args)
# dropout
hidden_states = attn.to_out[1](hidden_states)
if input_ndim == 4:
hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
if attn.residual_connection:
hidden_states = hidden_states + residual
hidden_states = hidden_states / attn.rescale_output_factor
return hidden_states
class IPAttnProcessor_referencenet_2_0(torch.nn.Module):
r"""
Attention processor for IP-Adapater for PyTorch 2.0.
Args:
hidden_size (`int`):
The hidden size of the attention layer.
cross_attention_dim (`int`):
The number of channels in the `encoder_hidden_states`.
scale (`float`, defaults to 1.0):
the weight scale of image prompt.
num_tokens (`int`, defaults to 4 when do ip_adapter_plus it should be 16):
The context length of the image features.
"""
def __init__(self, hidden_size, cross_attention_dim=None, scale=1.0, num_tokens=4,attn_head_dim=10):
super().__init__()
if not hasattr(F, "scaled_dot_product_attention"):
raise ImportError("AttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.")
self.hidden_size = hidden_size
self.cross_attention_dim = cross_attention_dim
self.scale = scale
self.num_tokens = num_tokens
self.attn_head_dim=attn_head_dim
self.to_k_ip = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)
self.to_v_ip = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)
def __call__(
self,
attn,
hidden_states,
encoder_hidden_states=None,
attention_mask=None,
temb=None,
):
residual = hidden_states
if attn.spatial_norm is not None:
hidden_states = attn.spatial_norm(hidden_states, temb)
input_ndim = hidden_states.ndim
if input_ndim == 4:
batch_size, channel, height, width = hidden_states.shape
hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
batch_size, sequence_length, _ = (
hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
)
if attention_mask is not None:
attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
# scaled_dot_product_attention expects attention_mask shape to be
# (batch, heads, source_length, target_length)
attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1])
if attn.group_norm is not None:
hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
query = attn.to_q(hidden_states)
if encoder_hidden_states is None:
encoder_hidden_states = hidden_states
else:
# get encoder_hidden_states, ip_hidden_states
end_pos = encoder_hidden_states.shape[1] - self.num_tokens
encoder_hidden_states, ip_hidden_states = (
encoder_hidden_states[:, :end_pos, :],
encoder_hidden_states[:, end_pos:, :],
)
if attn.norm_cross:
encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
key = attn.to_k(encoder_hidden_states)
value = attn.to_v(encoder_hidden_states)
inner_dim = key.shape[-1]
head_dim = inner_dim // attn.heads
query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
# the output of sdp = (batch, num_heads, seq_len, head_dim)
# TODO: add support for attn.scale when we move to Torch 2.1
hidden_states = F.scaled_dot_product_attention(
query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
)
hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
hidden_states = hidden_states.to(query.dtype)
# for ip-adapter
ip_key = self.to_k_ip(ip_hidden_states)
ip_value = self.to_v_ip(ip_hidden_states)
ip_key = ip_key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
ip_value = ip_value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
# the output of sdp = (batch, num_heads, seq_len, head_dim)
# TODO: add support for attn.scale when we move to Torch 2.1
ip_hidden_states = F.scaled_dot_product_attention(
query, ip_key, ip_value, attn_mask=None, dropout_p=0.0, is_causal=False
)
with torch.no_grad():
self.attn_map = query @ ip_key.transpose(-2, -1).softmax(dim=-1)
ip_hidden_states = ip_hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
ip_hidden_states = ip_hidden_states.to(query.dtype)
hidden_states = hidden_states + self.scale * ip_hidden_states
# linear proj
hidden_states = attn.to_out[0](hidden_states)
# dropout
hidden_states = attn.to_out[1](hidden_states)
if input_ndim == 4:
hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
if attn.residual_connection:
hidden_states = hidden_states + residual
hidden_states = hidden_states / attn.rescale_output_factor
return hidden_states
class IPAttnProcessor2_0_Lora(torch.nn.Module):
r"""
Attention processor for IP-Adapater for PyTorch 2.0.
Args:
hidden_size (`int`):
The hidden size of the attention layer.
cross_attention_dim (`int`):
The number of channels in the `encoder_hidden_states`.
scale (`float`, defaults to 1.0):
the weight scale of image prompt.
num_tokens (`int`, defaults to 4 when do ip_adapter_plus it should be 16):
The context length of the image features.
"""
def __init__(self, hidden_size, cross_attention_dim=None, scale=1.0, scale_lora=1.0, rank = 4,num_tokens=4):
super().__init__()
if not hasattr(F, "scaled_dot_product_attention"):
raise ImportError("AttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.")
self.hidden_size = hidden_size
self.cross_attention_dim = cross_attention_dim
self.scale = scale
self.scale_lora = scale_lora
self.num_tokens = num_tokens
self.to_k_ip = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)
self.to_v_ip = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)
self.to_k_ip_lora = LoRALinearLayer(in_features=self.to_k_ip.in_features, out_features=self.to_k_ip.out_features, rank=rank)
self.to_v_ip_lora =LoRALinearLayer(in_features=self.to_v_ip.in_features, out_features=self.to_v_ip.out_features, rank=rank)
def __call__(
self,
attn,
hidden_states,
encoder_hidden_states=None,
attention_mask=None,
temb=None,
):
residual = hidden_states
if attn.spatial_norm is not None:
hidden_states = attn.spatial_norm(hidden_states, temb)
input_ndim = hidden_states.ndim
if input_ndim == 4:
batch_size, channel, height, width = hidden_states.shape
hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
batch_size, sequence_length, _ = (
hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
)
if attention_mask is not None:
attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
# scaled_dot_product_attention expects attention_mask shape to be
# (batch, heads, source_length, target_length)
attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1])
if attn.group_norm is not None:
hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
if hasattr(attn,'q_lora'):
query = attn.to_q(hidden_states)
q_lora = attn.q_lora(hidden_states)
query = query + self.scale_lora * q_lora
else:
query = attn.to_q(hidden_states)
if encoder_hidden_states is None:
encoder_hidden_states = hidden_states
else:
# get encoder_hidden_states, ip_hidden_states
end_pos = encoder_hidden_states.shape[1] - self.num_tokens
encoder_hidden_states, ip_hidden_states = (
encoder_hidden_states[:, :end_pos, :],
encoder_hidden_states[:, end_pos:, :],
)
if attn.norm_cross:
encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
if hasattr(attn,'k_lora'):
key = attn.to_k(encoder_hidden_states)
k_lora = attn.k_lora(encoder_hidden_states)
key = key + self.scale_lora * k_lora
else:
key = attn.to_k(encoder_hidden_states)
if hasattr(attn,'v_lora'):
value = attn.to_v(encoder_hidden_states)
v_lora = attn.v_lora(encoder_hidden_states)
value = value + self.scale_lora * v_lora
else:
value = attn.to_v(encoder_hidden_states)
inner_dim = key.shape[-1]
head_dim = inner_dim // attn.heads
query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
# the output of sdp = (batch, num_heads, seq_len, head_dim)
# TODO: add support for attn.scale when we move to Torch 2.1
hidden_states = F.scaled_dot_product_attention(
query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
)
hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
hidden_states = hidden_states.to(query.dtype)
# for ip-adapter
ip_key = self.to_k_ip(ip_hidden_states)
ip_key_lora = self.to_k_ip_lora(ip_hidden_states)
ip_key = ip_key + self.scale_lora * ip_key_lora
ip_value = self.to_v_ip(ip_hidden_states)
ip_value_lora = self.to_v_ip_lora(ip_hidden_states)
ip_value = ip_value + self.scale_lora * ip_value_lora
ip_key = ip_key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
ip_value = ip_value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
# the output of sdp = (batch, num_heads, seq_len, head_dim)
# TODO: add support for attn.scale when we move to Torch 2.1
ip_hidden_states = F.scaled_dot_product_attention(
query, ip_key, ip_value, attn_mask=None, dropout_p=0.0, is_causal=False
)
ip_hidden_states = ip_hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
ip_hidden_states = ip_hidden_states.to(query.dtype)
hidden_states = hidden_states + self.scale * ip_hidden_states
# linear proj
if hasattr(attn,'out_lora'):
hidden_states = attn.to_out[0](hidden_states)
out_lora = attn.out_lora(hidden_states)
hidden_states = hidden_states+ self.scale_lora*out_lora
else:
hidden_states = attn.to_out[0](hidden_states)
# dropout
hidden_states = attn.to_out[1](hidden_states)
if input_ndim == 4:
hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
if attn.residual_connection:
hidden_states = hidden_states + residual
hidden_states = hidden_states / attn.rescale_output_factor
return hidden_states
## for controlnet
class CNAttnProcessor:
r"""
Default processor for performing attention-related computations.
"""
def __init__(self, num_tokens=4):
self.num_tokens = num_tokens
def __call__(self, attn, hidden_states, encoder_hidden_states=None, attention_mask=None, temb=None):
residual = hidden_states
if attn.spatial_norm is not None:
hidden_states = attn.spatial_norm(hidden_states, temb)
input_ndim = hidden_states.ndim
if input_ndim == 4:
batch_size, channel, height, width = hidden_states.shape
hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
batch_size, sequence_length, _ = (
hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
)
attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
if attn.group_norm is not None:
hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
query = attn.to_q(hidden_states)
if encoder_hidden_states is None:
encoder_hidden_states = hidden_states
else:
end_pos = encoder_hidden_states.shape[1] - self.num_tokens
encoder_hidden_states = encoder_hidden_states[:, :end_pos] # only use text
if attn.norm_cross:
encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
key = attn.to_k(encoder_hidden_states)
value = attn.to_v(encoder_hidden_states)
query = attn.head_to_batch_dim(query)
key = attn.head_to_batch_dim(key)
value = attn.head_to_batch_dim(value)
attention_probs = attn.get_attention_scores(query, key, attention_mask)
hidden_states = torch.bmm(attention_probs, value)
hidden_states = attn.batch_to_head_dim(hidden_states)
# linear proj
hidden_states = attn.to_out[0](hidden_states)
# dropout
hidden_states = attn.to_out[1](hidden_states)
if input_ndim == 4:
hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
if attn.residual_connection:
hidden_states = hidden_states + residual
hidden_states = hidden_states / attn.rescale_output_factor
return hidden_states
class CNAttnProcessor2_0:
r"""
Processor for implementing scaled dot-product attention (enabled by default if you're using PyTorch 2.0).
"""
def __init__(self, num_tokens=4):
if not hasattr(F, "scaled_dot_product_attention"):
raise ImportError("AttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.")
self.num_tokens = num_tokens
def __call__(
self,
attn,
hidden_states,
encoder_hidden_states=None,
attention_mask=None,
temb=None,
):
residual = hidden_states
if attn.spatial_norm is not None:
hidden_states = attn.spatial_norm(hidden_states, temb)
input_ndim = hidden_states.ndim
if input_ndim == 4:
batch_size, channel, height, width = hidden_states.shape
hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
batch_size, sequence_length, _ = (
hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
)
if attention_mask is not None:
attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
# scaled_dot_product_attention expects attention_mask shape to be
# (batch, heads, source_length, target_length)
attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1])
if attn.group_norm is not None:
hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
query = attn.to_q(hidden_states)
if encoder_hidden_states is None:
encoder_hidden_states = hidden_states
else:
end_pos = encoder_hidden_states.shape[1] - self.num_tokens
encoder_hidden_states = encoder_hidden_states[:, :end_pos] # only use text
if attn.norm_cross:
encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
key = attn.to_k(encoder_hidden_states)
value = attn.to_v(encoder_hidden_states)
inner_dim = key.shape[-1]
head_dim = inner_dim // attn.heads
query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
# the output of sdp = (batch, num_heads, seq_len, head_dim)
# TODO: add support for attn.scale when we move to Torch 2.1
hidden_states = F.scaled_dot_product_attention(
query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
)
hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
hidden_states = hidden_states.to(query.dtype)
# linear proj
hidden_states = attn.to_out[0](hidden_states)
# dropout
hidden_states = attn.to_out[1](hidden_states)
if input_ndim == 4:
hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
if attn.residual_connection:
hidden_states = hidden_states + residual
hidden_states = hidden_states / attn.rescale_output_factor
return hidden_states