detail_encoder/attention_processor.py

# 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.utils.import_utils import is_xformers_available
from torchvision import transforms
if is_xformers_available():
    import xformers
    import xformers.ops
else:
    xformers = None

class SSRAttnProcessor(nn.Module):
    r"""
    Attention processor for SSR-Adapater.
    """

    def __init__(self, hidden_size, cross_attention_dim=None, scale=1):
        super().__init__()
        self.hidden_size = hidden_size
        self.cross_attention_dim = cross_attention_dim
        self.scale = scale
        # self.to_q_SSR = nn.Linear(hidden_size, hidden_size, bias=False)
        self.to_k_SSR = nn.Linear(cross_attention_dim, hidden_size, bias=False)
        self.to_v_SSR = nn.Linear(cross_attention_dim, 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 = self.to_q_SSR(hidden_states)
        query = attn.to_q(hidden_states)
        query = attn.head_to_batch_dim(query)

        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)

        _hidden_states = encoder_hidden_states
        _key = self.to_k_SSR(_hidden_states)
        _value = self.to_v_SSR(_hidden_states)
        _key = attn.head_to_batch_dim(_key)
        _value = attn.head_to_batch_dim(_value)
        _attention_probs = attn.get_attention_scores(query, _key, None)
        _hidden_states = torch.bmm(_attention_probs, _value)
        _hidden_states = attn.batch_to_head_dim(_hidden_states)
        hidden_states = self.scale * _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 SSRAttnProcessor2_0(torch.nn.Module):
    r"""
    Attention processor for SSR-Adapater for PyTorch 2.0.
    """

    def __init__(self, hidden_size, cross_attention_dim=None, scale=1.0):
        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.to_q_SSR = nn.Linear(hidden_size, hidden_size, bias=False)
        self.to_k_SSR = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)
        self.to_v_SSR = 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 = self.to_q_SSR(hidden_states)
        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)

        # split hidden states
        _hidden_states = encoder_hidden_states

        _key = self.to_k_SSR(_hidden_states)
        _value = self.to_v_SSR(_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=None, 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)

        hidden_states = self.scale * _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,
    ):
        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 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 ConvAttnProcessor:
    def __call__(
        self,
        attn,
        hidden_states,
        encoder_hidden_states=None,
        attention_mask=None,
    ):
        ## map to 2D
        if len(hidden_states.shape) == 4:
            shape = hidden_states.shape
            hidden_states = torch.reshape(hidden_states, (shape[0], shape[1], shape[2] * shape[3]))
            hidden_states = hidden_states.permute(0, 2, 1)
        if encoder_hidden_states is not None:
            if len(encoder_hidden_states.shape) == 4:
                kv_shape = encoder_hidden_states.shape
                encoder_hidden_states = torch.reshape(
                    encoder_hidden_states, (kv_shape[0], kv_shape[1], kv_shape[2] * kv_shape[3])
                )
                encoder_hidden_states = encoder_hidden_states.permute(0, 2, 1)

        # the same to standard attn
        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)
        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_cross(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)

        # map back to 4D
        if len(hidden_states.shape) == 3:
            hidden_states = hidden_states.permute(0, 2, 1)
            hidden_states = torch.reshape(hidden_states, (shape[0], shape[1], shape[2], shape[3]))

        return hidden_states


class SSRAttnProcessor_text(nn.Module):
    r"""
    Attention processor for SSR-Adapater.
    """

    def __init__(self, hidden_size, cross_attention_dim=None, scale=1):
        super().__init__()
        self.text_context_len = 77
        self.hidden_size = hidden_size
        self.cross_attention_dim = cross_attention_dim
        self.scale = scale
        self.to_k_SSR = nn.Linear(cross_attention_dim, hidden_size, bias=False)
        self.to_v_SSR = nn.Linear(cross_attention_dim, 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)
        query = attn.head_to_batch_dim(query)

        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)

        # split hidden states
        encoder_hidden_states, _hidden_states = encoder_hidden_states[:, :self.text_context_len,
                                                :], encoder_hidden_states[:, self.text_context_len:, :]
        encoder_hidden_states = encoder_hidden_states[:, :, :768]
        # for text
        key = attn.to_k(encoder_hidden_states)
        value = attn.to_v(encoder_hidden_states)

        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 image
        _key = self.to_k_SSR(_hidden_states)
        _value = self.to_v_SSR(_hidden_states)
        _key = attn.head_to_batch_dim(_key)
        _value = attn.head_to_batch_dim(_value)
        _attention_probs = attn.get_attention_scores(query, _key, None)
        _hidden_states = torch.bmm(_attention_probs, _value)
        _hidden_states = attn.batch_to_head_dim(_hidden_states)
        hidden_states = self.scale * _hidden_states + 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 SSRAttnProcessor2_0_text(torch.nn.Module):
    r"""
    Attention processor for SSR-Adapater for PyTorch 2.0.
    """

    def __init__(self, hidden_size, cross_attention_dim=None, scale=1.0):
        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.text_context_len = 77
        self.hidden_size = hidden_size
        self.cross_attention_dim = cross_attention_dim
        self.scale = scale
        self.to_k_SSR = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)
        self.to_v_SSR = 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)

        # split hidden states
        encoder_hidden_states, _hidden_states = encoder_hidden_states[:, :self.text_context_len,
                                                :], encoder_hidden_states[:, self.text_context_len:, :]

        encoder_hidden_states = encoder_hidden_states[:, :, :768]
        # for text
        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 image
        _key = self.to_k_SSR(_hidden_states)
        _value = self.to_v_SSR(_hidden_states)
        inner_dim = _key.shape[-1]
        head_dim = inner_dim // attn.heads

        _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=None, 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)

        hidden_states = self.scale * _hidden_states + 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 SSRAttnProcessor_visual(nn.Module):
    r"""
    Attention processor for attn visualization.
    """

    def __init__(self, hidden_size, cross_attention_dim=None, scale=1, attnstore=None, place_in_unet=None):
        super().__init__()
        self.hidden_size = hidden_size
        self.cross_attention_dim = cross_attention_dim
        self.scale = scale
        self.to_k_SSR = nn.Linear(cross_attention_dim, hidden_size, bias=False)
        self.to_v_SSR = nn.Linear(cross_attention_dim, hidden_size, bias=False)
        self.attnstore = attnstore
        self.place_in_unet = place_in_unet

    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 = self.to_q_SSR(hidden_states)
        query = attn.to_q(hidden_states)
        query = attn.head_to_batch_dim(query)

        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)

        _hidden_states = encoder_hidden_states
        _key = self.to_k_SSR(_hidden_states)
        _value = self.to_v_SSR(_hidden_states)
        _key = attn.head_to_batch_dim(_key)
        _value = attn.head_to_batch_dim(_value)
        _attention_probs = attn.get_attention_scores(query, _key, None)

        # store attention maps
        is_cross = encoder_hidden_states is not None
        self.attnstore(_attention_probs, is_cross, self.place_in_unet)

        _hidden_states = torch.bmm(_attention_probs, _value)
        _hidden_states = attn.batch_to_head_dim(_hidden_states)
        hidden_states = self.scale * _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