from typing import Any, Dict, Optional, Tuple
import numpy as np
import torch
import torch.nn.functional as F
from diffusers.models.activations import get_activation
from diffusers.models.attention_processor import (Attention,
AttnAddedKVProcessor,
AttnAddedKVProcessor2_0)
from diffusers.models.normalization import AdaGroupNorm
from diffusers.models.resnet import (Downsample2D, FirDownsample2D,
FirUpsample2D, KDownsample2D, KUpsample2D,
ResnetBlock2D, Upsample2D)
from diffusers.models.transformer_2d import Transformer2DModel
from diffusers.utils import is_torch_version, logging
from einops import rearrange
from torch import nn
from .motion_module import get_motion_module
TORCH_CLEAR_CACHE = False
def clear_cache():
if TORCH_CLEAR_CACHE:
import inspect
callerframerecord = inspect.stack()[1]
frame = callerframerecord[0]
info = inspect.getframeinfo(frame)
import time
import GPUtil
torch.cuda.synchronize()
logger.info(f"{info.filename}/{info.lineno}/before clear cache")
GPUtil.showUtilization()
torch.cuda.empty_cache()
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,
transformer_layers_per_block=1,
num_attention_heads=None,
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",
attention_type="default",
resnet_skip_time_act=False,
resnet_out_scale_factor=1.0,
cross_attention_norm=None,
attention_head_dim=None,
downsample_type=None,
use_motion_module=None,
motion_module_type=None,
motion_module_kwargs=None,
):
# If attn head dim is not defined, we default it to the number of heads
if attention_head_dim is None:
logger.warn(
f"It is recommended to provide `attention_head_dim` when calling `get_down_block`. Defaulting `attention_head_dim` to {num_attention_heads}."
)
attention_head_dim = num_attention_heads
down_block_type = down_block_type[7:] if down_block_type.startswith("UNetRes") else down_block_type
if down_block_type == "DownBlock3D":
return DownBlock3D(
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,
use_motion_module=use_motion_module,
motion_module_type=motion_module_type,
motion_module_kwargs=motion_module_kwargs,
)
elif down_block_type == "CrossAttnDownBlock3D":
if cross_attention_dim is None:
raise ValueError("cross_attention_dim must be specified for CrossAttnDownBlock2D")
return CrossAttnDownBlock3D(
num_layers=num_layers,
transformer_layers_per_block=transformer_layers_per_block,
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,
num_attention_heads=num_attention_heads,
dual_cross_attention=dual_cross_attention,
use_linear_projection=use_linear_projection,
only_cross_attention=only_cross_attention,
upcast_attention=upcast_attention,
resnet_time_scale_shift=resnet_time_scale_shift,
attention_type=attention_type,
use_motion_module=use_motion_module,
motion_module_type=motion_module_type,
motion_module_kwargs=motion_module_kwargs,
)
raise ValueError(f"{down_block_type} does not exist.")
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,
transformer_layers_per_block=1,
num_attention_heads=None,
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",
attention_type="default",
resnet_skip_time_act=False,
resnet_out_scale_factor=1.0,
cross_attention_norm=None,
attention_head_dim=None,
upsample_type=None,
use_motion_module=None,
motion_module_type=None,
motion_module_kwargs=None,
):
# If attn head dim is not defined, we default it to the number of heads
if attention_head_dim is None:
logger.warn(
f"It is recommended to provide `attention_head_dim` when calling `get_up_block`. Defaulting `attention_head_dim` to {num_attention_heads}."
)
attention_head_dim = num_attention_heads
up_block_type = up_block_type[7:] if up_block_type.startswith("UNetRes") else up_block_type
if up_block_type == "UpBlock3D":
return UpBlock3D(
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,
use_motion_module=use_motion_module,
motion_module_type=motion_module_type,
motion_module_kwargs=motion_module_kwargs,
)
elif up_block_type == "CrossAttnUpBlock3D":
if cross_attention_dim is None:
raise ValueError("cross_attention_dim must be specified for CrossAttnUpBlock2D")
return CrossAttnUpBlock3D(
num_layers=num_layers,
transformer_layers_per_block=transformer_layers_per_block,
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,
num_attention_heads=num_attention_heads,
dual_cross_attention=dual_cross_attention,
use_linear_projection=use_linear_projection,
only_cross_attention=only_cross_attention,
upcast_attention=upcast_attention,
resnet_time_scale_shift=resnet_time_scale_shift,
attention_type=attention_type,
use_motion_module=use_motion_module,
motion_module_type=motion_module_type,
motion_module_kwargs=motion_module_kwargs,
)
raise ValueError(f"{up_block_type} does not exist.")
class UNetMidBlock3DCrossAttn(nn.Module):
def __init__(
self,
in_channels: int,
temb_channels: int,
dropout: float = 0.0,
num_layers: int = 1,
transformer_layers_per_block: 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,
num_attention_heads=1,
output_scale_factor=1.0,
cross_attention_dim=1280,
dual_cross_attention=False,
use_linear_projection=False,
upcast_attention=False,
attention_type="default",
use_motion_module=None,
motion_module_type=None,
motion_module_kwargs=None,
):
super().__init__()
self.has_cross_attention = True
self.num_attention_heads = num_attention_heads
resnet_groups = resnet_groups if resnet_groups is not None else min(in_channels // 4, 32)
# there is always at least one resnet
resnets = [
ResnetBlock2D(
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 = []
motion_modules = []
for _ in range(num_layers):
if not dual_cross_attention:
attentions.append(
Transformer2DModel(
num_attention_heads,
in_channels // num_attention_heads,
in_channels=in_channels,
num_layers=transformer_layers_per_block,
cross_attention_dim=cross_attention_dim,
norm_num_groups=resnet_groups,
use_linear_projection=use_linear_projection,
upcast_attention=upcast_attention,
attention_type=attention_type,
)
)
else:
attentions.append(
DualTransformer2DModel(
num_attention_heads,
in_channels // num_attention_heads,
in_channels=in_channels,
num_layers=1,
cross_attention_dim=cross_attention_dim,
norm_num_groups=resnet_groups,
)
)
motion_modules.append(
get_motion_module(
in_channels=in_channels,
motion_module_type=motion_module_type,
motion_module_kwargs=motion_module_kwargs,
) if use_motion_module else None
)
resnets.append(
ResnetBlock2D(
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)
self.motion_modules = nn.ModuleList(motion_modules)
self.gradient_checkpointing = False
def forward(
self,
hidden_states: torch.FloatTensor,
temb: Optional[torch.FloatTensor] = None,
encoder_hidden_states: Optional[torch.FloatTensor] = None,
attention_mask: Optional[torch.FloatTensor] = None,
cross_attention_kwargs: Optional[Dict[str, Any]] = None,
encoder_attention_mask: Optional[torch.FloatTensor] = None,
) -> torch.FloatTensor:
video_length = hidden_states.shape[2]
hidden_states = rearrange(hidden_states, "b c f h w -> (b f) c h w")
hidden_states = self.resnets[0](hidden_states, temb)
hidden_states = rearrange(hidden_states, "(b f) c h w -> b c f h w", f=video_length)
for attn, motion_module, resnet in zip(self.attentions, self.motion_modules, self.resnets[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 = rearrange(hidden_states, "b c f h w -> (b f) c h w")
ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
hidden_states = attn(
hidden_states,
encoder_hidden_states=encoder_hidden_states,
cross_attention_kwargs=cross_attention_kwargs,
attention_mask=attention_mask,
encoder_attention_mask=encoder_attention_mask,
return_dict=False,
)[0]
hidden_states = rearrange(hidden_states, "(b f) c h w -> b c f h w", f=video_length)
hidden_states = torch.utils.checkpoint.checkpoint(
create_custom_forward(motion_module), hidden_states, temb,
encoder_hidden_states) if motion_module is not None else hidden_states
hidden_states = rearrange(hidden_states, "b c f h w -> (b f) c h w")
hidden_states = torch.utils.checkpoint.checkpoint(
create_custom_forward(resnet),
hidden_states,
temb,
**ckpt_kwargs,
)
hidden_states = rearrange(hidden_states, "(b f) c h w -> b c f h w", f=video_length)
else:
hidden_states = rearrange(hidden_states, "b c f h w -> (b f) c h w")
hidden_states = attn(
hidden_states,
encoder_hidden_states=encoder_hidden_states,
cross_attention_kwargs=cross_attention_kwargs,
attention_mask=attention_mask,
encoder_attention_mask=encoder_attention_mask,
return_dict=False,
)[0]
hidden_states = rearrange(hidden_states, "(b f) c h w -> b c f h w", f=video_length)
hidden_states = motion_module(hidden_states, temb, encoder_hidden_states=encoder_hidden_states) if motion_module is not None else hidden_states
hidden_states = rearrange(hidden_states, "b c f h w -> (b f) c h w")
hidden_states = resnet(hidden_states, temb)
hidden_states = rearrange(hidden_states, "(b f) c h w -> b c f h w", f=video_length)
return hidden_states
class DownBlock3D(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,
use_motion_module=None,
motion_module_type=None,
motion_module_kwargs=None,
):
super().__init__()
resnets = []
motion_modules = []
for i in range(num_layers):
in_channels = in_channels if i == 0 else out_channels
resnets.append(
ResnetBlock2D(
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,
)
)
motion_modules.append(
get_motion_module(
in_channels=out_channels,
motion_module_type=motion_module_type,
motion_module_kwargs=motion_module_kwargs,
) if use_motion_module else None
)
self.resnets = nn.ModuleList(resnets)
self.motion_modules = nn.ModuleList(motion_modules)
if add_downsample:
self.downsamplers = nn.ModuleList(
[
Downsample2D(
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):
output_states = ()
for resnet, motion_module in zip(self.resnets, self.motion_modules):
if self.training and self.gradient_checkpointing:
def create_custom_forward(module):
def custom_forward(*inputs):
return module(*inputs)
return custom_forward
video_length = hidden_states.shape[2]
hidden_states = rearrange(hidden_states, "b c f h w -> (b f) c h w")
if is_torch_version(">=", "1.11.0"):
hidden_states = torch.utils.checkpoint.checkpoint(
create_custom_forward(resnet), hidden_states, temb, use_reentrant=False
)
else:
hidden_states = torch.utils.checkpoint.checkpoint(
create_custom_forward(resnet), hidden_states, temb
)
hidden_states = rearrange(hidden_states, "(b f) c h w -> b c f h w", f=video_length)
hidden_states = torch.utils.checkpoint.checkpoint(
create_custom_forward(motion_module), hidden_states, temb, encoder_hidden_states, use_reentrant=False) if motion_module is not None else hidden_states
else:
video_length = hidden_states.shape[2]
hidden_states = rearrange(hidden_states, "b c f h w -> (b f) c h w")
hidden_states = resnet(hidden_states, temb)
hidden_states = rearrange(hidden_states, "(b f) c h w -> b c f h w", f=video_length)
hidden_states = motion_module(hidden_states, temb, encoder_hidden_states) if motion_module is not None else hidden_states
output_states = output_states + (hidden_states,)
if self.downsamplers is not None:
for downsampler in self.downsamplers:
hidden_states = rearrange(hidden_states, "b c f h w -> (b f) c h w")
hidden_states = downsampler(hidden_states)
hidden_states = rearrange(hidden_states, "(b f) c h w -> b c f h w", f=video_length)
output_states = output_states + (hidden_states,)
return hidden_states, output_states
class UpBlock3D(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,
use_motion_module=None,
motion_module_type=None,
motion_module_kwargs=None,
):
super().__init__()
resnets = []
motion_modules = []
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(
ResnetBlock2D(
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,
)
)
motion_modules.append(
get_motion_module(
in_channels=out_channels,
motion_module_type=motion_module_type,
motion_module_kwargs=motion_module_kwargs,
) if use_motion_module else None
)
self.resnets = nn.ModuleList(resnets)
self.motion_modules = nn.ModuleList(motion_modules)
if add_upsample:
self.upsamplers = nn.ModuleList([Upsample2D(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, encoder_hidden_states=None, temb=None, upsample_size=None):
for (resnet, motion_module) in zip(self.resnets, self.motion_modules):
# 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
video_length = hidden_states.shape[2]
hidden_states = rearrange(hidden_states, "b c f h w -> (b f) c h w")
if is_torch_version(">=", "1.11.0"):
hidden_states = torch.utils.checkpoint.checkpoint(
create_custom_forward(resnet), hidden_states, temb, use_reentrant=False
)
else:
hidden_states = torch.utils.checkpoint.checkpoint(
create_custom_forward(resnet), hidden_states, temb
)
hidden_states = rearrange(hidden_states, "(b f) c h w -> b c f h w", f=video_length)
hidden_states = torch.utils.checkpoint.checkpoint(
create_custom_forward(motion_module), hidden_states,
temb, encoder_hidden_states, use_reentrant=False) if motion_module is not None else hidden_states
else:
video_length = hidden_states.shape[2]
hidden_states = rearrange(hidden_states, "b c f h w -> (b f) c h w")
hidden_states = resnet(hidden_states, temb)
hidden_states = rearrange(hidden_states, "(b f) c h w -> b c f h w", f=video_length)
hidden_states = motion_module(hidden_states, temb, encoder_hidden_states) if motion_module is not None else hidden_states
if self.upsamplers is not None:
for upsampler in self.upsamplers:
hidden_states = rearrange(hidden_states, "b c f h w -> (b f) c h w")
hidden_states = upsampler(hidden_states, upsample_size)
hidden_states = rearrange(hidden_states, "(b f) c h w -> b c f h w", f=video_length)
return hidden_states
class CrossAttnDownBlock3D(nn.Module):
def __init__(
self,
in_channels: int,
out_channels: int,
temb_channels: int,
dropout: float = 0.0,
num_layers: int = 1,
transformer_layers_per_block: 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,
num_attention_heads=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,
attention_type="default",
use_motion_module=None,
motion_module_type=None,
motion_module_kwargs=None,
):
super().__init__()
resnets = []
attentions = []
motion_modules = []
self.has_cross_attention = True
self.num_attention_heads = num_attention_heads
for i in range(num_layers):
in_channels = in_channels if i == 0 else out_channels
resnets.append(
ResnetBlock2D(
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(
num_attention_heads,
out_channels // num_attention_heads,
in_channels=out_channels,
num_layers=transformer_layers_per_block,
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,
attention_type=attention_type,
)
)
else:
attentions.append(
DualTransformer2DModel(
num_attention_heads,
out_channels // num_attention_heads,
in_channels=out_channels,
num_layers=1,
cross_attention_dim=cross_attention_dim,
norm_num_groups=resnet_groups,
)
)
motion_modules.append(
get_motion_module(
in_channels=out_channels,
motion_module_type=motion_module_type,
motion_module_kwargs=motion_module_kwargs,
) if use_motion_module else None
)
self.attentions = nn.ModuleList(attentions)
self.resnets = nn.ModuleList(resnets)
self.motion_modules = nn.ModuleList(motion_modules)
if add_downsample:
self.downsamplers = nn.ModuleList(
[
Downsample2D(
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: torch.FloatTensor,
temb: Optional[torch.FloatTensor] = None,
encoder_hidden_states: Optional[torch.FloatTensor] = None,
attention_mask: Optional[torch.FloatTensor] = None,
cross_attention_kwargs: Optional[Dict[str, Any]] = None,
encoder_attention_mask: Optional[torch.FloatTensor] = None,
additional_residuals=None,
):
output_states = ()
blocks = list(zip(self.resnets, self.attentions, self.motion_modules))
for i, (resnet, attn, motion_module) in enumerate(blocks):
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
ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
video_length = hidden_states.shape[2]
hidden_states = rearrange(hidden_states, "b c f h w -> (b f) c h w")
hidden_states = torch.utils.checkpoint.checkpoint(
create_custom_forward(resnet),
hidden_states,
temb,
**ckpt_kwargs,
)
hidden_states = attn(
hidden_states,
encoder_hidden_states=encoder_hidden_states,
cross_attention_kwargs=cross_attention_kwargs,
attention_mask=attention_mask,
encoder_attention_mask=encoder_attention_mask,
return_dict=False,
)[0]
hidden_states = rearrange(hidden_states, "(b f) c h w -> b c f h w", f=video_length)
hidden_states = torch.utils.checkpoint.checkpoint(
create_custom_forward(motion_module),
hidden_states, temb,
encoder_hidden_states, use_reentrant=False) if motion_module is not None else hidden_states
else:
video_length = hidden_states.shape[2]
hidden_states = rearrange(hidden_states, "b c f h w -> (b f) c h w")
hidden_states = resnet(hidden_states, temb)
hidden_states = attn(
hidden_states,
encoder_hidden_states=encoder_hidden_states,
cross_attention_kwargs=cross_attention_kwargs,
attention_mask=attention_mask,
encoder_attention_mask=encoder_attention_mask,
return_dict=False,
)[0]
hidden_states = rearrange(hidden_states, "(b f) c h w -> b c f h w", f=video_length)
hidden_states = motion_module(hidden_states, temb, encoder_hidden_states=encoder_hidden_states) if motion_module is not None else hidden_states
# apply additional residuals to the output of the last pair of resnet and attention blocks
if i == len(blocks) - 1 and additional_residuals is not None:
hidden_states = hidden_states + additional_residuals
output_states = output_states + (hidden_states,)
if self.downsamplers is not None:
for downsampler in self.downsamplers:
hidden_states = rearrange(hidden_states, "b c f h w -> (b f) c h w")
hidden_states = downsampler(hidden_states)
hidden_states = rearrange(hidden_states, "(b f) c h w -> b c f h w", f=video_length)
output_states = output_states + (hidden_states,)
return hidden_states, output_states
class CrossAttnUpBlock3D(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,
transformer_layers_per_block: 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,
num_attention_heads=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,
attention_type="default",
use_motion_module=None,
motion_module_type=None,
motion_module_kwargs=None,
):
super().__init__()
resnets = []
attentions = []
motion_modules = []
self.has_cross_attention = True
self.num_attention_heads = num_attention_heads
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(
ResnetBlock2D(
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(
num_attention_heads,
out_channels // num_attention_heads,
in_channels=out_channels,
num_layers=transformer_layers_per_block,
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,
attention_type=attention_type,
)
)
else:
attentions.append(
DualTransformer2DModel(
num_attention_heads,
out_channels // num_attention_heads,
in_channels=out_channels,
num_layers=1,
cross_attention_dim=cross_attention_dim,
norm_num_groups=resnet_groups,
)
)
motion_modules.append(
get_motion_module(
in_channels=out_channels,
motion_module_type=motion_module_type,
motion_module_kwargs=motion_module_kwargs,
) if use_motion_module else None
)
self.attentions = nn.ModuleList(attentions)
self.resnets = nn.ModuleList(resnets)
self.motion_modules = nn.ModuleList(motion_modules)
if add_upsample:
self.upsamplers = nn.ModuleList([Upsample2D(out_channels, use_conv=True, out_channels=out_channels)])
else:
self.upsamplers = None
self.gradient_checkpointing = False
def forward(
self,
hidden_states: torch.FloatTensor,
res_hidden_states_tuple: Tuple[torch.FloatTensor, ...],
temb: Optional[torch.FloatTensor] = None,
encoder_hidden_states: Optional[torch.FloatTensor] = None,
cross_attention_kwargs: Optional[Dict[str, Any]] = None,
upsample_size: Optional[int] = None,
attention_mask: Optional[torch.FloatTensor] = None,
encoder_attention_mask: Optional[torch.FloatTensor] = None,
):
for resnet, attn, motion_module in zip(self.resnets, self.attentions, self.motion_modules):
# 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
ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
video_length = hidden_states.shape[2]
hidden_states = rearrange(hidden_states, "b c f h w -> (b f) c h w")
hidden_states = torch.utils.checkpoint.checkpoint(
create_custom_forward(resnet),
hidden_states,
temb,
**ckpt_kwargs,
)
hidden_states = attn(
hidden_states,
encoder_hidden_states=encoder_hidden_states,
cross_attention_kwargs=cross_attention_kwargs,
attention_mask=attention_mask,
encoder_attention_mask=encoder_attention_mask,
return_dict=False,
)[0]
hidden_states = rearrange(hidden_states, "(b f) c h w -> b c f h w", f=video_length)
hidden_states = torch.utils.checkpoint.checkpoint(
create_custom_forward(motion_module),
hidden_states, temb, encoder_hidden_states,
use_reentrant=False) if motion_module is not None else hidden_states
else:
video_length = hidden_states.shape[2]
hidden_states = rearrange(hidden_states, "b c f h w -> (b f) c h w")
hidden_states = resnet(hidden_states, temb)
hidden_states = attn(
hidden_states,
encoder_hidden_states=encoder_hidden_states,
cross_attention_kwargs=cross_attention_kwargs,
attention_mask=attention_mask,
encoder_attention_mask=encoder_attention_mask,
return_dict=False,
)[0]
hidden_states = rearrange(hidden_states, "(b f) c h w -> b c f h w", f=video_length)
hidden_states = motion_module(hidden_states, temb, encoder_hidden_states=encoder_hidden_states) if motion_module is not None else hidden_states
if self.upsamplers is not None:
for upsampler in self.upsamplers:
hidden_states = rearrange(hidden_states, "b c f h w -> (b f) c h w")
hidden_states = upsampler(hidden_states, upsample_size)
hidden_states = rearrange(hidden_states, "(b f) c h w -> b c f h w", f=video_length)
return hidden_states