You cannot select more than 25 topics
Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
433 lines
16 KiB
Python
433 lines
16 KiB
Python
"""Classes for controlled image generation"""
|
|
|
|
import torch
|
|
from torch import nn
|
|
|
|
from imaginairy.modules.attention import SpatialTransformer
|
|
from imaginairy.modules.diffusion.ddpm import LatentDiffusion # type: ignore
|
|
from imaginairy.modules.diffusion.openaimodel import (
|
|
AttentionBlock,
|
|
Downsample,
|
|
ResBlock,
|
|
TimestepEmbedSequential,
|
|
UNetModel,
|
|
)
|
|
from imaginairy.modules.diffusion.util import (
|
|
conv_nd,
|
|
linear,
|
|
timestep_embedding,
|
|
zero_module,
|
|
)
|
|
|
|
|
|
class ControlledUnetModel(UNetModel):
|
|
def forward(
|
|
self,
|
|
x,
|
|
timesteps=None,
|
|
context=None,
|
|
control=None,
|
|
only_mid_control=False,
|
|
**kwargs,
|
|
):
|
|
hs = []
|
|
with torch.no_grad():
|
|
t_emb = timestep_embedding(
|
|
timesteps, self.model_channels, repeat_only=False
|
|
)
|
|
emb = self.time_embed(t_emb)
|
|
h = x.type(self.dtype)
|
|
for module in self.input_blocks:
|
|
h = module(h, emb, context)
|
|
hs.append(h)
|
|
h = self.middle_block(h, emb, context)
|
|
if control is not None:
|
|
h += control.pop()
|
|
|
|
for i, module in enumerate(self.output_blocks):
|
|
# allows us to work with multiples of 8 instead of just 32
|
|
if h.shape[-2:] != hs[-1].shape[-2:]:
|
|
h = nn.functional.interpolate(h, hs[-1].shape[-2:], mode="nearest")
|
|
if only_mid_control or control is None:
|
|
h = torch.cat([h, hs.pop()], dim=1)
|
|
else:
|
|
ctrl = control.pop()
|
|
if ctrl.shape[-2:] != hs[-1].shape[-2:]:
|
|
ctrl = nn.functional.interpolate(
|
|
ctrl, hs[-1].shape[-2:], mode="nearest"
|
|
)
|
|
h = torch.cat([h, hs.pop() + ctrl], dim=1)
|
|
del ctrl
|
|
h = module(h, emb, context)
|
|
|
|
h = h.type(x.dtype)
|
|
return self.out(h)
|
|
|
|
|
|
class ControlNet(nn.Module):
|
|
def __init__(
|
|
self,
|
|
image_size,
|
|
in_channels,
|
|
model_channels,
|
|
hint_channels,
|
|
num_res_blocks,
|
|
attention_resolutions,
|
|
dropout=0,
|
|
channel_mult=(1, 2, 4, 8),
|
|
conv_resample=True,
|
|
dims=2,
|
|
use_checkpoint=False,
|
|
use_fp16=False,
|
|
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
|
|
n_embed=None, # custom support for prediction of discrete ids into codebook of first stage vq model
|
|
legacy=True,
|
|
disable_self_attentions=None,
|
|
num_attention_blocks=None,
|
|
disable_middle_self_attn=False,
|
|
use_linear_in_transformer=False,
|
|
):
|
|
super().__init__()
|
|
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 isinstance(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):
|
|
msg = "provide num_res_blocks either as an int (globally constant) or as a list/tuple (per-level) with the same length as channel_mult"
|
|
raise ValueError(msg)
|
|
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(
|
|
self.num_res_blocks[i] >= num_attention_blocks[i]
|
|
for i in range(len(num_attention_blocks))
|
|
)
|
|
print(
|
|
f"Constructor of UNetModel received num_attention_blocks={num_attention_blocks}. "
|
|
f"This option has LESS priority than attention_resolutions {attention_resolutions}, "
|
|
f"i.e., in cases where num_attention_blocks[i] > 0 but 2**i not in attention_resolutions, "
|
|
f"attention will still not be set."
|
|
)
|
|
|
|
self.attention_resolutions = attention_resolutions
|
|
self.dropout = dropout
|
|
self.channel_mult = channel_mult
|
|
self.conv_resample = conv_resample
|
|
self.use_checkpoint = use_checkpoint
|
|
self.dtype = torch.float16 if use_fp16 else torch.float32
|
|
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(
|
|
linear(model_channels, time_embed_dim),
|
|
nn.SiLU(),
|
|
linear(time_embed_dim, time_embed_dim),
|
|
)
|
|
|
|
self.input_blocks = nn.ModuleList(
|
|
[
|
|
TimestepEmbedSequential(
|
|
conv_nd(dims, in_channels, model_channels, 3, padding=1)
|
|
)
|
|
]
|
|
)
|
|
self.zero_convs = nn.ModuleList([self.make_zero_conv(model_channels)])
|
|
|
|
self.input_hint_block = TimestepEmbedSequential(
|
|
conv_nd(dims, hint_channels, 16, 3, padding=1),
|
|
nn.SiLU(),
|
|
conv_nd(dims, 16, 16, 3, padding=1),
|
|
nn.SiLU(),
|
|
conv_nd(dims, 16, 32, 3, padding=1, stride=2),
|
|
nn.SiLU(),
|
|
conv_nd(dims, 32, 32, 3, padding=1),
|
|
nn.SiLU(),
|
|
conv_nd(dims, 32, 96, 3, padding=1, stride=2),
|
|
nn.SiLU(),
|
|
conv_nd(dims, 96, 96, 3, padding=1),
|
|
nn.SiLU(),
|
|
conv_nd(dims, 96, 256, 3, padding=1, stride=2),
|
|
nn.SiLU(),
|
|
zero_module(conv_nd(dims, 256, model_channels, 3, padding=1)),
|
|
)
|
|
|
|
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 = [
|
|
ResBlock(
|
|
ch,
|
|
time_embed_dim,
|
|
dropout,
|
|
out_channels=mult * model_channels,
|
|
dims=dims,
|
|
use_checkpoint=use_checkpoint,
|
|
use_scale_shift_norm=use_scale_shift_norm,
|
|
)
|
|
]
|
|
ch = mult * model_channels
|
|
if ds in attention_resolutions:
|
|
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 disable_self_attentions is not None:
|
|
disabled_sa = disable_self_attentions[level]
|
|
else:
|
|
disabled_sa = False
|
|
|
|
if num_attention_blocks is None or nr < num_attention_blocks[level]:
|
|
layers.append(
|
|
AttentionBlock(
|
|
ch,
|
|
use_checkpoint=use_checkpoint,
|
|
num_heads=num_heads,
|
|
num_head_channels=dim_head,
|
|
use_new_attention_order=use_new_attention_order,
|
|
)
|
|
if not use_spatial_transformer
|
|
else SpatialTransformer(
|
|
ch,
|
|
num_heads,
|
|
dim_head,
|
|
depth=transformer_depth,
|
|
context_dim=context_dim,
|
|
disable_self_attn=disabled_sa,
|
|
use_linear=use_linear_in_transformer,
|
|
use_checkpoint=use_checkpoint,
|
|
)
|
|
)
|
|
self.input_blocks.append(TimestepEmbedSequential(*layers))
|
|
self.zero_convs.append(self.make_zero_conv(ch))
|
|
self._feature_size += ch
|
|
input_block_chans.append(ch)
|
|
if level != len(channel_mult) - 1:
|
|
out_ch = ch
|
|
self.input_blocks.append(
|
|
TimestepEmbedSequential(
|
|
ResBlock(
|
|
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,
|
|
)
|
|
if resblock_updown
|
|
else Downsample(
|
|
ch, conv_resample, dims=dims, out_channels=out_ch
|
|
)
|
|
)
|
|
)
|
|
ch = out_ch
|
|
input_block_chans.append(ch)
|
|
self.zero_convs.append(self.make_zero_conv(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
|
|
self.middle_block = TimestepEmbedSequential(
|
|
ResBlock(
|
|
ch,
|
|
time_embed_dim,
|
|
dropout,
|
|
dims=dims,
|
|
use_checkpoint=use_checkpoint,
|
|
use_scale_shift_norm=use_scale_shift_norm,
|
|
),
|
|
AttentionBlock(
|
|
ch,
|
|
use_checkpoint=use_checkpoint,
|
|
num_heads=num_heads,
|
|
num_head_channels=dim_head,
|
|
use_new_attention_order=use_new_attention_order,
|
|
)
|
|
if not use_spatial_transformer
|
|
else SpatialTransformer( # always uses a self-attn
|
|
ch,
|
|
num_heads,
|
|
dim_head,
|
|
depth=transformer_depth,
|
|
context_dim=context_dim,
|
|
disable_self_attn=disable_middle_self_attn,
|
|
use_linear=use_linear_in_transformer,
|
|
use_checkpoint=use_checkpoint,
|
|
),
|
|
ResBlock(
|
|
ch,
|
|
time_embed_dim,
|
|
dropout,
|
|
dims=dims,
|
|
use_checkpoint=use_checkpoint,
|
|
use_scale_shift_norm=use_scale_shift_norm,
|
|
),
|
|
)
|
|
self.middle_block_out = self.make_zero_conv(ch)
|
|
self._feature_size += ch
|
|
|
|
def make_zero_conv(self, channels):
|
|
return TimestepEmbedSequential(
|
|
zero_module(conv_nd(self.dims, channels, channels, 1, padding=0))
|
|
)
|
|
|
|
def forward(self, x, hint, timesteps, context, **kwargs):
|
|
t_emb = timestep_embedding(timesteps, self.model_channels, repeat_only=False)
|
|
emb = self.time_embed(t_emb)
|
|
hint = hint.to(dtype=emb.dtype).to(device=emb.device)
|
|
|
|
guided_hint = self.input_hint_block(hint, emb, context)
|
|
|
|
outs = []
|
|
|
|
h = x.type(self.dtype)
|
|
for module, zero_conv in zip(self.input_blocks, self.zero_convs):
|
|
if guided_hint is not None:
|
|
h = module(h, emb, context)
|
|
# for wider img resolution handling?
|
|
if h.shape[-2:] != guided_hint[-1].shape[-2:]:
|
|
guided_hint = nn.functional.interpolate(
|
|
guided_hint, h[-1].shape[-2:], mode="nearest"
|
|
)
|
|
h += guided_hint
|
|
guided_hint = None
|
|
else:
|
|
h = module(h, emb, context)
|
|
outs.append(zero_conv(h, emb, context))
|
|
|
|
h = self.middle_block(h, emb, context)
|
|
outs.append(self.middle_block_out(h, emb, context))
|
|
|
|
return outs
|
|
|
|
|
|
class ControlLDM(LatentDiffusion):
|
|
def __init__(
|
|
self,
|
|
control_stage_config,
|
|
control_key,
|
|
only_mid_control,
|
|
*args,
|
|
global_average_pooling=False,
|
|
**kwargs,
|
|
):
|
|
super().__init__(*args, **kwargs)
|
|
self.control_stage_config = control_stage_config
|
|
# self.control_model = instantiate_from_config(control_stage_config)
|
|
self.control_models = []
|
|
self.control_key = control_key
|
|
self.only_mid_control = only_mid_control
|
|
self.control_scales = [1.0] * 13
|
|
self.global_average_pooling = global_average_pooling
|
|
|
|
def set_control_models(self, control_models):
|
|
self.control_models = control_models
|
|
|
|
def apply_model(self, x_noisy, t, cond, *args, **kwargs):
|
|
assert isinstance(cond, dict)
|
|
diffusion_model = self.model.diffusion_model
|
|
merged_control = None
|
|
cond_txt = torch.cat(cond["c_crossattn"], 1)
|
|
|
|
for control_model, c_concat, control_strength in zip(
|
|
self.control_models, cond["c_concat"], cond["control_strengths"]
|
|
):
|
|
cond_hint = torch.cat([c_concat], 1)
|
|
|
|
control = control_model(
|
|
x=x_noisy, hint=cond_hint, timesteps=t, context=cond_txt
|
|
)
|
|
control_scales = [control_strength] * 13
|
|
control = [c * scale for c, scale in zip(control, control_scales)]
|
|
if self.global_average_pooling:
|
|
control = [torch.mean(c, dim=(2, 3), keepdim=True) for c in control]
|
|
if merged_control is None:
|
|
merged_control = control
|
|
else:
|
|
merged_control = [mc + c for mc, c in zip(merged_control, control)]
|
|
|
|
eps = diffusion_model(
|
|
x=x_noisy,
|
|
timesteps=t,
|
|
context=cond_txt,
|
|
control=merged_control,
|
|
only_mid_control=self.only_mid_control,
|
|
)
|
|
|
|
return eps
|
|
|
|
def low_vram_shift(self, is_diffusing):
|
|
if is_diffusing:
|
|
self.model = self.model.cuda()
|
|
self.control_models = [cm.cuda() for cm in self.control_models]
|
|
self.first_stage_model = self.first_stage_model.cpu()
|
|
self.cond_stage_model = self.cond_stage_model.cpu()
|
|
else:
|
|
self.model = self.model.cpu()
|
|
self.control_models = [cm.cpu() for cm in self.control_models]
|
|
self.first_stage_model = self.first_stage_model.cuda()
|
|
self.cond_stage_model = self.cond_stage_model.cuda()
|