Spaces:
Running
on
Zero
Running
on
Zero
File size: 17,272 Bytes
703e263 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 |
import torch
from .sd_unet import ResnetBlock, DownSampler
from .sd_vae_decoder import VAEAttentionBlock
from .tiler import TileWorker
from einops import rearrange
class SDVAEEncoder(torch.nn.Module):
def __init__(self):
super().__init__()
self.scaling_factor = 0.18215
self.quant_conv = torch.nn.Conv2d(8, 8, kernel_size=1)
self.conv_in = torch.nn.Conv2d(3, 128, kernel_size=3, padding=1)
self.blocks = torch.nn.ModuleList([
# DownEncoderBlock2D
ResnetBlock(128, 128, eps=1e-6),
ResnetBlock(128, 128, eps=1e-6),
DownSampler(128, padding=0, extra_padding=True),
# DownEncoderBlock2D
ResnetBlock(128, 256, eps=1e-6),
ResnetBlock(256, 256, eps=1e-6),
DownSampler(256, padding=0, extra_padding=True),
# DownEncoderBlock2D
ResnetBlock(256, 512, eps=1e-6),
ResnetBlock(512, 512, eps=1e-6),
DownSampler(512, padding=0, extra_padding=True),
# DownEncoderBlock2D
ResnetBlock(512, 512, eps=1e-6),
ResnetBlock(512, 512, eps=1e-6),
# UNetMidBlock2D
ResnetBlock(512, 512, eps=1e-6),
VAEAttentionBlock(1, 512, 512, 1, eps=1e-6),
ResnetBlock(512, 512, eps=1e-6),
])
self.conv_norm_out = torch.nn.GroupNorm(num_channels=512, num_groups=32, eps=1e-6)
self.conv_act = torch.nn.SiLU()
self.conv_out = torch.nn.Conv2d(512, 8, kernel_size=3, padding=1)
def tiled_forward(self, sample, tile_size=64, tile_stride=32):
hidden_states = TileWorker().tiled_forward(
lambda x: self.forward(x),
sample,
tile_size,
tile_stride,
tile_device=sample.device,
tile_dtype=sample.dtype
)
return hidden_states
def forward(self, sample, tiled=False, tile_size=64, tile_stride=32, **kwargs):
original_dtype = sample.dtype
sample = sample.to(dtype=next(iter(self.parameters())).dtype)
# For VAE Decoder, we do not need to apply the tiler on each layer.
if tiled:
return self.tiled_forward(sample, tile_size=tile_size, tile_stride=tile_stride)
# 1. pre-process
hidden_states = self.conv_in(sample)
time_emb = None
text_emb = None
res_stack = None
# 2. blocks
for i, block in enumerate(self.blocks):
hidden_states, time_emb, text_emb, res_stack = block(hidden_states, time_emb, text_emb, res_stack)
# 3. output
hidden_states = self.conv_norm_out(hidden_states)
hidden_states = self.conv_act(hidden_states)
hidden_states = self.conv_out(hidden_states)
hidden_states = self.quant_conv(hidden_states)
hidden_states = hidden_states[:, :4]
hidden_states *= self.scaling_factor
hidden_states = hidden_states.to(original_dtype)
return hidden_states
def encode_video(self, sample, batch_size=8):
B = sample.shape[0]
hidden_states = []
for i in range(0, sample.shape[2], batch_size):
j = min(i + batch_size, sample.shape[2])
sample_batch = rearrange(sample[:,:,i:j], "B C T H W -> (B T) C H W")
hidden_states_batch = self(sample_batch)
hidden_states_batch = rearrange(hidden_states_batch, "(B T) C H W -> B C T H W", B=B)
hidden_states.append(hidden_states_batch)
hidden_states = torch.concat(hidden_states, dim=2)
return hidden_states
@staticmethod
def state_dict_converter():
return SDVAEEncoderStateDictConverter()
class SDVAEEncoderStateDictConverter:
def __init__(self):
pass
def from_diffusers(self, state_dict):
# architecture
block_types = [
'ResnetBlock', 'ResnetBlock', 'DownSampler',
'ResnetBlock', 'ResnetBlock', 'DownSampler',
'ResnetBlock', 'ResnetBlock', 'DownSampler',
'ResnetBlock', 'ResnetBlock',
'ResnetBlock', 'VAEAttentionBlock', 'ResnetBlock'
]
# Rename each parameter
local_rename_dict = {
"quant_conv": "quant_conv",
"encoder.conv_in": "conv_in",
"encoder.mid_block.attentions.0.group_norm": "blocks.12.norm",
"encoder.mid_block.attentions.0.to_q": "blocks.12.transformer_blocks.0.to_q",
"encoder.mid_block.attentions.0.to_k": "blocks.12.transformer_blocks.0.to_k",
"encoder.mid_block.attentions.0.to_v": "blocks.12.transformer_blocks.0.to_v",
"encoder.mid_block.attentions.0.to_out.0": "blocks.12.transformer_blocks.0.to_out",
"encoder.mid_block.resnets.0.norm1": "blocks.11.norm1",
"encoder.mid_block.resnets.0.conv1": "blocks.11.conv1",
"encoder.mid_block.resnets.0.norm2": "blocks.11.norm2",
"encoder.mid_block.resnets.0.conv2": "blocks.11.conv2",
"encoder.mid_block.resnets.1.norm1": "blocks.13.norm1",
"encoder.mid_block.resnets.1.conv1": "blocks.13.conv1",
"encoder.mid_block.resnets.1.norm2": "blocks.13.norm2",
"encoder.mid_block.resnets.1.conv2": "blocks.13.conv2",
"encoder.conv_norm_out": "conv_norm_out",
"encoder.conv_out": "conv_out",
}
name_list = sorted([name for name in state_dict])
rename_dict = {}
block_id = {"ResnetBlock": -1, "DownSampler": -1, "UpSampler": -1}
last_block_type_with_id = {"ResnetBlock": "", "DownSampler": "", "UpSampler": ""}
for name in name_list:
names = name.split(".")
name_prefix = ".".join(names[:-1])
if name_prefix in local_rename_dict:
rename_dict[name] = local_rename_dict[name_prefix] + "." + names[-1]
elif name.startswith("encoder.down_blocks"):
block_type = {"resnets": "ResnetBlock", "downsamplers": "DownSampler", "upsamplers": "UpSampler"}[names[3]]
block_type_with_id = ".".join(names[:5])
if block_type_with_id != last_block_type_with_id[block_type]:
block_id[block_type] += 1
last_block_type_with_id[block_type] = block_type_with_id
while block_id[block_type] < len(block_types) and block_types[block_id[block_type]] != block_type:
block_id[block_type] += 1
block_type_with_id = ".".join(names[:5])
names = ["blocks", str(block_id[block_type])] + names[5:]
rename_dict[name] = ".".join(names)
# Convert state_dict
state_dict_ = {}
for name, param in state_dict.items():
if name in rename_dict:
state_dict_[rename_dict[name]] = param
return state_dict_
def from_civitai(self, state_dict):
rename_dict = {
"first_stage_model.encoder.conv_in.bias": "conv_in.bias",
"first_stage_model.encoder.conv_in.weight": "conv_in.weight",
"first_stage_model.encoder.conv_out.bias": "conv_out.bias",
"first_stage_model.encoder.conv_out.weight": "conv_out.weight",
"first_stage_model.encoder.down.0.block.0.conv1.bias": "blocks.0.conv1.bias",
"first_stage_model.encoder.down.0.block.0.conv1.weight": "blocks.0.conv1.weight",
"first_stage_model.encoder.down.0.block.0.conv2.bias": "blocks.0.conv2.bias",
"first_stage_model.encoder.down.0.block.0.conv2.weight": "blocks.0.conv2.weight",
"first_stage_model.encoder.down.0.block.0.norm1.bias": "blocks.0.norm1.bias",
"first_stage_model.encoder.down.0.block.0.norm1.weight": "blocks.0.norm1.weight",
"first_stage_model.encoder.down.0.block.0.norm2.bias": "blocks.0.norm2.bias",
"first_stage_model.encoder.down.0.block.0.norm2.weight": "blocks.0.norm2.weight",
"first_stage_model.encoder.down.0.block.1.conv1.bias": "blocks.1.conv1.bias",
"first_stage_model.encoder.down.0.block.1.conv1.weight": "blocks.1.conv1.weight",
"first_stage_model.encoder.down.0.block.1.conv2.bias": "blocks.1.conv2.bias",
"first_stage_model.encoder.down.0.block.1.conv2.weight": "blocks.1.conv2.weight",
"first_stage_model.encoder.down.0.block.1.norm1.bias": "blocks.1.norm1.bias",
"first_stage_model.encoder.down.0.block.1.norm1.weight": "blocks.1.norm1.weight",
"first_stage_model.encoder.down.0.block.1.norm2.bias": "blocks.1.norm2.bias",
"first_stage_model.encoder.down.0.block.1.norm2.weight": "blocks.1.norm2.weight",
"first_stage_model.encoder.down.0.downsample.conv.bias": "blocks.2.conv.bias",
"first_stage_model.encoder.down.0.downsample.conv.weight": "blocks.2.conv.weight",
"first_stage_model.encoder.down.1.block.0.conv1.bias": "blocks.3.conv1.bias",
"first_stage_model.encoder.down.1.block.0.conv1.weight": "blocks.3.conv1.weight",
"first_stage_model.encoder.down.1.block.0.conv2.bias": "blocks.3.conv2.bias",
"first_stage_model.encoder.down.1.block.0.conv2.weight": "blocks.3.conv2.weight",
"first_stage_model.encoder.down.1.block.0.nin_shortcut.bias": "blocks.3.conv_shortcut.bias",
"first_stage_model.encoder.down.1.block.0.nin_shortcut.weight": "blocks.3.conv_shortcut.weight",
"first_stage_model.encoder.down.1.block.0.norm1.bias": "blocks.3.norm1.bias",
"first_stage_model.encoder.down.1.block.0.norm1.weight": "blocks.3.norm1.weight",
"first_stage_model.encoder.down.1.block.0.norm2.bias": "blocks.3.norm2.bias",
"first_stage_model.encoder.down.1.block.0.norm2.weight": "blocks.3.norm2.weight",
"first_stage_model.encoder.down.1.block.1.conv1.bias": "blocks.4.conv1.bias",
"first_stage_model.encoder.down.1.block.1.conv1.weight": "blocks.4.conv1.weight",
"first_stage_model.encoder.down.1.block.1.conv2.bias": "blocks.4.conv2.bias",
"first_stage_model.encoder.down.1.block.1.conv2.weight": "blocks.4.conv2.weight",
"first_stage_model.encoder.down.1.block.1.norm1.bias": "blocks.4.norm1.bias",
"first_stage_model.encoder.down.1.block.1.norm1.weight": "blocks.4.norm1.weight",
"first_stage_model.encoder.down.1.block.1.norm2.bias": "blocks.4.norm2.bias",
"first_stage_model.encoder.down.1.block.1.norm2.weight": "blocks.4.norm2.weight",
"first_stage_model.encoder.down.1.downsample.conv.bias": "blocks.5.conv.bias",
"first_stage_model.encoder.down.1.downsample.conv.weight": "blocks.5.conv.weight",
"first_stage_model.encoder.down.2.block.0.conv1.bias": "blocks.6.conv1.bias",
"first_stage_model.encoder.down.2.block.0.conv1.weight": "blocks.6.conv1.weight",
"first_stage_model.encoder.down.2.block.0.conv2.bias": "blocks.6.conv2.bias",
"first_stage_model.encoder.down.2.block.0.conv2.weight": "blocks.6.conv2.weight",
"first_stage_model.encoder.down.2.block.0.nin_shortcut.bias": "blocks.6.conv_shortcut.bias",
"first_stage_model.encoder.down.2.block.0.nin_shortcut.weight": "blocks.6.conv_shortcut.weight",
"first_stage_model.encoder.down.2.block.0.norm1.bias": "blocks.6.norm1.bias",
"first_stage_model.encoder.down.2.block.0.norm1.weight": "blocks.6.norm1.weight",
"first_stage_model.encoder.down.2.block.0.norm2.bias": "blocks.6.norm2.bias",
"first_stage_model.encoder.down.2.block.0.norm2.weight": "blocks.6.norm2.weight",
"first_stage_model.encoder.down.2.block.1.conv1.bias": "blocks.7.conv1.bias",
"first_stage_model.encoder.down.2.block.1.conv1.weight": "blocks.7.conv1.weight",
"first_stage_model.encoder.down.2.block.1.conv2.bias": "blocks.7.conv2.bias",
"first_stage_model.encoder.down.2.block.1.conv2.weight": "blocks.7.conv2.weight",
"first_stage_model.encoder.down.2.block.1.norm1.bias": "blocks.7.norm1.bias",
"first_stage_model.encoder.down.2.block.1.norm1.weight": "blocks.7.norm1.weight",
"first_stage_model.encoder.down.2.block.1.norm2.bias": "blocks.7.norm2.bias",
"first_stage_model.encoder.down.2.block.1.norm2.weight": "blocks.7.norm2.weight",
"first_stage_model.encoder.down.2.downsample.conv.bias": "blocks.8.conv.bias",
"first_stage_model.encoder.down.2.downsample.conv.weight": "blocks.8.conv.weight",
"first_stage_model.encoder.down.3.block.0.conv1.bias": "blocks.9.conv1.bias",
"first_stage_model.encoder.down.3.block.0.conv1.weight": "blocks.9.conv1.weight",
"first_stage_model.encoder.down.3.block.0.conv2.bias": "blocks.9.conv2.bias",
"first_stage_model.encoder.down.3.block.0.conv2.weight": "blocks.9.conv2.weight",
"first_stage_model.encoder.down.3.block.0.norm1.bias": "blocks.9.norm1.bias",
"first_stage_model.encoder.down.3.block.0.norm1.weight": "blocks.9.norm1.weight",
"first_stage_model.encoder.down.3.block.0.norm2.bias": "blocks.9.norm2.bias",
"first_stage_model.encoder.down.3.block.0.norm2.weight": "blocks.9.norm2.weight",
"first_stage_model.encoder.down.3.block.1.conv1.bias": "blocks.10.conv1.bias",
"first_stage_model.encoder.down.3.block.1.conv1.weight": "blocks.10.conv1.weight",
"first_stage_model.encoder.down.3.block.1.conv2.bias": "blocks.10.conv2.bias",
"first_stage_model.encoder.down.3.block.1.conv2.weight": "blocks.10.conv2.weight",
"first_stage_model.encoder.down.3.block.1.norm1.bias": "blocks.10.norm1.bias",
"first_stage_model.encoder.down.3.block.1.norm1.weight": "blocks.10.norm1.weight",
"first_stage_model.encoder.down.3.block.1.norm2.bias": "blocks.10.norm2.bias",
"first_stage_model.encoder.down.3.block.1.norm2.weight": "blocks.10.norm2.weight",
"first_stage_model.encoder.mid.attn_1.k.bias": "blocks.12.transformer_blocks.0.to_k.bias",
"first_stage_model.encoder.mid.attn_1.k.weight": "blocks.12.transformer_blocks.0.to_k.weight",
"first_stage_model.encoder.mid.attn_1.norm.bias": "blocks.12.norm.bias",
"first_stage_model.encoder.mid.attn_1.norm.weight": "blocks.12.norm.weight",
"first_stage_model.encoder.mid.attn_1.proj_out.bias": "blocks.12.transformer_blocks.0.to_out.bias",
"first_stage_model.encoder.mid.attn_1.proj_out.weight": "blocks.12.transformer_blocks.0.to_out.weight",
"first_stage_model.encoder.mid.attn_1.q.bias": "blocks.12.transformer_blocks.0.to_q.bias",
"first_stage_model.encoder.mid.attn_1.q.weight": "blocks.12.transformer_blocks.0.to_q.weight",
"first_stage_model.encoder.mid.attn_1.v.bias": "blocks.12.transformer_blocks.0.to_v.bias",
"first_stage_model.encoder.mid.attn_1.v.weight": "blocks.12.transformer_blocks.0.to_v.weight",
"first_stage_model.encoder.mid.block_1.conv1.bias": "blocks.11.conv1.bias",
"first_stage_model.encoder.mid.block_1.conv1.weight": "blocks.11.conv1.weight",
"first_stage_model.encoder.mid.block_1.conv2.bias": "blocks.11.conv2.bias",
"first_stage_model.encoder.mid.block_1.conv2.weight": "blocks.11.conv2.weight",
"first_stage_model.encoder.mid.block_1.norm1.bias": "blocks.11.norm1.bias",
"first_stage_model.encoder.mid.block_1.norm1.weight": "blocks.11.norm1.weight",
"first_stage_model.encoder.mid.block_1.norm2.bias": "blocks.11.norm2.bias",
"first_stage_model.encoder.mid.block_1.norm2.weight": "blocks.11.norm2.weight",
"first_stage_model.encoder.mid.block_2.conv1.bias": "blocks.13.conv1.bias",
"first_stage_model.encoder.mid.block_2.conv1.weight": "blocks.13.conv1.weight",
"first_stage_model.encoder.mid.block_2.conv2.bias": "blocks.13.conv2.bias",
"first_stage_model.encoder.mid.block_2.conv2.weight": "blocks.13.conv2.weight",
"first_stage_model.encoder.mid.block_2.norm1.bias": "blocks.13.norm1.bias",
"first_stage_model.encoder.mid.block_2.norm1.weight": "blocks.13.norm1.weight",
"first_stage_model.encoder.mid.block_2.norm2.bias": "blocks.13.norm2.bias",
"first_stage_model.encoder.mid.block_2.norm2.weight": "blocks.13.norm2.weight",
"first_stage_model.encoder.norm_out.bias": "conv_norm_out.bias",
"first_stage_model.encoder.norm_out.weight": "conv_norm_out.weight",
"first_stage_model.quant_conv.bias": "quant_conv.bias",
"first_stage_model.quant_conv.weight": "quant_conv.weight",
}
state_dict_ = {}
for name in state_dict:
if name in rename_dict:
param = state_dict[name]
if "transformer_blocks" in rename_dict[name]:
param = param.squeeze()
state_dict_[rename_dict[name]] = param
return state_dict_
|