mvdiffusion/models/unet_mv2d_blocks.py

# Copyright 2023 The HuggingFace Team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from typing import Any, Dict, Optional, Tuple

import numpy as np
import torch
import torch.nn.functional as F
from torch import nn

from diffusers.utils import is_torch_version, logging
from diffusers.models.normalization import AdaGroupNorm
from diffusers.models.attention_processor import Attention, AttnAddedKVProcessor, AttnAddedKVProcessor2_0
from diffusers.models.dual_transformer_2d import DualTransformer2DModel
from diffusers.models.resnet import Downsample2D, FirDownsample2D, FirUpsample2D, KDownsample2D, KUpsample2D, ResnetBlock2D, Upsample2D

from diffusers.models.unets.unet_2d_blocks import DownBlock2D, ResnetDownsampleBlock2D, AttnDownBlock2D, CrossAttnDownBlock2D, SimpleCrossAttnDownBlock2D, SkipDownBlock2D, AttnSkipDownBlock2D, DownEncoderBlock2D, AttnDownEncoderBlock2D, KDownBlock2D, KCrossAttnDownBlock2D
from diffusers.models.unets.unet_2d_blocks import UpBlock2D, ResnetUpsampleBlock2D, CrossAttnUpBlock2D, SimpleCrossAttnUpBlock2D, AttnUpBlock2D, SkipUpBlock2D, AttnSkipUpBlock2D, UpDecoderBlock2D, AttnUpDecoderBlock2D, KUpBlock2D, KCrossAttnUpBlock2D


logger = logging.get_logger(__name__)  # pylint: disable=invalid-name


def get_down_block(
    down_block_type,
    num_layers,
    in_channels,
    out_channels,
    temb_channels,
    add_downsample,
    resnet_eps,
    resnet_act_fn,
    transformer_layers_per_block=1,
    num_attention_heads=None,
    resnet_groups=None,
    cross_attention_dim=None,
    downsample_padding=None,
    dual_cross_attention=False,
    use_linear_projection=False,
    only_cross_attention=False,
    upcast_attention=False,
    resnet_time_scale_shift="default",
    resnet_skip_time_act=False,
    resnet_out_scale_factor=1.0,
    cross_attention_norm=None,
    attention_head_dim=None,
    downsample_type=None,
    num_views=1,
    cd_attention_last: bool = False,
    cd_attention_mid: bool = False,
    multiview_attention: bool = True,
    sparse_mv_attention: bool = False,
    selfattn_block: str = "custom",
    mvcd_attention: bool=False,
    use_dino: bool = False
):
    # If attn head dim is not defined, we default it to the number of heads
    if attention_head_dim is None:
        logger.warn(
            f"It is recommended to provide `attention_head_dim` when calling `get_down_block`. Defaulting `attention_head_dim` to {num_attention_heads}."
        )
        attention_head_dim = num_attention_heads

    down_block_type = down_block_type[7:] if down_block_type.startswith("UNetRes") else down_block_type
    if down_block_type == "DownBlock2D":
        return DownBlock2D(
            num_layers=num_layers,
            in_channels=in_channels,
            out_channels=out_channels,
            temb_channels=temb_channels,
            add_downsample=add_downsample,
            resnet_eps=resnet_eps,
            resnet_act_fn=resnet_act_fn,
            resnet_groups=resnet_groups,
            downsample_padding=downsample_padding,
            resnet_time_scale_shift=resnet_time_scale_shift,
        )
    elif down_block_type == "ResnetDownsampleBlock2D":
        return ResnetDownsampleBlock2D(
            num_layers=num_layers,
            in_channels=in_channels,
            out_channels=out_channels,
            temb_channels=temb_channels,
            add_downsample=add_downsample,
            resnet_eps=resnet_eps,
            resnet_act_fn=resnet_act_fn,
            resnet_groups=resnet_groups,
            resnet_time_scale_shift=resnet_time_scale_shift,
            skip_time_act=resnet_skip_time_act,
            output_scale_factor=resnet_out_scale_factor,
        )
    elif down_block_type == "AttnDownBlock2D":
        if add_downsample is False:
            downsample_type = None
        else:
            downsample_type = downsample_type or "conv"  # default to 'conv'
        return AttnDownBlock2D(
            num_layers=num_layers,
            in_channels=in_channels,
            out_channels=out_channels,
            temb_channels=temb_channels,
            resnet_eps=resnet_eps,
            resnet_act_fn=resnet_act_fn,
            resnet_groups=resnet_groups,
            downsample_padding=downsample_padding,
            attention_head_dim=attention_head_dim,
            resnet_time_scale_shift=resnet_time_scale_shift,
            downsample_type=downsample_type,
        )
    elif down_block_type == "CrossAttnDownBlock2D":
        if cross_attention_dim is None:
            raise ValueError("cross_attention_dim must be specified for CrossAttnDownBlock2D")
        return CrossAttnDownBlock2D(
            num_layers=num_layers,
            transformer_layers_per_block=transformer_layers_per_block,
            in_channels=in_channels,
            out_channels=out_channels,
            temb_channels=temb_channels,
            add_downsample=add_downsample,
            resnet_eps=resnet_eps,
            resnet_act_fn=resnet_act_fn,
            resnet_groups=resnet_groups,
            downsample_padding=downsample_padding,
            cross_attention_dim=cross_attention_dim,
            num_attention_heads=num_attention_heads,
            dual_cross_attention=dual_cross_attention,
            use_linear_projection=use_linear_projection,
            only_cross_attention=only_cross_attention,
            upcast_attention=upcast_attention,
            resnet_time_scale_shift=resnet_time_scale_shift,
        )
    # custom MV2D attention block
    elif down_block_type == "CrossAttnDownBlockMV2D":
        if cross_attention_dim is None:
            raise ValueError("cross_attention_dim must be specified for CrossAttnDownBlockMV2D")
        return CrossAttnDownBlockMV2D(
            num_layers=num_layers,
            transformer_layers_per_block=transformer_layers_per_block,
            in_channels=in_channels,
            out_channels=out_channels,
            temb_channels=temb_channels,
            add_downsample=add_downsample,
            resnet_eps=resnet_eps,
            resnet_act_fn=resnet_act_fn,
            resnet_groups=resnet_groups,
            downsample_padding=downsample_padding,
            cross_attention_dim=cross_attention_dim,
            num_attention_heads=num_attention_heads,
            dual_cross_attention=dual_cross_attention,
            use_linear_projection=use_linear_projection,
            only_cross_attention=only_cross_attention,
            upcast_attention=upcast_attention,
            resnet_time_scale_shift=resnet_time_scale_shift,
            num_views=num_views,
            cd_attention_last=cd_attention_last,
            cd_attention_mid=cd_attention_mid,
            multiview_attention=multiview_attention,
            sparse_mv_attention=sparse_mv_attention,
            selfattn_block=selfattn_block,
            mvcd_attention=mvcd_attention,
            use_dino=use_dino
        )
    elif down_block_type == "SimpleCrossAttnDownBlock2D":
        if cross_attention_dim is None:
            raise ValueError("cross_attention_dim must be specified for SimpleCrossAttnDownBlock2D")
        return SimpleCrossAttnDownBlock2D(
            num_layers=num_layers,
            in_channels=in_channels,
            out_channels=out_channels,
            temb_channels=temb_channels,
            add_downsample=add_downsample,
            resnet_eps=resnet_eps,
            resnet_act_fn=resnet_act_fn,
            resnet_groups=resnet_groups,
            cross_attention_dim=cross_attention_dim,
            attention_head_dim=attention_head_dim,
            resnet_time_scale_shift=resnet_time_scale_shift,
            skip_time_act=resnet_skip_time_act,
            output_scale_factor=resnet_out_scale_factor,
            only_cross_attention=only_cross_attention,
            cross_attention_norm=cross_attention_norm,
        )
    elif down_block_type == "SkipDownBlock2D":
        return SkipDownBlock2D(
            num_layers=num_layers,
            in_channels=in_channels,
            out_channels=out_channels,
            temb_channels=temb_channels,
            add_downsample=add_downsample,
            resnet_eps=resnet_eps,
            resnet_act_fn=resnet_act_fn,
            downsample_padding=downsample_padding,
            resnet_time_scale_shift=resnet_time_scale_shift,
        )
    elif down_block_type == "AttnSkipDownBlock2D":
        return AttnSkipDownBlock2D(
            num_layers=num_layers,
            in_channels=in_channels,
            out_channels=out_channels,
            temb_channels=temb_channels,
            add_downsample=add_downsample,
            resnet_eps=resnet_eps,
            resnet_act_fn=resnet_act_fn,
            attention_head_dim=attention_head_dim,
            resnet_time_scale_shift=resnet_time_scale_shift,
        )
    elif down_block_type == "DownEncoderBlock2D":
        return DownEncoderBlock2D(
            num_layers=num_layers,
            in_channels=in_channels,
            out_channels=out_channels,
            add_downsample=add_downsample,
            resnet_eps=resnet_eps,
            resnet_act_fn=resnet_act_fn,
            resnet_groups=resnet_groups,
            downsample_padding=downsample_padding,
            resnet_time_scale_shift=resnet_time_scale_shift,
        )
    elif down_block_type == "AttnDownEncoderBlock2D":
        return AttnDownEncoderBlock2D(
            num_layers=num_layers,
            in_channels=in_channels,
            out_channels=out_channels,
            add_downsample=add_downsample,
            resnet_eps=resnet_eps,
            resnet_act_fn=resnet_act_fn,
            resnet_groups=resnet_groups,
            downsample_padding=downsample_padding,
            attention_head_dim=attention_head_dim,
            resnet_time_scale_shift=resnet_time_scale_shift,
        )
    elif down_block_type == "KDownBlock2D":
        return KDownBlock2D(
            num_layers=num_layers,
            in_channels=in_channels,
            out_channels=out_channels,
            temb_channels=temb_channels,
            add_downsample=add_downsample,
            resnet_eps=resnet_eps,
            resnet_act_fn=resnet_act_fn,
        )
    elif down_block_type == "KCrossAttnDownBlock2D":
        return KCrossAttnDownBlock2D(
            num_layers=num_layers,
            in_channels=in_channels,
            out_channels=out_channels,
            temb_channels=temb_channels,
            add_downsample=add_downsample,
            resnet_eps=resnet_eps,
            resnet_act_fn=resnet_act_fn,
            cross_attention_dim=cross_attention_dim,
            attention_head_dim=attention_head_dim,
            add_self_attention=True if not add_downsample else False,
        )
    raise ValueError(f"{down_block_type} does not exist.")


def get_up_block(
    up_block_type,
    num_layers,
    in_channels,
    out_channels,
    prev_output_channel,
    temb_channels,
    add_upsample,
    resnet_eps,
    resnet_act_fn,
    transformer_layers_per_block=1,
    num_attention_heads=None,
    resnet_groups=None,
    cross_attention_dim=None,
    dual_cross_attention=False,
    use_linear_projection=False,
    only_cross_attention=False,
    upcast_attention=False,
    resnet_time_scale_shift="default",
    resnet_skip_time_act=False,
    resnet_out_scale_factor=1.0,
    cross_attention_norm=None,
    attention_head_dim=None,
    upsample_type=None,
    num_views=1,
    cd_attention_last: bool = False,
    cd_attention_mid: bool = False,
    multiview_attention: bool = True,
    sparse_mv_attention: bool = False,
    selfattn_block: str = "custom",
    mvcd_attention: bool=False,
    use_dino: bool = False  
):
    # If attn head dim is not defined, we default it to the number of heads
    if attention_head_dim is None:
        logger.warn(
            f"It is recommended to provide `attention_head_dim` when calling `get_up_block`. Defaulting `attention_head_dim` to {num_attention_heads}."
        )
        attention_head_dim = num_attention_heads

    up_block_type = up_block_type[7:] if up_block_type.startswith("UNetRes") else up_block_type
    if up_block_type == "UpBlock2D":
        return UpBlock2D(
            num_layers=num_layers,
            in_channels=in_channels,
            out_channels=out_channels,
            prev_output_channel=prev_output_channel,
            temb_channels=temb_channels,
            add_upsample=add_upsample,
            resnet_eps=resnet_eps,
            resnet_act_fn=resnet_act_fn,
            resnet_groups=resnet_groups,
            resnet_time_scale_shift=resnet_time_scale_shift,
        )
    elif up_block_type == "ResnetUpsampleBlock2D":
        return ResnetUpsampleBlock2D(
            num_layers=num_layers,
            in_channels=in_channels,
            out_channels=out_channels,
            prev_output_channel=prev_output_channel,
            temb_channels=temb_channels,
            add_upsample=add_upsample,
            resnet_eps=resnet_eps,
            resnet_act_fn=resnet_act_fn,
            resnet_groups=resnet_groups,
            resnet_time_scale_shift=resnet_time_scale_shift,
            skip_time_act=resnet_skip_time_act,
            output_scale_factor=resnet_out_scale_factor,
        )
    elif up_block_type == "CrossAttnUpBlock2D":
        if cross_attention_dim is None:
            raise ValueError("cross_attention_dim must be specified for CrossAttnUpBlock2D")
        return CrossAttnUpBlock2D(
            num_layers=num_layers,
            transformer_layers_per_block=transformer_layers_per_block,
            in_channels=in_channels,
            out_channels=out_channels,
            prev_output_channel=prev_output_channel,
            temb_channels=temb_channels,
            add_upsample=add_upsample,
            resnet_eps=resnet_eps,
            resnet_act_fn=resnet_act_fn,
            resnet_groups=resnet_groups,
            cross_attention_dim=cross_attention_dim,
            num_attention_heads=num_attention_heads,
            dual_cross_attention=dual_cross_attention,
            use_linear_projection=use_linear_projection,
            only_cross_attention=only_cross_attention,
            upcast_attention=upcast_attention,
            resnet_time_scale_shift=resnet_time_scale_shift,
        )
    # custom MV2D attention block
    elif up_block_type == "CrossAttnUpBlockMV2D":
        if cross_attention_dim is None:
            raise ValueError("cross_attention_dim must be specified for CrossAttnUpBlockMV2D")
        return CrossAttnUpBlockMV2D(
            num_layers=num_layers,
            transformer_layers_per_block=transformer_layers_per_block,
            in_channels=in_channels,
            out_channels=out_channels,
            prev_output_channel=prev_output_channel,
            temb_channels=temb_channels,
            add_upsample=add_upsample,
            resnet_eps=resnet_eps,
            resnet_act_fn=resnet_act_fn,
            resnet_groups=resnet_groups,
            cross_attention_dim=cross_attention_dim,
            num_attention_heads=num_attention_heads,
            dual_cross_attention=dual_cross_attention,
            use_linear_projection=use_linear_projection,
            only_cross_attention=only_cross_attention,
            upcast_attention=upcast_attention,
            resnet_time_scale_shift=resnet_time_scale_shift,
            num_views=num_views,
            cd_attention_last=cd_attention_last,
            cd_attention_mid=cd_attention_mid,
            multiview_attention=multiview_attention,
            sparse_mv_attention=sparse_mv_attention,
            selfattn_block=selfattn_block,
            mvcd_attention=mvcd_attention,
            use_dino=use_dino
        )    
    elif up_block_type == "SimpleCrossAttnUpBlock2D":
        if cross_attention_dim is None:
            raise ValueError("cross_attention_dim must be specified for SimpleCrossAttnUpBlock2D")
        return SimpleCrossAttnUpBlock2D(
            num_layers=num_layers,
            in_channels=in_channels,
            out_channels=out_channels,
            prev_output_channel=prev_output_channel,
            temb_channels=temb_channels,
            add_upsample=add_upsample,
            resnet_eps=resnet_eps,
            resnet_act_fn=resnet_act_fn,
            resnet_groups=resnet_groups,
            cross_attention_dim=cross_attention_dim,
            attention_head_dim=attention_head_dim,
            resnet_time_scale_shift=resnet_time_scale_shift,
            skip_time_act=resnet_skip_time_act,
            output_scale_factor=resnet_out_scale_factor,
            only_cross_attention=only_cross_attention,
            cross_attention_norm=cross_attention_norm,
        )
    elif up_block_type == "AttnUpBlock2D":
        if add_upsample is False:
            upsample_type = None
        else:
            upsample_type = upsample_type or "conv"  # default to 'conv'

        return AttnUpBlock2D(
            num_layers=num_layers,
            in_channels=in_channels,
            out_channels=out_channels,
            prev_output_channel=prev_output_channel,
            temb_channels=temb_channels,
            resnet_eps=resnet_eps,
            resnet_act_fn=resnet_act_fn,
            resnet_groups=resnet_groups,
            attention_head_dim=attention_head_dim,
            resnet_time_scale_shift=resnet_time_scale_shift,
            upsample_type=upsample_type,
        )
    elif up_block_type == "SkipUpBlock2D":
        return SkipUpBlock2D(
            num_layers=num_layers,
            in_channels=in_channels,
            out_channels=out_channels,
            prev_output_channel=prev_output_channel,
            temb_channels=temb_channels,
            add_upsample=add_upsample,
            resnet_eps=resnet_eps,
            resnet_act_fn=resnet_act_fn,
            resnet_time_scale_shift=resnet_time_scale_shift,
        )
    elif up_block_type == "AttnSkipUpBlock2D":
        return AttnSkipUpBlock2D(
            num_layers=num_layers,
            in_channels=in_channels,
            out_channels=out_channels,
            prev_output_channel=prev_output_channel,
            temb_channels=temb_channels,
            add_upsample=add_upsample,
            resnet_eps=resnet_eps,
            resnet_act_fn=resnet_act_fn,
            attention_head_dim=attention_head_dim,
            resnet_time_scale_shift=resnet_time_scale_shift,
        )
    elif up_block_type == "UpDecoderBlock2D":
        return UpDecoderBlock2D(
            num_layers=num_layers,
            in_channels=in_channels,
            out_channels=out_channels,
            add_upsample=add_upsample,
            resnet_eps=resnet_eps,
            resnet_act_fn=resnet_act_fn,
            resnet_groups=resnet_groups,
            resnet_time_scale_shift=resnet_time_scale_shift,
            temb_channels=temb_channels,
        )
    elif up_block_type == "AttnUpDecoderBlock2D":
        return AttnUpDecoderBlock2D(
            num_layers=num_layers,
            in_channels=in_channels,
            out_channels=out_channels,
            add_upsample=add_upsample,
            resnet_eps=resnet_eps,
            resnet_act_fn=resnet_act_fn,
            resnet_groups=resnet_groups,
            attention_head_dim=attention_head_dim,
            resnet_time_scale_shift=resnet_time_scale_shift,
            temb_channels=temb_channels,
        )
    elif up_block_type == "KUpBlock2D":
        return KUpBlock2D(
            num_layers=num_layers,
            in_channels=in_channels,
            out_channels=out_channels,
            temb_channels=temb_channels,
            add_upsample=add_upsample,
            resnet_eps=resnet_eps,
            resnet_act_fn=resnet_act_fn,
        )
    elif up_block_type == "KCrossAttnUpBlock2D":
        return KCrossAttnUpBlock2D(
            num_layers=num_layers,
            in_channels=in_channels,
            out_channels=out_channels,
            temb_channels=temb_channels,
            add_upsample=add_upsample,
            resnet_eps=resnet_eps,
            resnet_act_fn=resnet_act_fn,
            cross_attention_dim=cross_attention_dim,
            attention_head_dim=attention_head_dim,
        )

    raise ValueError(f"{up_block_type} does not exist.")


class UNetMidBlockMV2DCrossAttn(nn.Module):
    def __init__(
        self,
        in_channels: int,
        temb_channels: int,
        dropout: float = 0.0,
        num_layers: int = 1,
        transformer_layers_per_block: 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=1,
        output_scale_factor=1.0,
        cross_attention_dim=1280,
        dual_cross_attention=False,
        use_linear_projection=False,
        upcast_attention=False,
        num_views: int = 1,
        cd_attention_last: bool = False,
        cd_attention_mid: bool = False,
        multiview_attention: bool = True,
        sparse_mv_attention: bool = False,
        selfattn_block: str = "custom", 
        mvcd_attention: bool=False,
        use_dino: bool = False
    ):
        super().__init__()

        self.has_cross_attention = True
        self.num_attention_heads = num_attention_heads
        resnet_groups = resnet_groups if resnet_groups is not None else min(in_channels // 4, 32)
        if selfattn_block == "custom":
            from .transformer_mv2d import TransformerMV2DModel
        elif selfattn_block == "rowwise":
            from .transformer_mv2d_rowwise import TransformerMV2DModel
        elif selfattn_block == "self_rowwise":
            from .transformer_mv2d_self_rowwise import TransformerMV2DModel
        else:
            raise NotImplementedError
        
        # there is always at least one resnet
        resnets = [
            ResnetBlock2D(
                in_channels=in_channels,
                out_channels=in_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,
            )
        ]
        attentions = []

        for _ in range(num_layers):
            if not dual_cross_attention:
                attentions.append(
                    TransformerMV2DModel(
                        num_attention_heads,
                        in_channels // num_attention_heads,
                        in_channels=in_channels,
                        num_layers=transformer_layers_per_block,
                        cross_attention_dim=cross_attention_dim,
                        norm_num_groups=resnet_groups,
                        use_linear_projection=use_linear_projection,
                        upcast_attention=upcast_attention,
                        num_views=num_views,
                        cd_attention_last=cd_attention_last,
                        cd_attention_mid=cd_attention_mid,
                        multiview_attention=multiview_attention,
                        sparse_mv_attention=sparse_mv_attention,
                        mvcd_attention=mvcd_attention,
                        use_dino=use_dino
                    )
                )
            else:
                raise NotImplementedError
            resnets.append(
                ResnetBlock2D(
                    in_channels=in_channels,
                    out_channels=in_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,
                )
            )

        self.attentions = nn.ModuleList(attentions)
        self.resnets = nn.ModuleList(resnets)

    def forward(
        self,
        hidden_states: torch.FloatTensor,
        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,
        dino_feature: Optional[torch.FloatTensor] = None
    ) -> torch.FloatTensor:
        hidden_states = self.resnets[0](hidden_states, temb)
        for attn, resnet in zip(self.attentions, self.resnets[1:]):
            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,
                dino_feature=dino_feature,
                return_dict=False,
            )[0]
            hidden_states = resnet(hidden_states, temb)

        return hidden_states


class CrossAttnUpBlockMV2D(nn.Module):
    def __init__(
        self,
        in_channels: int,
        out_channels: int,
        prev_output_channel: int,
        temb_channels: int,
        dropout: float = 0.0,
        num_layers: int = 1,
        transformer_layers_per_block: 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=1,
        cross_attention_dim=1280,
        output_scale_factor=1.0,
        add_upsample=True,
        dual_cross_attention=False,
        use_linear_projection=False,
        only_cross_attention=False,
        upcast_attention=False,
        num_views: int = 1,
        cd_attention_last: bool = False,
        cd_attention_mid: bool = False,
        multiview_attention: bool = True,
        sparse_mv_attention: bool = False,
        selfattn_block: str = "custom",
        mvcd_attention: bool=False,
        use_dino: bool = False
    ):
        super().__init__()
        resnets = []
        attentions = []

        self.has_cross_attention = True
        self.num_attention_heads = num_attention_heads

        if selfattn_block == "custom":
            from .transformer_mv2d import TransformerMV2DModel
        elif selfattn_block == "rowwise":
            from .transformer_mv2d_rowwise import TransformerMV2DModel
        elif selfattn_block == "self_rowwise":
            from .transformer_mv2d_self_rowwise import TransformerMV2DModel
        else:
            raise NotImplementedError
        
        for i in range(num_layers):
            res_skip_channels = in_channels if (i == num_layers - 1) else out_channels
            resnet_in_channels = prev_output_channel if i == 0 else out_channels

            resnets.append(
                ResnetBlock2D(
                    in_channels=resnet_in_channels + res_skip_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,
                )
            )
            if not dual_cross_attention:
                attentions.append(
                    TransformerMV2DModel(
                        num_attention_heads,
                        out_channels // num_attention_heads,
                        in_channels=out_channels,
                        num_layers=transformer_layers_per_block,
                        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,
                        num_views=num_views,
                        cd_attention_last=cd_attention_last,
                        cd_attention_mid=cd_attention_mid,
                        multiview_attention=multiview_attention,
                        sparse_mv_attention=sparse_mv_attention,
                        mvcd_attention=mvcd_attention,
                        use_dino=use_dino
                    )
                )
            else:
                raise NotImplementedError
        self.attentions = nn.ModuleList(attentions)
        self.resnets = nn.ModuleList(resnets)

        if add_upsample:
            self.upsamplers = nn.ModuleList([Upsample2D(out_channels, use_conv=True, out_channels=out_channels)])
        else:
            self.upsamplers = None

        self.gradient_checkpointing = False

    def forward(
        self,
        hidden_states: torch.FloatTensor,
        res_hidden_states_tuple: Tuple[torch.FloatTensor, ...],
        temb: Optional[torch.FloatTensor] = None,
        encoder_hidden_states: Optional[torch.FloatTensor] = None,
        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
        upsample_size: Optional[int] = None,
        attention_mask: Optional[torch.FloatTensor] = None,
        encoder_attention_mask: Optional[torch.FloatTensor] = None,
        dino_feature: Optional[torch.FloatTensor] = None
    ):
        for resnet, attn in zip(self.resnets, self.attentions):
            # pop res hidden states
            res_hidden_states = res_hidden_states_tuple[-1]
            res_hidden_states_tuple = res_hidden_states_tuple[:-1]
            hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)

            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 = torch.utils.checkpoint.checkpoint(
                    create_custom_forward(attn, return_dict=False),
                    hidden_states,
                    encoder_hidden_states,
                    dino_feature,
                    None,  # timestep
                    None,  # class_labels
                    cross_attention_kwargs,
                    attention_mask,
                    encoder_attention_mask,
                    **ckpt_kwargs,
                )[0]
            else:
                hidden_states = resnet(hidden_states, temb)
                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,
                    dino_feature=dino_feature,
                    return_dict=False,
                )[0]

        if self.upsamplers is not None:
            for upsampler in self.upsamplers:
                hidden_states = upsampler(hidden_states, upsample_size)

        return hidden_states


class CrossAttnDownBlockMV2D(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: 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=1,
        cross_attention_dim=1280,
        output_scale_factor=1.0,
        downsample_padding=1,
        add_downsample=True,
        dual_cross_attention=False,
        use_linear_projection=False,
        only_cross_attention=False,
        upcast_attention=False,
        num_views: int = 1,
        cd_attention_last: bool = False,
        cd_attention_mid: bool = False,
        multiview_attention: bool = True,
        sparse_mv_attention: bool = False,
        selfattn_block: str = "custom",
        mvcd_attention: bool=False,
        use_dino: bool = False
    ):
        super().__init__()
        resnets = []
        attentions = []

        self.has_cross_attention = True
        self.num_attention_heads = num_attention_heads
        if selfattn_block == "custom":
            from .transformer_mv2d import TransformerMV2DModel
        elif selfattn_block == "rowwise":
            from .transformer_mv2d_rowwise import TransformerMV2DModel
        elif selfattn_block == "self_rowwise":
            from .transformer_mv2d_self_rowwise import TransformerMV2DModel
        else:
            raise NotImplementedError
        
        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,
                )
            )
            if not dual_cross_attention:
                attentions.append(
                    TransformerMV2DModel(
                        num_attention_heads,
                        out_channels // num_attention_heads,
                        in_channels=out_channels,
                        num_layers=transformer_layers_per_block,
                        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,
                        num_views=num_views,
                        cd_attention_last=cd_attention_last,
                        cd_attention_mid=cd_attention_mid,
                        multiview_attention=multiview_attention,
                        sparse_mv_attention=sparse_mv_attention,
                        mvcd_attention=mvcd_attention,
                        use_dino=use_dino
                    )
                )
            else:
                raise NotImplementedError
        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,
        temb: Optional[torch.FloatTensor] = None,
        encoder_hidden_states: Optional[torch.FloatTensor] = None,
        dino_feature: 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=None,
    ):
        output_states = ()

        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 = torch.utils.checkpoint.checkpoint(
                    create_custom_forward(attn, return_dict=False),
                    hidden_states,
                    encoder_hidden_states,
                    dino_feature,
                    None,  # timestep
                    None,  # class_labels
                    cross_attention_kwargs,
                    attention_mask,
                    encoder_attention_mask,
                    **ckpt_kwargs,
                )[0]
            else:
                hidden_states = resnet(hidden_states, temb)
                hidden_states = attn(
                    hidden_states,
                    encoder_hidden_states=encoder_hidden_states,
                    dino_feature=dino_feature,
                    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)

            output_states = output_states + (hidden_states,)

        return hidden_states, output_states