from typing import Any, Dict, Optional import torch import torch.nn.functional as F from torch import nn from diffusers.configuration_utils import LegacyConfigMixin, register_to_config from diffusers.utils import deprecate, logging from diffusers.utils.torch_utils import maybe_allow_in_graph from diffusers.models.attention import BasicTransformerBlock, FeedForward, _chunked_feed_forward, TemporalBasicTransformerBlock from diffusers.models.attention_processor import Attention from diffusers.models.embeddings import ImagePositionalEmbeddings, PatchEmbed, PixArtAlphaTextProjection from diffusers.models.modeling_outputs import Transformer2DModelOutput from diffusers.models.modeling_utils import LegacyModelMixin from diffusers.models.normalization import AdaLayerNormSingle logger = logging.get_logger(__name__) # pylint: disable=invalid-name @maybe_allow_in_graph class CrossFrameTransformerBlock(nn.Module): r""" modified from TemporalBasicTransformerBlock A basic Transformer block for video like data. Parameters: dim (`int`): The number of channels in the input and output. time_mix_inner_dim (`int`): The number of channels for temporal attention. num_attention_heads (`int`): The number of heads to use for multi-head attention. attention_head_dim (`int`): The number of channels in each head. cross_attention_dim (`int`, *optional*): The size of the encoder_hidden_states vector for cross attention. """ def __init__( self, dim: int, time_mix_inner_dim: int, num_attention_heads: int, attention_head_dim: int, cross_attention_dim: Optional[int] = None, ): super().__init__() self.is_res = dim == time_mix_inner_dim self.norm_in = nn.LayerNorm(dim) # Define 3 blocks. Each block has its own normalization layer. # 1. Self-Attn self.ff_in = FeedForward( dim, dim_out=time_mix_inner_dim, activation_fn="geglu", ) self.norm1 = nn.LayerNorm(time_mix_inner_dim) self.attn1 = Attention( query_dim=time_mix_inner_dim, heads=num_attention_heads, dim_head=attention_head_dim, cross_attention_dim=None, ) # 2. Cross-Attn if cross_attention_dim is not None: # We currently only use AdaLayerNormZero for self attention where there will only be one attention block. # I.e. the number of returned modulation chunks from AdaLayerZero would not make sense if returned during # the second cross attention block. self.norm2 = nn.LayerNorm(time_mix_inner_dim) self.attn2 = Attention( query_dim=time_mix_inner_dim, cross_attention_dim=cross_attention_dim, heads=num_attention_heads, dim_head=attention_head_dim, ) # is self-attn if encoder_hidden_states is none else: self.norm2 = None self.attn2 = None # 3. Feed-forward self.norm3 = nn.LayerNorm(time_mix_inner_dim) self.ff = FeedForward(time_mix_inner_dim, activation_fn="geglu") # let chunk size default to None self._chunk_size = None self._chunk_dim = None def set_chunk_feed_forward(self, chunk_size: Optional[int], **kwargs): # Sets chunk feed-forward self._chunk_size = chunk_size # chunk dim should be hardcoded to 1 to have better speed vs. memory trade-off self._chunk_dim = 1 def forward( self, hidden_states: torch.Tensor, num_frames: int, encoder_hidden_states: Optional[torch.Tensor] = None, ) -> torch.Tensor: # Notice that normalization is always applied before the real computation in the following blocks. # 0. Self-Attention batch_size = hidden_states.shape[0] batch_frames, seq_length, channels = hidden_states.shape batch_size = batch_frames // num_frames hidden_states = hidden_states[None, :].reshape(batch_size, num_frames, seq_length, channels) hidden_states = hidden_states.permute(0, 2, 1, 3) hidden_states = hidden_states.reshape(batch_size * seq_length, num_frames, channels) residual = hidden_states hidden_states = self.norm_in(hidden_states) if self._chunk_size is not None: hidden_states = _chunked_feed_forward(self.ff_in, hidden_states, self._chunk_dim, self._chunk_size) else: hidden_states = self.ff_in(hidden_states) if self.is_res: hidden_states = hidden_states + residual norm_hidden_states = self.norm1(hidden_states) attn_output = self.attn1(norm_hidden_states, encoder_hidden_states=None) hidden_states = attn_output + hidden_states # 3. Cross-Attention if self.attn2 is not None: norm_hidden_states = self.norm2(hidden_states) attn_output = self.attn2(norm_hidden_states, encoder_hidden_states=encoder_hidden_states) hidden_states = attn_output + hidden_states # 4. Feed-forward norm_hidden_states = self.norm3(hidden_states) if self._chunk_size is not None: ff_output = _chunked_feed_forward(self.ff, norm_hidden_states, self._chunk_dim, self._chunk_size) else: ff_output = self.ff(norm_hidden_states) # if self.is_res: # hidden_states = ff_output + hidden_states # else: hidden_states = ff_output hidden_states = hidden_states[None, :].reshape(batch_size, seq_length, num_frames, channels) hidden_states = hidden_states.permute(0, 2, 1, 3) hidden_states = hidden_states.reshape(batch_size * num_frames, seq_length, channels) return hidden_states class Transformer3DModel(LegacyModelMixin, LegacyConfigMixin): """ A 2D Transformer model for image-like data. Parameters: num_attention_heads (`int`, *optional*, defaults to 16): The number of heads to use for multi-head attention. attention_head_dim (`int`, *optional*, defaults to 88): The number of channels in each head. in_channels (`int`, *optional*): The number of channels in the input and output (specify if the input is **continuous**). num_layers (`int`, *optional*, defaults to 1): The number of layers of Transformer blocks to use. dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use. cross_attention_dim (`int`, *optional*): The number of `encoder_hidden_states` dimensions to use. sample_size (`int`, *optional*): The width of the latent images (specify if the input is **discrete**). This is fixed during training since it is used to learn a number of position embeddings. num_vector_embeds (`int`, *optional*): The number of classes of the vector embeddings of the latent pixels (specify if the input is **discrete**). Includes the class for the masked latent pixel. activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to use in feed-forward. num_embeds_ada_norm ( `int`, *optional*): The number of diffusion steps used during training. Pass if at least one of the norm_layers is `AdaLayerNorm`. This is fixed during training since it is used to learn a number of embeddings that are added to the hidden states. During inference, you can denoise for up to but not more steps than `num_embeds_ada_norm`. attention_bias (`bool`, *optional*): Configure if the `TransformerBlocks` attention should contain a bias parameter. """ _supports_gradient_checkpointing = True _no_split_modules = ["BasicTransformerBlock"] _skip_layerwise_casting_patterns = ["latent_image_embedding", "norm"] @register_to_config def __init__( self, num_attention_heads: int = 16, attention_head_dim: int = 88, in_channels: Optional[int] = None, out_channels: Optional[int] = None, num_layers: int = 1, dropout: float = 0.0, norm_num_groups: int = 32, cross_attention_dim: Optional[int] = None, attention_bias: bool = False, sample_size: Optional[int] = None, num_vector_embeds: Optional[int] = None, patch_size: Optional[int] = None, activation_fn: str = "geglu", num_embeds_ada_norm: Optional[int] = None, use_linear_projection: bool = False, only_cross_attention: bool = False, double_self_attention: bool = False, upcast_attention: bool = False, norm_type: str = "layer_norm", # 'layer_norm', 'ada_norm', 'ada_norm_zero', 'ada_norm_single', 'ada_norm_continuous', 'layer_norm_i2vgen' norm_elementwise_affine: bool = True, norm_eps: float = 1e-5, attention_type: str = "default", caption_channels: int = None, interpolation_scale: float = None, use_additional_conditions: Optional[bool] = None, ): super().__init__() # Validate inputs. if patch_size is not None: if norm_type not in ["ada_norm", "ada_norm_zero", "ada_norm_single"]: raise NotImplementedError( f"Forward pass is not implemented when `patch_size` is not None and `norm_type` is '{norm_type}'." ) elif norm_type in ["ada_norm", "ada_norm_zero"] and num_embeds_ada_norm is None: raise ValueError( f"When using a `patch_size` and this `norm_type` ({norm_type}), `num_embeds_ada_norm` cannot be None." ) if norm_type == "layer_norm" and num_embeds_ada_norm is not None: deprecation_message = ( f"The configuration file of this model: {self.__class__} is outdated. `norm_type` is either not set or" " incorrectly set to `'layer_norm'`. Make sure to set `norm_type` to `'ada_norm'` in the config." " Please make sure to update the config accordingly as leaving `norm_type` might led to incorrect" " results in future versions. If you have downloaded this checkpoint from the Hugging Face Hub, it" " would be very nice if you could open a Pull request for the `transformer/config.json` file" ) deprecate("norm_type!=num_embeds_ada_norm", "1.0.0", deprecation_message, standard_warn=False) norm_type = "ada_norm" # Set some common variables used across the board. self.use_linear_projection = use_linear_projection self.interpolation_scale = interpolation_scale self.caption_channels = caption_channels self.num_attention_heads = num_attention_heads self.attention_head_dim = attention_head_dim self.inner_dim = self.config.num_attention_heads * self.config.attention_head_dim self.in_channels = in_channels self.out_channels = in_channels if out_channels is None else out_channels self.gradient_checkpointing = False if use_additional_conditions is None: if norm_type == "ada_norm_single" and sample_size == 128: use_additional_conditions = True else: use_additional_conditions = False self.use_additional_conditions = use_additional_conditions self.norm = torch.nn.GroupNorm( num_groups=self.config.norm_num_groups, num_channels=self.in_channels, eps=1e-6, affine=True ) if self.use_linear_projection: self.proj_in = torch.nn.Linear(self.in_channels, self.inner_dim) else: self.proj_in = torch.nn.Conv2d(self.in_channels, self.inner_dim, kernel_size=1, stride=1, padding=0) self.transformer_blocks = nn.ModuleList( [ BasicTransformerBlock( self.inner_dim, self.config.num_attention_heads, self.config.attention_head_dim, dropout=self.config.dropout, cross_attention_dim=self.config.cross_attention_dim, activation_fn=self.config.activation_fn, num_embeds_ada_norm=self.config.num_embeds_ada_norm, attention_bias=self.config.attention_bias, only_cross_attention=self.config.only_cross_attention, double_self_attention=self.config.double_self_attention, upcast_attention=self.config.upcast_attention, norm_type=norm_type, norm_elementwise_affine=self.config.norm_elementwise_affine, norm_eps=self.config.norm_eps, attention_type=self.config.attention_type, ) for _ in range(self.config.num_layers) ] ) if self.use_linear_projection: self.proj_out = torch.nn.Linear(self.inner_dim, self.out_channels) else: self.proj_out = torch.nn.Conv2d(self.inner_dim, self.out_channels, kernel_size=1, stride=1, padding=0) time_mix_inner_dim = self.inner_dim self.temporal_block_stride = 1 temporal_transformer_blocks = [] if self.config.num_layers >= 3: self.temporal_block_stride = 2 for ii in range(self.config.num_layers): if (ii + 1) % self.temporal_block_stride == 0: temporal_transformer_blocks.append( CrossFrameTransformerBlock( self.inner_dim, time_mix_inner_dim, num_attention_heads, attention_head_dim, cross_attention_dim=None, ) ) # else: # print('skip!') self.temporal_transformer_blocks = nn.ModuleList(temporal_transformer_blocks) def forward( self, hidden_states: torch.Tensor, encoder_hidden_states: Optional[torch.Tensor] = None, timestep: Optional[torch.LongTensor] = None, added_cond_kwargs: Dict[str, torch.Tensor] = None, class_labels: Optional[torch.LongTensor] = None, cross_attention_kwargs: Dict[str, Any] = None, attention_mask: Optional[torch.Tensor] = None, encoder_attention_mask: Optional[torch.Tensor] = None, return_dict: bool = True, num_frames=1 ): """ The [`Transformer2DModel`] forward method. Args: hidden_states (`torch.LongTensor` of shape `(batch size, num latent pixels)` if discrete, `torch.Tensor` of shape `(batch size, channel, height, width)` if continuous): Input `hidden_states`. encoder_hidden_states ( `torch.Tensor` of shape `(batch size, sequence len, embed dims)`, *optional*): Conditional embeddings for cross attention layer. If not given, cross-attention defaults to self-attention. timestep ( `torch.LongTensor`, *optional*): Used to indicate denoising step. Optional timestep to be applied as an embedding in `AdaLayerNorm`. class_labels ( `torch.LongTensor` of shape `(batch size, num classes)`, *optional*): Used to indicate class labels conditioning. Optional class labels to be applied as an embedding in `AdaLayerZeroNorm`. cross_attention_kwargs ( `Dict[str, Any]`, *optional*): A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under `self.processor` in [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). attention_mask ( `torch.Tensor`, *optional*): An attention mask of shape `(batch, key_tokens)` is applied to `encoder_hidden_states`. If `1` the mask is kept, otherwise if `0` it is discarded. Mask will be converted into a bias, which adds large negative values to the attention scores corresponding to "discard" tokens. encoder_attention_mask ( `torch.Tensor`, *optional*): Cross-attention mask applied to `encoder_hidden_states`. Two formats supported: * Mask `(batch, sequence_length)` True = keep, False = discard. * Bias `(batch, 1, sequence_length)` 0 = keep, -10000 = discard. If `ndim == 2`: will be interpreted as a mask, then converted into a bias consistent with the format above. This bias will be added to the cross-attention scores. return_dict (`bool`, *optional*, defaults to `True`): Whether or not to return a [`~models.unets.unet_2d_condition.UNet2DConditionOutput`] instead of a plain tuple. Returns: If `return_dict` is True, an [`~models.transformers.transformer_2d.Transformer2DModelOutput`] is returned, otherwise a `tuple` where the first element is the sample tensor. """ if cross_attention_kwargs is not None: if cross_attention_kwargs.get("scale", None) is not None: logger.warning("Passing `scale` to `cross_attention_kwargs` is deprecated. `scale` will be ignored.") # ensure attention_mask is a bias, and give it a singleton query_tokens dimension. # we may have done this conversion already, e.g. if we came here via UNet2DConditionModel#forward. # we can tell by counting dims; if ndim == 2: it's a mask rather than a bias. # expects mask of shape: # [batch, key_tokens] # adds singleton query_tokens dimension: # [batch, 1, key_tokens] # this helps to broadcast it as a bias over attention scores, which will be in one of the following shapes: # [batch, heads, query_tokens, key_tokens] (e.g. torch sdp attn) # [batch * heads, query_tokens, key_tokens] (e.g. xformers or classic attn) if attention_mask is not None and attention_mask.ndim == 2: # assume that mask is expressed as: # (1 = keep, 0 = discard) # convert mask into a bias that can be added to attention scores: # (keep = +0, discard = -10000.0) attention_mask = (1 - attention_mask.to(hidden_states.dtype)) * -10000.0 attention_mask = attention_mask.unsqueeze(1) # convert encoder_attention_mask to a bias the same way we do for attention_mask if encoder_attention_mask is not None and encoder_attention_mask.ndim == 2: encoder_attention_mask = (1 - encoder_attention_mask.to(hidden_states.dtype)) * -10000.0 encoder_attention_mask = encoder_attention_mask.unsqueeze(1) batch_size, _, height, width = hidden_states.shape residual = hidden_states hidden_states = self.norm(hidden_states) if not self.use_linear_projection: hidden_states = self.proj_in(hidden_states) inner_dim = hidden_states.shape[1] hidden_states = hidden_states.permute(0, 2, 3, 1).reshape(batch_size, height * width, inner_dim) else: inner_dim = hidden_states.shape[1] hidden_states = hidden_states.permute(0, 2, 3, 1).reshape(batch_size, height * width, inner_dim) hidden_states = self.proj_in(hidden_states) # 2. Blocks n_temporal_iters = 0 for ii, block in enumerate(self.transformer_blocks): if torch.is_grad_enabled() and self.gradient_checkpointing: hidden_states = self._gradient_checkpointing_func( block, hidden_states, attention_mask, encoder_hidden_states, encoder_attention_mask, timestep, cross_attention_kwargs, class_labels, ) if (ii + 1) % self.temporal_block_stride == 0: temporal_block = self.temporal_transformer_blocks[n_temporal_iters] hidden_states_mix = hidden_states hidden_states_mix = self._gradient_checkpointing_func( temporal_block, hidden_states_mix, num_frames, encoder_hidden_states ) hidden_states = hidden_states + hidden_states_mix n_temporal_iters += 1 else: hidden_states = block( hidden_states, attention_mask=attention_mask, encoder_hidden_states=encoder_hidden_states, encoder_attention_mask=encoder_attention_mask, timestep=timestep, cross_attention_kwargs=cross_attention_kwargs, class_labels=class_labels, ) if (ii + 1) % self.temporal_block_stride == 0: temporal_block = self.temporal_transformer_blocks[n_temporal_iters] hidden_states_mix = hidden_states hidden_states_mix = temporal_block( hidden_states_mix, num_frames=num_frames, encoder_hidden_states=encoder_hidden_states ) hidden_states = hidden_states + hidden_states_mix n_temporal_iters += 1 # 3. Output if not self.use_linear_projection: hidden_states = ( hidden_states.reshape(batch_size, height, width, inner_dim).permute(0, 3, 1, 2).contiguous() ) hidden_states = self.proj_out(hidden_states) else: hidden_states = self.proj_out(hidden_states) hidden_states = ( hidden_states.reshape(batch_size, height, width, inner_dim).permute(0, 3, 1, 2).contiguous() ) output = hidden_states + residual if not return_dict: return (output,) return Transformer2DModelOutput(sample=output)