Upload cldm.py
Browse files
cldm.py
ADDED
|
@@ -0,0 +1,456 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import os
|
| 2 |
+
import einops
|
| 3 |
+
from omegaconf import OmegaConf
|
| 4 |
+
import torch
|
| 5 |
+
import torch as th
|
| 6 |
+
import torch.nn as nn
|
| 7 |
+
from modules import devices, lowvram, shared
|
| 8 |
+
|
| 9 |
+
from ldm.modules.diffusionmodules.util import (
|
| 10 |
+
conv_nd,
|
| 11 |
+
linear,
|
| 12 |
+
zero_module,
|
| 13 |
+
timestep_embedding,
|
| 14 |
+
)
|
| 15 |
+
|
| 16 |
+
from ldm.modules.attention import SpatialTransformer
|
| 17 |
+
from ldm.modules.diffusionmodules.openaimodel import UNetModel, TimestepEmbedSequential, ResBlock, Downsample, AttentionBlock
|
| 18 |
+
from ldm.util import exists
|
| 19 |
+
|
| 20 |
+
|
| 21 |
+
def load_state_dict(ckpt_path, location='cpu'):
|
| 22 |
+
_, extension = os.path.splitext(ckpt_path)
|
| 23 |
+
if extension.lower() == ".safetensors":
|
| 24 |
+
import safetensors.torch
|
| 25 |
+
state_dict = safetensors.torch.load_file(ckpt_path, device=location)
|
| 26 |
+
else:
|
| 27 |
+
state_dict = get_state_dict(torch.load(
|
| 28 |
+
ckpt_path, map_location=torch.device(location)))
|
| 29 |
+
state_dict = get_state_dict(state_dict)
|
| 30 |
+
print(f'Loaded state_dict from [{ckpt_path}]')
|
| 31 |
+
return state_dict
|
| 32 |
+
|
| 33 |
+
|
| 34 |
+
def get_state_dict(d):
|
| 35 |
+
return d.get('state_dict', d)
|
| 36 |
+
|
| 37 |
+
|
| 38 |
+
def align(hint, size):
|
| 39 |
+
b, c, h1, w1 = hint.shape
|
| 40 |
+
h, w = size
|
| 41 |
+
if h != h1 or w != w1:
|
| 42 |
+
hint = torch.nn.functional.interpolate(hint, size=size, mode="nearest")
|
| 43 |
+
return hint
|
| 44 |
+
|
| 45 |
+
|
| 46 |
+
def get_node_name(name, parent_name):
|
| 47 |
+
if len(name) <= len(parent_name):
|
| 48 |
+
return False, ''
|
| 49 |
+
p = name[:len(parent_name)]
|
| 50 |
+
if p != parent_name:
|
| 51 |
+
return False, ''
|
| 52 |
+
return True, name[len(parent_name):]
|
| 53 |
+
|
| 54 |
+
|
| 55 |
+
class PlugableControlModel(nn.Module):
|
| 56 |
+
def __init__(self, model_path, config_path, weight=1.0, lowvram=False, base_model=None) -> None:
|
| 57 |
+
super().__init__()
|
| 58 |
+
# temp code
|
| 59 |
+
config_path = "/mnt/workspace/stable-diffusion-webui/extensions/sd-webui-controlnet/models/cldm_v15.yaml"
|
| 60 |
+
print(config_path)
|
| 61 |
+
config = OmegaConf.load(config_path)
|
| 62 |
+
|
| 63 |
+
self.control_model = ControlNet(**config.model.params.control_stage_config.params)
|
| 64 |
+
state_dict = load_state_dict(model_path)
|
| 65 |
+
|
| 66 |
+
if any([k.startswith("control_model.") for k, v in state_dict.items()]):
|
| 67 |
+
|
| 68 |
+
is_diff_model = 'difference' in state_dict
|
| 69 |
+
transfer_ctrl_opt = shared.opts.data.get("control_net_control_transfer", False) and \
|
| 70 |
+
any([k.startswith("model.diffusion_model.") for k, v in state_dict.items()])
|
| 71 |
+
|
| 72 |
+
if (is_diff_model or transfer_ctrl_opt) and base_model is not None:
|
| 73 |
+
# apply transfer control - https://github.com/lllyasviel/ControlNet/blob/main/tool_transfer_control.py
|
| 74 |
+
|
| 75 |
+
unet_state_dict = base_model.state_dict()
|
| 76 |
+
unet_state_dict_keys = unet_state_dict.keys()
|
| 77 |
+
final_state_dict = {}
|
| 78 |
+
counter = 0
|
| 79 |
+
for key in state_dict.keys():
|
| 80 |
+
if not key.startswith("control_model."):
|
| 81 |
+
continue
|
| 82 |
+
|
| 83 |
+
p = state_dict[key]
|
| 84 |
+
is_control, node_name = get_node_name(key, 'control_')
|
| 85 |
+
key_name = node_name.replace("model.", "") if is_control else key
|
| 86 |
+
|
| 87 |
+
if key_name in unet_state_dict_keys:
|
| 88 |
+
if is_diff_model:
|
| 89 |
+
# transfer control by make difference in advance
|
| 90 |
+
p_new = p + unet_state_dict[key_name].clone().cpu()
|
| 91 |
+
else:
|
| 92 |
+
# transfer control by calculate offsets from (delta = p + current_unet_encoder - frozen_unet_encoder)
|
| 93 |
+
p_new = p + unet_state_dict[key_name].clone().cpu() - state_dict["model.diffusion_model."+key_name]
|
| 94 |
+
counter += 1
|
| 95 |
+
else:
|
| 96 |
+
p_new = p
|
| 97 |
+
final_state_dict[key] = p_new
|
| 98 |
+
|
| 99 |
+
print(f'Offset cloned: {counter} values')
|
| 100 |
+
state_dict = final_state_dict
|
| 101 |
+
|
| 102 |
+
state_dict = {k.replace("control_model.", ""): v for k, v in state_dict.items() if k.startswith("control_model.")}
|
| 103 |
+
else:
|
| 104 |
+
# assume that model is done by user
|
| 105 |
+
pass
|
| 106 |
+
|
| 107 |
+
self.control_model.load_state_dict(state_dict)
|
| 108 |
+
self.lowvram = lowvram
|
| 109 |
+
self.weight = weight
|
| 110 |
+
self.only_mid_control = False
|
| 111 |
+
self.control = None
|
| 112 |
+
self.hint_cond = None
|
| 113 |
+
|
| 114 |
+
if not self.lowvram:
|
| 115 |
+
self.control_model.to(devices.get_device_for("controlnet"))
|
| 116 |
+
|
| 117 |
+
def hook(self, model, parent_model):
|
| 118 |
+
outer = self
|
| 119 |
+
|
| 120 |
+
def forward(self, x, timesteps=None, context=None, **kwargs):
|
| 121 |
+
only_mid_control = outer.only_mid_control
|
| 122 |
+
|
| 123 |
+
# hires stuffs
|
| 124 |
+
# note that this method may not works if hr_scale < 1.1
|
| 125 |
+
if abs(x.shape[-1] - outer.hint_cond.shape[-1] // 8) > 8:
|
| 126 |
+
only_mid_control = shared.opts.data.get("control_net_only_midctrl_hires", True)
|
| 127 |
+
# If you want to completely disable control net, uncomment this.
|
| 128 |
+
# return self._original_forward(x, timesteps=timesteps, context=context, **kwargs)
|
| 129 |
+
|
| 130 |
+
control = outer.control_model(x=x, hint=outer.hint_cond, timesteps=timesteps, context=context)
|
| 131 |
+
assert timesteps is not None, ValueError(f"insufficient timestep: {timesteps}")
|
| 132 |
+
hs = []
|
| 133 |
+
with torch.no_grad():
|
| 134 |
+
t_emb = timestep_embedding(
|
| 135 |
+
timesteps, self.model_channels, repeat_only=False)
|
| 136 |
+
emb = self.time_embed(t_emb)
|
| 137 |
+
h = x.type(self.dtype)
|
| 138 |
+
for module in self.input_blocks:
|
| 139 |
+
h = module(h, emb, context)
|
| 140 |
+
hs.append(h)
|
| 141 |
+
h = self.middle_block(h, emb, context)
|
| 142 |
+
|
| 143 |
+
h += control.pop()
|
| 144 |
+
|
| 145 |
+
for i, module in enumerate(self.output_blocks):
|
| 146 |
+
if only_mid_control:
|
| 147 |
+
h = torch.cat([h, hs.pop()], dim=1)
|
| 148 |
+
else:
|
| 149 |
+
hs_input, control_input = hs.pop(), control.pop()
|
| 150 |
+
h = align(h, hs_input.shape[-2:])
|
| 151 |
+
h = torch.cat([h, hs_input + control_input * outer.weight], dim=1)
|
| 152 |
+
h = module(h, emb, context)
|
| 153 |
+
|
| 154 |
+
h = h.type(x.dtype)
|
| 155 |
+
return self.out(h)
|
| 156 |
+
|
| 157 |
+
def forward2(*args, **kwargs):
|
| 158 |
+
# webui will handle other compoments
|
| 159 |
+
try:
|
| 160 |
+
if shared.cmd_opts.lowvram:
|
| 161 |
+
lowvram.send_everything_to_cpu()
|
| 162 |
+
if self.lowvram:
|
| 163 |
+
self.control_model.to(devices.get_device_for("controlnet"))
|
| 164 |
+
return forward(*args, **kwargs)
|
| 165 |
+
finally:
|
| 166 |
+
if self.lowvram:
|
| 167 |
+
self.control_model.cpu()
|
| 168 |
+
|
| 169 |
+
model._original_forward = model.forward
|
| 170 |
+
model.forward = forward2.__get__(model, UNetModel)
|
| 171 |
+
|
| 172 |
+
def notify(self, cond_like, weight):
|
| 173 |
+
self.hint_cond = cond_like
|
| 174 |
+
self.weight = weight
|
| 175 |
+
# print(self.hint_cond.shape)
|
| 176 |
+
|
| 177 |
+
def restore(self, model):
|
| 178 |
+
if not hasattr(model, "_original_forward"):
|
| 179 |
+
# no such handle, ignore
|
| 180 |
+
return
|
| 181 |
+
|
| 182 |
+
model.forward = model._original_forward
|
| 183 |
+
del model._original_forward
|
| 184 |
+
|
| 185 |
+
|
| 186 |
+
class ControlNet(nn.Module):
|
| 187 |
+
def __init__(
|
| 188 |
+
self,
|
| 189 |
+
image_size,
|
| 190 |
+
in_channels,
|
| 191 |
+
model_channels,
|
| 192 |
+
hint_channels,
|
| 193 |
+
num_res_blocks,
|
| 194 |
+
attention_resolutions,
|
| 195 |
+
dropout=0,
|
| 196 |
+
channel_mult=(1, 2, 4, 8),
|
| 197 |
+
conv_resample=True,
|
| 198 |
+
dims=2,
|
| 199 |
+
use_checkpoint=False,
|
| 200 |
+
use_fp16=False,
|
| 201 |
+
num_heads=-1,
|
| 202 |
+
num_head_channels=-1,
|
| 203 |
+
num_heads_upsample=-1,
|
| 204 |
+
use_scale_shift_norm=False,
|
| 205 |
+
resblock_updown=False,
|
| 206 |
+
use_new_attention_order=False,
|
| 207 |
+
use_spatial_transformer=False, # custom transformer support
|
| 208 |
+
transformer_depth=1, # custom transformer support
|
| 209 |
+
context_dim=None, # custom transformer support
|
| 210 |
+
# custom support for prediction of discrete ids into codebook of first stage vq model
|
| 211 |
+
n_embed=None,
|
| 212 |
+
legacy=True,
|
| 213 |
+
disable_self_attentions=None,
|
| 214 |
+
num_attention_blocks=None,
|
| 215 |
+
disable_middle_self_attn=False,
|
| 216 |
+
use_linear_in_transformer=False,
|
| 217 |
+
):
|
| 218 |
+
super().__init__()
|
| 219 |
+
if use_spatial_transformer:
|
| 220 |
+
assert context_dim is not None, 'Fool!! You forgot to include the dimension of your cross-attention conditioning...'
|
| 221 |
+
|
| 222 |
+
if context_dim is not None:
|
| 223 |
+
assert use_spatial_transformer, 'Fool!! You forgot to use the spatial transformer for your cross-attention conditioning...'
|
| 224 |
+
from omegaconf.listconfig import ListConfig
|
| 225 |
+
if type(context_dim) == ListConfig:
|
| 226 |
+
context_dim = list(context_dim)
|
| 227 |
+
|
| 228 |
+
if num_heads_upsample == -1:
|
| 229 |
+
num_heads_upsample = num_heads
|
| 230 |
+
|
| 231 |
+
if num_heads == -1:
|
| 232 |
+
assert num_head_channels != -1, 'Either num_heads or num_head_channels has to be set'
|
| 233 |
+
|
| 234 |
+
if num_head_channels == -1:
|
| 235 |
+
assert num_heads != -1, 'Either num_heads or num_head_channels has to be set'
|
| 236 |
+
|
| 237 |
+
self.dims = dims
|
| 238 |
+
self.image_size = image_size
|
| 239 |
+
self.in_channels = in_channels
|
| 240 |
+
self.model_channels = model_channels
|
| 241 |
+
if isinstance(num_res_blocks, int):
|
| 242 |
+
self.num_res_blocks = len(channel_mult) * [num_res_blocks]
|
| 243 |
+
else:
|
| 244 |
+
if len(num_res_blocks) != len(channel_mult):
|
| 245 |
+
raise ValueError("provide num_res_blocks either as an int (globally constant) or "
|
| 246 |
+
"as a list/tuple (per-level) with the same length as channel_mult")
|
| 247 |
+
self.num_res_blocks = num_res_blocks
|
| 248 |
+
if disable_self_attentions is not None:
|
| 249 |
+
# should be a list of booleans, indicating whether to disable self-attention in TransformerBlocks or not
|
| 250 |
+
assert len(disable_self_attentions) == len(channel_mult)
|
| 251 |
+
if num_attention_blocks is not None:
|
| 252 |
+
assert len(num_attention_blocks) == len(self.num_res_blocks)
|
| 253 |
+
assert all(map(lambda i: self.num_res_blocks[i] >= num_attention_blocks[i], range(
|
| 254 |
+
len(num_attention_blocks))))
|
| 255 |
+
print(f"Constructor of UNetModel received num_attention_blocks={num_attention_blocks}. "
|
| 256 |
+
f"This option has LESS priority than attention_resolutions {attention_resolutions}, "
|
| 257 |
+
f"i.e., in cases where num_attention_blocks[i] > 0 but 2**i not in attention_resolutions, "
|
| 258 |
+
f"attention will still not be set.")
|
| 259 |
+
|
| 260 |
+
self.attention_resolutions = attention_resolutions
|
| 261 |
+
self.dropout = dropout
|
| 262 |
+
self.channel_mult = channel_mult
|
| 263 |
+
self.conv_resample = conv_resample
|
| 264 |
+
self.use_checkpoint = use_checkpoint
|
| 265 |
+
self.dtype = th.float16 if use_fp16 else th.float32
|
| 266 |
+
self.num_heads = num_heads
|
| 267 |
+
self.num_head_channels = num_head_channels
|
| 268 |
+
self.num_heads_upsample = num_heads_upsample
|
| 269 |
+
self.predict_codebook_ids = n_embed is not None
|
| 270 |
+
|
| 271 |
+
time_embed_dim = model_channels * 4
|
| 272 |
+
self.time_embed = nn.Sequential(
|
| 273 |
+
linear(model_channels, time_embed_dim),
|
| 274 |
+
nn.SiLU(),
|
| 275 |
+
linear(time_embed_dim, time_embed_dim),
|
| 276 |
+
)
|
| 277 |
+
|
| 278 |
+
self.input_blocks = nn.ModuleList(
|
| 279 |
+
[
|
| 280 |
+
TimestepEmbedSequential(
|
| 281 |
+
conv_nd(dims, in_channels, model_channels, 3, padding=1)
|
| 282 |
+
)
|
| 283 |
+
]
|
| 284 |
+
)
|
| 285 |
+
self.zero_convs = nn.ModuleList([self.make_zero_conv(model_channels)])
|
| 286 |
+
|
| 287 |
+
self.input_hint_block = TimestepEmbedSequential(
|
| 288 |
+
conv_nd(dims, hint_channels, 16, 3, padding=1),
|
| 289 |
+
nn.SiLU(),
|
| 290 |
+
conv_nd(dims, 16, 16, 3, padding=1),
|
| 291 |
+
nn.SiLU(),
|
| 292 |
+
conv_nd(dims, 16, 32, 3, padding=1, stride=2),
|
| 293 |
+
nn.SiLU(),
|
| 294 |
+
conv_nd(dims, 32, 32, 3, padding=1),
|
| 295 |
+
nn.SiLU(),
|
| 296 |
+
conv_nd(dims, 32, 96, 3, padding=1, stride=2),
|
| 297 |
+
nn.SiLU(),
|
| 298 |
+
conv_nd(dims, 96, 96, 3, padding=1),
|
| 299 |
+
nn.SiLU(),
|
| 300 |
+
conv_nd(dims, 96, 256, 3, padding=1, stride=2),
|
| 301 |
+
nn.SiLU(),
|
| 302 |
+
zero_module(conv_nd(dims, 256, model_channels, 3, padding=1))
|
| 303 |
+
)
|
| 304 |
+
|
| 305 |
+
self._feature_size = model_channels
|
| 306 |
+
input_block_chans = [model_channels]
|
| 307 |
+
ch = model_channels
|
| 308 |
+
ds = 1
|
| 309 |
+
for level, mult in enumerate(channel_mult):
|
| 310 |
+
for nr in range(self.num_res_blocks[level]):
|
| 311 |
+
layers = [
|
| 312 |
+
ResBlock(
|
| 313 |
+
ch,
|
| 314 |
+
time_embed_dim,
|
| 315 |
+
dropout,
|
| 316 |
+
out_channels=mult * model_channels,
|
| 317 |
+
dims=dims,
|
| 318 |
+
use_checkpoint=use_checkpoint,
|
| 319 |
+
use_scale_shift_norm=use_scale_shift_norm,
|
| 320 |
+
)
|
| 321 |
+
]
|
| 322 |
+
ch = mult * model_channels
|
| 323 |
+
if ds in attention_resolutions:
|
| 324 |
+
if num_head_channels == -1:
|
| 325 |
+
dim_head = ch // num_heads
|
| 326 |
+
else:
|
| 327 |
+
num_heads = ch // num_head_channels
|
| 328 |
+
dim_head = num_head_channels
|
| 329 |
+
if legacy:
|
| 330 |
+
#num_heads = 1
|
| 331 |
+
dim_head = ch // num_heads if use_spatial_transformer else num_head_channels
|
| 332 |
+
if exists(disable_self_attentions):
|
| 333 |
+
disabled_sa = disable_self_attentions[level]
|
| 334 |
+
else:
|
| 335 |
+
disabled_sa = False
|
| 336 |
+
|
| 337 |
+
if not exists(num_attention_blocks) or nr < num_attention_blocks[level]:
|
| 338 |
+
layers.append(
|
| 339 |
+
AttentionBlock(
|
| 340 |
+
ch,
|
| 341 |
+
use_checkpoint=use_checkpoint,
|
| 342 |
+
num_heads=num_heads,
|
| 343 |
+
num_head_channels=dim_head,
|
| 344 |
+
use_new_attention_order=use_new_attention_order,
|
| 345 |
+
) if not use_spatial_transformer else SpatialTransformer(
|
| 346 |
+
ch, num_heads, dim_head, depth=transformer_depth, context_dim=context_dim,
|
| 347 |
+
disable_self_attn=disabled_sa, use_linear=use_linear_in_transformer,
|
| 348 |
+
use_checkpoint=use_checkpoint
|
| 349 |
+
)
|
| 350 |
+
)
|
| 351 |
+
self.input_blocks.append(TimestepEmbedSequential(*layers))
|
| 352 |
+
self.zero_convs.append(self.make_zero_conv(ch))
|
| 353 |
+
self._feature_size += ch
|
| 354 |
+
input_block_chans.append(ch)
|
| 355 |
+
if level != len(channel_mult) - 1:
|
| 356 |
+
out_ch = ch
|
| 357 |
+
self.input_blocks.append(
|
| 358 |
+
TimestepEmbedSequential(
|
| 359 |
+
ResBlock(
|
| 360 |
+
ch,
|
| 361 |
+
time_embed_dim,
|
| 362 |
+
dropout,
|
| 363 |
+
out_channels=out_ch,
|
| 364 |
+
dims=dims,
|
| 365 |
+
use_checkpoint=use_checkpoint,
|
| 366 |
+
use_scale_shift_norm=use_scale_shift_norm,
|
| 367 |
+
down=True,
|
| 368 |
+
)
|
| 369 |
+
if resblock_updown
|
| 370 |
+
else Downsample(
|
| 371 |
+
ch, conv_resample, dims=dims, out_channels=out_ch
|
| 372 |
+
)
|
| 373 |
+
)
|
| 374 |
+
)
|
| 375 |
+
ch = out_ch
|
| 376 |
+
input_block_chans.append(ch)
|
| 377 |
+
self.zero_convs.append(self.make_zero_conv(ch))
|
| 378 |
+
ds *= 2
|
| 379 |
+
self._feature_size += ch
|
| 380 |
+
|
| 381 |
+
if num_head_channels == -1:
|
| 382 |
+
dim_head = ch // num_heads
|
| 383 |
+
else:
|
| 384 |
+
num_heads = ch // num_head_channels
|
| 385 |
+
dim_head = num_head_channels
|
| 386 |
+
if legacy:
|
| 387 |
+
#num_heads = 1
|
| 388 |
+
dim_head = ch // num_heads if use_spatial_transformer else num_head_channels
|
| 389 |
+
self.middle_block = TimestepEmbedSequential(
|
| 390 |
+
ResBlock(
|
| 391 |
+
ch,
|
| 392 |
+
time_embed_dim,
|
| 393 |
+
dropout,
|
| 394 |
+
dims=dims,
|
| 395 |
+
use_checkpoint=use_checkpoint,
|
| 396 |
+
use_scale_shift_norm=use_scale_shift_norm,
|
| 397 |
+
),
|
| 398 |
+
AttentionBlock(
|
| 399 |
+
ch,
|
| 400 |
+
use_checkpoint=use_checkpoint,
|
| 401 |
+
num_heads=num_heads,
|
| 402 |
+
num_head_channels=dim_head,
|
| 403 |
+
use_new_attention_order=use_new_attention_order,
|
| 404 |
+
# always uses a self-attn
|
| 405 |
+
) if not use_spatial_transformer else SpatialTransformer(
|
| 406 |
+
ch, num_heads, dim_head, depth=transformer_depth, context_dim=context_dim,
|
| 407 |
+
disable_self_attn=disable_middle_self_attn, use_linear=use_linear_in_transformer,
|
| 408 |
+
use_checkpoint=use_checkpoint
|
| 409 |
+
),
|
| 410 |
+
ResBlock(
|
| 411 |
+
ch,
|
| 412 |
+
time_embed_dim,
|
| 413 |
+
dropout,
|
| 414 |
+
dims=dims,
|
| 415 |
+
use_checkpoint=use_checkpoint,
|
| 416 |
+
use_scale_shift_norm=use_scale_shift_norm,
|
| 417 |
+
),
|
| 418 |
+
)
|
| 419 |
+
self.middle_block_out = self.make_zero_conv(ch)
|
| 420 |
+
self._feature_size += ch
|
| 421 |
+
|
| 422 |
+
def make_zero_conv(self, channels):
|
| 423 |
+
return TimestepEmbedSequential(zero_module(conv_nd(self.dims, channels, channels, 1, padding=0)))
|
| 424 |
+
|
| 425 |
+
def align(self, hint, h, w):
|
| 426 |
+
c, h1, w1 = hint.shape
|
| 427 |
+
if h != h1 or w != w1:
|
| 428 |
+
hint = align(hint.unsqueeze(0), (h, w))
|
| 429 |
+
return hint.squeeze(0)
|
| 430 |
+
return hint
|
| 431 |
+
|
| 432 |
+
def forward(self, x, hint, timesteps, context, **kwargs):
|
| 433 |
+
t_emb = timestep_embedding(
|
| 434 |
+
timesteps, self.model_channels, repeat_only=False)
|
| 435 |
+
emb = self.time_embed(t_emb)
|
| 436 |
+
|
| 437 |
+
guided_hint = self.input_hint_block(hint, emb, context)
|
| 438 |
+
outs = []
|
| 439 |
+
|
| 440 |
+
h1, w1 = x.shape[-2:]
|
| 441 |
+
guided_hint = self.align(guided_hint, h1, w1)
|
| 442 |
+
|
| 443 |
+
h = x.type(self.dtype)
|
| 444 |
+
for module, zero_conv in zip(self.input_blocks, self.zero_convs):
|
| 445 |
+
if guided_hint is not None:
|
| 446 |
+
h = module(h, emb, context)
|
| 447 |
+
h += guided_hint
|
| 448 |
+
guided_hint = None
|
| 449 |
+
else:
|
| 450 |
+
h = module(h, emb, context)
|
| 451 |
+
outs.append(zero_conv(h, emb, context))
|
| 452 |
+
|
| 453 |
+
h = self.middle_block(h, emb, context)
|
| 454 |
+
outs.append(self.middle_block_out(h, emb, context))
|
| 455 |
+
|
| 456 |
+
return outs
|