MuseV/musev/models/transformer_2d.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 __future__ import annotations
from dataclasses import dataclass
from typing import Any, Dict, List, Literal, Optional
import logging

from einops import rearrange

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

from diffusers.models.transformer_2d import (
    Transformer2DModelOutput,
    Transformer2DModel as DiffusersTransformer2DModel,
)

from diffusers.configuration_utils import ConfigMixin, register_to_config
from diffusers.models.embeddings import ImagePositionalEmbeddings
from diffusers.utils import BaseOutput, deprecate
from diffusers.models.attention import (
    BasicTransformerBlock as DiffusersBasicTransformerBlock,
)
from diffusers.models.embeddings import PatchEmbed
from diffusers.models.lora import LoRACompatibleConv, LoRACompatibleLinear
from diffusers.models.modeling_utils import ModelMixin
from diffusers.utils.constants import USE_PEFT_BACKEND

from .attention import BasicTransformerBlock

logger = logging.getLogger(__name__)

# 本部分 与 diffusers/models/transformer_2d.py 几乎一样
# 更新部分
# 1. 替换自定义 BasicTransformerBlock 类
# 2. 在forward 里增加了 self_attn_block_embs 用于 提取 self_attn 中的emb

# this module is same as diffusers/models/transformer_2d.py. The update part is
# 1 redefine BasicTransformerBlock
# 2. add self_attn_block_embs in forward to extract emb from self_attn


class Transformer2DModel(DiffusersTransformer2DModel):
    """
    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: int | None = None,
        out_channels: int | None = None,
        num_layers: int = 1,
        dropout: float = 0,
        norm_num_groups: int = 32,
        cross_attention_dim: int | None = None,
        attention_bias: bool = False,
        sample_size: int | None = None,
        num_vector_embeds: int | None = None,
        patch_size: int | None = None,
        activation_fn: str = "geglu",
        num_embeds_ada_norm: int | None = 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",
        norm_elementwise_affine: bool = True,
        attention_type: str = "default",
        cross_attn_temporal_cond: bool = False,
        ip_adapter_cross_attn: bool = False,
        need_t2i_facein: bool = False,
        need_t2i_ip_adapter_face: bool = False,
        image_scale: float = 1.0,
    ):
        super().__init__(
            num_attention_heads,
            attention_head_dim,
            in_channels,
            out_channels,
            num_layers,
            dropout,
            norm_num_groups,
            cross_attention_dim,
            attention_bias,
            sample_size,
            num_vector_embeds,
            patch_size,
            activation_fn,
            num_embeds_ada_norm,
            use_linear_projection,
            only_cross_attention,
            double_self_attention,
            upcast_attention,
            norm_type,
            norm_elementwise_affine,
            attention_type,
        )
        inner_dim = num_attention_heads * attention_head_dim
        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,
                    double_self_attention=double_self_attention,
                    upcast_attention=upcast_attention,
                    norm_type=norm_type,
                    norm_elementwise_affine=norm_elementwise_affine,
                    attention_type=attention_type,
                    cross_attn_temporal_cond=cross_attn_temporal_cond,
                    ip_adapter_cross_attn=ip_adapter_cross_attn,
                    need_t2i_facein=need_t2i_facein,
                    need_t2i_ip_adapter_face=need_t2i_ip_adapter_face,
                    image_scale=image_scale,
                )
                for d in range(num_layers)
            ]
        )
        self.num_layers = num_layers
        self.cross_attn_temporal_cond = cross_attn_temporal_cond
        self.ip_adapter_cross_attn = ip_adapter_cross_attn

        self.need_t2i_facein = need_t2i_facein
        self.need_t2i_ip_adapter_face = need_t2i_ip_adapter_face
        self.image_scale = image_scale
        self.print_idx = 0

    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,
        self_attn_block_embs: Optional[List[torch.Tensor]] = None,
        self_attn_block_embs_mode: Literal["read", "write"] = "write",
        return_dict: bool = True,
    ):
        """
        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`.
            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.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)

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

        # 1. Input
        if self.is_input_continuous:
            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, scale=lora_scale)
                    if not USE_PEFT_BACKEND
                    else 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, scale=lora_scale)
                    if not USE_PEFT_BACKEND
                    else self.proj_in(hidden_states)
                )

        elif self.is_input_vectorized:
            hidden_states = self.latent_image_embedding(hidden_states)
        elif self.is_input_patches:
            height, width = (
                hidden_states.shape[-2] // self.patch_size,
                hidden_states.shape[-1] // self.patch_size,
            )
            hidden_states = self.pos_embed(hidden_states)

            if self.adaln_single is not None:
                if self.use_additional_conditions and added_cond_kwargs is None:
                    raise ValueError(
                        "`added_cond_kwargs` cannot be None when using additional conditions for `adaln_single`."
                    )
                batch_size = hidden_states.shape[0]
                timestep, embedded_timestep = self.adaln_single(
                    timestep,
                    added_cond_kwargs,
                    batch_size=batch_size,
                    hidden_dtype=hidden_states.dtype,
                )

        # 2. Blocks
        if self.caption_projection is not None:
            batch_size = hidden_states.shape[0]
            encoder_hidden_states = self.caption_projection(encoder_hidden_states)
            encoder_hidden_states = encoder_hidden_states.view(
                batch_size, -1, hidden_states.shape[-1]
            )

        for block in self.transformer_blocks:
            if self.training and self.gradient_checkpointing:
                hidden_states = torch.utils.checkpoint.checkpoint(
                    block,
                    hidden_states,
                    attention_mask,
                    encoder_hidden_states,
                    encoder_attention_mask,
                    timestep,
                    cross_attention_kwargs,
                    class_labels,
                    self_attn_block_embs,
                    self_attn_block_embs_mode,
                    use_reentrant=False,
                )
            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,
                    self_attn_block_embs=self_attn_block_embs,
                    self_attn_block_embs_mode=self_attn_block_embs_mode,
                )
            # 将 转换 self_attn_emb的尺寸
            if (
                self_attn_block_embs is not None
                and self_attn_block_embs_mode.lower() == "write"
            ):
                self_attn_idx = block.spatial_self_attn_idx
                if self.print_idx == 0:
                    logger.debug(
                        f"self_attn_block_embs, num={len(self_attn_block_embs)}, before, shape={self_attn_block_embs[self_attn_idx].shape}, height={height}, width={width}"
                    )
                self_attn_block_embs[self_attn_idx] = rearrange(
                    self_attn_block_embs[self_attn_idx],
                    "bt (h w) c->bt c h w",
                    h=height,
                    w=width,
                )
                if self.print_idx == 0:
                    logger.debug(
                        f"self_attn_block_embs, num={len(self_attn_block_embs)},  after ,shape={self_attn_block_embs[self_attn_idx].shape}, height={height}, width={width}"
                    )

        if self.proj_out is None:
            return hidden_states

        # 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, scale=lora_scale)
                    if not USE_PEFT_BACKEND
                    else self.proj_out(hidden_states)
                )
            else:
                hidden_states = (
                    self.proj_out(hidden_states, scale=lora_scale)
                    if not USE_PEFT_BACKEND
                    else 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()

        if self.is_input_patches:
            if self.config.norm_type != "ada_norm_single":
                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)
            elif self.config.norm_type == "ada_norm_single":
                shift, scale = (
                    self.scale_shift_table[None] + embedded_timestep[:, None]
                ).chunk(2, dim=1)
                hidden_states = self.norm_out(hidden_states)
                # Modulation
                hidden_states = hidden_states * (1 + scale) + shift
                hidden_states = self.proj_out(hidden_states)
                hidden_states = hidden_states.squeeze(1)

            # unpatchify
            if self.adaln_single is None:
                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,
                )
            )
        self.print_idx += 1
        if not return_dict:
            return (output,)

        return Transformer2DModelOutput(sample=output)