model/blocks/CrossAttnDownBlock2D.py

from typing import Any, Dict, Tuple,Optional, Union
import torch
from torch import nn
from diffusers.utils import is_torch_version
from .ResnetBlock2D import ResnetBlock2D
from diffusers.models.transformer_2d import Transformer2DModel
from diffusers.models.dual_transformer_2d import DualTransformer2DModel
from diffusers.models.resnet import Downsample2D

class CrossAttnDownBlock2D_SPADE(nn.Module):
    def __init__(
        self,
        in_channels: int,
        out_channels: int,
        temb_channels: int,
        dropout: float = 0.0,
        num_layers: int = 1,
        transformer_layers_per_block: Union[int, Tuple[int]] = 1,
        resnet_eps: float = 1e-6,
        resnet_time_scale_shift: str = "default",
        resnet_act_fn: str = "swish",
        resnet_groups: int = 32,
        resnet_pre_norm: bool = True,
        num_attention_heads: int = 1,
        cross_attention_dim: int = 1280,
        output_scale_factor: float = 1.0,
        downsample_padding: int = 1,
        add_downsample: bool = True,
        dual_cross_attention: bool = False,
        use_linear_projection: bool = False,
        only_cross_attention: bool = False,
        upcast_attention: bool = False,
        attention_type: str = "default",
        SPADE_chs = (320, 640, 1280, 1280),
        normalization_type = None,
        is_first = False,
        is_crossAttn = True,
    ):
        super().__init__()
        resnets = []
        attentions = []

        self.has_cross_attention = True
        self.num_attention_heads = num_attention_heads
        if isinstance(transformer_layers_per_block, int):
            transformer_layers_per_block = [transformer_layers_per_block] * num_layers

        for i in range(num_layers):
            in_channels = in_channels if i == 0 else out_channels

            resnets.append(
                ResnetBlock2D(
                in_channels=in_channels,
                out_channels=out_channels,
                temb_channels=temb_channels,
                eps=resnet_eps,
                groups=resnet_groups,
                dropout=dropout,
                time_embedding_norm=resnet_time_scale_shift,
                non_linearity=resnet_act_fn,
                output_scale_factor=output_scale_factor,
                pre_norm=resnet_pre_norm,
                SPADE_chs = SPADE_chs,
                normalization_type=normalization_type,
                half_the_ch=True if not is_first and (i == 0) else False,
                is_crossAttn = is_crossAttn
                )
            )
           

            if not dual_cross_attention:
                attentions.append(
                    Transformer2DModel(
                        num_attention_heads,
                        out_channels // num_attention_heads,
                        in_channels=out_channels,
                        num_layers=transformer_layers_per_block[i],
                        cross_attention_dim=cross_attention_dim,
                        norm_num_groups=resnet_groups,
                        use_linear_projection=use_linear_projection,
                        only_cross_attention=only_cross_attention,
                        upcast_attention=upcast_attention,
                        attention_type=attention_type,
                    )
                )
            else:
                attentions.append(
                    DualTransformer2DModel(
                        num_attention_heads,
                        out_channels // num_attention_heads,
                        in_channels=out_channels,
                        num_layers=1,
                        cross_attention_dim=cross_attention_dim,
                        norm_num_groups=resnet_groups,
                    )
                )
        self.attentions = nn.ModuleList(attentions)
        self.resnets = nn.ModuleList(resnets)

        if add_downsample:
            self.downsamplers = nn.ModuleList(
                [
                    Downsample2D(
                        out_channels, use_conv=True, out_channels=out_channels, padding=downsample_padding, name="op"
                    )
                ]
            )
        else:
            self.downsamplers = None

        self.gradient_checkpointing = False

    def forward(
        self,
        hidden_states: torch.FloatTensor,
        segmap,
        temb: Optional[torch.FloatTensor] = None,
        encoder_hidden_states: Optional[torch.FloatTensor] = None,
        attention_mask: Optional[torch.FloatTensor] = None,
        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
        encoder_attention_mask: Optional[torch.FloatTensor] = None,
        additional_residuals: Optional[torch.FloatTensor] = None,
    ) -> Tuple[torch.FloatTensor, Tuple[torch.FloatTensor, ...]]:
        output_states = ()

        lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0

        blocks = list(zip(self.resnets, self.attentions))

        for i, (resnet, attn) in enumerate(blocks):
            if self.training and self.gradient_checkpointing:

                def create_custom_forward(module, return_dict=None):
                    def custom_forward(*inputs):
                        if return_dict is not None:
                            return module(*inputs, return_dict=return_dict)
                        else:
                            return module(*inputs)

                    return custom_forward

                ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
                hidden_states = torch.utils.checkpoint.checkpoint(
                    create_custom_forward(resnet),
                    hidden_states,
                    temb,
                    **ckpt_kwargs,
                )
                hidden_states = attn(
                    hidden_states,
                    encoder_hidden_states=encoder_hidden_states,
                    cross_attention_kwargs=cross_attention_kwargs,
                    attention_mask=attention_mask,
                    encoder_attention_mask=encoder_attention_mask,
                    return_dict=False,
                )[0]
            else:
                hidden_states = resnet(hidden_states, temb, segmaps=segmap,scale=lora_scale)
                hidden_states = attn(
                    hidden_states,
                    encoder_hidden_states=encoder_hidden_states,
                    cross_attention_kwargs=cross_attention_kwargs,
                    attention_mask=attention_mask,
                    encoder_attention_mask=encoder_attention_mask,
                    return_dict=False,
                )[0]

            # apply additional residuals to the output of the last pair of resnet and attention blocks
            if i == len(blocks) - 1 and additional_residuals is not None:
                hidden_states = hidden_states + additional_residuals

            output_states = output_states + (hidden_states,)

        if self.downsamplers is not None:
            for downsampler in self.downsamplers:
                hidden_states = downsampler(hidden_states, scale=lora_scale)

            output_states = output_states + (hidden_states,)

        return hidden_states, output_states