# 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 from diffusers.models.dual_transformer_2d import DualTransformer2DModel from diffusers.models.resnet import Downsample2D, ResnetBlock2D, Upsample2D from .transformer_2d import Transformer2DModel def get_down_block( down_block_type, num_layers, in_channels, out_channels, temb_channels, add_downsample, resnet_eps, resnet_act_fn, attn_num_head_channels, 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, use_gated_attention=False, ): 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 == "CrossAttnDownBlock2D": if cross_attention_dim is None: raise ValueError( "cross_attention_dim must be specified for CrossAttnDownBlock2D") return CrossAttnDownBlock2D( 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, cross_attention_dim=cross_attention_dim, attn_num_head_channels=attn_num_head_channels, 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, use_gated_attention=use_gated_attention, ) 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, attn_num_head_channels, 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, use_gated_attention=False, ): 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 == "CrossAttnUpBlock2D": if cross_attention_dim is None: raise ValueError( "cross_attention_dim must be specified for CrossAttnUpBlock2D") return CrossAttnUpBlock2D( 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, attn_num_head_channels=attn_num_head_channels, 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, use_gated_attention=use_gated_attention, ) raise ValueError(f"{up_block_type} does not exist.") class UNetMidBlock2DCrossAttn(nn.Module): def __init__( self, in_channels: int, temb_channels: int, dropout: float = 0.0, num_layers: 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, attn_num_head_channels=1, output_scale_factor=1.0, cross_attention_dim=1280, dual_cross_attention=False, use_linear_projection=False, upcast_attention=False, use_gated_attention=False, ): super().__init__() self.has_cross_attention = True self.attn_num_head_channels = attn_num_head_channels resnet_groups = resnet_groups if resnet_groups is not None else min( in_channels // 4, 32) # 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( Transformer2DModel( attn_num_head_channels, in_channels // attn_num_head_channels, in_channels=in_channels, num_layers=1, cross_attention_dim=cross_attention_dim, norm_num_groups=resnet_groups, use_linear_projection=use_linear_projection, upcast_attention=upcast_attention, use_gated_attention=use_gated_attention, ) ) else: attentions.append( DualTransformer2DModel( attn_num_head_channels, in_channels // attn_num_head_channels, in_channels=in_channels, num_layers=1, cross_attention_dim=cross_attention_dim, norm_num_groups=resnet_groups, ) ) 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, return_cross_attention_probs: bool = False, ) -> torch.FloatTensor: hidden_states = self.resnets[0](hidden_states, temb) cross_attention_probs_all = [] base_attn_key = cross_attention_kwargs["attn_key"] for attn_key, (attn, resnet) in enumerate(zip(self.attentions, self.resnets[1:])): cross_attention_kwargs["attn_key"] = base_attn_key + [attn_key] 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, return_cross_attention_probs=return_cross_attention_probs, ) if return_cross_attention_probs: hidden_states, cross_attention_probs = hidden_states cross_attention_probs_all.append(cross_attention_probs) else: hidden_states = hidden_states[0] hidden_states = resnet(hidden_states, temb) if return_cross_attention_probs: return hidden_states, cross_attention_probs_all return hidden_states class CrossAttnDownBlock2D(nn.Module): def __init__( self, in_channels: int, out_channels: int, temb_channels: int, dropout: float = 0.0, num_layers: 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, attn_num_head_channels=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, use_gated_attention=False, ): super().__init__() resnets = [] attentions = [] self.has_cross_attention = True self.attn_num_head_channels = attn_num_head_channels 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( Transformer2DModel( attn_num_head_channels, out_channels // attn_num_head_channels, in_channels=out_channels, num_layers=1, 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, use_gated_attention=use_gated_attention ) ) else: attentions.append( DualTransformer2DModel( attn_num_head_channels, out_channels // attn_num_head_channels, 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, 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, return_cross_attention_probs: bool = False, ): output_states = () cross_attention_probs_all = [] base_attn_key = cross_attention_kwargs["attn_key"] for attn_key, (resnet, attn) in enumerate(zip(self.resnets, self.attentions)): cross_attention_kwargs["attn_key"] = base_attn_key + [attn_key] 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, None, # timestep None, # class_labels cross_attention_kwargs, attention_mask, encoder_attention_mask, return_cross_attention_probs=return_cross_attention_probs, **ckpt_kwargs, ) if return_cross_attention_probs: hidden_states, cross_attention_probs = hidden_states cross_attention_probs_all.append(cross_attention_probs) else: hidden_states = hidden_states[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, return_dict=False, return_cross_attention_probs=return_cross_attention_probs, ) if return_cross_attention_probs: hidden_states, cross_attention_probs = hidden_states cross_attention_probs_all.append(cross_attention_probs) else: hidden_states = hidden_states[0] 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,) if return_cross_attention_probs: return hidden_states, output_states, cross_attention_probs_all return hidden_states, output_states class DownBlock2D(nn.Module): def __init__( self, in_channels: int, out_channels: int, temb_channels: int, dropout: float = 0.0, num_layers: 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, output_scale_factor=1.0, add_downsample=True, downsample_padding=1, ): super().__init__() resnets = [] 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, ) ) 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, temb=None): output_states = () for resnet in self.resnets: if self.training and self.gradient_checkpointing: def create_custom_forward(module): def custom_forward(*inputs): return module(*inputs) return custom_forward if is_torch_version(">=", "1.11.0"): hidden_states = torch.utils.checkpoint.checkpoint( create_custom_forward(resnet), hidden_states, temb, use_reentrant=False ) else: hidden_states = torch.utils.checkpoint.checkpoint( create_custom_forward(resnet), hidden_states, temb ) else: hidden_states = resnet(hidden_states, temb) 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 class CrossAttnUpBlock2D(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, 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, attn_num_head_channels=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, use_gated_attention=False, ): super().__init__() resnets = [] attentions = [] self.has_cross_attention = True self.attn_num_head_channels = attn_num_head_channels 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( Transformer2DModel( attn_num_head_channels, out_channels // attn_num_head_channels, in_channels=out_channels, num_layers=1, 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, use_gated_attention=use_gated_attention, ) ) else: attentions.append( DualTransformer2DModel( attn_num_head_channels, out_channels // attn_num_head_channels, 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_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, return_cross_attention_probs: bool = False, ): cross_attention_probs_all = [] base_attn_key = cross_attention_kwargs["attn_key"] for attn_key, (resnet, attn) in enumerate(zip(self.resnets, self.attentions)): cross_attention_kwargs["attn_key"] = base_attn_key + [attn_key] # 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, None, # timestep None, # class_labels cross_attention_kwargs, attention_mask, encoder_attention_mask, **ckpt_kwargs, ) if return_cross_attention_probs: hidden_states, cross_attention_probs = hidden_states cross_attention_probs_all.append(cross_attention_probs) else: hidden_states = hidden_states[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, return_dict=False, return_cross_attention_probs=return_cross_attention_probs, ) if return_cross_attention_probs: hidden_states, cross_attention_probs = hidden_states cross_attention_probs_all.append(cross_attention_probs) else: hidden_states = hidden_states[0] if self.upsamplers is not None: for upsampler in self.upsamplers: hidden_states = upsampler(hidden_states, upsample_size) if return_cross_attention_probs: return hidden_states, cross_attention_probs_all return hidden_states class UpBlock2D(nn.Module): def __init__( self, in_channels: int, prev_output_channel: int, out_channels: int, temb_channels: int, dropout: float = 0.0, num_layers: 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, output_scale_factor=1.0, add_upsample=True, ): super().__init__() resnets = [] 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, ) ) 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, res_hidden_states_tuple, temb=None, upsample_size=None): for resnet in self.resnets: # 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): def custom_forward(*inputs): return module(*inputs) return custom_forward if is_torch_version(">=", "1.11.0"): hidden_states = torch.utils.checkpoint.checkpoint( create_custom_forward(resnet), hidden_states, temb, use_reentrant=False ) else: hidden_states = torch.utils.checkpoint.checkpoint( create_custom_forward(resnet), hidden_states, temb ) else: hidden_states = resnet(hidden_states, temb) if self.upsamplers is not None: for upsampler in self.upsamplers: hidden_states = upsampler(hidden_states, upsample_size) return hidden_states