Create conditional_unet_model.py

unet/conditional_unet_model.py

@@ -0,0 +1,862 @@
+import random
+from typing import List, Optional, Tuple, Union
+
+import torch
+from dataclasses import dataclass
+from typing import Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+
+from diffusers.configuration_utils import ConfigMixin, register_to_config
+from diffusers.utils import BaseOutput
+from diffusers.models.embeddings import GaussianFourierProjection, TimestepEmbedding, Timesteps
+from diffusers.models.modeling_utils import ModelMixin
+from diffusers.models.unets.unet_2d_blocks import UNetMidBlock2D, get_down_block, get_up_block
+from torch import tensor
+
+
+@dataclass
+class UNet2DOutput(BaseOutput):
+    """
+    The output of [`UNet2DModel`].
+
+    Args:
+        sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`):
+            The hidden states output from the last layer of the model.
+    """
+
+    sample: torch.FloatTensor
+
+
+class UNet2DModel(ModelMixin, ConfigMixin):
+    r"""
+    A 2D UNet model that takes a noisy sample and a timestep and returns a sample shaped output.
+
+    This model inherits from [`ModelMixin`]. Check the superclass documentation for it's generic methods implemented
+    for all models (such as downloading or saving).
+
+    Parameters:
+        sample_size (`int` or `Tuple[int, int]`, *optional*, defaults to `None`):
+            Height and width of input/output sample. Dimensions must be a multiple of `2 ** (len(block_out_channels) -
+            1)`.
+        in_channels (`int`, *optional*, defaults to 3): Number of channels in the input sample.
+        out_channels (`int`, *optional*, defaults to 3): Number of channels in the output.
+        center_input_sample (`bool`, *optional*, defaults to `False`): Whether to center the input sample.
+        time_embedding_type (`str`, *optional*, defaults to `"positional"`): Type of time embedding to use.
+        freq_shift (`int`, *optional*, defaults to 0): Frequency shift for Fourier time embedding.
+        flip_sin_to_cos (`bool`, *optional*, defaults to `True`):
+            Whether to flip sin to cos for Fourier time embedding.
+        down_block_types (`Tuple[str]`, *optional*, defaults to `("DownBlock2D", "AttnDownBlock2D", "AttnDownBlock2D", "AttnDownBlock2D")`):
+            Tuple of downsample block types.
+        mid_block_type (`str`, *optional*, defaults to `"UNetMidBlock2D"`):
+            Block type for middle of UNet, it can be either `UNetMidBlock2D` or `UnCLIPUNetMidBlock2D`.
+        up_block_types (`Tuple[str]`, *optional*, defaults to `("AttnUpBlock2D", "AttnUpBlock2D", "AttnUpBlock2D", "UpBlock2D")`):
+            Tuple of upsample block types.
+        block_out_channels (`Tuple[int]`, *optional*, defaults to `(224, 448, 672, 896)`):
+            Tuple of block output channels.
+        layers_per_block (`int`, *optional*, defaults to `2`): The number of layers per block.
+        mid_block_scale_factor (`float`, *optional*, defaults to `1`): The scale factor for the mid block.
+        downsample_padding (`int`, *optional*, defaults to `1`): The padding for the downsample convolution.
+        downsample_type (`str`, *optional*, defaults to `conv`):
+            The downsample type for downsampling layers. Choose between "conv" and "resnet"
+        upsample_type (`str`, *optional*, defaults to `conv`):
+            The upsample type for upsampling layers. Choose between "conv" and "resnet"
+        dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
+        act_fn (`str`, *optional*, defaults to `"silu"`): The activation function to use.
+        attention_head_dim (`int`, *optional*, defaults to `8`): The attention head dimension.
+        norm_num_groups (`int`, *optional*, defaults to `32`): The number of groups for normalization.
+        attn_norm_num_groups (`int`, *optional*, defaults to `None`):
+            If set to an integer, a group norm layer will be created in the mid block's [`Attention`] layer with the
+            given number of groups. If left as `None`, the group norm layer will only be created if
+            `resnet_time_scale_shift` is set to `default`, and if created will have `norm_num_groups` groups.
+        norm_eps (`float`, *optional*, defaults to `1e-5`): The epsilon for normalization.
+        resnet_time_scale_shift (`str`, *optional*, defaults to `"default"`): Time scale shift config
+            for ResNet blocks (see [`~models.resnet.ResnetBlock2D`]). Choose from `default` or `scale_shift`.
+        class_embed_type (`str`, *optional*, defaults to `None`):
+            The type of class embedding to use which is ultimately summed with the time embeddings. Choose from `None`,
+            `"timestep"`, or `"identity"`.
+        num_class_embeds (`int`, *optional*, defaults to `None`):
+            Input dimension of the learnable embedding matrix to be projected to `time_embed_dim` when performing class
+            conditioning with `class_embed_type` equal to `None`.
+    """
+
+    @register_to_config
+    def __init__(
+        self,
+        sample_size: Optional[Union[int, Tuple[int, int]]] = None,
+        in_channels: int = 3,
+        out_channels: int = 3,
+        center_input_sample: bool = False,
+        time_embedding_type: str = "positional",
+        freq_shift: int = 0,
+        flip_sin_to_cos: bool = True,
+        down_block_types: Tuple[str, ...] = ("DownBlock2D", "AttnDownBlock2D", "AttnDownBlock2D", "AttnDownBlock2D"),
+        up_block_types: Tuple[str, ...] = ("AttnUpBlock2D", "AttnUpBlock2D", "AttnUpBlock2D", "UpBlock2D"),
+        block_out_channels: Tuple[int, ...] = (224, 448, 672, 896),
+        layers_per_block: int = 2,
+        mid_block_scale_factor: float = 1,
+        downsample_padding: int = 1,
+        downsample_type: str = "conv",
+        upsample_type: str = "conv",
+        dropout: float = 0.0,
+        act_fn: str = "silu",
+        attention_head_dim: Optional[int] = 8,
+        norm_num_groups: int = 32,
+        attn_norm_num_groups: Optional[int] = None,
+        norm_eps: float = 1e-5,
+        resnet_time_scale_shift: str = "default",
+        add_attention: bool = True,
+        class_embed_type: Optional[str] = None,
+        num_class_embeds: Optional[int] = None,
+        num_train_timesteps: Optional[int] = None,
+        set_W_to_weight: Optional[bool] = True
+    ):
+        super().__init__()
+
+        self.sample_size = sample_size
+        time_embed_dim = block_out_channels[0] * 4
+
+        # Check inputs
+        if len(down_block_types) != len(up_block_types):
+            raise ValueError(
+                f"Must provide the same number of `down_block_types` as `up_block_types`. `down_block_types`: {down_block_types}. `up_block_types`: {up_block_types}."
+            )
+
+        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}."
+            )
+
+        # input
+        self.conv_in = nn.Conv2d(in_channels, block_out_channels[0], kernel_size=3, padding=(1, 1))
+
+        # time
+        if time_embedding_type == "fourier":
+            self.time_proj = GaussianFourierProjection(embedding_size=block_out_channels[0], scale=16, set_W_to_weight=set_W_to_weight)
+            timestep_input_dim = 2 * block_out_channels[0]
+        elif time_embedding_type == "positional":
+            self.time_proj = Timesteps(block_out_channels[0], flip_sin_to_cos, freq_shift)
+            timestep_input_dim = block_out_channels[0]
+        elif time_embedding_type == "learned":
+            self.time_proj = nn.Embedding(num_train_timesteps, block_out_channels[0])
+            timestep_input_dim = block_out_channels[0]
+
+        self.time_embedding = TimestepEmbedding(timestep_input_dim, time_embed_dim)
+
+        # class embedding
+        if class_embed_type is None and num_class_embeds is not None:
+            self.class_embedding = nn.Embedding(num_class_embeds, time_embed_dim)
+        elif class_embed_type == "timestep":
+            self.class_embedding = TimestepEmbedding(timestep_input_dim, time_embed_dim)
+        elif class_embed_type == "identity":
+            self.class_embedding = nn.Identity(time_embed_dim, time_embed_dim)
+        else:
+            self.class_embedding = None
+
+        self.down_blocks = nn.ModuleList([])
+        self.mid_block = None
+        self.up_blocks = nn.ModuleList([])
+
+        # down
+        output_channel = block_out_channels[0]
+        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,
+                downsample_type=downsample_type,
+                dropout=dropout,
+            )
+            self.down_blocks.append(down_block)
+
+        # mid
+        self.mid_block = UNetMidBlock2D(
+            in_channels=block_out_channels[-1],
+            temb_channels=time_embed_dim,
+            dropout=dropout,
+            resnet_eps=norm_eps,
+            resnet_act_fn=act_fn,
+            output_scale_factor=mid_block_scale_factor,
+            resnet_time_scale_shift=resnet_time_scale_shift,
+            attention_head_dim=attention_head_dim if attention_head_dim is not None else block_out_channels[-1],
+            resnet_groups=norm_num_groups,
+            attn_groups=attn_norm_num_groups,
+            add_attention=add_attention,
+        )
+
+        # up
+        reversed_block_out_channels = list(reversed(block_out_channels))
+        output_channel = reversed_block_out_channels[0]
+        for i, up_block_type in enumerate(up_block_types):
+            prev_output_channel = output_channel
+            output_channel = reversed_block_out_channels[i]
+            input_channel = reversed_block_out_channels[min(i + 1, len(block_out_channels) - 1)]
+
+            is_final_block = i == len(block_out_channels) - 1
+
+            up_block = get_up_block(
+                up_block_type,
+                num_layers=layers_per_block + 1,
+                in_channels=input_channel,
+                out_channels=output_channel,
+                prev_output_channel=prev_output_channel,
+                temb_channels=time_embed_dim,
+                add_upsample=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,
+                resnet_time_scale_shift=resnet_time_scale_shift,
+                upsample_type=upsample_type,
+                dropout=dropout,
+            )
+            self.up_blocks.append(up_block)
+            prev_output_channel = output_channel
+
+        # out
+        num_groups_out = norm_num_groups if norm_num_groups is not None else min(block_out_channels[0] // 4, 32)
+        self.conv_norm_out = nn.GroupNorm(num_channels=block_out_channels[0], num_groups=num_groups_out, eps=norm_eps)
+        self.conv_act = nn.SiLU()
+        self.conv_out = nn.Conv2d(block_out_channels[0], out_channels, kernel_size=3, padding=1)
+
+    def forward(
+        self,
+        sample: torch.FloatTensor,
+        timestep: Union[torch.Tensor, float, int],
+        class_labels: Optional[torch.Tensor] = None,
+        return_dict: bool = True,
+    ) -> Union[UNet2DOutput, Tuple]:
+        r"""
+        The [`UNet2DModel`] forward method.
+
+        Args:
+            sample (`torch.FloatTensor`):
+                The noisy input tensor with the following shape `(batch, channel, height, width)`.
+            timestep (`torch.FloatTensor` or `float` or `int`): The number of timesteps to denoise an input.
+            class_labels (`torch.FloatTensor`, *optional*, defaults to `None`):
+                Optional class labels for conditioning. Their embeddings will be summed with the timestep embeddings.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~models.unet_2d.UNet2DOutput`] instead of a plain tuple.
+
+        Returns:
+            [`~models.unet_2d.UNet2DOutput`] or `tuple`:
+                If `return_dict` is True, an [`~models.unet_2d.UNet2DOutput`] is returned, otherwise a `tuple` is
+                returned where the first element is the sample tensor.
+        """
+        # 0. center input if necessary
+        if self.config.center_input_sample:
+            sample = 2 * sample - 1.0
+
+        # 1. time
+        timesteps = timestep
+        if not torch.is_tensor(timesteps):
+            timesteps = torch.tensor([timesteps], dtype=torch.long, device=sample.device)
+        elif torch.is_tensor(timesteps) and 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 * torch.ones(sample.shape[0], dtype=timesteps.dtype, device=timesteps.device)
+
+        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=self.dtype)
+        emb = self.time_embedding(t_emb)
+
+        if self.class_embedding is not None:
+            if class_labels is None:
+                raise ValueError("class_labels should be provided when doing class conditioning")
+
+            if self.config.class_embed_type == "timestep":
+                class_labels = self.time_proj(class_labels)
+
+            class_emb = self.class_embedding(class_labels).to(dtype=self.dtype)
+            emb = emb + class_emb
+        elif self.class_embedding is None and class_labels is not None:
+            raise ValueError("class_embedding needs to be initialized in order to use class conditioning")
+
+        # 2. pre-process
+        skip_sample = sample
+        sample = self.conv_in(sample)
+
+        # 3. down
+        down_block_res_samples = (sample,)
+        for downsample_block in self.down_blocks:
+            if hasattr(downsample_block, "skip_conv"):
+                sample, res_samples, skip_sample = downsample_block(
+                    hidden_states=sample, temb=emb, skip_sample=skip_sample
+                )
+            else:
+                sample, res_samples = downsample_block(hidden_states=sample, temb=emb)
+
+            down_block_res_samples += res_samples
+
+        # 4. mid
+        sample = self.mid_block(sample, emb)
+
+        # 5. up
+        skip_sample = None
+        for upsample_block in self.up_blocks:
+            res_samples = down_block_res_samples[-len(upsample_block.resnets) :]
+            down_block_res_samples = down_block_res_samples[: -len(upsample_block.resnets)]
+
+            if hasattr(upsample_block, "skip_conv"):
+                sample, skip_sample = upsample_block(sample, res_samples, emb, skip_sample)
+            else:
+                sample = upsample_block(sample, res_samples, emb)
+
+        # 6. post-process
+        sample = self.conv_norm_out(sample)
+        sample = self.conv_act(sample)
+        sample = self.conv_out(sample)
+
+        if skip_sample is not None:
+            sample += skip_sample
+
+        if self.config.time_embedding_type == "fourier":
+            timesteps = timesteps.reshape((sample.shape[0], *([1] * len(sample.shape[1:]))))
+            sample = sample / timesteps
+
+        if not return_dict:
+            return (sample,)
+
+        return UNet2DOutput(sample=sample)
+
+
+class MultiLabelConditionalUNet2DModelForCelebaHQ(ModelMixin, ConfigMixin):
+    r"""
+    A 2D UNet model that takes a noisy sample and a timestep and returns a sample shaped output.
+
+    This model inherits from [`ModelMixin`]. Check the superclass documentation for it's generic methods implemented
+    for all models (such as downloading or saving).
+
+    Parameters:
+        sample_size (`int` or `Tuple[int, int]`, *optional*, defaults to `None`):
+            Height and width of input/output sample. Dimensions must be a multiple of `2 ** (len(block_out_channels) -
+            1)`.
+        in_channels (`int`, *optional*, defaults to 3): Number of channels in the input sample.
+        out_channels (`int`, *optional*, defaults to 3): Number of channels in the output.
+        center_input_sample (`bool`, *optional*, defaults to `False`): Whether to center the input sample.
+        time_embedding_type (`str`, *optional*, defaults to `"positional"`): Type of time embedding to use.
+        freq_shift (`int`, *optional*, defaults to 0): Frequency shift for Fourier time embedding.
+        flip_sin_to_cos (`bool`, *optional*, defaults to `True`):
+            Whether to flip sin to cos for Fourier time embedding.
+        down_block_types (`Tuple[str]`, *optional*, defaults to `("DownBlock2D", "AttnDownBlock2D", "AttnDownBlock2D", "AttnDownBlock2D")`):
+            Tuple of downsample block types.
+        mid_block_type (`str`, *optional*, defaults to `"UNetMidBlock2D"`):
+            Block type for middle of UNet, it can be either `UNetMidBlock2D` or `UnCLIPUNetMidBlock2D`.
+        up_block_types (`Tuple[str]`, *optional*, defaults to `("AttnUpBlock2D", "AttnUpBlock2D", "AttnUpBlock2D", "UpBlock2D")`):
+            Tuple of upsample block types.
+        block_out_channels (`Tuple[int]`, *optional*, defaults to `(224, 448, 672, 896)`):
+            Tuple of block output channels.
+        layers_per_block (`int`, *optional*, defaults to `2`): The number of layers per block.
+        mid_block_scale_factor (`float`, *optional*, defaults to `1`): The scale factor for the mid block.
+        downsample_padding (`int`, *optional*, defaults to `1`): The padding for the downsample convolution.
+        downsample_type (`str`, *optional*, defaults to `conv`):
+            The downsample type for downsampling layers. Choose between "conv" and "resnet"
+        upsample_type (`str`, *optional*, defaults to `conv`):
+            The upsample type for upsampling layers. Choose between "conv" and "resnet"
+        dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
+        act_fn (`str`, *optional*, defaults to `"silu"`): The activation function to use.
+        attention_head_dim (`int`, *optional*, defaults to `8`): The attention head dimension.
+        norm_num_groups (`int`, *optional*, defaults to `32`): The number of groups for normalization.
+        attn_norm_num_groups (`int`, *optional*, defaults to `None`):
+            If set to an integer, a group norm layer will be created in the mid block's [`Attention`] layer with the
+            given number of groups. If left as `None`, the group norm layer will only be created if
+            `resnet_time_scale_shift` is set to `default`, and if created will have `norm_num_groups` groups.
+        norm_eps (`float`, *optional*, defaults to `1e-5`): The epsilon for normalization.
+        resnet_time_scale_shift (`str`, *optional*, defaults to `"default"`): Time scale shift config
+            for ResNet blocks (see [`~models.resnet.ResnetBlock2D`]). Choose from `default` or `scale_shift`.
+        class_embed_type (`str`, *optional*, defaults to `None`):
+            The type of class embedding to use which is ultimately summed with the time embeddings. Choose from `None`,
+            `"timestep"`, or `"identity"`.
+        num_class_embeds (`int`, *optional*, defaults to `None`):
+            Input dimension of the learnable embedding matrix to be projected to `time_embed_dim` when performing class
+            conditioning with `class_embed_type` equal to `None`.
+    """
+
+    @register_to_config
+    def __init__(
+        self,
+        sample_size: Optional[Union[int, Tuple[int, int]]] = None,
+        in_channels: int = 3,
+        out_channels: int = 3,
+        center_input_sample: bool = False,
+        time_embedding_type: str = "positional",
+        freq_shift: int = 0,
+        flip_sin_to_cos: bool = True,
+        down_block_types: Tuple[str, ...] = ("DownBlock2D", "AttnDownBlock2D", "AttnDownBlock2D", "AttnDownBlock2D"),
+        up_block_types: Tuple[str, ...] = ("AttnUpBlock2D", "AttnUpBlock2D", "AttnUpBlock2D", "UpBlock2D"),
+        block_out_channels: Tuple[int, ...] = (224, 448, 672, 896),
+        layers_per_block: int = 2,
+        mid_block_scale_factor: float = 1,
+        downsample_padding: int = 1,
+        downsample_type: str = "conv",
+        upsample_type: str = "conv",
+        dropout: float = 0.0,
+        act_fn: str = "silu",
+        attention_head_dim: Optional[int] = 8,
+        norm_num_groups: int = 32,
+        attn_norm_num_groups: Optional[int] = None,
+        norm_eps: float = 1e-5,
+        resnet_time_scale_shift: str = "default",
+        add_attention: bool = True,
+        class_embed_type: Optional[str] = None,
+        num_train_timesteps: Optional[int] = None,
+        set_W_to_weight: Optional[bool] = True
+    ):
+        super().__init__()
+
+        self.sample_size = sample_size
+        time_embed_dim = block_out_channels[0] * 4
+
+        # Check inputs
+        if len(down_block_types) != len(up_block_types):
+            raise ValueError(
+                f"Must provide the same number of `down_block_types` as `up_block_types`. `down_block_types`: {down_block_types}. `up_block_types`: {up_block_types}."
+            )
+
+        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}."
+            )
+
+        # input
+        self.conv_in = nn.Conv2d(in_channels, block_out_channels[0], kernel_size=3, padding=(1, 1))
+
+        # time
+        if time_embedding_type == "fourier":
+            self.time_proj = GaussianFourierProjection(embedding_size=block_out_channels[0], scale=16, set_W_to_weight=set_W_to_weight)
+            timestep_input_dim = 2 * block_out_channels[0]
+        elif time_embedding_type == "positional":
+            self.time_proj = Timesteps(block_out_channels[0], flip_sin_to_cos, freq_shift)
+            timestep_input_dim = block_out_channels[0]
+        elif time_embedding_type == "learned":
+            self.time_proj = nn.Embedding(num_train_timesteps, block_out_channels[0])
+            timestep_input_dim = block_out_channels[0]
+
+        self.time_embedding = TimestepEmbedding(timestep_input_dim, time_embed_dim)
+
+        # class embedding
+        if class_embed_type is None:
+            # self.class_embedding_5_o_Clock_Shadow = nn.Embedding(2, time_embed_dim)
+            # self.class_embedding_Arched_Eyebrows = nn.Embedding(2, time_embed_dim)
+            # self.class_embedding_Attractive = nn.Embedding(2, time_embed_dim)
+            # self.class_embedding_Bags_Under_Eyes = nn.Embedding(2, time_embed_dim)
+            # self.class_embedding_Bald = nn.Embedding(2, time_embed_dim)
+            self.class_embedding_Bangs = nn.Embedding(3, time_embed_dim)
+            # self.class_embedding_Big_Lips = nn.Embedding(2, time_embed_dim)
+            # self.class_embedding_Big_Nose = nn.Embedding(2, time_embed_dim)
+            # self.class_embedding_Black_Hair = nn.Embedding(2, time_embed_dim)
+            # self.class_embedding_Blond_Hair = nn.Embedding(2, time_embed_dim)
+            # self.class_embedding_Blurry = nn.Embedding(2, time_embed_dim)
+            # self.class_embedding_Brown_Hair = nn.Embedding(2, time_embed_dim)
+            # self.class_embedding_Bushy_Eyebrows = nn.Embedding(2, time_embed_dim)
+            # self.class_embedding_Chubby = nn.Embedding(2, time_embed_dim)
+            # self.class_embedding_Double_Chin = nn.Embedding(2, time_embed_dim)
+            self.class_embedding_Eyeglasses = nn.Embedding(3, time_embed_dim)
+            # self.class_embedding_Goatee = nn.Embedding(2, time_embed_dim)
+            # self.class_embedding_Gray_Hair = nn.Embedding(2, time_embed_dim)
+            # self.class_embedding_Heavy_Makeup = nn.Embedding(2, time_embed_dim)
+            # self.class_embedding_High_Cheekbones = nn.Embedding(2, time_embed_dim)
+            self.class_embedding_Male = nn.Embedding(3, time_embed_dim)
+            # self.class_embedding_Mouth_Slightly_Open = nn.Embedding(2, time_embed_dim)
+            self.class_embedding_Mustache = nn.Embedding(3, time_embed_dim)
+            # self.class_embedding_Narrow_Eyes = nn.Embedding(2, time_embed_dim)
+            # self.class_embedding_No_Beard = nn.Embedding(2, time_embed_dim)
+            # self.class_embedding_Oval_Face = nn.Embedding(2, time_embed_dim)
+            # self.class_embedding_Pale_Skin = nn.Embedding(2, time_embed_dim)
+            self.class_embedding_Pointy_Nose = nn.Embedding(3, time_embed_dim)
+            # self.class_embedding_Receding_Hairline = nn.Embedding(2, time_embed_dim)
+            # self.class_embedding_Rosy_Cheeks = nn.Embedding(2, time_embed_dim)
+            # self.class_embedding_Sideburns = nn.Embedding(2, time_embed_dim)
+            self.class_embedding_Smiling = nn.Embedding(3, time_embed_dim)
+            # self.class_embedding_Straight_Hair = nn.Embedding(2, time_embed_dim)
+            # self.class_embedding_Wavy_Hair = nn.Embedding(2, time_embed_dim)
+            # self.class_embedding_Wearing_Earrings = nn.Embedding(2, time_embed_dim)
+            self.class_embedding_Wearing_Hat = nn.Embedding(3, time_embed_dim)
+            # self.class_embedding_Wearing_Lipstick = nn.Embedding(2, time_embed_dim)
+            # self.class_embedding_Wearing_Necklace = nn.Embedding(2, time_embed_dim)
+            # self.class_embedding_Wearing_Necktie = nn.Embedding(2, time_embed_dim)
+            self.class_embedding_Young = nn.Embedding(3, time_embed_dim)
+            self.apply_class_emb = True
+        elif class_embed_type == "timestep":
+            # self.class_embedding_5_o_Clock_Shadow = TimestepEmbedding(timestep_input_dim, time_embed_dim)
+            # self.class_embedding_Arched_Eyebrows = TimestepEmbedding(timestep_input_dim, time_embed_dim)
+            # self.class_embedding_Attractive = TimestepEmbedding(timestep_input_dim, time_embed_dim)
+            # self.class_embedding_Bags_Under_Eyes = TimestepEmbedding(timestep_input_dim, time_embed_dim)
+            # self.class_embedding_Bald = TimestepEmbedding(timestep_input_dim, time_embed_dim)
+            self.class_embedding_Bangs = TimestepEmbedding(timestep_input_dim, time_embed_dim)
+            # self.class_embedding_Big_Lips = TimestepEmbedding(timestep_input_dim, time_embed_dim)
+            # self.class_embedding_Big_Nose = TimestepEmbedding(timestep_input_dim, time_embed_dim)
+            # self.class_embedding_Black_Hair = TimestepEmbedding(timestep_input_dim, time_embed_dim)
+            # self.class_embedding_Blond_Hair = TimestepEmbedding(timestep_input_dim, time_embed_dim)
+            # self.class_embedding_Blurry = TimestepEmbedding(timestep_input_dim, time_embed_dim)
+            # self.class_embedding_Brown_Hair = TimestepEmbedding(timestep_input_dim, time_embed_dim)
+            # self.class_embedding_Bushy_Eyebrows = TimestepEmbedding(timestep_input_dim, time_embed_dim)
+            # self.class_embedding_Chubby = TimestepEmbedding(timestep_input_dim, time_embed_dim)
+            # self.class_embedding_Double_Chin = TimestepEmbedding(timestep_input_dim, time_embed_dim)
+            self.class_embedding_Eyeglasses = TimestepEmbedding(timestep_input_dim, time_embed_dim)
+            # self.class_embedding_Goatee = TimestepEmbedding(timestep_input_dim, time_embed_dim)
+            # self.class_embedding_Gray_Hair = TimestepEmbedding(timestep_input_dim, time_embed_dim)
+            # self.class_embedding_Heavy_Makeup = TimestepEmbedding(timestep_input_dim, time_embed_dim)
+            # self.class_embedding_High_Cheekbones = TimestepEmbedding(timestep_input_dim, time_embed_dim)
+            self.class_embedding_Male = TimestepEmbedding(timestep_input_dim, time_embed_dim)
+            # self.class_embedding_Mouth_Slightly_Open = TimestepEmbedding(timestep_input_dim, time_embed_dim)
+            self.class_embedding_Mustache = TimestepEmbedding(timestep_input_dim, time_embed_dim)
+            # self.class_embedding_Narrow_Eyes = TimestepEmbedding(timestep_input_dim, time_embed_dim)
+            # self.class_embedding_No_Beard = TimestepEmbedding(timestep_input_dim, time_embed_dim)
+            # self.class_embedding_Oval_Face = TimestepEmbedding(timestep_input_dim, time_embed_dim)
+            # self.class_embedding_Pale_Skin = TimestepEmbedding(timestep_input_dim, time_embed_dim)
+            self.class_embedding_Pointy_Nose = TimestepEmbedding(timestep_input_dim, time_embed_dim)
+            # self.class_embedding_Receding_Hairline = TimestepEmbedding(timestep_input_dim, time_embed_dim)
+            # self.class_embedding_Rosy_Cheeks = TimestepEmbedding(timestep_input_dim, time_embed_dim)
+            # self.class_embedding_Sideburns = TimestepEmbedding(timestep_input_dim, time_embed_dim)
+            self.class_embedding_Smiling = TimestepEmbedding(timestep_input_dim, time_embed_dim)
+            # self.class_embedding_Straight_Hair = TimestepEmbedding(timestep_input_dim, time_embed_dim)
+            # self.class_embedding_Wavy_Hair = TimestepEmbedding(timestep_input_dim, time_embed_dim)
+            # self.class_embedding_Wearing_Earrings = TimestepEmbedding(timestep_input_dim, time_embed_dim)
+            self.class_embedding_Wearing_Hat = TimestepEmbedding(timestep_input_dim, time_embed_dim)
+            # self.class_embedding_Wearing_Lipstick = TimestepEmbedding(timestep_input_dim, time_embed_dim)
+            # self.class_embedding_Wearing_Necklace = TimestepEmbedding(timestep_input_dim, time_embed_dim)
+            # self.class_embedding_Wearing_Necktie = TimestepEmbedding(timestep_input_dim, time_embed_dim)
+            self.class_embedding_Young = TimestepEmbedding(timestep_input_dim, time_embed_dim)
+            self.apply_class_emb = True
+        elif class_embed_type == "identity":
+            # self.class_embedding_5_o_Clock_Shadow = nn.Identity(time_embed_dim)
+            # self.class_embedding_Arched_Eyebrows = nn.Identity(time_embed_dim)
+            # self.class_embedding_Attractive = nn.Identity(time_embed_dim)
+            # self.class_embedding_Bags_Under_Eyes = nn.Identity(time_embed_dim)
+            # self.class_embedding_Bald = nn.Identity(time_embed_dim)
+            self.class_embedding_Bangs = nn.Identity(time_embed_dim)
+            # self.class_embedding_Big_Lips = nn.Identity(time_embed_dim)
+            # self.class_embedding_Big_Nose = nn.Identity(time_embed_dim)
+            # self.class_embedding_Black_Hair = nn.Identity(time_embed_dim)
+            # self.class_embedding_Blond_Hair = nn.Identity(time_embed_dim)
+            # self.class_embedding_Blurry = nn.Identity(time_embed_dim)
+            # self.class_embedding_Brown_Hair = nn.Identity(time_embed_dim)
+            # self.class_embedding_Bushy_Eyebrows = nn.Identity(time_embed_dim)
+            # self.class_embedding_Chubby = nn.Identity(time_embed_dim)
+            # self.class_embedding_Double_Chin = nn.Identity(time_embed_dim)
+            self.class_embedding_Eyeglasses = nn.Identity(time_embed_dim)
+            # self.class_embedding_Goatee = nn.Identity(time_embed_dim)
+            # self.class_embedding_Gray_Hair = nn.Identity(time_embed_dim)
+            # self.class_embedding_Heavy_Makeup = nn.Identity(time_embed_dim)
+            # self.class_embedding_High_Cheekbones = nn.Identity(time_embed_dim)
+            self.class_embedding_Male = nn.Identity(time_embed_dim)
+            # self.class_embedding_Mouth_Slightly_Open = nn.Identity(time_embed_dim)
+            self.class_embedding_Mustache = nn.Identity(time_embed_dim)
+            # self.class_embedding_Narrow_Eyes = nn.Identity(time_embed_dim)
+            # self.class_embedding_No_Beard = nn.Identity(time_embed_dim)
+            # self.class_embedding_Oval_Face = nn.Identity(time_embed_dim)
+            # self.class_embedding_Pale_Skin = nn.Identity(time_embed_dim)
+            self.class_embedding_Pointy_Nose = nn.Identity(time_embed_dim)
+            # self.class_embedding_Receding_Hairline = nn.Identity(time_embed_dim)
+            # self.class_embedding_Rosy_Cheeks = nn.Identity(time_embed_dim)
+            # self.class_embedding_Sideburns = nn.Identity(time_embed_dim)
+            self.class_embedding_Smiling = nn.Identity(time_embed_dim)
+            # self.class_embedding_Straight_Hair = nn.Identity(time_embed_dim)
+            # self.class_embedding_Wavy_Hair = nn.Identity(time_embed_dim)
+            # self.class_embedding_Wearing_Earrings = nn.Identity(time_embed_dim)
+            self.class_embedding_Wearing_Hat = nn.Identity(time_embed_dim)
+            # self.class_embedding_Wearing_Lipstick = nn.Identity(time_embed_dim)
+            # self.class_embedding_Wearing_Necklace = nn.Identity(time_embed_dim)
+            # self.class_embedding_Wearing_Necktie = nn.Identity(time_embed_dim)
+            self.class_embedding_Young = nn.Identity(time_embed_dim)
+            self.apply_class_emb = True
+        else:
+            self.apply_class_emb = False
+
+        self.down_blocks = nn.ModuleList([])
+        self.mid_block = None
+        self.up_blocks = nn.ModuleList([])
+
+        # down
+        output_channel = block_out_channels[0]
+        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,
+                downsample_type=downsample_type,
+                dropout=dropout,
+            )
+            self.down_blocks.append(down_block)
+
+        # mid
+        self.mid_block = UNetMidBlock2D(
+            in_channels=block_out_channels[-1],
+            temb_channels=time_embed_dim,
+            dropout=dropout,
+            resnet_eps=norm_eps,
+            resnet_act_fn=act_fn,
+            output_scale_factor=mid_block_scale_factor,
+            resnet_time_scale_shift=resnet_time_scale_shift,
+            attention_head_dim=attention_head_dim if attention_head_dim is not None else block_out_channels[-1],
+            resnet_groups=norm_num_groups,
+            attn_groups=attn_norm_num_groups,
+            add_attention=add_attention,
+        )
+
+        # up
+        reversed_block_out_channels = list(reversed(block_out_channels))
+        output_channel = reversed_block_out_channels[0]
+        for i, up_block_type in enumerate(up_block_types):
+            prev_output_channel = output_channel
+            output_channel = reversed_block_out_channels[i]
+            input_channel = reversed_block_out_channels[min(i + 1, len(block_out_channels) - 1)]
+
+            is_final_block = i == len(block_out_channels) - 1
+
+            up_block = get_up_block(
+                up_block_type,
+                num_layers=layers_per_block + 1,
+                in_channels=input_channel,
+                out_channels=output_channel,
+                prev_output_channel=prev_output_channel,
+                temb_channels=time_embed_dim,
+                add_upsample=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,
+                resnet_time_scale_shift=resnet_time_scale_shift,
+                upsample_type=upsample_type,
+                dropout=dropout,
+            )
+            self.up_blocks.append(up_block)
+            prev_output_channel = output_channel
+
+        # out
+        num_groups_out = norm_num_groups if norm_num_groups is not None else min(block_out_channels[0] // 4, 32)
+        self.conv_norm_out = nn.GroupNorm(num_channels=block_out_channels[0], num_groups=num_groups_out, eps=norm_eps)
+        self.conv_act = nn.SiLU()
+        self.conv_out = nn.Conv2d(block_out_channels[0], out_channels, kernel_size=3, padding=1)
+
+    def forward(
+        self,
+        sample: torch.FloatTensor,
+        timestep: Union[torch.Tensor, float, int],
+        class_labels: Optional[torch.Tensor] = None,
+        return_dict: bool = True,
+    ) -> Union[UNet2DOutput, Tuple]:
+        r"""
+        The [`UNet2DModel`] forward method.
+
+        Args:
+            sample (`torch.FloatTensor`):
+                The noisy input tensor with the following shape `(batch, channel, height, width)`.
+            timestep (`torch.FloatTensor` or `float` or `int`): The number of timesteps to denoise an input.
+            class_labels (`torch.FloatTensor`, *optional*, defaults to `None`):
+                Optional class labels for conditioning. Their embeddings will be summed with the timestep embeddings.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~models.unet_2d.UNet2DOutput`] instead of a plain tuple.
+
+        Returns:
+            [`~models.unet_2d.UNet2DOutput`] or `tuple`:
+                If `return_dict` is True, an [`~models.unet_2d.UNet2DOutput`] is returned, otherwise a `tuple` is
+                returned where the first element is the sample tensor.
+        """
+        # 0. center input if necessary
+        if self.config.center_input_sample:
+            sample = 2 * sample - 1.0
+
+        # 1. time
+        timesteps = timestep
+        if not torch.is_tensor(timesteps):
+            timesteps = torch.tensor([timesteps], dtype=torch.long, device=sample.device)
+        elif torch.is_tensor(timesteps) and 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 * torch.ones(sample.shape[0], dtype=timesteps.dtype, device=timesteps.device)
+
+        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=self.dtype)
+        emb = self.time_embedding(t_emb)
+
+        if self.apply_class_emb:
+            if class_labels is None:
+                raise ValueError("class_labels should be provided when doing class conditioning")
+            # class_labels_5_o_Clock_Shadow = class_labels[:, 0]
+            # class_labels_Arched_Eyebrows = class_labels[:, 1]
+            # class_labels_Attractive = class_labels[:, 2]
+            # class_labels_Bags_Under_Eyes = class_labels[:, 3]
+            # class_labels_Bald = class_labels[:, 4]
+            class_labels_Bangs = class_labels[:, 5]
+            # class_labels_Big_Lips = class_labels[:, 6]
+            # class_labels_Big_Nose = class_labels[:, 7]
+            # class_labels_Black_Hair = class_labels[:, 8]
+            # class_labels_Blond_Hair = class_labels[:, 9]
+            # class_labels_Blurry = class_labels[:, 10]
+            # class_labels_Brown_Hair = class_labels[:, 11]
+            # class_labels_Bushy_Eyebrows = class_labels[:, 12]
+            # class_labels_Chubby = class_labels[:, 13]
+            # class_labels_Double_Chin = class_labels[:, 14]
+            class_labels_Eyeglasses = class_labels[:, 15]
+            # class_labels_Goatee = class_labels[:, 16]
+            # class_labels_Gray_Hair = class_labels[:, 17]
+            # class_labels_Heavy_Makeup = class_labels[:, 18]
+            # class_labels_High_Cheekbones = class_labels[:, 19]
+            class_labels_Male = class_labels[:, 20]
+            # class_labels_Mouth_Slightly_Open = class_labels[:, 21]
+            class_labels_Mustache = class_labels[:, 22]
+            # class_labels_Narrow_Eyes = class_labels[:, 23]
+            # class_labels_No_Beard = class_labels[:, 24]
+            # class_labels_Oval_Face = class_labels[:, 25]
+            # class_labels_Pale_Skin = class_labels[:, 26]
+            class_labels_Pointy_Nose = class_labels[:, 27]
+            # class_labels_Receding_Hairline = class_labels[:, 28]
+            # class_labels_Rosy_Cheeks = class_labels[:, 29]
+            # class_labels_Sideburns = class_labels[:, 30]
+            class_labels_Smiling = class_labels[:, 31]
+            # class_labels_Straight_Hair = class_labels[:, 32]
+            # class_labels_Wavy_Hair = class_labels[:, 33]
+            # class_labels_Wearing_Earrings = class_labels[:, 34]
+            class_labels_Wearing_Hat = class_labels[:, 35]
+            # class_labels_Wearing_Lipstick = class_labels[:, 36]
+            # class_labels_Wearing_Necklace = class_labels[:, 37]
+            # class_labels_Wearing_Necktie = class_labels[:, 38]
+            class_labels_Young = class_labels[:, 39]
+
+            # Apply time projection if configured
+            if self.config.class_embed_type == "timestep":
+                # class_labels_5_o_Clock_Shadow = self.time_proj(class_labels_5_o_Clock_Shadow)
+                # class_labels_Arched_Eyebrows = self.time_proj(class_labels_Arched_Eyebrows)
+                # class_labels_Attractive = self.time_proj(class_labels_Attractive)
+                # class_labels_Bags_Under_Eyes = self.time_proj(class_labels_Bags_Under_Eyes)
+                # class_labels_Bald = self.time_proj(class_labels_Bald)
+                class_labels_Bangs = self.time_proj(class_labels_Bangs)
+                # class_labels_Big_Lips = self.time_proj(class_labels_Big_Lips)
+                # class_labels_Big_Nose = self.time_proj(class_labels_Big_Nose)
+                # class_labels_Black_Hair = self.time_proj(class_labels_Black_Hair)
+                # class_labels_Blond_Hair = self.time_proj(class_labels_Blond_Hair)
+                # class_labels_Blurry = self.time_proj(class_labels_Blurry)
+                # class_labels_Brown_Hair = self.time_proj(class_labels_Brown_Hair)
+                # class_labels_Bushy_Eyebrows = self.time_proj(class_labels_Bushy_Eyebrows)
+                # class_labels_Chubby = self.time_proj(class_labels_Chubby)
+                # class_labels_Double_Chin = self.time_proj(class_labels_Double_Chin)
+                class_labels_Eyeglasses = self.time_proj(class_labels_Eyeglasses)
+                # class_labels_Goatee = self.time_proj(class_labels_Goatee)
+                # class_labels_Gray_Hair = self.time_proj(class_labels_Gray_Hair)
+                # class_labels_Heavy_Makeup = self.time_proj(class_labels_Heavy_Makeup)
+                # class_labels_High_Cheekbones = self.time_proj(class_labels_High_Cheekbones)
+                class_labels_Male = self.time_proj(class_labels_Male)
+                # class_labels_Mouth_Slightly_Open = self.time_proj(class_labels_Mouth_Slightly_Open)
+                class_labels_Mustache = self.time_proj(class_labels_Mustache)
+                # class_labels_Narrow_Eyes = self.time_proj(class_labels_Narrow_Eyes)
+                # class_labels_No_Beard = self.time_proj(class_labels_No_Beard)
+                # class_labels_Oval_Face = self.time_proj(class_labels_Oval_Face)
+                # class_labels_Pale_Skin = self.time_proj(class_labels_Pale_Skin)
+                class_labels_Pointy_Nose = self.time_proj(class_labels_Pointy_Nose)
+                # class_labels_Receding_Hairline = self.time_proj(class_labels_Receding_Hairline)
+                # class_labels_Rosy_Cheeks = self.time_proj(class_labels_Rosy_Cheeks)
+                # class_labels_Sideburns = self.time_proj(class_labels_Sideburns)
+                class_labels_Smiling = self.time_proj(class_labels_Smiling)
+                # class_labels_Straight_Hair = self.time_proj(class_labels_Straight_Hair)
+                # class_labels_Wavy_Hair = self.time_proj(class_labels_Wavy_Hair)
+                # class_labels_Wearing_Earrings = self.time_proj(class_labels_Wearing_Earrings)
+                class_labels_Wearing_Hat = self.time_proj(class_labels_Wearing_Hat)
+                # class_labels_Wearing_Lipstick = self.time_proj(class_labels_Wearing_Lipstick)
+                # class_labels_Wearing_Necklace = self.time_proj(class_labels_Wearing_Necklace)
+                # class_labels_Wearing_Necktie = self.time_proj(class_labels_Wearing_Necktie)
+                class_labels_Young = self.time_proj(class_labels_Young)
+
+            if self.class_embedding_Bangs:
+                emb += self.class_embedding_Bangs(class_labels_Bangs + 1)
+            if self.class_embedding_Male:
+                emb += self.class_embedding_Male(class_labels_Male + 1)
+            if self.class_embedding_Eyeglasses:
+                emb += self.class_embedding_Eyeglasses(class_labels_Eyeglasses + 1)
+            if self.class_embedding_Mustache:
+                emb += self.class_embedding_Mustache(class_labels_Mustache + 1)
+            if self.class_embedding_Pointy_Nose:
+                emb += self.class_embedding_Pointy_Nose(class_labels_Pointy_Nose + 1)
+            if self.class_embedding_Smiling:
+                emb += self.class_embedding_Smiling(class_labels_Smiling + 1)
+            if self.class_embedding_Wearing_Hat:
+                emb += self.class_embedding_Wearing_Hat(class_labels_Wearing_Hat + 1)
+            if self.class_embedding_Young:
+                emb += self.class_embedding_Young(class_labels_Young + 1)
+        elif not self.apply_class_emb and class_labels is not None:
+            raise ValueError("class_embedding needs to be initialized in order to use class conditioning")
+
+        # 2. pre-process
+        skip_sample = sample
+        sample = self.conv_in(sample)
+
+        # 3. down
+        down_block_res_samples = (sample,)
+        for downsample_block in self.down_blocks:
+            if hasattr(downsample_block, "skip_conv"):
+                sample, res_samples, skip_sample = downsample_block(
+                    hidden_states=sample, temb=emb, skip_sample=skip_sample
+                )
+            else:
+                sample, res_samples = downsample_block(hidden_states=sample, temb=emb)
+
+            down_block_res_samples += res_samples
+
+        # 4. mid
+        sample = self.mid_block(sample, emb)
+
+        # 5. up
+        skip_sample = None
+        for upsample_block in self.up_blocks:
+            res_samples = down_block_res_samples[-len(upsample_block.resnets) :]
+            down_block_res_samples = down_block_res_samples[: -len(upsample_block.resnets)]
+
+            if hasattr(upsample_block, "skip_conv"):
+                sample, skip_sample = upsample_block(sample, res_samples, emb, skip_sample)
+            else:
+                sample = upsample_block(sample, res_samples, emb)
+
+        # 6. post-process
+        sample = self.conv_norm_out(sample)
+        sample = self.conv_act(sample)
+        sample = self.conv_out(sample)
+
+        if skip_sample is not None:
+            sample += skip_sample
+
+        if self.config.time_embedding_type == "fourier":
+            timesteps = timesteps.reshape((sample.shape[0], *([1] * len(sample.shape[1:]))))
+            sample = sample / timesteps
+
+        if not return_dict:
+            return (sample,)
+
+        return UNet2DOutput(sample=sample)