404 lines
17 KiB
Python
404 lines
17 KiB
Python
# taken from: https://github.com/lllyasviel/ControlNet
|
|
# and modified with Flatten modules
|
|
# Mostly an experiment
|
|
|
|
from einops import rearrange
|
|
import torch
|
|
import torch as th
|
|
import torch.nn as nn
|
|
|
|
import comfy.ops
|
|
from comfy.ldm.modules.diffusionmodules.util import (
|
|
zero_module,
|
|
timestep_embedding,
|
|
)
|
|
from comfy.ldm.util import exists
|
|
|
|
from .unet import TimestepEmbedSequential
|
|
from .convs import InflatedConv3d
|
|
from .transformer3d import Transformer3DModel
|
|
from .downsample3d import Downsample3D
|
|
from .resnet_block3d import ResnetBlock3D
|
|
|
|
|
|
class FlattenControlNet(nn.Module):
|
|
def __init__(
|
|
self,
|
|
image_size,
|
|
in_channels,
|
|
model_channels,
|
|
hint_channels,
|
|
num_res_blocks,
|
|
dropout=0,
|
|
channel_mult=(1, 2, 4, 8),
|
|
conv_resample=True,
|
|
dims=2,
|
|
num_classes=None,
|
|
use_checkpoint=False,
|
|
dtype=torch.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,
|
|
device=None,
|
|
operations=comfy.ops.disable_weight_init,
|
|
**kwargs,
|
|
):
|
|
super().__init__()
|
|
assert use_spatial_transformer == True, "use_spatial_transformer has to be true"
|
|
if use_spatial_transformer:
|
|
assert context_dim is not None, 'Fool!! You forgot to include the dimension of your cross-attention conditioning...'
|
|
|
|
if context_dim is not None:
|
|
assert use_spatial_transformer, 'Fool!! You forgot to use the spatial transformer for your cross-attention conditioning...'
|
|
# from omegaconf.listconfig import ListConfig
|
|
# if type(context_dim) == ListConfig:
|
|
# context_dim = list(context_dim)
|
|
|
|
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.dims = dims
|
|
self.image_size = image_size
|
|
self.in_channels = in_channels
|
|
self.model_channels = model_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)
|
|
assert all(map(lambda i: self.num_res_blocks[i] >= num_attention_blocks[i], range(
|
|
len(num_attention_blocks))))
|
|
|
|
transformer_depth = transformer_depth[:]
|
|
|
|
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.predict_codebook_ids = n_embed is not 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)
|
|
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,
|
|
3, padding=1).half()
|
|
)
|
|
]
|
|
)
|
|
self.zero_convs = nn.ModuleList([self.make_zero_conv(
|
|
model_channels, operations=operations, dtype=self.dtype, device=device)])
|
|
|
|
self.input_hint_block = TimestepEmbedSequential(
|
|
InflatedConv3d(hint_channels, 16, 3, padding=1).half(),
|
|
nn.SiLU(),
|
|
InflatedConv3d(16, 16, 3, padding=1).half(),
|
|
nn.SiLU(),
|
|
InflatedConv3d(16, 32, 3, padding=1, stride=2).half(),
|
|
nn.SiLU(),
|
|
InflatedConv3d(32, 32, 3, padding=1).half(),
|
|
nn.SiLU(),
|
|
InflatedConv3d(32, 96, 3, padding=1, stride=2).half(),
|
|
nn.SiLU(),
|
|
InflatedConv3d(96, 96, 3, padding=1).half(),
|
|
nn.SiLU(),
|
|
InflatedConv3d(96, 256, 3, padding=1, stride=2).half(),
|
|
nn.SiLU(),
|
|
InflatedConv3d(256, model_channels, 3, padding=1).half(),
|
|
)
|
|
|
|
self._feature_size = model_channels
|
|
input_block_chans = [model_channels]
|
|
ch = model_channels
|
|
ds = 1
|
|
for level, mult in enumerate(channel_mult):
|
|
for nr in range(self.num_res_blocks[level]):
|
|
layers = [
|
|
ResnetBlock3D(
|
|
ch,
|
|
time_embed_dim,
|
|
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(
|
|
Transformer3DModel(
|
|
ch, num_heads, dim_head, depth=num_transformers, context_dim=context_dim,
|
|
disable_self_attn=disabled_sa, use_linear=use_linear_in_transformer,
|
|
use_checkpoint=use_checkpoint, dtype=self.dtype, device=device, operations=operations
|
|
)
|
|
)
|
|
self.input_blocks.append(TimestepEmbedSequential(*layers))
|
|
self.zero_convs.append(self.make_zero_conv(
|
|
ch, operations=operations, dtype=self.dtype, device=device))
|
|
self._feature_size += ch
|
|
input_block_chans.append(ch)
|
|
if level != len(channel_mult) - 1:
|
|
out_ch = ch
|
|
self.input_blocks.append(
|
|
TimestepEmbedSequential(
|
|
ResnetBlock3D(
|
|
ch,
|
|
time_embed_dim,
|
|
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)
|
|
self.zero_convs.append(self.make_zero_conv(
|
|
ch, operations=operations, dtype=self.dtype, device=device))
|
|
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 = [
|
|
ResnetBlock3D(
|
|
ch,
|
|
time_embed_dim,
|
|
dropout,
|
|
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 += [Transformer3DModel( # 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_linear=use_linear_in_transformer,
|
|
use_checkpoint=use_checkpoint, dtype=self.dtype, device=device, operations=operations
|
|
),
|
|
ResnetBlock3D(
|
|
ch,
|
|
time_embed_dim,
|
|
dropout,
|
|
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.middle_block_out = self.make_zero_conv(
|
|
ch, operations=operations, dtype=self.dtype, device=device)
|
|
self._feature_size += ch
|
|
self.injection = {}
|
|
self.timestep = -1
|
|
self.step = 0
|
|
|
|
def make_zero_conv(self, channels, operations=None, dtype=None, device=None):
|
|
return TimestepEmbedSequential(InflatedConv3d(channels, channels, 1, padding=0).half())
|
|
|
|
def forward(self, x, hint, timesteps, context, y=None, model_options={}, batched_number=1, **kwargs):
|
|
transformer_options = model_options['transformer_options']
|
|
original_shape = transformer_options['flatten']['original_shape']
|
|
video_length = original_shape[0]
|
|
|
|
stage = transformer_options['flatten'].get('stage', None)
|
|
injection_steps = transformer_options['flatten'].get(
|
|
'injection_steps', 0)
|
|
step = None
|
|
if stage == 'inversion' or stage is None:
|
|
self.injection = {
|
|
'input': {},
|
|
'middle': {}
|
|
}
|
|
for i in range(len(self.input_blocks)):
|
|
self.injection['input'][i] = {}
|
|
elif stage == 'sampling':
|
|
if int(timesteps[0]) > self.timestep:
|
|
self.timestep = int(timesteps[0])
|
|
self.step = 0
|
|
elif int(timesteps[0]) < self.timestep:
|
|
self.timestep = int(timesteps[0])
|
|
self.step += 1
|
|
step = self.step
|
|
|
|
incoming_shape = x.shape
|
|
if len(incoming_shape) == 4:
|
|
# TODO get video_length in here
|
|
x = rearrange(x, '(b f) c h w -> b c f h w', f=video_length)
|
|
if len(hint.shape) == 4:
|
|
hint = rearrange(hint, '(b f) c h w -> b c f h w', f=video_length)
|
|
|
|
cond_length = x.shape[0]
|
|
t_emb = timestep_embedding(
|
|
timesteps, self.model_channels, repeat_only=False).to(x.dtype)
|
|
emb = self.time_embed(t_emb)
|
|
|
|
guided_hint = self.input_hint_block(hint, emb, context)
|
|
|
|
outs = []
|
|
|
|
hs = []
|
|
if self.num_classes is not None:
|
|
assert y.shape[0] == x.shape[0]
|
|
emb = emb + self.label_emb(y)
|
|
|
|
h = x
|
|
i = 0
|
|
for k, attn in enumerate(self.named_modules()):
|
|
if hasattr(attn, 'inject_q') and hasattr(attn, 'inject_k'):
|
|
attn.inject_q = None
|
|
attn.inject_k = None
|
|
|
|
for module, zero_conv in zip(self.input_blocks, self.zero_convs):
|
|
if stage == 'sampling':
|
|
for k, attn in enumerate(self.named_modules()):
|
|
if hasattr(attn, 'inject_q') and hasattr(attn, 'inject_k'):
|
|
if step < injection_steps:
|
|
attn.inject_q = torch.cat(
|
|
[self.injection['input'][i][k]['q'][step]]*cond_length).to('cuda')
|
|
attn.inject_k = torch.cat(
|
|
[self.injection['input'][i][k]['k'][step]]*cond_length).to('cuda')
|
|
if guided_hint is not None:
|
|
h = module(h, emb, context,
|
|
transformer_options=transformer_options)
|
|
h += guided_hint
|
|
guided_hint = None
|
|
else:
|
|
h = module(h, emb, context,
|
|
transformer_options=transformer_options)
|
|
result = rearrange(zero_conv(h, emb, context),
|
|
'b c f h w -> (b f) c h w')
|
|
outs.append(result)
|
|
if stage == 'inversion':
|
|
for k, attn in enumerate(self.input_blocks.named_modules()):
|
|
if hasattr(attn, 'q') and hasattr(attn, 'k'):
|
|
if k not in self.injection['input'][i]:
|
|
self.injection['input'][i][k] = {'q': [], 'k': []}
|
|
self.injection['input'][i][k]['q'].append(attn.q.cpu())
|
|
self.injection['input'][i][k]['k'].append(attn.k.cpu())
|
|
i += 1
|
|
|
|
if stage == 'sampling' and step < injection_steps:
|
|
for k, attn in enumerate(self.middle_block.named_modules()):
|
|
if hasattr(attn, 'inject_q') and hasattr(attn, 'inject_k'):
|
|
attn.inject_q = torch.cat(
|
|
[self.injection['middle'][k]['q'][step]]*cond_length).to('cuda')
|
|
attn.inject_k = torch.cat(
|
|
[self.injection['middle'][k]['k'][step]]*cond_length).to('cuda')
|
|
h = self.middle_block(
|
|
h, emb, context, transformer_options=transformer_options)
|
|
if stage == 'inversion':
|
|
for k, attn in enumerate(self.named_modules()):
|
|
if hasattr(attn, 'q') and hasattr(attn, 'k'):
|
|
if k not in self.injection['middle']:
|
|
self.injection['middle'][k] = {'q': [], 'k': []}
|
|
self.injection['middle'][k]['q'].append(attn.q.cpu())
|
|
self.injection['middle'][k]['k'].append(attn.k.cpu())
|
|
result = rearrange(self.middle_block_out(
|
|
h, emb, context), 'b c f h w -> (b f) c h w')
|
|
outs.append(result)
|
|
|
|
return outs
|