from dataclasses import dataclass from typing import Optional, Tuple, Union import torch import torch.nn as nn from ..configuration_utils import ConfigMixin, register_to_config from ..modeling_utils import ModelMixin from ..utils import BaseOutput from .embeddings import GaussianFourierProjection, TimestepEmbedding, Timesteps from .unet_blocks import UNetMidBlock2D, get_down_block, get_up_block @dataclass class UNet2DOutput(BaseOutput): """ Args: sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`): Hidden states output. Output of last layer of model. """ sample: torch.DoubleTensor class UNet2DModel(ModelMixin, ConfigMixin): r""" UNet2DModel is a 2D UNet model that takes in a noisy sample and a timestep and returns sample shaped output. This model inherits from [`ModelMixin`]. Check the superclass documentation for the generic methods the library implements for all the model (such as downloading or saving, etc.) Parameters: sample_size (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`, *optional*): Input sample size. in_channels (`int`, *optional*, defaults to 3): Number of channels in the input image. 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 : obj:`False`): Whether to flip sin to cos for fourier time embedding. down_block_types (`Tuple[str]`, *optional*, defaults to : obj:`("DownBlock2D", "AttnDownBlock2D", "AttnDownBlock2D", "AttnDownBlock2D")`): Tuple of downsample block types. up_block_types (`Tuple[str]`, *optional*, defaults to : obj:`("AttnUpBlock2D", "AttnUpBlock2D", "AttnUpBlock2D", "UpBlock2D")`): Tuple of upsample block types. block_out_channels (`Tuple[int]`, *optional*, defaults to : obj:`(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. 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 the normalization. norm_eps (`float`, *optional*, defaults to `1e-5`): The epsilon for the normalization. """ @register_to_config def __init__( self, sample_size: Optional[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 = 1, downsample_padding: int = 1, act_fn: str = "silu", attention_head_dim: int = 8, norm_num_groups: int = 32, norm_eps = 1e-5, ): super().__init__() self.sample_size = sample_size time_embed_dim = block_out_channels[0] * 4 # 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) 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] self.time_embedding = TimestepEmbedding(timestep_input_dim, time_embed_dim) 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, attn_num_head_channels=attention_head_dim, downsample_padding=downsample_padding, ) self.down_blocks.append(down_block) # mid 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_time_scale_shift="default", attn_num_head_channels=attention_head_dim, resnet_groups=norm_num_groups, ) # 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, attn_num_head_channels=attention_head_dim, ) 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, 3, padding=1) def forward( self, sample: torch.DoubleTensor, timestep: Union[torch.Tensor, float, int], return_dict: bool = True, ) -> Union[UNet2DOutput, Tuple]: """r Args: sample (`torch.FloatTensor`): (batch, channel, height, width) noisy inputs tensor timestep (`torch.FloatTensor` or `float` or `int): (batch) timesteps 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`: [`~models.unet_2d.UNet2DOutput`] if `return_dict` is True, otherwise a `tuple`. When returning a tuple, 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) emb = self.time_embedding(t_emb) # 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 # make sure hidden states is in float32 # when running in half-precision sample = self.conv_norm_out(sample.double()).type(sample.dtype) 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)