Spaces:
Running
Running
File size: 12,338 Bytes
6b448ad |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 |
import argparse
import huggingface_hub
import k_diffusion as K
import torch
from diffusers import UNet2DConditionModel
UPSCALER_REPO = "pcuenq/k-upscaler"
def resnet_to_diffusers_checkpoint(resnet, checkpoint, *, diffusers_resnet_prefix, resnet_prefix):
rv = {
# norm1
f"{diffusers_resnet_prefix}.norm1.linear.weight": checkpoint[f"{resnet_prefix}.main.0.mapper.weight"],
f"{diffusers_resnet_prefix}.norm1.linear.bias": checkpoint[f"{resnet_prefix}.main.0.mapper.bias"],
# conv1
f"{diffusers_resnet_prefix}.conv1.weight": checkpoint[f"{resnet_prefix}.main.2.weight"],
f"{diffusers_resnet_prefix}.conv1.bias": checkpoint[f"{resnet_prefix}.main.2.bias"],
# norm2
f"{diffusers_resnet_prefix}.norm2.linear.weight": checkpoint[f"{resnet_prefix}.main.4.mapper.weight"],
f"{diffusers_resnet_prefix}.norm2.linear.bias": checkpoint[f"{resnet_prefix}.main.4.mapper.bias"],
# conv2
f"{diffusers_resnet_prefix}.conv2.weight": checkpoint[f"{resnet_prefix}.main.6.weight"],
f"{diffusers_resnet_prefix}.conv2.bias": checkpoint[f"{resnet_prefix}.main.6.bias"],
}
if resnet.conv_shortcut is not None:
rv.update(
{
f"{diffusers_resnet_prefix}.conv_shortcut.weight": checkpoint[f"{resnet_prefix}.skip.weight"],
}
)
return rv
def self_attn_to_diffusers_checkpoint(checkpoint, *, diffusers_attention_prefix, attention_prefix):
weight_q, weight_k, weight_v = checkpoint[f"{attention_prefix}.qkv_proj.weight"].chunk(3, dim=0)
bias_q, bias_k, bias_v = checkpoint[f"{attention_prefix}.qkv_proj.bias"].chunk(3, dim=0)
rv = {
# norm
f"{diffusers_attention_prefix}.norm1.linear.weight": checkpoint[f"{attention_prefix}.norm_in.mapper.weight"],
f"{diffusers_attention_prefix}.norm1.linear.bias": checkpoint[f"{attention_prefix}.norm_in.mapper.bias"],
# to_q
f"{diffusers_attention_prefix}.attn1.to_q.weight": weight_q.squeeze(-1).squeeze(-1),
f"{diffusers_attention_prefix}.attn1.to_q.bias": bias_q,
# to_k
f"{diffusers_attention_prefix}.attn1.to_k.weight": weight_k.squeeze(-1).squeeze(-1),
f"{diffusers_attention_prefix}.attn1.to_k.bias": bias_k,
# to_v
f"{diffusers_attention_prefix}.attn1.to_v.weight": weight_v.squeeze(-1).squeeze(-1),
f"{diffusers_attention_prefix}.attn1.to_v.bias": bias_v,
# to_out
f"{diffusers_attention_prefix}.attn1.to_out.0.weight": checkpoint[f"{attention_prefix}.out_proj.weight"]
.squeeze(-1)
.squeeze(-1),
f"{diffusers_attention_prefix}.attn1.to_out.0.bias": checkpoint[f"{attention_prefix}.out_proj.bias"],
}
return rv
def cross_attn_to_diffusers_checkpoint(
checkpoint, *, diffusers_attention_prefix, diffusers_attention_index, attention_prefix
):
weight_k, weight_v = checkpoint[f"{attention_prefix}.kv_proj.weight"].chunk(2, dim=0)
bias_k, bias_v = checkpoint[f"{attention_prefix}.kv_proj.bias"].chunk(2, dim=0)
rv = {
# norm2 (ada groupnorm)
f"{diffusers_attention_prefix}.norm{diffusers_attention_index}.linear.weight": checkpoint[
f"{attention_prefix}.norm_dec.mapper.weight"
],
f"{diffusers_attention_prefix}.norm{diffusers_attention_index}.linear.bias": checkpoint[
f"{attention_prefix}.norm_dec.mapper.bias"
],
# layernorm on encoder_hidden_state
f"{diffusers_attention_prefix}.attn{diffusers_attention_index}.norm_cross.weight": checkpoint[
f"{attention_prefix}.norm_enc.weight"
],
f"{diffusers_attention_prefix}.attn{diffusers_attention_index}.norm_cross.bias": checkpoint[
f"{attention_prefix}.norm_enc.bias"
],
# to_q
f"{diffusers_attention_prefix}.attn{diffusers_attention_index}.to_q.weight": checkpoint[
f"{attention_prefix}.q_proj.weight"
]
.squeeze(-1)
.squeeze(-1),
f"{diffusers_attention_prefix}.attn{diffusers_attention_index}.to_q.bias": checkpoint[
f"{attention_prefix}.q_proj.bias"
],
# to_k
f"{diffusers_attention_prefix}.attn{diffusers_attention_index}.to_k.weight": weight_k.squeeze(-1).squeeze(-1),
f"{diffusers_attention_prefix}.attn{diffusers_attention_index}.to_k.bias": bias_k,
# to_v
f"{diffusers_attention_prefix}.attn{diffusers_attention_index}.to_v.weight": weight_v.squeeze(-1).squeeze(-1),
f"{diffusers_attention_prefix}.attn{diffusers_attention_index}.to_v.bias": bias_v,
# to_out
f"{diffusers_attention_prefix}.attn{diffusers_attention_index}.to_out.0.weight": checkpoint[
f"{attention_prefix}.out_proj.weight"
]
.squeeze(-1)
.squeeze(-1),
f"{diffusers_attention_prefix}.attn{diffusers_attention_index}.to_out.0.bias": checkpoint[
f"{attention_prefix}.out_proj.bias"
],
}
return rv
def block_to_diffusers_checkpoint(block, checkpoint, block_idx, block_type):
block_prefix = "inner_model.u_net.u_blocks" if block_type == "up" else "inner_model.u_net.d_blocks"
block_prefix = f"{block_prefix}.{block_idx}"
diffusers_checkpoint = {}
if not hasattr(block, "attentions"):
n = 1 # resnet only
elif not block.attentions[0].add_self_attention:
n = 2 # resnet -> cross-attention
else:
n = 3 # resnet -> self-attention -> cross-attention)
for resnet_idx, resnet in enumerate(block.resnets):
# diffusers_resnet_prefix = f"{diffusers_up_block_prefix}.resnets.{resnet_idx}"
diffusers_resnet_prefix = f"{block_type}_blocks.{block_idx}.resnets.{resnet_idx}"
idx = n * resnet_idx if block_type == "up" else n * resnet_idx + 1
resnet_prefix = f"{block_prefix}.{idx}" if block_type == "up" else f"{block_prefix}.{idx}"
diffusers_checkpoint.update(
resnet_to_diffusers_checkpoint(
resnet, checkpoint, diffusers_resnet_prefix=diffusers_resnet_prefix, resnet_prefix=resnet_prefix
)
)
if hasattr(block, "attentions"):
for attention_idx, attention in enumerate(block.attentions):
diffusers_attention_prefix = f"{block_type}_blocks.{block_idx}.attentions.{attention_idx}"
idx = n * attention_idx + 1 if block_type == "up" else n * attention_idx + 2
self_attention_prefix = f"{block_prefix}.{idx}"
cross_attention_prefix = f"{block_prefix}.{idx }"
cross_attention_index = 1 if not attention.add_self_attention else 2
idx = (
n * attention_idx + cross_attention_index
if block_type == "up"
else n * attention_idx + cross_attention_index + 1
)
cross_attention_prefix = f"{block_prefix}.{idx }"
diffusers_checkpoint.update(
cross_attn_to_diffusers_checkpoint(
checkpoint,
diffusers_attention_prefix=diffusers_attention_prefix,
diffusers_attention_index=2,
attention_prefix=cross_attention_prefix,
)
)
if attention.add_self_attention is True:
diffusers_checkpoint.update(
self_attn_to_diffusers_checkpoint(
checkpoint,
diffusers_attention_prefix=diffusers_attention_prefix,
attention_prefix=self_attention_prefix,
)
)
return diffusers_checkpoint
def unet_to_diffusers_checkpoint(model, checkpoint):
diffusers_checkpoint = {}
# pre-processing
diffusers_checkpoint.update(
{
"conv_in.weight": checkpoint["inner_model.proj_in.weight"],
"conv_in.bias": checkpoint["inner_model.proj_in.bias"],
}
)
# timestep and class embedding
diffusers_checkpoint.update(
{
"time_proj.weight": checkpoint["inner_model.timestep_embed.weight"].squeeze(-1),
"time_embedding.linear_1.weight": checkpoint["inner_model.mapping.0.weight"],
"time_embedding.linear_1.bias": checkpoint["inner_model.mapping.0.bias"],
"time_embedding.linear_2.weight": checkpoint["inner_model.mapping.2.weight"],
"time_embedding.linear_2.bias": checkpoint["inner_model.mapping.2.bias"],
"time_embedding.cond_proj.weight": checkpoint["inner_model.mapping_cond.weight"],
}
)
# down_blocks
for down_block_idx, down_block in enumerate(model.down_blocks):
diffusers_checkpoint.update(block_to_diffusers_checkpoint(down_block, checkpoint, down_block_idx, "down"))
# up_blocks
for up_block_idx, up_block in enumerate(model.up_blocks):
diffusers_checkpoint.update(block_to_diffusers_checkpoint(up_block, checkpoint, up_block_idx, "up"))
# post-processing
diffusers_checkpoint.update(
{
"conv_out.weight": checkpoint["inner_model.proj_out.weight"],
"conv_out.bias": checkpoint["inner_model.proj_out.bias"],
}
)
return diffusers_checkpoint
def unet_model_from_original_config(original_config):
in_channels = original_config["input_channels"] + original_config["unet_cond_dim"]
out_channels = original_config["input_channels"] + (1 if original_config["has_variance"] else 0)
block_out_channels = original_config["channels"]
assert (
len(set(original_config["depths"])) == 1
), "UNet2DConditionModel currently do not support blocks with different number of layers"
layers_per_block = original_config["depths"][0]
class_labels_dim = original_config["mapping_cond_dim"]
cross_attention_dim = original_config["cross_cond_dim"]
attn1_types = []
attn2_types = []
for s, c in zip(original_config["self_attn_depths"], original_config["cross_attn_depths"]):
if s:
a1 = "self"
a2 = "cross" if c else None
elif c:
a1 = "cross"
a2 = None
else:
a1 = None
a2 = None
attn1_types.append(a1)
attn2_types.append(a2)
unet = UNet2DConditionModel(
in_channels=in_channels,
out_channels=out_channels,
down_block_types=("KDownBlock2D", "KCrossAttnDownBlock2D", "KCrossAttnDownBlock2D", "KCrossAttnDownBlock2D"),
mid_block_type=None,
up_block_types=("KCrossAttnUpBlock2D", "KCrossAttnUpBlock2D", "KCrossAttnUpBlock2D", "KUpBlock2D"),
block_out_channels=block_out_channels,
layers_per_block=layers_per_block,
act_fn="gelu",
norm_num_groups=None,
cross_attention_dim=cross_attention_dim,
attention_head_dim=64,
time_cond_proj_dim=class_labels_dim,
resnet_time_scale_shift="scale_shift",
time_embedding_type="fourier",
timestep_post_act="gelu",
conv_in_kernel=1,
conv_out_kernel=1,
)
return unet
def main(args):
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
orig_config_path = huggingface_hub.hf_hub_download(UPSCALER_REPO, "config_laion_text_cond_latent_upscaler_2.json")
orig_weights_path = huggingface_hub.hf_hub_download(
UPSCALER_REPO, "laion_text_cond_latent_upscaler_2_1_00470000_slim.pth"
)
print(f"loading original model configuration from {orig_config_path}")
print(f"loading original model checkpoint from {orig_weights_path}")
print("converting to diffusers unet")
orig_config = K.config.load_config(open(orig_config_path))["model"]
model = unet_model_from_original_config(orig_config)
orig_checkpoint = torch.load(orig_weights_path, map_location=device)["model_ema"]
converted_checkpoint = unet_to_diffusers_checkpoint(model, orig_checkpoint)
model.load_state_dict(converted_checkpoint, strict=True)
model.save_pretrained(args.dump_path)
print(f"saving converted unet model in {args.dump_path}")
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument("--dump_path", default=None, type=str, required=True, help="Path to the output model.")
args = parser.parse_args()
main(args)
|