1271 lines
51 KiB
Python
1271 lines
51 KiB
Python
from typing import List, Optional, Tuple, Union
|
|
|
|
import numpy as np
|
|
import torch
|
|
import torch.nn as nn
|
|
|
|
from ...configuration_utils import ConfigMixin, register_to_config
|
|
from ...modeling_utils import ModelMixin
|
|
from ...models.attention import CrossAttention, DualTransformer2DModel, Transformer2DModel
|
|
from ...models.embeddings import TimestepEmbedding, Timesteps
|
|
from ...models.unet_2d_blocks import UNetMidBlock2DSimpleCrossAttn as UNetMidBlockFlatSimpleCrossAttn
|
|
from ...models.unet_2d_condition import UNet2DConditionOutput
|
|
from ...utils import logging
|
|
|
|
|
|
logger = logging.get_logger(__name__) # pylint: disable=invalid-name
|
|
|
|
|
|
def get_down_block(
|
|
down_block_type,
|
|
num_layers,
|
|
in_channels,
|
|
out_channels,
|
|
temb_channels,
|
|
add_downsample,
|
|
resnet_eps,
|
|
resnet_act_fn,
|
|
attn_num_head_channels,
|
|
resnet_groups=None,
|
|
cross_attention_dim=None,
|
|
downsample_padding=None,
|
|
dual_cross_attention=False,
|
|
use_linear_projection=False,
|
|
only_cross_attention=False,
|
|
upcast_attention=False,
|
|
resnet_time_scale_shift="default",
|
|
):
|
|
down_block_type = down_block_type[7:] if down_block_type.startswith("UNetRes") else down_block_type
|
|
if down_block_type == "DownBlockFlat":
|
|
return DownBlockFlat(
|
|
num_layers=num_layers,
|
|
in_channels=in_channels,
|
|
out_channels=out_channels,
|
|
temb_channels=temb_channels,
|
|
add_downsample=add_downsample,
|
|
resnet_eps=resnet_eps,
|
|
resnet_act_fn=resnet_act_fn,
|
|
resnet_groups=resnet_groups,
|
|
downsample_padding=downsample_padding,
|
|
resnet_time_scale_shift=resnet_time_scale_shift,
|
|
)
|
|
elif down_block_type == "CrossAttnDownBlockFlat":
|
|
if cross_attention_dim is None:
|
|
raise ValueError("cross_attention_dim must be specified for CrossAttnDownBlockFlat")
|
|
return CrossAttnDownBlockFlat(
|
|
num_layers=num_layers,
|
|
in_channels=in_channels,
|
|
out_channels=out_channels,
|
|
temb_channels=temb_channels,
|
|
add_downsample=add_downsample,
|
|
resnet_eps=resnet_eps,
|
|
resnet_act_fn=resnet_act_fn,
|
|
resnet_groups=resnet_groups,
|
|
downsample_padding=downsample_padding,
|
|
cross_attention_dim=cross_attention_dim,
|
|
attn_num_head_channels=attn_num_head_channels,
|
|
dual_cross_attention=dual_cross_attention,
|
|
use_linear_projection=use_linear_projection,
|
|
only_cross_attention=only_cross_attention,
|
|
resnet_time_scale_shift=resnet_time_scale_shift,
|
|
)
|
|
raise ValueError(f"{down_block_type} is not supported.")
|
|
|
|
|
|
def get_up_block(
|
|
up_block_type,
|
|
num_layers,
|
|
in_channels,
|
|
out_channels,
|
|
prev_output_channel,
|
|
temb_channels,
|
|
add_upsample,
|
|
resnet_eps,
|
|
resnet_act_fn,
|
|
attn_num_head_channels,
|
|
resnet_groups=None,
|
|
cross_attention_dim=None,
|
|
dual_cross_attention=False,
|
|
use_linear_projection=False,
|
|
only_cross_attention=False,
|
|
upcast_attention=False,
|
|
resnet_time_scale_shift="default",
|
|
):
|
|
up_block_type = up_block_type[7:] if up_block_type.startswith("UNetRes") else up_block_type
|
|
if up_block_type == "UpBlockFlat":
|
|
return UpBlockFlat(
|
|
num_layers=num_layers,
|
|
in_channels=in_channels,
|
|
out_channels=out_channels,
|
|
prev_output_channel=prev_output_channel,
|
|
temb_channels=temb_channels,
|
|
add_upsample=add_upsample,
|
|
resnet_eps=resnet_eps,
|
|
resnet_act_fn=resnet_act_fn,
|
|
resnet_groups=resnet_groups,
|
|
resnet_time_scale_shift=resnet_time_scale_shift,
|
|
)
|
|
elif up_block_type == "CrossAttnUpBlockFlat":
|
|
if cross_attention_dim is None:
|
|
raise ValueError("cross_attention_dim must be specified for CrossAttnUpBlockFlat")
|
|
return CrossAttnUpBlockFlat(
|
|
num_layers=num_layers,
|
|
in_channels=in_channels,
|
|
out_channels=out_channels,
|
|
prev_output_channel=prev_output_channel,
|
|
temb_channels=temb_channels,
|
|
add_upsample=add_upsample,
|
|
resnet_eps=resnet_eps,
|
|
resnet_act_fn=resnet_act_fn,
|
|
resnet_groups=resnet_groups,
|
|
cross_attention_dim=cross_attention_dim,
|
|
attn_num_head_channels=attn_num_head_channels,
|
|
dual_cross_attention=dual_cross_attention,
|
|
use_linear_projection=use_linear_projection,
|
|
only_cross_attention=only_cross_attention,
|
|
resnet_time_scale_shift=resnet_time_scale_shift,
|
|
)
|
|
raise ValueError(f"{up_block_type} is not supported.")
|
|
|
|
|
|
# Copied from diffusers.models.unet_2d_condition.UNet2DConditionModel with UNet2DConditionModel->UNetFlatConditionModel, nn.Conv2d->LinearMultiDim, Block2D->BlockFlat
|
|
class UNetFlatConditionModel(ModelMixin, ConfigMixin):
|
|
r"""
|
|
UNetFlatConditionModel is a conditional 2D UNet model that takes in a noisy sample, conditional state, 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 models (such as downloading or saving, etc.)
|
|
|
|
Parameters:
|
|
sample_size (`int` or `Tuple[int, int]`, *optional*, defaults to `None`):
|
|
Height and width of input/output sample.
|
|
in_channels (`int`, *optional*, defaults to 4): The number of channels in the input sample.
|
|
out_channels (`int`, *optional*, defaults to 4): The number of channels in the output.
|
|
center_input_sample (`bool`, *optional*, defaults to `False`): Whether to center the input sample.
|
|
flip_sin_to_cos (`bool`, *optional*, defaults to `False`):
|
|
Whether to flip the sin to cos in the time embedding.
|
|
freq_shift (`int`, *optional*, defaults to 0): The frequency shift to apply to the time embedding.
|
|
down_block_types (`Tuple[str]`, *optional*, defaults to `("CrossAttnDownBlockFlat", "CrossAttnDownBlockFlat", "CrossAttnDownBlockFlat", "DownBlockFlat")`):
|
|
The tuple of downsample blocks to use.
|
|
mid_block_type (`str`, *optional*, defaults to `"UNetMidBlockFlatCrossAttn"`):
|
|
The mid block type. Choose from `UNetMidBlockFlatCrossAttn` or `UNetMidBlockFlatSimpleCrossAttn`.
|
|
up_block_types (`Tuple[str]`, *optional*, defaults to `("UpBlockFlat", "CrossAttnUpBlockFlat", "CrossAttnUpBlockFlat", "CrossAttnUpBlockFlat",)`):
|
|
The tuple of upsample blocks to use.
|
|
block_out_channels (`Tuple[int]`, *optional*, defaults to `(320, 640, 1280, 1280)`):
|
|
The tuple of output channels for each block.
|
|
layers_per_block (`int`, *optional*, defaults to 2): The number of layers per block.
|
|
downsample_padding (`int`, *optional*, defaults to 1): The padding to use for the downsampling convolution.
|
|
mid_block_scale_factor (`float`, *optional*, defaults to 1.0): The scale factor to use for the mid block.
|
|
act_fn (`str`, *optional*, 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.
|
|
norm_eps (`float`, *optional*, defaults to 1e-5): The epsilon to use for the normalization.
|
|
cross_attention_dim (`int`, *optional*, defaults to 1280): The dimension of the cross attention features.
|
|
attention_head_dim (`int`, *optional*, defaults to 8): The dimension of the attention heads.
|
|
resnet_time_scale_shift (`str`, *optional*, defaults to `"default"`): Time scale shift config
|
|
for resnet blocks, see [`~models.resnet.ResnetBlockFlat`]. 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"`.
|
|
"""
|
|
|
|
_supports_gradient_checkpointing = True
|
|
|
|
@register_to_config
|
|
def __init__(
|
|
self,
|
|
sample_size: Optional[int] = None,
|
|
in_channels: int = 4,
|
|
out_channels: int = 4,
|
|
center_input_sample: bool = False,
|
|
flip_sin_to_cos: bool = True,
|
|
freq_shift: int = 0,
|
|
down_block_types: Tuple[str] = (
|
|
"CrossAttnDownBlockFlat",
|
|
"CrossAttnDownBlockFlat",
|
|
"CrossAttnDownBlockFlat",
|
|
"DownBlockFlat",
|
|
),
|
|
mid_block_type: str = "UNetMidBlockFlatCrossAttn",
|
|
up_block_types: Tuple[str] = (
|
|
"UpBlockFlat",
|
|
"CrossAttnUpBlockFlat",
|
|
"CrossAttnUpBlockFlat",
|
|
"CrossAttnUpBlockFlat",
|
|
),
|
|
only_cross_attention: Union[bool, Tuple[bool]] = False,
|
|
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: int = 32,
|
|
norm_eps: float = 1e-5,
|
|
cross_attention_dim: int = 1280,
|
|
attention_head_dim: Union[int, Tuple[int]] = 8,
|
|
dual_cross_attention: bool = False,
|
|
use_linear_projection: bool = False,
|
|
class_embed_type: Optional[str] = None,
|
|
num_class_embeds: Optional[int] = None,
|
|
upcast_attention: bool = False,
|
|
resnet_time_scale_shift: str = "default",
|
|
):
|
|
super().__init__()
|
|
|
|
self.sample_size = sample_size
|
|
time_embed_dim = block_out_channels[0] * 4
|
|
|
|
# input
|
|
self.conv_in = LinearMultiDim(in_channels, block_out_channels[0], kernel_size=3, padding=(1, 1))
|
|
|
|
# time
|
|
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)
|
|
|
|
# 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([])
|
|
|
|
if isinstance(only_cross_attention, bool):
|
|
only_cross_attention = [only_cross_attention] * len(down_block_types)
|
|
|
|
if isinstance(attention_head_dim, int):
|
|
attention_head_dim = (attention_head_dim,) * len(down_block_types)
|
|
|
|
# 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,
|
|
cross_attention_dim=cross_attention_dim,
|
|
attn_num_head_channels=attention_head_dim[i],
|
|
downsample_padding=downsample_padding,
|
|
dual_cross_attention=dual_cross_attention,
|
|
use_linear_projection=use_linear_projection,
|
|
only_cross_attention=only_cross_attention[i],
|
|
upcast_attention=upcast_attention,
|
|
resnet_time_scale_shift=resnet_time_scale_shift,
|
|
)
|
|
self.down_blocks.append(down_block)
|
|
|
|
# mid
|
|
if mid_block_type == "UNetMidBlockFlatCrossAttn":
|
|
self.mid_block = UNetMidBlockFlatCrossAttn(
|
|
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=resnet_time_scale_shift,
|
|
cross_attention_dim=cross_attention_dim,
|
|
attn_num_head_channels=attention_head_dim[-1],
|
|
resnet_groups=norm_num_groups,
|
|
dual_cross_attention=dual_cross_attention,
|
|
use_linear_projection=use_linear_projection,
|
|
upcast_attention=upcast_attention,
|
|
)
|
|
elif mid_block_type == "UNetMidBlockFlatSimpleCrossAttn":
|
|
self.mid_block = UNetMidBlockFlatSimpleCrossAttn(
|
|
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,
|
|
cross_attention_dim=cross_attention_dim,
|
|
attn_num_head_channels=attention_head_dim[-1],
|
|
resnet_groups=norm_num_groups,
|
|
resnet_time_scale_shift=resnet_time_scale_shift,
|
|
)
|
|
else:
|
|
raise ValueError(f"unknown mid_block_type : {mid_block_type}")
|
|
|
|
# count how many layers upsample the images
|
|
self.num_upsamplers = 0
|
|
|
|
# up
|
|
reversed_block_out_channels = list(reversed(block_out_channels))
|
|
reversed_attention_head_dim = list(reversed(attention_head_dim))
|
|
only_cross_attention = list(reversed(only_cross_attention))
|
|
output_channel = reversed_block_out_channels[0]
|
|
for i, up_block_type in enumerate(up_block_types):
|
|
is_final_block = i == len(block_out_channels) - 1
|
|
|
|
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)]
|
|
|
|
# add upsample block for all BUT final layer
|
|
if not is_final_block:
|
|
add_upsample = True
|
|
self.num_upsamplers += 1
|
|
else:
|
|
add_upsample = False
|
|
|
|
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=add_upsample,
|
|
resnet_eps=norm_eps,
|
|
resnet_act_fn=act_fn,
|
|
resnet_groups=norm_num_groups,
|
|
cross_attention_dim=cross_attention_dim,
|
|
attn_num_head_channels=reversed_attention_head_dim[i],
|
|
dual_cross_attention=dual_cross_attention,
|
|
use_linear_projection=use_linear_projection,
|
|
only_cross_attention=only_cross_attention[i],
|
|
upcast_attention=upcast_attention,
|
|
resnet_time_scale_shift=resnet_time_scale_shift,
|
|
)
|
|
self.up_blocks.append(up_block)
|
|
prev_output_channel = output_channel
|
|
|
|
# out
|
|
self.conv_norm_out = nn.GroupNorm(num_channels=block_out_channels[0], num_groups=norm_num_groups, eps=norm_eps)
|
|
self.conv_act = nn.SiLU()
|
|
self.conv_out = LinearMultiDim(block_out_channels[0], out_channels, kernel_size=3, padding=1)
|
|
|
|
def set_attention_slice(self, slice_size):
|
|
r"""
|
|
Enable sliced attention computation.
|
|
|
|
When this option is enabled, the attention module will split the input tensor in slices, to compute attention
|
|
in several steps. This is useful to save some memory in exchange for a small speed decrease.
|
|
|
|
Args:
|
|
slice_size (`str` or `int` or `list(int)`, *optional*, defaults to `"auto"`):
|
|
When `"auto"`, halves the input to the attention heads, so attention will be computed in two steps. If
|
|
`"max"`, maxium amount of memory will be saved by running only one slice at a time. If a number is
|
|
provided, uses as many slices as `attention_head_dim // slice_size`. In this case, `attention_head_dim`
|
|
must be a multiple of `slice_size`.
|
|
"""
|
|
sliceable_head_dims = []
|
|
|
|
def fn_recursive_retrieve_slicable_dims(module: torch.nn.Module):
|
|
if hasattr(module, "set_attention_slice"):
|
|
sliceable_head_dims.append(module.sliceable_head_dim)
|
|
|
|
for child in module.children():
|
|
fn_recursive_retrieve_slicable_dims(child)
|
|
|
|
# retrieve number of attention layers
|
|
for module in self.children():
|
|
fn_recursive_retrieve_slicable_dims(module)
|
|
|
|
num_slicable_layers = len(sliceable_head_dims)
|
|
|
|
if slice_size == "auto":
|
|
# half the attention head size is usually a good trade-off between
|
|
# speed and memory
|
|
slice_size = [dim // 2 for dim in sliceable_head_dims]
|
|
elif slice_size == "max":
|
|
# make smallest slice possible
|
|
slice_size = num_slicable_layers * [1]
|
|
|
|
slice_size = num_slicable_layers * [slice_size] if not isinstance(slice_size, list) else slice_size
|
|
|
|
if len(slice_size) != len(sliceable_head_dims):
|
|
raise ValueError(
|
|
f"You have provided {len(slice_size)}, but {self.config} has {len(sliceable_head_dims)} different"
|
|
f" attention layers. Make sure to match `len(slice_size)` to be {len(sliceable_head_dims)}."
|
|
)
|
|
|
|
for i in range(len(slice_size)):
|
|
size = slice_size[i]
|
|
dim = sliceable_head_dims[i]
|
|
if size is not None and size > dim:
|
|
raise ValueError(f"size {size} has to be smaller or equal to {dim}.")
|
|
|
|
# Recursively walk through all the children.
|
|
# Any children which exposes the set_attention_slice method
|
|
# gets the message
|
|
def fn_recursive_set_attention_slice(module: torch.nn.Module, slice_size: List[int]):
|
|
if hasattr(module, "set_attention_slice"):
|
|
module.set_attention_slice(slice_size.pop())
|
|
|
|
for child in module.children():
|
|
fn_recursive_set_attention_slice(child, slice_size)
|
|
|
|
reversed_slice_size = list(reversed(slice_size))
|
|
for module in self.children():
|
|
fn_recursive_set_attention_slice(module, reversed_slice_size)
|
|
|
|
def _set_gradient_checkpointing(self, module, value=False):
|
|
if isinstance(module, (CrossAttnDownBlockFlat, DownBlockFlat, CrossAttnUpBlockFlat, UpBlockFlat)):
|
|
module.gradient_checkpointing = value
|
|
|
|
def forward(
|
|
self,
|
|
sample: torch.FloatTensor,
|
|
timestep: Union[torch.Tensor, float, int],
|
|
encoder_hidden_states: torch.Tensor,
|
|
class_labels: Optional[torch.Tensor] = None,
|
|
attention_mask: Optional[torch.Tensor] = None,
|
|
return_dict: bool = True,
|
|
) -> Union[UNet2DConditionOutput, Tuple]:
|
|
r"""
|
|
Args:
|
|
sample (`torch.FloatTensor`): (batch, channel, height, width) noisy inputs tensor
|
|
timestep (`torch.FloatTensor` or `float` or `int`): (batch) timesteps
|
|
encoder_hidden_states (`torch.FloatTensor`): (batch, sequence_length, feature_dim) encoder hidden states
|
|
return_dict (`bool`, *optional*, defaults to `True`):
|
|
Whether or not to return a [`models.unet_2d_condition.UNet2DConditionOutput`] instead of a plain tuple.
|
|
|
|
Returns:
|
|
[`~models.unet_2d_condition.UNet2DConditionOutput`] or `tuple`:
|
|
[`~models.unet_2d_condition.UNet2DConditionOutput`] if `return_dict` is True, otherwise a `tuple`. When
|
|
returning a tuple, the first element is the sample tensor.
|
|
"""
|
|
# By default samples have to be AT least a multiple of the overall upsampling factor.
|
|
# The overall upsampling factor is equal to 2 ** (# num of upsampling layears).
|
|
# However, the upsampling interpolation output size can be forced to fit any upsampling size
|
|
# on the fly if necessary.
|
|
default_overall_up_factor = 2**self.num_upsamplers
|
|
|
|
# upsample size should be forwarded when sample is not a multiple of `default_overall_up_factor`
|
|
forward_upsample_size = False
|
|
upsample_size = None
|
|
|
|
if any(s % default_overall_up_factor != 0 for s in sample.shape[-2:]):
|
|
logger.info("Forward upsample size to force interpolation output size.")
|
|
forward_upsample_size = True
|
|
|
|
# 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)
|
|
|
|
# 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):
|
|
# 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=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 num_class_embeds > 0")
|
|
|
|
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
|
|
|
|
# 2. pre-process
|
|
sample = self.conv_in(sample)
|
|
|
|
# 3. down
|
|
down_block_res_samples = (sample,)
|
|
for downsample_block in self.down_blocks:
|
|
if hasattr(downsample_block, "has_cross_attention") and downsample_block.has_cross_attention:
|
|
sample, res_samples = downsample_block(
|
|
hidden_states=sample,
|
|
temb=emb,
|
|
encoder_hidden_states=encoder_hidden_states,
|
|
attention_mask=attention_mask,
|
|
)
|
|
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, encoder_hidden_states=encoder_hidden_states, attention_mask=attention_mask
|
|
)
|
|
|
|
# 5. up
|
|
for i, upsample_block in enumerate(self.up_blocks):
|
|
is_final_block = i == len(self.up_blocks) - 1
|
|
|
|
res_samples = down_block_res_samples[-len(upsample_block.resnets) :]
|
|
down_block_res_samples = down_block_res_samples[: -len(upsample_block.resnets)]
|
|
|
|
# if we have not reached the final block and need to forward the
|
|
# upsample size, we do it here
|
|
if not is_final_block and forward_upsample_size:
|
|
upsample_size = down_block_res_samples[-1].shape[2:]
|
|
|
|
if hasattr(upsample_block, "has_cross_attention") and upsample_block.has_cross_attention:
|
|
sample = upsample_block(
|
|
hidden_states=sample,
|
|
temb=emb,
|
|
res_hidden_states_tuple=res_samples,
|
|
encoder_hidden_states=encoder_hidden_states,
|
|
upsample_size=upsample_size,
|
|
attention_mask=attention_mask,
|
|
)
|
|
else:
|
|
sample = upsample_block(
|
|
hidden_states=sample, temb=emb, res_hidden_states_tuple=res_samples, upsample_size=upsample_size
|
|
)
|
|
# 6. post-process
|
|
sample = self.conv_norm_out(sample)
|
|
sample = self.conv_act(sample)
|
|
sample = self.conv_out(sample)
|
|
|
|
if not return_dict:
|
|
return (sample,)
|
|
|
|
return UNet2DConditionOutput(sample=sample)
|
|
|
|
|
|
class LinearMultiDim(nn.Linear):
|
|
def __init__(self, in_features, out_features=None, second_dim=4, *args, **kwargs):
|
|
in_features = [in_features, second_dim, 1] if isinstance(in_features, int) else list(in_features)
|
|
if out_features is None:
|
|
out_features = in_features
|
|
out_features = [out_features, second_dim, 1] if isinstance(out_features, int) else list(out_features)
|
|
self.in_features_multidim = in_features
|
|
self.out_features_multidim = out_features
|
|
super().__init__(np.array(in_features).prod(), np.array(out_features).prod())
|
|
|
|
def forward(self, input_tensor, *args, **kwargs):
|
|
shape = input_tensor.shape
|
|
n_dim = len(self.in_features_multidim)
|
|
input_tensor = input_tensor.reshape(*shape[0:-n_dim], self.in_features)
|
|
output_tensor = super().forward(input_tensor)
|
|
output_tensor = output_tensor.view(*shape[0:-n_dim], *self.out_features_multidim)
|
|
return output_tensor
|
|
|
|
|
|
class ResnetBlockFlat(nn.Module):
|
|
def __init__(
|
|
self,
|
|
*,
|
|
in_channels,
|
|
out_channels=None,
|
|
dropout=0.0,
|
|
temb_channels=512,
|
|
groups=32,
|
|
groups_out=None,
|
|
pre_norm=True,
|
|
eps=1e-6,
|
|
time_embedding_norm="default",
|
|
use_in_shortcut=None,
|
|
second_dim=4,
|
|
**kwargs,
|
|
):
|
|
super().__init__()
|
|
self.pre_norm = pre_norm
|
|
self.pre_norm = True
|
|
|
|
in_channels = [in_channels, second_dim, 1] if isinstance(in_channels, int) else list(in_channels)
|
|
self.in_channels_prod = np.array(in_channels).prod()
|
|
self.channels_multidim = in_channels
|
|
|
|
if out_channels is not None:
|
|
out_channels = [out_channels, second_dim, 1] if isinstance(out_channels, int) else list(out_channels)
|
|
out_channels_prod = np.array(out_channels).prod()
|
|
self.out_channels_multidim = out_channels
|
|
else:
|
|
out_channels_prod = self.in_channels_prod
|
|
self.out_channels_multidim = self.channels_multidim
|
|
self.time_embedding_norm = time_embedding_norm
|
|
|
|
if groups_out is None:
|
|
groups_out = groups
|
|
|
|
self.norm1 = torch.nn.GroupNorm(num_groups=groups, num_channels=self.in_channels_prod, eps=eps, affine=True)
|
|
self.conv1 = torch.nn.Conv2d(self.in_channels_prod, out_channels_prod, kernel_size=1, padding=0)
|
|
|
|
if temb_channels is not None:
|
|
self.time_emb_proj = torch.nn.Linear(temb_channels, out_channels_prod)
|
|
else:
|
|
self.time_emb_proj = None
|
|
|
|
self.norm2 = torch.nn.GroupNorm(num_groups=groups_out, num_channels=out_channels_prod, eps=eps, affine=True)
|
|
self.dropout = torch.nn.Dropout(dropout)
|
|
self.conv2 = torch.nn.Conv2d(out_channels_prod, out_channels_prod, kernel_size=1, padding=0)
|
|
|
|
self.nonlinearity = nn.SiLU()
|
|
|
|
self.use_in_shortcut = (
|
|
self.in_channels_prod != out_channels_prod if use_in_shortcut is None else use_in_shortcut
|
|
)
|
|
|
|
self.conv_shortcut = None
|
|
if self.use_in_shortcut:
|
|
self.conv_shortcut = torch.nn.Conv2d(
|
|
self.in_channels_prod, out_channels_prod, kernel_size=1, stride=1, padding=0
|
|
)
|
|
|
|
def forward(self, input_tensor, temb):
|
|
shape = input_tensor.shape
|
|
n_dim = len(self.channels_multidim)
|
|
input_tensor = input_tensor.reshape(*shape[0:-n_dim], self.in_channels_prod, 1, 1)
|
|
input_tensor = input_tensor.view(-1, self.in_channels_prod, 1, 1)
|
|
|
|
hidden_states = input_tensor
|
|
|
|
hidden_states = self.norm1(hidden_states)
|
|
hidden_states = self.nonlinearity(hidden_states)
|
|
hidden_states = self.conv1(hidden_states)
|
|
|
|
if temb is not None:
|
|
temb = self.time_emb_proj(self.nonlinearity(temb))[:, :, None, None]
|
|
hidden_states = hidden_states + temb
|
|
|
|
hidden_states = self.norm2(hidden_states)
|
|
hidden_states = self.nonlinearity(hidden_states)
|
|
|
|
hidden_states = self.dropout(hidden_states)
|
|
hidden_states = self.conv2(hidden_states)
|
|
|
|
if self.conv_shortcut is not None:
|
|
input_tensor = self.conv_shortcut(input_tensor)
|
|
|
|
output_tensor = input_tensor + hidden_states
|
|
|
|
output_tensor = output_tensor.view(*shape[0:-n_dim], -1)
|
|
output_tensor = output_tensor.view(*shape[0:-n_dim], *self.out_channels_multidim)
|
|
|
|
return output_tensor
|
|
|
|
|
|
# Copied from diffusers.models.unet_2d_blocks.DownBlock2D with DownBlock2D->DownBlockFlat, ResnetBlock2D->ResnetBlockFlat, Downsample2D->LinearMultiDim
|
|
class DownBlockFlat(nn.Module):
|
|
def __init__(
|
|
self,
|
|
in_channels: int,
|
|
out_channels: int,
|
|
temb_channels: int,
|
|
dropout: float = 0.0,
|
|
num_layers: int = 1,
|
|
resnet_eps: float = 1e-6,
|
|
resnet_time_scale_shift: str = "default",
|
|
resnet_act_fn: str = "swish",
|
|
resnet_groups: int = 32,
|
|
resnet_pre_norm: bool = True,
|
|
output_scale_factor=1.0,
|
|
add_downsample=True,
|
|
downsample_padding=1,
|
|
):
|
|
super().__init__()
|
|
resnets = []
|
|
|
|
for i in range(num_layers):
|
|
in_channels = in_channels if i == 0 else out_channels
|
|
resnets.append(
|
|
ResnetBlockFlat(
|
|
in_channels=in_channels,
|
|
out_channels=out_channels,
|
|
temb_channels=temb_channels,
|
|
eps=resnet_eps,
|
|
groups=resnet_groups,
|
|
dropout=dropout,
|
|
time_embedding_norm=resnet_time_scale_shift,
|
|
non_linearity=resnet_act_fn,
|
|
output_scale_factor=output_scale_factor,
|
|
pre_norm=resnet_pre_norm,
|
|
)
|
|
)
|
|
|
|
self.resnets = nn.ModuleList(resnets)
|
|
|
|
if add_downsample:
|
|
self.downsamplers = nn.ModuleList(
|
|
[
|
|
LinearMultiDim(
|
|
out_channels, use_conv=True, out_channels=out_channels, padding=downsample_padding, name="op"
|
|
)
|
|
]
|
|
)
|
|
else:
|
|
self.downsamplers = None
|
|
|
|
self.gradient_checkpointing = False
|
|
|
|
def forward(self, hidden_states, temb=None):
|
|
output_states = ()
|
|
|
|
for resnet in self.resnets:
|
|
if self.training and self.gradient_checkpointing:
|
|
|
|
def create_custom_forward(module):
|
|
def custom_forward(*inputs):
|
|
return module(*inputs)
|
|
|
|
return custom_forward
|
|
|
|
hidden_states = torch.utils.checkpoint.checkpoint(create_custom_forward(resnet), hidden_states, temb)
|
|
else:
|
|
hidden_states = resnet(hidden_states, temb)
|
|
|
|
output_states += (hidden_states,)
|
|
|
|
if self.downsamplers is not None:
|
|
for downsampler in self.downsamplers:
|
|
hidden_states = downsampler(hidden_states)
|
|
|
|
output_states += (hidden_states,)
|
|
|
|
return hidden_states, output_states
|
|
|
|
|
|
# Copied from diffusers.models.unet_2d_blocks.CrossAttnDownBlock2D with CrossAttnDownBlock2D->CrossAttnDownBlockFlat, ResnetBlock2D->ResnetBlockFlat, Downsample2D->LinearMultiDim
|
|
class CrossAttnDownBlockFlat(nn.Module):
|
|
def __init__(
|
|
self,
|
|
in_channels: int,
|
|
out_channels: int,
|
|
temb_channels: int,
|
|
dropout: float = 0.0,
|
|
num_layers: int = 1,
|
|
resnet_eps: float = 1e-6,
|
|
resnet_time_scale_shift: str = "default",
|
|
resnet_act_fn: str = "swish",
|
|
resnet_groups: int = 32,
|
|
resnet_pre_norm: bool = True,
|
|
attn_num_head_channels=1,
|
|
cross_attention_dim=1280,
|
|
output_scale_factor=1.0,
|
|
downsample_padding=1,
|
|
add_downsample=True,
|
|
dual_cross_attention=False,
|
|
use_linear_projection=False,
|
|
only_cross_attention=False,
|
|
upcast_attention=False,
|
|
):
|
|
super().__init__()
|
|
resnets = []
|
|
attentions = []
|
|
|
|
self.has_cross_attention = True
|
|
self.attn_num_head_channels = attn_num_head_channels
|
|
|
|
for i in range(num_layers):
|
|
in_channels = in_channels if i == 0 else out_channels
|
|
resnets.append(
|
|
ResnetBlockFlat(
|
|
in_channels=in_channels,
|
|
out_channels=out_channels,
|
|
temb_channels=temb_channels,
|
|
eps=resnet_eps,
|
|
groups=resnet_groups,
|
|
dropout=dropout,
|
|
time_embedding_norm=resnet_time_scale_shift,
|
|
non_linearity=resnet_act_fn,
|
|
output_scale_factor=output_scale_factor,
|
|
pre_norm=resnet_pre_norm,
|
|
)
|
|
)
|
|
if not dual_cross_attention:
|
|
attentions.append(
|
|
Transformer2DModel(
|
|
attn_num_head_channels,
|
|
out_channels // attn_num_head_channels,
|
|
in_channels=out_channels,
|
|
num_layers=1,
|
|
cross_attention_dim=cross_attention_dim,
|
|
norm_num_groups=resnet_groups,
|
|
use_linear_projection=use_linear_projection,
|
|
only_cross_attention=only_cross_attention,
|
|
upcast_attention=upcast_attention,
|
|
)
|
|
)
|
|
else:
|
|
attentions.append(
|
|
DualTransformer2DModel(
|
|
attn_num_head_channels,
|
|
out_channels // attn_num_head_channels,
|
|
in_channels=out_channels,
|
|
num_layers=1,
|
|
cross_attention_dim=cross_attention_dim,
|
|
norm_num_groups=resnet_groups,
|
|
)
|
|
)
|
|
self.attentions = nn.ModuleList(attentions)
|
|
self.resnets = nn.ModuleList(resnets)
|
|
|
|
if add_downsample:
|
|
self.downsamplers = nn.ModuleList(
|
|
[
|
|
LinearMultiDim(
|
|
out_channels, use_conv=True, out_channels=out_channels, padding=downsample_padding, name="op"
|
|
)
|
|
]
|
|
)
|
|
else:
|
|
self.downsamplers = None
|
|
|
|
self.gradient_checkpointing = False
|
|
|
|
def forward(self, hidden_states, temb=None, encoder_hidden_states=None, attention_mask=None):
|
|
# TODO(Patrick, William) - attention mask is not used
|
|
output_states = ()
|
|
|
|
for resnet, attn in zip(self.resnets, self.attentions):
|
|
if self.training and self.gradient_checkpointing:
|
|
|
|
def create_custom_forward(module, return_dict=None):
|
|
def custom_forward(*inputs):
|
|
if return_dict is not None:
|
|
return module(*inputs, return_dict=return_dict)
|
|
else:
|
|
return module(*inputs)
|
|
|
|
return custom_forward
|
|
|
|
hidden_states = torch.utils.checkpoint.checkpoint(create_custom_forward(resnet), hidden_states, temb)
|
|
hidden_states = torch.utils.checkpoint.checkpoint(
|
|
create_custom_forward(attn, return_dict=False),
|
|
hidden_states,
|
|
encoder_hidden_states,
|
|
)[0]
|
|
else:
|
|
hidden_states = resnet(hidden_states, temb)
|
|
hidden_states = attn(hidden_states, encoder_hidden_states=encoder_hidden_states).sample
|
|
|
|
output_states += (hidden_states,)
|
|
|
|
if self.downsamplers is not None:
|
|
for downsampler in self.downsamplers:
|
|
hidden_states = downsampler(hidden_states)
|
|
|
|
output_states += (hidden_states,)
|
|
|
|
return hidden_states, output_states
|
|
|
|
|
|
# Copied from diffusers.models.unet_2d_blocks.UpBlock2D with UpBlock2D->UpBlockFlat, ResnetBlock2D->ResnetBlockFlat, Upsample2D->LinearMultiDim
|
|
class UpBlockFlat(nn.Module):
|
|
def __init__(
|
|
self,
|
|
in_channels: int,
|
|
prev_output_channel: int,
|
|
out_channels: int,
|
|
temb_channels: int,
|
|
dropout: float = 0.0,
|
|
num_layers: int = 1,
|
|
resnet_eps: float = 1e-6,
|
|
resnet_time_scale_shift: str = "default",
|
|
resnet_act_fn: str = "swish",
|
|
resnet_groups: int = 32,
|
|
resnet_pre_norm: bool = True,
|
|
output_scale_factor=1.0,
|
|
add_upsample=True,
|
|
):
|
|
super().__init__()
|
|
resnets = []
|
|
|
|
for i in range(num_layers):
|
|
res_skip_channels = in_channels if (i == num_layers - 1) else out_channels
|
|
resnet_in_channels = prev_output_channel if i == 0 else out_channels
|
|
|
|
resnets.append(
|
|
ResnetBlockFlat(
|
|
in_channels=resnet_in_channels + res_skip_channels,
|
|
out_channels=out_channels,
|
|
temb_channels=temb_channels,
|
|
eps=resnet_eps,
|
|
groups=resnet_groups,
|
|
dropout=dropout,
|
|
time_embedding_norm=resnet_time_scale_shift,
|
|
non_linearity=resnet_act_fn,
|
|
output_scale_factor=output_scale_factor,
|
|
pre_norm=resnet_pre_norm,
|
|
)
|
|
)
|
|
|
|
self.resnets = nn.ModuleList(resnets)
|
|
|
|
if add_upsample:
|
|
self.upsamplers = nn.ModuleList([LinearMultiDim(out_channels, use_conv=True, out_channels=out_channels)])
|
|
else:
|
|
self.upsamplers = None
|
|
|
|
self.gradient_checkpointing = False
|
|
|
|
def forward(self, hidden_states, res_hidden_states_tuple, temb=None, upsample_size=None):
|
|
for resnet in self.resnets:
|
|
# pop res hidden states
|
|
res_hidden_states = res_hidden_states_tuple[-1]
|
|
res_hidden_states_tuple = res_hidden_states_tuple[:-1]
|
|
hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)
|
|
|
|
if self.training and self.gradient_checkpointing:
|
|
|
|
def create_custom_forward(module):
|
|
def custom_forward(*inputs):
|
|
return module(*inputs)
|
|
|
|
return custom_forward
|
|
|
|
hidden_states = torch.utils.checkpoint.checkpoint(create_custom_forward(resnet), hidden_states, temb)
|
|
else:
|
|
hidden_states = resnet(hidden_states, temb)
|
|
|
|
if self.upsamplers is not None:
|
|
for upsampler in self.upsamplers:
|
|
hidden_states = upsampler(hidden_states, upsample_size)
|
|
|
|
return hidden_states
|
|
|
|
|
|
# Copied from diffusers.models.unet_2d_blocks.CrossAttnUpBlock2D with CrossAttnUpBlock2D->CrossAttnUpBlockFlat, ResnetBlock2D->ResnetBlockFlat, Upsample2D->LinearMultiDim
|
|
class CrossAttnUpBlockFlat(nn.Module):
|
|
def __init__(
|
|
self,
|
|
in_channels: int,
|
|
out_channels: int,
|
|
prev_output_channel: int,
|
|
temb_channels: int,
|
|
dropout: float = 0.0,
|
|
num_layers: int = 1,
|
|
resnet_eps: float = 1e-6,
|
|
resnet_time_scale_shift: str = "default",
|
|
resnet_act_fn: str = "swish",
|
|
resnet_groups: int = 32,
|
|
resnet_pre_norm: bool = True,
|
|
attn_num_head_channels=1,
|
|
cross_attention_dim=1280,
|
|
output_scale_factor=1.0,
|
|
add_upsample=True,
|
|
dual_cross_attention=False,
|
|
use_linear_projection=False,
|
|
only_cross_attention=False,
|
|
upcast_attention=False,
|
|
):
|
|
super().__init__()
|
|
resnets = []
|
|
attentions = []
|
|
|
|
self.has_cross_attention = True
|
|
self.attn_num_head_channels = attn_num_head_channels
|
|
|
|
for i in range(num_layers):
|
|
res_skip_channels = in_channels if (i == num_layers - 1) else out_channels
|
|
resnet_in_channels = prev_output_channel if i == 0 else out_channels
|
|
|
|
resnets.append(
|
|
ResnetBlockFlat(
|
|
in_channels=resnet_in_channels + res_skip_channels,
|
|
out_channels=out_channels,
|
|
temb_channels=temb_channels,
|
|
eps=resnet_eps,
|
|
groups=resnet_groups,
|
|
dropout=dropout,
|
|
time_embedding_norm=resnet_time_scale_shift,
|
|
non_linearity=resnet_act_fn,
|
|
output_scale_factor=output_scale_factor,
|
|
pre_norm=resnet_pre_norm,
|
|
)
|
|
)
|
|
if not dual_cross_attention:
|
|
attentions.append(
|
|
Transformer2DModel(
|
|
attn_num_head_channels,
|
|
out_channels // attn_num_head_channels,
|
|
in_channels=out_channels,
|
|
num_layers=1,
|
|
cross_attention_dim=cross_attention_dim,
|
|
norm_num_groups=resnet_groups,
|
|
use_linear_projection=use_linear_projection,
|
|
only_cross_attention=only_cross_attention,
|
|
upcast_attention=upcast_attention,
|
|
)
|
|
)
|
|
else:
|
|
attentions.append(
|
|
DualTransformer2DModel(
|
|
attn_num_head_channels,
|
|
out_channels // attn_num_head_channels,
|
|
in_channels=out_channels,
|
|
num_layers=1,
|
|
cross_attention_dim=cross_attention_dim,
|
|
norm_num_groups=resnet_groups,
|
|
)
|
|
)
|
|
self.attentions = nn.ModuleList(attentions)
|
|
self.resnets = nn.ModuleList(resnets)
|
|
|
|
if add_upsample:
|
|
self.upsamplers = nn.ModuleList([LinearMultiDim(out_channels, use_conv=True, out_channels=out_channels)])
|
|
else:
|
|
self.upsamplers = None
|
|
|
|
self.gradient_checkpointing = False
|
|
|
|
def forward(
|
|
self,
|
|
hidden_states,
|
|
res_hidden_states_tuple,
|
|
temb=None,
|
|
encoder_hidden_states=None,
|
|
upsample_size=None,
|
|
attention_mask=None,
|
|
):
|
|
# TODO(Patrick, William) - attention mask is not used
|
|
for resnet, attn in zip(self.resnets, self.attentions):
|
|
# pop res hidden states
|
|
res_hidden_states = res_hidden_states_tuple[-1]
|
|
res_hidden_states_tuple = res_hidden_states_tuple[:-1]
|
|
hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)
|
|
|
|
if self.training and self.gradient_checkpointing:
|
|
|
|
def create_custom_forward(module, return_dict=None):
|
|
def custom_forward(*inputs):
|
|
if return_dict is not None:
|
|
return module(*inputs, return_dict=return_dict)
|
|
else:
|
|
return module(*inputs)
|
|
|
|
return custom_forward
|
|
|
|
hidden_states = torch.utils.checkpoint.checkpoint(create_custom_forward(resnet), hidden_states, temb)
|
|
hidden_states = torch.utils.checkpoint.checkpoint(
|
|
create_custom_forward(attn, return_dict=False),
|
|
hidden_states,
|
|
encoder_hidden_states,
|
|
)[0]
|
|
else:
|
|
hidden_states = resnet(hidden_states, temb)
|
|
hidden_states = attn(hidden_states, encoder_hidden_states=encoder_hidden_states).sample
|
|
|
|
if self.upsamplers is not None:
|
|
for upsampler in self.upsamplers:
|
|
hidden_states = upsampler(hidden_states, upsample_size)
|
|
|
|
return hidden_states
|
|
|
|
|
|
# Copied from diffusers.models.unet_2d_blocks.UNetMidBlock2DCrossAttn with UNetMidBlock2DCrossAttn->UNetMidBlockFlatCrossAttn, ResnetBlock2D->ResnetBlockFlat
|
|
class UNetMidBlockFlatCrossAttn(nn.Module):
|
|
def __init__(
|
|
self,
|
|
in_channels: int,
|
|
temb_channels: int,
|
|
dropout: float = 0.0,
|
|
num_layers: int = 1,
|
|
resnet_eps: float = 1e-6,
|
|
resnet_time_scale_shift: str = "default",
|
|
resnet_act_fn: str = "swish",
|
|
resnet_groups: int = 32,
|
|
resnet_pre_norm: bool = True,
|
|
attn_num_head_channels=1,
|
|
output_scale_factor=1.0,
|
|
cross_attention_dim=1280,
|
|
dual_cross_attention=False,
|
|
use_linear_projection=False,
|
|
upcast_attention=False,
|
|
):
|
|
super().__init__()
|
|
|
|
self.has_cross_attention = True
|
|
self.attn_num_head_channels = attn_num_head_channels
|
|
resnet_groups = resnet_groups if resnet_groups is not None else min(in_channels // 4, 32)
|
|
|
|
# there is always at least one resnet
|
|
resnets = [
|
|
ResnetBlockFlat(
|
|
in_channels=in_channels,
|
|
out_channels=in_channels,
|
|
temb_channels=temb_channels,
|
|
eps=resnet_eps,
|
|
groups=resnet_groups,
|
|
dropout=dropout,
|
|
time_embedding_norm=resnet_time_scale_shift,
|
|
non_linearity=resnet_act_fn,
|
|
output_scale_factor=output_scale_factor,
|
|
pre_norm=resnet_pre_norm,
|
|
)
|
|
]
|
|
attentions = []
|
|
|
|
for _ in range(num_layers):
|
|
if not dual_cross_attention:
|
|
attentions.append(
|
|
Transformer2DModel(
|
|
attn_num_head_channels,
|
|
in_channels // attn_num_head_channels,
|
|
in_channels=in_channels,
|
|
num_layers=1,
|
|
cross_attention_dim=cross_attention_dim,
|
|
norm_num_groups=resnet_groups,
|
|
use_linear_projection=use_linear_projection,
|
|
upcast_attention=upcast_attention,
|
|
)
|
|
)
|
|
else:
|
|
attentions.append(
|
|
DualTransformer2DModel(
|
|
attn_num_head_channels,
|
|
in_channels // attn_num_head_channels,
|
|
in_channels=in_channels,
|
|
num_layers=1,
|
|
cross_attention_dim=cross_attention_dim,
|
|
norm_num_groups=resnet_groups,
|
|
)
|
|
)
|
|
resnets.append(
|
|
ResnetBlockFlat(
|
|
in_channels=in_channels,
|
|
out_channels=in_channels,
|
|
temb_channels=temb_channels,
|
|
eps=resnet_eps,
|
|
groups=resnet_groups,
|
|
dropout=dropout,
|
|
time_embedding_norm=resnet_time_scale_shift,
|
|
non_linearity=resnet_act_fn,
|
|
output_scale_factor=output_scale_factor,
|
|
pre_norm=resnet_pre_norm,
|
|
)
|
|
)
|
|
|
|
self.attentions = nn.ModuleList(attentions)
|
|
self.resnets = nn.ModuleList(resnets)
|
|
|
|
def forward(self, hidden_states, temb=None, encoder_hidden_states=None, attention_mask=None):
|
|
# TODO(Patrick, William) - attention_mask is currently not used. Implement once used
|
|
hidden_states = self.resnets[0](hidden_states, temb)
|
|
for attn, resnet in zip(self.attentions, self.resnets[1:]):
|
|
hidden_states = attn(hidden_states, encoder_hidden_states=encoder_hidden_states).sample
|
|
hidden_states = resnet(hidden_states, temb)
|
|
|
|
return hidden_states
|
|
|
|
|
|
# Copied from diffusers.models.unet_2d_blocks.UNetMidBlock2DSimpleCrossAttn with UNetMidBlock2DSimpleCrossAttn->UNetMidBlockFlatCrossAttn, ResnetBlock2D->ResnetBlockFlat
|
|
class UnCLIPUNetMidBlockFlatCrossAttn(nn.Module):
|
|
def __init__(
|
|
self,
|
|
in_channels: int,
|
|
temb_channels: int,
|
|
dropout: float = 0.0,
|
|
num_layers: int = 1,
|
|
resnet_eps: float = 1e-6,
|
|
resnet_time_scale_shift: str = "default",
|
|
resnet_act_fn: str = "swish",
|
|
resnet_groups: int = 32,
|
|
resnet_pre_norm: bool = True,
|
|
attn_num_head_channels=1,
|
|
output_scale_factor=1.0,
|
|
cross_attention_dim=1280,
|
|
):
|
|
super().__init__()
|
|
|
|
self.has_cross_attention = True
|
|
|
|
self.attn_num_head_channels = attn_num_head_channels
|
|
resnet_groups = resnet_groups if resnet_groups is not None else min(in_channels // 4, 32)
|
|
|
|
self.num_heads = in_channels // self.attn_num_head_channels
|
|
|
|
# there is always at least one resnet
|
|
resnets = [
|
|
ResnetBlockFlat(
|
|
in_channels=in_channels,
|
|
out_channels=in_channels,
|
|
temb_channels=temb_channels,
|
|
eps=resnet_eps,
|
|
groups=resnet_groups,
|
|
dropout=dropout,
|
|
time_embedding_norm=resnet_time_scale_shift,
|
|
non_linearity=resnet_act_fn,
|
|
output_scale_factor=output_scale_factor,
|
|
pre_norm=resnet_pre_norm,
|
|
)
|
|
]
|
|
attentions = []
|
|
|
|
for _ in range(num_layers):
|
|
attentions.append(
|
|
CrossAttention(
|
|
query_dim=in_channels,
|
|
cross_attention_dim=in_channels,
|
|
heads=self.num_heads,
|
|
dim_head=attn_num_head_channels,
|
|
added_kv_proj_dim=cross_attention_dim,
|
|
norm_num_groups=resnet_groups,
|
|
bias=True,
|
|
upcast_softmax=True,
|
|
)
|
|
)
|
|
resnets.append(
|
|
ResnetBlockFlat(
|
|
in_channels=in_channels,
|
|
out_channels=in_channels,
|
|
temb_channels=temb_channels,
|
|
eps=resnet_eps,
|
|
groups=resnet_groups,
|
|
dropout=dropout,
|
|
time_embedding_norm=resnet_time_scale_shift,
|
|
non_linearity=resnet_act_fn,
|
|
output_scale_factor=output_scale_factor,
|
|
pre_norm=resnet_pre_norm,
|
|
)
|
|
)
|
|
|
|
self.attentions = nn.ModuleList(attentions)
|
|
self.resnets = nn.ModuleList(resnets)
|
|
|
|
def forward(self, hidden_states, temb=None, encoder_hidden_states=None, attention_mask=None):
|
|
hidden_states = self.resnets[0](hidden_states, temb)
|
|
for attn, resnet in zip(self.attentions, self.resnets[1:]):
|
|
# attn
|
|
residual = hidden_states
|
|
hidden_states = hidden_states.view(hidden_states.shape[0], hidden_states.shape[1], -1).transpose(1, 2)
|
|
hidden_states = attn(
|
|
hidden_states,
|
|
encoder_hidden_states=encoder_hidden_states.transpose(1, 2),
|
|
attention_mask=attention_mask,
|
|
)
|
|
hidden_states = hidden_states.transpose(-1, -2).reshape(residual.shape)
|
|
hidden_states = hidden_states + residual
|
|
|
|
# resnet
|
|
hidden_states = resnet(hidden_states, temb)
|
|
|
|
return hidden_states
|