base/models/transformer_3d.py

# 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 dataclasses import dataclass
from typing import Any, Dict, Optional

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

from diffusers.configuration_utils import ConfigMixin, register_to_config
from diffusers.models.embeddings import ImagePositionalEmbeddings
from diffusers.utils import BaseOutput, deprecate
from diffusers.models.embeddings import PatchEmbed
from diffusers.models.lora import LoRACompatibleConv, LoRACompatibleLinear
from diffusers.models.modeling_utils import ModelMixin
from einops import rearrange, repeat

try:
    from attention import BasicTransformerBlock
except:
    from .attention import BasicTransformerBlock

@dataclass
class Transformer3DModelOutput(BaseOutput):
    """
    The output of [`Transformer2DModel`].

    Args:
        sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)` or `(batch size, num_vector_embeds - 1, num_latent_pixels)` if [`Transformer2DModel`] is discrete):
            The hidden states output conditioned on the `encoder_hidden_states` input. If discrete, returns probability
            distributions for the unnoised latent pixels.
    """

    sample: torch.FloatTensor


class Transformer3DModel(ModelMixin, ConfigMixin):
    """
    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.
    """

    @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,
        upcast_attention: bool = False,
        norm_type: str = "layer_norm",
        norm_elementwise_affine: bool = True,
        rotary_emb=None,
    ):
        super().__init__()
        self.use_linear_projection = use_linear_projection
        self.num_attention_heads = num_attention_heads
        self.attention_head_dim = attention_head_dim
        inner_dim = num_attention_heads * attention_head_dim

        # 1. Transformer2DModel can process both standard continuous images of shape `(batch_size, num_channels, width, height)` as well as quantized image embeddings of shape `(batch_size, num_image_vectors)`
        # Define whether input is continuous or discrete depending on configuration
        self.is_input_continuous = (in_channels is not None) and (patch_size is None)
        self.is_input_vectorized = num_vector_embeds is not None
        self.is_input_patches = in_channels is not None and patch_size is not 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"

        if self.is_input_continuous and self.is_input_vectorized:
            raise ValueError(
                f"Cannot define both `in_channels`: {in_channels} and `num_vector_embeds`: {num_vector_embeds}. Make"
                " sure that either `in_channels` or `num_vector_embeds` is None."
            )
        elif self.is_input_vectorized and self.is_input_patches:
            raise ValueError(
                f"Cannot define both `num_vector_embeds`: {num_vector_embeds} and `patch_size`: {patch_size}. Make"
                " sure that either `num_vector_embeds` or `num_patches` is None."
            )
        elif not self.is_input_continuous and not self.is_input_vectorized and not self.is_input_patches:
            raise ValueError(
                f"Has to define `in_channels`: {in_channels}, `num_vector_embeds`: {num_vector_embeds}, or patch_size:"
                f" {patch_size}. Make sure that `in_channels`, `num_vector_embeds` or `num_patches` is not None."
            )

        # 2. Define input layers
        if self.is_input_continuous:
            self.in_channels = in_channels

            self.norm = torch.nn.GroupNorm(num_groups=norm_num_groups, num_channels=in_channels, eps=1e-6, affine=True)
            if use_linear_projection:
                self.proj_in = LoRACompatibleLinear(in_channels, inner_dim)
            else:
                self.proj_in = LoRACompatibleConv(in_channels, inner_dim, kernel_size=1, stride=1, padding=0)
        elif self.is_input_vectorized:
            assert sample_size is not None, "Transformer2DModel over discrete input must provide sample_size"
            assert num_vector_embeds is not None, "Transformer2DModel over discrete input must provide num_embed"

            self.height = sample_size
            self.width = sample_size
            self.num_vector_embeds = num_vector_embeds
            self.num_latent_pixels = self.height * self.width

            self.latent_image_embedding = ImagePositionalEmbeddings(
                num_embed=num_vector_embeds, embed_dim=inner_dim, height=self.height, width=self.width
            )
        elif self.is_input_patches:
            assert sample_size is not None, "Transformer2DModel over patched input must provide sample_size"

            self.height = sample_size
            self.width = sample_size

            self.patch_size = patch_size
            self.pos_embed = PatchEmbed(
                height=sample_size,
                width=sample_size,
                patch_size=patch_size,
                in_channels=in_channels,
                embed_dim=inner_dim,
            )

        # 3. Define transformers blocks
        self.transformer_blocks = nn.ModuleList(
            [
                BasicTransformerBlock(
                    inner_dim,
                    num_attention_heads,
                    attention_head_dim,
                    dropout=dropout,
                    cross_attention_dim=cross_attention_dim,
                    activation_fn=activation_fn,
                    num_embeds_ada_norm=num_embeds_ada_norm,
                    attention_bias=attention_bias,
                    only_cross_attention=only_cross_attention,
                    upcast_attention=upcast_attention,
                    norm_type=norm_type,
                    norm_elementwise_affine=norm_elementwise_affine,
                    rotary_emb=rotary_emb,
                )
                for d in range(num_layers)
            ]
        )

        # 4. Define output layers
        self.out_channels = in_channels if out_channels is None else out_channels
        if self.is_input_continuous:
            # TODO: should use out_channels for continuous projections
            if use_linear_projection:
                self.proj_out = LoRACompatibleLinear(inner_dim, in_channels)
            else:
                self.proj_out = LoRACompatibleConv(inner_dim, in_channels, kernel_size=1, stride=1, padding=0)
        elif self.is_input_vectorized:
            self.norm_out = nn.LayerNorm(inner_dim)
            self.out = nn.Linear(inner_dim, self.num_vector_embeds - 1)
        elif self.is_input_patches:
            self.norm_out = nn.LayerNorm(inner_dim, elementwise_affine=False, eps=1e-6)
            self.proj_out_1 = nn.Linear(inner_dim, 2 * inner_dim)
            self.proj_out_2 = nn.Linear(inner_dim, patch_size * patch_size * self.out_channels)

    def forward(
        self,
        hidden_states: torch.Tensor,
        encoder_hidden_states: Optional[torch.Tensor] = None,
        timestep: Optional[torch.LongTensor] = 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,
        use_image_num=None,
    ):
        """
        The [`Transformer2DModel`] forward method.

        Args:
            hidden_states (`torch.LongTensor` of shape `(batch size, num latent pixels)` if discrete, `torch.FloatTensor` of shape `(batch size, channel, height, width)` if continuous):
                Input `hidden_states`.
            encoder_hidden_states ( `torch.FloatTensor` 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`.
            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.unet_2d_condition.UNet2DConditionOutput`] instead of a plain
                tuple.

        Returns:
            If `return_dict` is True, an [`~models.transformer_2d.Transformer2DModelOutput`] is returned, otherwise a
            `tuple` where the first element is the sample tensor.
        """
        # 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)

        # 1. Input
        if self.is_input_continuous: # True

            assert hidden_states.dim() == 5, f"Expected hidden_states to have ndim=5, but got ndim={hidden_states.dim()}."
            if self.training:
                video_length = hidden_states.shape[2] - use_image_num
                hidden_states = rearrange(hidden_states, "b c f h w -> (b f) c h w").contiguous()
                encoder_hidden_states_length = encoder_hidden_states.shape[1]
                encoder_hidden_states_video = encoder_hidden_states[:, :encoder_hidden_states_length - use_image_num, ...]
                encoder_hidden_states_video = repeat(encoder_hidden_states_video, 'b m n c -> b (m f) n c', f=video_length).contiguous()
                encoder_hidden_states_image = encoder_hidden_states[:, encoder_hidden_states_length - use_image_num:, ...]
                encoder_hidden_states = torch.cat([encoder_hidden_states_video, encoder_hidden_states_image], dim=1)
                encoder_hidden_states = rearrange(encoder_hidden_states, 'b m n c -> (b m) n c').contiguous()
            else:
                video_length = hidden_states.shape[2]
                hidden_states = rearrange(hidden_states, "b c f h w -> (b f) c h w").contiguous()
                encoder_hidden_states = repeat(encoder_hidden_states, 'b n c -> (b f) n c', f=video_length).contiguous()

            batch, _, 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, height * width, inner_dim)
            else:
                inner_dim = hidden_states.shape[1]
                hidden_states = hidden_states.permute(0, 2, 3, 1).reshape(batch, height * width, inner_dim)
                hidden_states = self.proj_in(hidden_states)
        elif self.is_input_vectorized:
            hidden_states = self.latent_image_embedding(hidden_states)
        elif self.is_input_patches:
            hidden_states = self.pos_embed(hidden_states)

        # 2. Blocks
        for block in self.transformer_blocks:
            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,
                video_length=video_length,
                use_image_num=use_image_num,
            )

        # 3. Output
        if self.is_input_continuous:
            if not self.use_linear_projection:
                hidden_states = hidden_states.reshape(batch, 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, height, width, inner_dim).permute(0, 3, 1, 2).contiguous()

            output = hidden_states + residual
        elif self.is_input_vectorized:
            hidden_states = self.norm_out(hidden_states)
            logits = self.out(hidden_states)
            # (batch, self.num_vector_embeds - 1, self.num_latent_pixels)
            logits = logits.permute(0, 2, 1)

            # log(p(x_0))
            output = F.log_softmax(logits.double(), dim=1).float()
        elif self.is_input_patches:
            # TODO: cleanup!
            conditioning = self.transformer_blocks[0].norm1.emb(
                timestep, class_labels, hidden_dtype=hidden_states.dtype
            )
            shift, scale = self.proj_out_1(F.silu(conditioning)).chunk(2, dim=1)
            hidden_states = self.norm_out(hidden_states) * (1 + scale[:, None]) + shift[:, None]
            hidden_states = self.proj_out_2(hidden_states)

            # unpatchify
            height = width = int(hidden_states.shape[1] ** 0.5)
            hidden_states = hidden_states.reshape(
                shape=(-1, height, width, self.patch_size, self.patch_size, self.out_channels)
            )
            hidden_states = torch.einsum("nhwpqc->nchpwq", hidden_states)
            output = hidden_states.reshape(
                shape=(-1, self.out_channels, height * self.patch_size, width * self.patch_size)
            )

        output = rearrange(output, "(b f) c h w -> b c f h w", f=video_length + use_image_num).contiguous()
        if not return_dict:
            return (output,)

        return Transformer3DModelOutput(sample=output)