531 lines
22 KiB
Python
531 lines
22 KiB
Python
from .resnet_block3d import ResnetBlock3D
|
|
from .downsample3d import Downsample3D
|
|
from .upsample3d import Upsample3D
|
|
from .convs import InflatedConv3d
|
|
from .transformer3d import Transformer3DModel
|
|
|
|
|
|
import torch as th
|
|
import torch.nn as nn
|
|
|
|
from comfy.ldm.modules.diffusionmodules.util import (
|
|
zero_module,
|
|
timestep_embedding,
|
|
)
|
|
from comfy.ldm.util import exists
|
|
from comfy.ldm.modules.diffusionmodules.openaimodel import TimestepBlock, apply_control
|
|
import comfy.ops
|
|
ops = comfy.ops.disable_weight_init
|
|
|
|
|
|
def forward_timestep_embed(ts, x, emb, context=None, transformer_options={}, output_shape=None, time_context=None, num_video_frames=None, image_only_indicator=None):
|
|
for layer in ts:
|
|
if isinstance(layer, TimestepBlock) or isinstance(layer, ResnetBlock3D):
|
|
x = layer(x, emb)
|
|
elif isinstance(layer, Transformer3DModel):
|
|
x = layer(x, context, transformer_options)
|
|
if "transformer_index" in transformer_options:
|
|
transformer_options["transformer_index"] += 1
|
|
elif isinstance(layer, Upsample3D):
|
|
x = layer(x, output_shape=output_shape)
|
|
else:
|
|
x = layer(x)
|
|
return x
|
|
|
|
|
|
class TimestepEmbedSequential(nn.Sequential, TimestepBlock):
|
|
"""
|
|
A sequential module that passes timestep embeddings to the children that
|
|
support it as an extra input.
|
|
"""
|
|
|
|
def forward(self, *args, **kwargs):
|
|
return forward_timestep_embed(self, *args, **kwargs)
|
|
|
|
|
|
class UNetModel(nn.Module):
|
|
"""
|
|
The full UNet model with attention and timestep embedding.
|
|
:param in_channels: channels in the input Tensor.
|
|
:param model_channels: base channel count for the model.
|
|
:param out_channels: channels in the output Tensor.
|
|
:param num_res_blocks: number of residual blocks per downsample.
|
|
:param dropout: the dropout probability.
|
|
:param channel_mult: channel multiplier for each level of the UNet.
|
|
:param conv_resample: if True, use learned convolutions for upsampling and
|
|
downsampling.
|
|
:param dims: determines if the signal is 1D, 2D, or 3D.
|
|
:param num_classes: if specified (as an int), then this model will be
|
|
class-conditional with `num_classes` classes.
|
|
:param use_checkpoint: use gradient checkpointing to reduce memory usage.
|
|
:param num_heads: the number of attention heads in each attention layer.
|
|
:param num_heads_channels: if specified, ignore num_heads and instead use
|
|
a fixed channel width per attention head.
|
|
:param num_heads_upsample: works with num_heads to set a different number
|
|
of heads for upsampling. Deprecated.
|
|
:param use_scale_shift_norm: use a FiLM-like conditioning mechanism.
|
|
:param resblock_updown: use residual blocks for up/downsampling.
|
|
:param use_new_attention_order: use a different attention pattern for potentially
|
|
increased efficiency.
|
|
"""
|
|
|
|
def __init__(
|
|
self,
|
|
image_size,
|
|
in_channels,
|
|
model_channels,
|
|
out_channels,
|
|
num_res_blocks,
|
|
dropout=0,
|
|
channel_mult=(1, 2, 4, 8),
|
|
conv_resample=True,
|
|
dims=2,
|
|
num_classes=None,
|
|
use_checkpoint=False,
|
|
dtype=th.float32,
|
|
num_heads=-1,
|
|
num_head_channels=-1,
|
|
num_heads_upsample=-1,
|
|
use_scale_shift_norm=False,
|
|
resblock_updown=False,
|
|
use_new_attention_order=False,
|
|
use_spatial_transformer=False, # custom transformer support
|
|
transformer_depth=1, # custom transformer support
|
|
context_dim=None, # custom transformer support
|
|
# custom support for prediction of discrete ids into codebook of first stage vq model
|
|
n_embed=None,
|
|
legacy=True,
|
|
disable_self_attentions=None,
|
|
num_attention_blocks=None,
|
|
disable_middle_self_attn=False,
|
|
use_linear_in_transformer=False,
|
|
adm_in_channels=None,
|
|
transformer_depth_middle=None,
|
|
transformer_depth_output=None,
|
|
use_temporal_resblock=False,
|
|
use_temporal_attention=False,
|
|
time_context_dim=None,
|
|
extra_ff_mix_layer=False,
|
|
use_spatial_context=False,
|
|
merge_strategy=None,
|
|
merge_factor=0.0,
|
|
video_kernel_size=None,
|
|
disable_temporal_crossattention=False,
|
|
max_ddpm_temb_period=10000,
|
|
device=None,
|
|
operations=ops,
|
|
):
|
|
super().__init__()
|
|
|
|
if context_dim is not None:
|
|
assert use_spatial_transformer, 'You forgot to use the spatial transformer for your cross-attention conditioning...'
|
|
|
|
if num_heads_upsample == -1:
|
|
num_heads_upsample = num_heads
|
|
|
|
if num_heads == -1:
|
|
assert num_head_channels != -1, 'Either num_heads or num_head_channels has to be set'
|
|
|
|
if num_head_channels == -1:
|
|
assert num_heads != -1, 'Either num_heads or num_head_channels has to be set'
|
|
|
|
self.in_channels = in_channels
|
|
self.model_channels = model_channels
|
|
self.out_channels = out_channels
|
|
|
|
if isinstance(num_res_blocks, int):
|
|
self.num_res_blocks = len(channel_mult) * [num_res_blocks]
|
|
else:
|
|
if len(num_res_blocks) != len(channel_mult):
|
|
raise ValueError("provide num_res_blocks either as an int (globally constant) or "
|
|
"as a list/tuple (per-level) with the same length as channel_mult")
|
|
self.num_res_blocks = num_res_blocks
|
|
|
|
if disable_self_attentions is not None:
|
|
# should be a list of booleans, indicating whether to disable self-attention in TransformerBlocks or not
|
|
assert len(disable_self_attentions) == len(channel_mult)
|
|
if num_attention_blocks is not None:
|
|
assert len(num_attention_blocks) == len(self.num_res_blocks)
|
|
|
|
transformer_depth = transformer_depth[:]
|
|
transformer_depth_output = transformer_depth_output[:]
|
|
|
|
self.dropout = dropout
|
|
self.channel_mult = channel_mult
|
|
self.conv_resample = conv_resample
|
|
self.num_classes = num_classes
|
|
self.use_checkpoint = use_checkpoint
|
|
self.dtype = dtype
|
|
self.num_heads = num_heads
|
|
self.num_head_channels = num_head_channels
|
|
self.num_heads_upsample = num_heads_upsample
|
|
self.use_temporal_resblocks = use_temporal_resblock
|
|
self.predict_codebook_ids = n_embed is not None
|
|
|
|
self.default_num_video_frames = None
|
|
self.default_image_only_indicator = None
|
|
|
|
time_embed_dim = model_channels * 4
|
|
self.time_embed = nn.Sequential(
|
|
operations.Linear(model_channels, time_embed_dim,
|
|
dtype=self.dtype, device=device),
|
|
nn.SiLU(),
|
|
operations.Linear(time_embed_dim, time_embed_dim,
|
|
dtype=self.dtype, device=device),
|
|
)
|
|
|
|
if self.num_classes is not None:
|
|
if isinstance(self.num_classes, int):
|
|
self.label_emb = nn.Embedding(
|
|
num_classes, time_embed_dim, dtype=self.dtype, device=device)
|
|
elif self.num_classes == "continuous":
|
|
print("setting up linear c_adm embedding layer")
|
|
self.label_emb = nn.Linear(1, time_embed_dim)
|
|
elif self.num_classes == "sequential":
|
|
assert adm_in_channels is not None
|
|
self.label_emb = nn.Sequential(
|
|
nn.Sequential(
|
|
operations.Linear(
|
|
adm_in_channels, time_embed_dim, dtype=self.dtype, device=device),
|
|
nn.SiLU(),
|
|
operations.Linear(
|
|
time_embed_dim, time_embed_dim, dtype=self.dtype, device=device),
|
|
)
|
|
)
|
|
else:
|
|
raise ValueError()
|
|
|
|
self.input_blocks = nn.ModuleList(
|
|
[
|
|
TimestepEmbedSequential(
|
|
InflatedConv3d(
|
|
in_channels, model_channels, kernel_size=3, padding=(1, 1))
|
|
).half()
|
|
]
|
|
)
|
|
self._feature_size = model_channels
|
|
input_block_chans = [model_channels]
|
|
ch = model_channels
|
|
ds = 1
|
|
|
|
def get_attention_layer(
|
|
ch,
|
|
num_heads,
|
|
dim_head,
|
|
depth=1,
|
|
context_dim=None,
|
|
use_checkpoint=False,
|
|
disable_self_attn=False,
|
|
):
|
|
|
|
return Transformer3DModel(
|
|
ch, num_heads, dim_head, depth=depth, context_dim=context_dim,
|
|
disable_self_attn=disable_self_attn, use_linear=use_linear_in_transformer,
|
|
use_checkpoint=use_checkpoint, dtype=self.dtype, device=device, operations=operations
|
|
)
|
|
|
|
def get_resblock(
|
|
ch,
|
|
time_embed_dim,
|
|
dropout,
|
|
out_channels,
|
|
dims,
|
|
use_checkpoint,
|
|
use_scale_shift_norm,
|
|
down=False,
|
|
up=False,
|
|
dtype=None,
|
|
device=None,
|
|
operations=ops
|
|
):
|
|
return ResnetBlock3D(
|
|
channels=ch,
|
|
emb_channels=time_embed_dim,
|
|
dropout=dropout,
|
|
out_channels=out_channels,
|
|
use_checkpoint=use_checkpoint,
|
|
dims=dims,
|
|
use_scale_shift_norm=use_scale_shift_norm,
|
|
down=down,
|
|
up=up,
|
|
dtype=dtype,
|
|
device=device,
|
|
operations=operations
|
|
)
|
|
|
|
for level, mult in enumerate(channel_mult):
|
|
for nr in range(self.num_res_blocks[level]):
|
|
layers = [
|
|
get_resblock(
|
|
ch=ch,
|
|
time_embed_dim=time_embed_dim,
|
|
dropout=dropout,
|
|
out_channels=mult * model_channels,
|
|
dims=dims,
|
|
use_checkpoint=use_checkpoint,
|
|
use_scale_shift_norm=use_scale_shift_norm,
|
|
dtype=self.dtype,
|
|
device=device,
|
|
operations=operations,
|
|
)
|
|
]
|
|
ch = mult * model_channels
|
|
num_transformers = transformer_depth.pop(0)
|
|
if num_transformers > 0:
|
|
if num_head_channels == -1:
|
|
dim_head = ch // num_heads
|
|
else:
|
|
num_heads = ch // num_head_channels
|
|
dim_head = num_head_channels
|
|
if legacy:
|
|
# num_heads = 1
|
|
dim_head = ch // num_heads if use_spatial_transformer else num_head_channels
|
|
if exists(disable_self_attentions):
|
|
disabled_sa = disable_self_attentions[level]
|
|
else:
|
|
disabled_sa = False
|
|
|
|
if not exists(num_attention_blocks) or nr < num_attention_blocks[level]:
|
|
layers.append(get_attention_layer(
|
|
ch, num_heads, dim_head, depth=num_transformers, context_dim=context_dim,
|
|
disable_self_attn=disabled_sa, use_checkpoint=use_checkpoint)
|
|
)
|
|
self.input_blocks.append(TimestepEmbedSequential(*layers))
|
|
self._feature_size += ch
|
|
input_block_chans.append(ch)
|
|
if level != len(channel_mult) - 1:
|
|
out_ch = ch
|
|
self.input_blocks.append(
|
|
TimestepEmbedSequential(
|
|
get_resblock(
|
|
ch=ch,
|
|
time_embed_dim=time_embed_dim,
|
|
dropout=dropout,
|
|
out_channels=out_ch,
|
|
dims=dims,
|
|
use_checkpoint=use_checkpoint,
|
|
use_scale_shift_norm=use_scale_shift_norm,
|
|
down=True,
|
|
dtype=self.dtype,
|
|
device=device,
|
|
operations=operations
|
|
)
|
|
if resblock_updown
|
|
else Downsample3D(
|
|
ch, conv_resample, dims=dims, out_channels=out_ch, dtype=self.dtype, device=device, operations=operations
|
|
)
|
|
)
|
|
)
|
|
ch = out_ch
|
|
input_block_chans.append(ch)
|
|
ds *= 2
|
|
self._feature_size += ch
|
|
|
|
if num_head_channels == -1:
|
|
dim_head = ch // num_heads
|
|
else:
|
|
num_heads = ch // num_head_channels
|
|
dim_head = num_head_channels
|
|
if legacy:
|
|
# num_heads = 1
|
|
dim_head = ch // num_heads if use_spatial_transformer else num_head_channels
|
|
mid_block = [
|
|
get_resblock(
|
|
ch=ch,
|
|
time_embed_dim=time_embed_dim,
|
|
dropout=dropout,
|
|
out_channels=None,
|
|
dims=dims,
|
|
use_checkpoint=use_checkpoint,
|
|
use_scale_shift_norm=use_scale_shift_norm,
|
|
dtype=self.dtype,
|
|
device=device,
|
|
operations=operations
|
|
)]
|
|
if transformer_depth_middle >= 0:
|
|
mid_block += [get_attention_layer( # always uses a self-attn
|
|
ch, num_heads, dim_head, depth=transformer_depth_middle, context_dim=context_dim,
|
|
disable_self_attn=disable_middle_self_attn, use_checkpoint=use_checkpoint
|
|
),
|
|
get_resblock(
|
|
ch=ch,
|
|
time_embed_dim=time_embed_dim,
|
|
dropout=dropout,
|
|
out_channels=None,
|
|
dims=dims,
|
|
use_checkpoint=use_checkpoint,
|
|
use_scale_shift_norm=use_scale_shift_norm,
|
|
dtype=self.dtype,
|
|
device=device,
|
|
operations=operations
|
|
)]
|
|
self.middle_block = TimestepEmbedSequential(*mid_block)
|
|
self._feature_size += ch
|
|
|
|
self.num_upsamplers = 0
|
|
self.output_blocks = nn.ModuleList([])
|
|
for level, mult in list(enumerate(channel_mult))[::-1]:
|
|
for i in range(self.num_res_blocks[level] + 1):
|
|
ich = input_block_chans.pop()
|
|
layers = [
|
|
get_resblock(
|
|
ch=ch + ich,
|
|
time_embed_dim=time_embed_dim,
|
|
dropout=dropout,
|
|
out_channels=model_channels * mult,
|
|
dims=dims,
|
|
use_checkpoint=use_checkpoint,
|
|
use_scale_shift_norm=use_scale_shift_norm,
|
|
dtype=self.dtype,
|
|
device=device,
|
|
operations=operations
|
|
)
|
|
]
|
|
ch = model_channels * mult
|
|
num_transformers = transformer_depth_output.pop()
|
|
if num_transformers > 0:
|
|
if num_head_channels == -1:
|
|
dim_head = ch // num_heads
|
|
else:
|
|
num_heads = ch // num_head_channels
|
|
dim_head = num_head_channels
|
|
if legacy:
|
|
# num_heads = 1
|
|
dim_head = ch // num_heads if use_spatial_transformer else num_head_channels
|
|
if exists(disable_self_attentions):
|
|
disabled_sa = disable_self_attentions[level]
|
|
else:
|
|
disabled_sa = False
|
|
|
|
if not exists(num_attention_blocks) or i < num_attention_blocks[level]:
|
|
layers.append(
|
|
get_attention_layer(
|
|
ch, num_heads, dim_head, depth=num_transformers, context_dim=context_dim,
|
|
disable_self_attn=disabled_sa, use_checkpoint=use_checkpoint
|
|
)
|
|
)
|
|
if level and i == self.num_res_blocks[level]:
|
|
out_ch = ch
|
|
layers.append(
|
|
get_resblock(
|
|
ch=ch,
|
|
time_embed_dim=time_embed_dim,
|
|
dropout=dropout,
|
|
out_channels=out_ch,
|
|
dims=dims,
|
|
use_checkpoint=use_checkpoint,
|
|
use_scale_shift_norm=use_scale_shift_norm,
|
|
up=True,
|
|
dtype=self.dtype,
|
|
device=device,
|
|
operations=operations
|
|
)
|
|
if resblock_updown
|
|
else Upsample3D(ch, conv_resample, dims=dims, out_channels=out_ch, dtype=self.dtype, device=device, operations=operations)
|
|
)
|
|
if not resblock_updown:
|
|
self.num_upsamplers += 1
|
|
ds //= 2
|
|
self.output_blocks.append(TimestepEmbedSequential(*layers))
|
|
self._feature_size += ch
|
|
|
|
self.out = nn.Sequential(
|
|
operations.GroupNorm(32, ch, dtype=self.dtype, device=device),
|
|
nn.SiLU(),
|
|
zero_module(InflatedConv3d(model_channels,
|
|
out_channels, 3, padding=1, dtype=self.dtype, device=device).half()),
|
|
)
|
|
if self.predict_codebook_ids:
|
|
self.id_predictor = nn.Sequential(
|
|
operations.GroupNorm(32, ch, dtype=self.dtype, device=device),
|
|
InflatedConv3d(model_channels, n_embed,
|
|
1, dtype=self.dtype, device=device).half(),
|
|
# nn.LogSoftmax(dim=1) # change to cross_entropy and produce non-normalized logits
|
|
)
|
|
|
|
def forward(self, x, timesteps=None, context=None, y=None, control=None, transformer_options={}, **kwargs):
|
|
"""
|
|
Apply the model to an input batch.
|
|
:param x: an [N x C x ...] Tensor of inputs.
|
|
:param timesteps: a 1-D batch of timesteps.
|
|
:param context: conditioning plugged in via crossattn
|
|
:param y: an [N] Tensor of labels, if class-conditional.
|
|
:return: an [N x C x ...] Tensor of outputs.
|
|
"""
|
|
default_overall_up_factor = 2**self.num_upsamplers
|
|
forward_upsample_size = False
|
|
upsample_size = None
|
|
|
|
if any(s % default_overall_up_factor != 0 for s in x.shape[-2:]):
|
|
print('Upsampling True')
|
|
forward_upsample_size = True
|
|
|
|
transformer_options["original_shape"] = list(x.shape)
|
|
transformer_options["transformer_index"] = 0
|
|
transformer_patches = transformer_options.get("patches", {})
|
|
|
|
num_video_frames = kwargs.get(
|
|
"num_video_frames", self.default_num_video_frames)
|
|
image_only_indicator = kwargs.get(
|
|
"image_only_indicator", self.default_image_only_indicator)
|
|
time_context = kwargs.get("time_context", None)
|
|
|
|
assert (y is not None) == (
|
|
self.num_classes is not None
|
|
), "must specify y if and only if the model is class-conditional"
|
|
hs = []
|
|
t_emb = timestep_embedding(
|
|
timesteps, self.model_channels, repeat_only=False).to(x.dtype)
|
|
emb = self.time_embed(t_emb)
|
|
|
|
if self.num_classes is not None:
|
|
assert y.shape[0] == x.shape[0]
|
|
emb = emb + self.label_emb(y)
|
|
|
|
h = x
|
|
for id, module in enumerate(self.input_blocks):
|
|
transformer_options["block"] = ("input", id)
|
|
h = forward_timestep_embed(module, h, emb, context, transformer_options, time_context=time_context,
|
|
num_video_frames=num_video_frames, image_only_indicator=image_only_indicator)
|
|
h = apply_control(h, control, 'input')
|
|
if "input_block_patch" in transformer_patches:
|
|
patch = transformer_patches["input_block_patch"]
|
|
for p in patch:
|
|
h = p(h, transformer_options)
|
|
|
|
hs.append(h)
|
|
if "input_block_patch_after_skip" in transformer_patches:
|
|
patch = transformer_patches["input_block_patch_after_skip"]
|
|
for p in patch:
|
|
h = p(h, transformer_options)
|
|
|
|
transformer_options["block"] = ("middle", 0)
|
|
h = forward_timestep_embed(self.middle_block, h, emb, context, transformer_options, time_context=time_context,
|
|
num_video_frames=num_video_frames, image_only_indicator=image_only_indicator)
|
|
h = apply_control(h, control, 'middle')
|
|
|
|
for id, module in enumerate(self.output_blocks):
|
|
transformer_options["block"] = ("output", id)
|
|
hsp = hs.pop()
|
|
hsp = apply_control(hsp, control, 'output')
|
|
|
|
if "output_block_patch" in transformer_patches:
|
|
patch = transformer_patches["output_block_patch"]
|
|
for p in patch:
|
|
h, hsp = p(h, hsp, transformer_options)
|
|
|
|
h = th.cat([h, hsp], dim=1)
|
|
del hsp
|
|
if len(hs) > 0 and forward_upsample_size:
|
|
output_shape = hs[-1].shape[2:]
|
|
else:
|
|
output_shape = None
|
|
|
|
h = forward_timestep_embed(module, h, emb, context, transformer_options, output_shape,
|
|
time_context=time_context, num_video_frames=num_video_frames, image_only_indicator=image_only_indicator)
|
|
h = h.type(x.dtype)
|
|
if self.predict_codebook_ids:
|
|
return self.id_predictor(h)
|
|
else:
|
|
return self.out(h)
|