models/controlnet.py

# Copyright 2024 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 dataclasses import dataclass
import pdb
from typing import Any, Dict, List, Optional, Tuple, Union

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

from diffusers.configuration_utils import ConfigMixin, register_to_config
from diffusers.loaders.single_file_model import FromOriginalModelMixin
from diffusers.utils import BaseOutput, logging
from diffusers.models.attention_processor import (
    ADDED_KV_ATTENTION_PROCESSORS,
    CROSS_ATTENTION_PROCESSORS,
    AttentionProcessor,
    AttnAddedKVProcessor,
    AttnProcessor,
)
from diffusers.models.embeddings import TextImageProjection, TextImageTimeEmbedding, TextTimeEmbedding, TimestepEmbedding, Timesteps
from diffusers.models.modeling_utils import ModelMixin
from diffusers.models.unets.unet_2d_blocks import (
    CrossAttnDownBlock2D,
    DownBlock2D,
    UNetMidBlock2D,
    UNetMidBlock2DCrossAttn,
    get_down_block,
)
from models.unet_2d_condition import UNet2DConditionModel


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


@dataclass
class ControlNetOutput(BaseOutput):
    """
    The output of [`ControlNetModel`].

    Args:
        down_block_res_samples (`tuple[torch.Tensor]`):
            A tuple of downsample activations at different resolutions for each downsampling block. Each tensor should
            be of shape `(batch_size, channel * resolution, height //resolution, width // resolution)`. Output can be
            used to condition the original UNet's downsampling activations.
        mid_down_block_re_sample (`torch.Tensor`):
            The activation of the middle block (the lowest sample resolution). Each tensor should be of shape
            `(batch_size, channel * lowest_resolution, height // lowest_resolution, width // lowest_resolution)`.
            Output can be used to condition the original UNet's middle block activation.
    """

    down_block_res_samples: Tuple[torch.Tensor]
    mid_block_res_sample: torch.Tensor


class ControlNetConditioningEmbedding(nn.Module):
    """
    Quoting from https://arxiv.org/abs/2302.05543: "Stable Diffusion uses a pre-processing method similar to VQ-GAN
    [11] to convert the entire dataset of 512 × 512 images into smaller 64 × 64 “latent images” for stabilized
    training. This requires ControlNets to convert image-based conditions to 64 × 64 feature space to match the
    convolution size. We use a tiny network E(·) of four convolution layers with 4 × 4 kernels and 2 × 2 strides
    (activated by ReLU, channels are 16, 32, 64, 128, initialized with Gaussian weights, trained jointly with the full
    model) to encode image-space conditions ... into feature maps ..."
    """

    def __init__(
        self,
        conditioning_embedding_channels: int,
        conditioning_channels: int = 3,
        block_out_channels: Tuple[int, ...] = (16, 32, 96, 256),
    ):
        super().__init__()

        self.conv_in = nn.Conv2d(conditioning_channels, block_out_channels[0], kernel_size=3, padding=1)

        self.blocks = nn.ModuleList([])

        for i in range(len(block_out_channels) - 1):
            channel_in = block_out_channels[i]
            channel_out = block_out_channels[i + 1]
            self.blocks.append(nn.Conv2d(channel_in, channel_in, kernel_size=3, padding=1))
            self.blocks.append(nn.Conv2d(channel_in, channel_out, kernel_size=3, padding=1, stride=1))

        self.conv_out = zero_module(
            nn.Conv2d(block_out_channels[-1], conditioning_embedding_channels, kernel_size=3, padding=1)
        )

    def forward(self, conditioning):
        embedding = self.conv_in(conditioning)
        embedding = F.silu(embedding)

        for block in self.blocks:
            embedding = block(embedding)
            embedding = F.silu(embedding)

        embedding = self.conv_out(embedding)

        return embedding


class ControlNetModel(ModelMixin, ConfigMixin, FromOriginalModelMixin):
    """
    A ControlNet model.

    Args:
        in_channels (`int`, defaults to 4):
            The number of channels in the input sample.
        flip_sin_to_cos (`bool`, defaults to `True`):
            Whether to flip the sin to cos in the time embedding.
        freq_shift (`int`, defaults to 0):
            The frequency shift to apply to the time embedding.
        down_block_types (`tuple[str]`, defaults to `("CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "DownBlock2D")`):
            The tuple of downsample blocks to use.
        block_out_channels (`tuple[int]`, defaults to `(320, 640, 1280, 1280)`):
            The tuple of output channels for each block.
        layers_per_block (`int`, defaults to 2):
            The number of layers per block.
        downsample_padding (`int`, defaults to 1):
            The padding to use for the downsampling convolution.
        mid_block_scale_factor (`float`, defaults to 1):
            The scale factor to use for the mid block.
        act_fn (`str`, defaults to "silu"):
            The activation function to use.
        norm_num_groups (`int`, *optional*, defaults to 32):
            The number of groups to use for the normalization. If None, normalization and activation layers is skipped
            in post-processing.
        norm_eps (`float`, defaults to 1e-5):
            The epsilon to use for the normalization.
        transformer_layers_per_block (`int` or `Tuple[int]`, *optional*, defaults to 1):
            The number of transformer blocks of type [`~models.attention.BasicTransformerBlock`]. Only relevant for
            [`~models.unet_2d_blocks.CrossAttnDownBlock2D`], [`~models.unet_2d_blocks.CrossAttnUpBlock2D`],
            [`~models.unet_2d_blocks.UNetMidBlock2DCrossAttn`].
        attention_head_dim (`Union[int, Tuple[int]]`, defaults to 8):
            The dimension of the attention heads.
        resnet_time_scale_shift (`str`, defaults to `"default"`):
            Time scale shift config for ResNet blocks (see `ResnetBlock2D`). Choose from `default` or `scale_shift`.
        conditioning_embedding_out_channels (`tuple[int]`, *optional*, defaults to `(16, 32, 96, 256)`):
            The tuple of output channel for each block in the `conditioning_embedding` layer.
    """

    _supports_gradient_checkpointing = True

    @register_to_config
    def __init__(
        self,
        in_channels: int = 4,
        conditioning_channels: int = 3,
        flip_sin_to_cos: bool = True,
        freq_shift: int = 0,
        down_block_types: Tuple[str, ...] = (
            "CrossAttnDownBlock2D",
            "CrossAttnDownBlock2D",
            "CrossAttnDownBlock2D",
            "DownBlock2D",
        ),
        mid_block_type: Optional[str] = "UNetMidBlock2D",
        block_out_channels: Tuple[int, ...] = (320, 640, 1280, 1280),
        layers_per_block: int = 2,
        downsample_padding: int = 1,
        mid_block_scale_factor: float = 1,
        act_fn: str = "silu",
        norm_num_groups: Optional[int] = 32,
        norm_eps: float = 1e-5,
        transformer_layers_per_block: Union[int, Tuple[int, ...]] = 1,
        attention_head_dim: Union[int] = 8,
        resnet_time_scale_shift: str = "default",
        add_attention: bool = True,
        conditioning_embedding_out_channels: Optional[Tuple[int, ...]] = (16, 32, 96, 256),
        global_pool_conditions: bool = False,
        use_prompt: bool = False,
        encoder_size: Optional[int] = None,
    ):
        super().__init__()

        # If `num_attention_heads` is not defined (which is the case for most models)
        # it will default to `attention_head_dim`. This looks weird upon first reading it and it is.
        # The reason for this behavior is to correct for incorrectly named variables that were introduced
        # when this library was created. The incorrect naming was only discovered much later in https://github.com/huggingface/diffusers/issues/2011#issuecomment-1547958131
        # Changing `attention_head_dim` to `num_attention_heads` for 40,000+ configurations is too backwards breaking
        # which is why we correct for the naming here.

        # Check inputs
        if len(block_out_channels) != len(down_block_types):
            raise ValueError(
                f"Must provide the same number of `block_out_channels` as `down_block_types`. `block_out_channels`: {block_out_channels}. `down_block_types`: {down_block_types}."
            )

        if isinstance(transformer_layers_per_block, int):
            transformer_layers_per_block = [transformer_layers_per_block] * len(down_block_types)

        # input
        conv_in_kernel = 3
        conv_in_padding = (conv_in_kernel - 1) // 2
        self.conv_in = nn.Conv2d(
            in_channels, block_out_channels[0], kernel_size=conv_in_kernel, padding=conv_in_padding
        )

        # time
        time_embed_dim = block_out_channels[0] * 4
        self.time_proj = Timesteps(block_out_channels[0], flip_sin_to_cos, freq_shift)
        timestep_input_dim = block_out_channels[0]
        self.time_embedding = TimestepEmbedding(
            timestep_input_dim,
            time_embed_dim,
            act_fn=act_fn,
        )

        # prompt
        if use_prompt:
            self.prompt_embedding = nn.Sequential(
                nn.Linear(encoder_size, time_embed_dim),
                nn.SiLU(), 
                nn.Linear(time_embed_dim, time_embed_dim)
            )
        else:
            self.prompt_embedding = None
        
        # control net conditioning embedding
        self.controlnet_cond_embedding = ControlNetConditioningEmbedding(
            conditioning_embedding_channels=block_out_channels[0],
            block_out_channels=conditioning_embedding_out_channels,
            conditioning_channels=conditioning_channels,
        )

        self.down_blocks = nn.ModuleList([])
        self.controlnet_down_blocks = nn.ModuleList([])

        # down
        output_channel = block_out_channels[0]

        controlnet_block = nn.Conv2d(output_channel, output_channel, kernel_size=1)
        controlnet_block = zero_module(controlnet_block)
        self.controlnet_down_blocks.append(controlnet_block)

        for i, down_block_type in enumerate(down_block_types):
            input_channel = output_channel
            output_channel = block_out_channels[i]
            is_final_block = i == len(block_out_channels) - 1

            down_block = get_down_block(
                down_block_type,
                num_layers=layers_per_block,
                in_channels=input_channel,
                out_channels=output_channel,
                temb_channels=time_embed_dim,
                add_downsample=not is_final_block,
                resnet_eps=norm_eps,
                resnet_act_fn=act_fn,
                resnet_groups=norm_num_groups,
                attention_head_dim=attention_head_dim if attention_head_dim is not None else output_channel,
                downsample_padding=downsample_padding,
                resnet_time_scale_shift=resnet_time_scale_shift,
            )
            self.down_blocks.append(down_block)

            for _ in range(layers_per_block):
                controlnet_block = nn.Conv2d(output_channel, output_channel, kernel_size=1)
                controlnet_block = zero_module(controlnet_block)
                self.controlnet_down_blocks.append(controlnet_block)

            if not is_final_block:
                controlnet_block = nn.Conv2d(output_channel, output_channel, kernel_size=1)
                controlnet_block = zero_module(controlnet_block)
                self.controlnet_down_blocks.append(controlnet_block)

        # mid
        mid_block_channel = block_out_channels[-1]

        controlnet_block = nn.Conv2d(mid_block_channel, mid_block_channel, kernel_size=1)
        controlnet_block = zero_module(controlnet_block)
        self.controlnet_mid_block = controlnet_block

        if mid_block_type == "UNetMidBlock2D":
            self.mid_block = UNetMidBlock2D(
                in_channels=block_out_channels[-1],
                temb_channels=time_embed_dim,
                resnet_eps=norm_eps,
                resnet_act_fn=act_fn,
                output_scale_factor=mid_block_scale_factor,
                resnet_groups=norm_num_groups,
                resnet_time_scale_shift=resnet_time_scale_shift,
                add_attention=add_attention,
            )
        else:
            raise ValueError(f"unknown mid_block_type : {mid_block_type}")

    @classmethod
    def from_unet(
        cls,
        unet: UNet2DConditionModel,
        controlnet_conditioning_channel_order: str = "rgb",
        conditioning_embedding_out_channels: Optional[Tuple[int, ...]] = (16, 32, 96, 256),
        load_weights_from_unet: bool = True,
        conditioning_channels: int = 3,
    ):
        r"""
        Instantiate a [`ControlNetModel`] from [`UNet2DConditionModel`].

        Parameters:
            unet (`UNet2DConditionModel`):
                The UNet model weights to copy to the [`ControlNetModel`]. All configuration options are also copied
                where applicable.
        """
        mid_block_type = unet.config.mid_block_type if "mid_block_type" in unet.config else "UNetMidBlock2D"
        transformer_layers_per_block = (
            unet.config.transformer_layers_per_block if "transformer_layers_per_block" in unet.config else 1
        )
        addition_embed_type = unet.config.addition_embed_type if "addition_embed_type" in unet.config else None
        addition_time_embed_dim = (
            unet.config.addition_time_embed_dim if "addition_time_embed_dim" in unet.config else None
        )

        controlnet = cls(
            in_channels=unet.config.in_channels,
            conditioning_channels=conditioning_channels,
            flip_sin_to_cos=unet.config.flip_sin_to_cos,
            freq_shift=unet.config.freq_shift,
            down_block_types=unet.config.down_block_types,
            mid_block_type=mid_block_type,
            block_out_channels=unet.config.block_out_channels,
            layers_per_block=unet.config.layers_per_block,
            downsample_padding=unet.config.downsample_padding,
            mid_block_scale_factor=unet.config.mid_block_scale_factor,
            act_fn=unet.config.act_fn,
            norm_num_groups=unet.config.norm_num_groups,
            norm_eps=unet.config.norm_eps,
            transformer_layers_per_block=transformer_layers_per_block,
            attention_head_dim=unet.config.attention_head_dim,
            resnet_time_scale_shift=unet.config.resnet_time_scale_shift,
            use_prompt=unet.config.use_prompt,
            encoder_size=unet.config.encoder_size,
            conditioning_embedding_out_channels=conditioning_embedding_out_channels,
        )

        if load_weights_from_unet:
            controlnet.conv_in.load_state_dict(unet.conv_in.state_dict())
            controlnet.time_proj.load_state_dict(unet.time_proj.state_dict())
            controlnet.time_embedding.load_state_dict(unet.time_embedding.state_dict())
            controlnet.prompt_embedding.load_state_dict(unet.prompt_embedding.state_dict())

            if hasattr(controlnet, "add_embedding"):
                controlnet.add_embedding.load_state_dict(unet.add_embedding.state_dict())

            controlnet.down_blocks.load_state_dict(unet.down_blocks.state_dict())
            controlnet.mid_block.load_state_dict(unet.mid_block.state_dict())

        return controlnet
    

    def forward(
        self,
        sample: torch.Tensor,
        timestep: Union[torch.Tensor, float, int],
        controlnet_cond: torch.Tensor,
        encoder_hidden_states: Optional[torch.Tensor] = None,
        conditioning_scale: float = 1.0,
        timestep_cond: Optional[torch.Tensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
        guess_mode: bool = False,
        return_dict: bool = True,
    ) -> Union[ControlNetOutput, Tuple[Tuple[torch.Tensor, ...], torch.Tensor]]:
        """
        The [`ControlNetModel`] forward method.

        Args:
            sample (`torch.Tensor`):
                The noisy input tensor.
            timestep (`Union[torch.Tensor, float, int]`):
                The number of timesteps to denoise an input.
            encoder_hidden_states (`torch.Tensor`):
                The encoder hidden states.
            controlnet_cond (`torch.Tensor`):
                The conditional input tensor of shape `(batch_size, sequence_length, hidden_size)`.
            conditioning_scale (`float`, defaults to `1.0`):
                The scale factor for ControlNet outputs.
            timestep_cond (`torch.Tensor`, *optional*, defaults to `None`):
                Additional conditional embeddings for timestep. If provided, the embeddings will be summed with the
                timestep_embedding passed through the `self.time_embedding` layer to obtain the final timestep
                embeddings.
            attention_mask (`torch.Tensor`, *optional*, defaults to `None`):
                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.
            guess_mode (`bool`, defaults to `False`):
                In this mode, the ControlNet encoder tries its best to recognize the input content of the input even if
                you remove all prompts. A `guidance_scale` between 3.0 and 5.0 is recommended.
            return_dict (`bool`, defaults to `True`):
                Whether or not to return a [`~models.controlnet.ControlNetOutput`] instead of a plain tuple.

        Returns:
            [`~models.controlnet.ControlNetOutput`] **or** `tuple`:
                If `return_dict` is `True`, a [`~models.controlnet.ControlNetOutput`] is returned, otherwise a tuple is
                returned where the first element is the sample tensor.
        """
        # prepare attention_mask
        if attention_mask is not None:
            attention_mask = (1 - attention_mask.to(sample.dtype)) * -10000.0
            attention_mask = attention_mask.unsqueeze(1)
        # 1. time
        timesteps = timestep
        if not torch.is_tensor(timesteps):
            # TODO: this requires sync between CPU and GPU. So try to pass timesteps as tensors if you can
            # This would be a good case for the `match` statement (Python 3.10+)
            is_mps = sample.device.type == "mps"
            if isinstance(timestep, float):
                dtype = torch.float32 if is_mps else torch.float64
            else:
                dtype = torch.int32 if is_mps else torch.int64
            timesteps = torch.tensor([timesteps], dtype=dtype, device=sample.device)
        elif len(timesteps.shape) == 0:
            timesteps = timesteps[None].to(sample.device)

        # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
        timesteps = timesteps.expand(sample.shape[0])

        t_emb = self.time_proj(timesteps)

        # timesteps does not contain any weights and will always return f32 tensors
        # but time_embedding might actually be running in fp16. so we need to cast here.
        # there might be better ways to encapsulate this.
        t_emb = t_emb.to(dtype=sample.dtype)
        emb = self.time_embedding(t_emb, timestep_cond)

        if self.prompt_embedding is not None:
            encoder_hidden_states = encoder_hidden_states.reshape(sample.shape[0], -1).contiguous()
            prompt_emb = self.prompt_embedding(encoder_hidden_states)
            emb = emb + prompt_emb
        
        # 2. pre-process
        sample = self.conv_in(sample)
        
        controlnet_cond = self.controlnet_cond_embedding(controlnet_cond)
        sample = sample + controlnet_cond

        # 3. down
        down_block_res_samples = (sample,)
        for downsample_block in self.down_blocks:
            sample, res_samples = downsample_block(hidden_states=sample, temb=emb)
            down_block_res_samples += res_samples

        # 4. mid
        if self.mid_block is not None:
            sample = self.mid_block(sample, emb)

        # 5. Control net blocks
        controlnet_down_block_res_samples = ()

        for down_block_res_sample, controlnet_block in zip(down_block_res_samples, self.controlnet_down_blocks):
            down_block_res_sample = controlnet_block(down_block_res_sample)
            controlnet_down_block_res_samples = controlnet_down_block_res_samples + (down_block_res_sample,)

        down_block_res_samples = controlnet_down_block_res_samples

        mid_block_res_sample = self.controlnet_mid_block(sample)

        # 6. scaling
        if guess_mode and not self.config.global_pool_conditions:
            scales = torch.logspace(-1, 0, len(down_block_res_samples) + 1, device=sample.device)  # 0.1 to 1.0
            scales = scales * conditioning_scale
            down_block_res_samples = [sample * scale for sample, scale in zip(down_block_res_samples, scales)]
            mid_block_res_sample = mid_block_res_sample * scales[-1]  # last one
        else:
            down_block_res_samples = [sample * conditioning_scale for sample in down_block_res_samples]
            mid_block_res_sample = mid_block_res_sample * conditioning_scale

        if self.config.global_pool_conditions:
            down_block_res_samples = [
                torch.mean(sample, dim=(2, 3), keepdim=True) for sample in down_block_res_samples
            ]
            mid_block_res_sample = torch.mean(mid_block_res_sample, dim=(2, 3), keepdim=True)

        if not return_dict:
            return (down_block_res_samples, mid_block_res_sample)

        return ControlNetOutput(
            down_block_res_samples=down_block_res_samples, mid_block_res_sample=mid_block_res_sample
        )


def zero_module(module):
    for p in module.parameters():
        nn.init.zeros_(p)
    return module