File size: 10,380 Bytes
43b7e92 |
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 |
import argparse
import os
import torch
from transformers import T5EncoderModel, T5Tokenizer
from diffusers import AutoencoderKL, DPMSolverMultistepScheduler, PixArtSigmaPipeline, Transformer2DModel
ckpt_id = "PixArt-alpha"
# https://github.com/PixArt-alpha/PixArt-sigma/blob/dd087141864e30ec44f12cb7448dd654be065e88/scripts/inference.py#L158
interpolation_scale = {256: 0.5, 512: 1, 1024: 2, 2048: 4}
def main(args):
all_state_dict = torch.load(args.orig_ckpt_path)
state_dict = all_state_dict.pop("state_dict")
converted_state_dict = {}
# Patch embeddings.
converted_state_dict["pos_embed.proj.weight"] = state_dict.pop("x_embedder.proj.weight")
converted_state_dict["pos_embed.proj.bias"] = state_dict.pop("x_embedder.proj.bias")
# Caption projection.
converted_state_dict["caption_projection.linear_1.weight"] = state_dict.pop("y_embedder.y_proj.fc1.weight")
converted_state_dict["caption_projection.linear_1.bias"] = state_dict.pop("y_embedder.y_proj.fc1.bias")
converted_state_dict["caption_projection.linear_2.weight"] = state_dict.pop("y_embedder.y_proj.fc2.weight")
converted_state_dict["caption_projection.linear_2.bias"] = state_dict.pop("y_embedder.y_proj.fc2.bias")
# AdaLN-single LN
converted_state_dict["adaln_single.emb.timestep_embedder.linear_1.weight"] = state_dict.pop(
"t_embedder.mlp.0.weight"
)
converted_state_dict["adaln_single.emb.timestep_embedder.linear_1.bias"] = state_dict.pop("t_embedder.mlp.0.bias")
converted_state_dict["adaln_single.emb.timestep_embedder.linear_2.weight"] = state_dict.pop(
"t_embedder.mlp.2.weight"
)
converted_state_dict["adaln_single.emb.timestep_embedder.linear_2.bias"] = state_dict.pop("t_embedder.mlp.2.bias")
if args.micro_condition:
# Resolution.
converted_state_dict["adaln_single.emb.resolution_embedder.linear_1.weight"] = state_dict.pop(
"csize_embedder.mlp.0.weight"
)
converted_state_dict["adaln_single.emb.resolution_embedder.linear_1.bias"] = state_dict.pop(
"csize_embedder.mlp.0.bias"
)
converted_state_dict["adaln_single.emb.resolution_embedder.linear_2.weight"] = state_dict.pop(
"csize_embedder.mlp.2.weight"
)
converted_state_dict["adaln_single.emb.resolution_embedder.linear_2.bias"] = state_dict.pop(
"csize_embedder.mlp.2.bias"
)
# Aspect ratio.
converted_state_dict["adaln_single.emb.aspect_ratio_embedder.linear_1.weight"] = state_dict.pop(
"ar_embedder.mlp.0.weight"
)
converted_state_dict["adaln_single.emb.aspect_ratio_embedder.linear_1.bias"] = state_dict.pop(
"ar_embedder.mlp.0.bias"
)
converted_state_dict["adaln_single.emb.aspect_ratio_embedder.linear_2.weight"] = state_dict.pop(
"ar_embedder.mlp.2.weight"
)
converted_state_dict["adaln_single.emb.aspect_ratio_embedder.linear_2.bias"] = state_dict.pop(
"ar_embedder.mlp.2.bias"
)
# Shared norm.
converted_state_dict["adaln_single.linear.weight"] = state_dict.pop("t_block.1.weight")
converted_state_dict["adaln_single.linear.bias"] = state_dict.pop("t_block.1.bias")
for depth in range(28):
# Transformer blocks.
converted_state_dict[f"transformer_blocks.{depth}.scale_shift_table"] = state_dict.pop(
f"blocks.{depth}.scale_shift_table"
)
# Attention is all you need 🤘
# Self attention.
q, k, v = torch.chunk(state_dict.pop(f"blocks.{depth}.attn.qkv.weight"), 3, dim=0)
q_bias, k_bias, v_bias = torch.chunk(state_dict.pop(f"blocks.{depth}.attn.qkv.bias"), 3, dim=0)
converted_state_dict[f"transformer_blocks.{depth}.attn1.to_q.weight"] = q
converted_state_dict[f"transformer_blocks.{depth}.attn1.to_q.bias"] = q_bias
converted_state_dict[f"transformer_blocks.{depth}.attn1.to_k.weight"] = k
converted_state_dict[f"transformer_blocks.{depth}.attn1.to_k.bias"] = k_bias
converted_state_dict[f"transformer_blocks.{depth}.attn1.to_v.weight"] = v
converted_state_dict[f"transformer_blocks.{depth}.attn1.to_v.bias"] = v_bias
# Projection.
converted_state_dict[f"transformer_blocks.{depth}.attn1.to_out.0.weight"] = state_dict.pop(
f"blocks.{depth}.attn.proj.weight"
)
converted_state_dict[f"transformer_blocks.{depth}.attn1.to_out.0.bias"] = state_dict.pop(
f"blocks.{depth}.attn.proj.bias"
)
if args.qk_norm:
converted_state_dict[f"transformer_blocks.{depth}.attn1.q_norm.weight"] = state_dict.pop(
f"blocks.{depth}.attn.q_norm.weight"
)
converted_state_dict[f"transformer_blocks.{depth}.attn1.q_norm.bias"] = state_dict.pop(
f"blocks.{depth}.attn.q_norm.bias"
)
converted_state_dict[f"transformer_blocks.{depth}.attn1.k_norm.weight"] = state_dict.pop(
f"blocks.{depth}.attn.k_norm.weight"
)
converted_state_dict[f"transformer_blocks.{depth}.attn1.k_norm.bias"] = state_dict.pop(
f"blocks.{depth}.attn.k_norm.bias"
)
# Feed-forward.
converted_state_dict[f"transformer_blocks.{depth}.ff.net.0.proj.weight"] = state_dict.pop(
f"blocks.{depth}.mlp.fc1.weight"
)
converted_state_dict[f"transformer_blocks.{depth}.ff.net.0.proj.bias"] = state_dict.pop(
f"blocks.{depth}.mlp.fc1.bias"
)
converted_state_dict[f"transformer_blocks.{depth}.ff.net.2.weight"] = state_dict.pop(
f"blocks.{depth}.mlp.fc2.weight"
)
converted_state_dict[f"transformer_blocks.{depth}.ff.net.2.bias"] = state_dict.pop(
f"blocks.{depth}.mlp.fc2.bias"
)
# Cross-attention.
q = state_dict.pop(f"blocks.{depth}.cross_attn.q_linear.weight")
q_bias = state_dict.pop(f"blocks.{depth}.cross_attn.q_linear.bias")
k, v = torch.chunk(state_dict.pop(f"blocks.{depth}.cross_attn.kv_linear.weight"), 2, dim=0)
k_bias, v_bias = torch.chunk(state_dict.pop(f"blocks.{depth}.cross_attn.kv_linear.bias"), 2, dim=0)
converted_state_dict[f"transformer_blocks.{depth}.attn2.to_q.weight"] = q
converted_state_dict[f"transformer_blocks.{depth}.attn2.to_q.bias"] = q_bias
converted_state_dict[f"transformer_blocks.{depth}.attn2.to_k.weight"] = k
converted_state_dict[f"transformer_blocks.{depth}.attn2.to_k.bias"] = k_bias
converted_state_dict[f"transformer_blocks.{depth}.attn2.to_v.weight"] = v
converted_state_dict[f"transformer_blocks.{depth}.attn2.to_v.bias"] = v_bias
converted_state_dict[f"transformer_blocks.{depth}.attn2.to_out.0.weight"] = state_dict.pop(
f"blocks.{depth}.cross_attn.proj.weight"
)
converted_state_dict[f"transformer_blocks.{depth}.attn2.to_out.0.bias"] = state_dict.pop(
f"blocks.{depth}.cross_attn.proj.bias"
)
# Final block.
converted_state_dict["proj_out.weight"] = state_dict.pop("final_layer.linear.weight")
converted_state_dict["proj_out.bias"] = state_dict.pop("final_layer.linear.bias")
converted_state_dict["scale_shift_table"] = state_dict.pop("final_layer.scale_shift_table")
# PixArt XL/2
transformer = Transformer2DModel(
sample_size=args.image_size // 8,
num_layers=28,
attention_head_dim=72,
in_channels=4,
out_channels=8,
patch_size=2,
attention_bias=True,
num_attention_heads=16,
cross_attention_dim=1152,
activation_fn="gelu-approximate",
num_embeds_ada_norm=1000,
norm_type="ada_norm_single",
norm_elementwise_affine=False,
norm_eps=1e-6,
caption_channels=4096,
interpolation_scale=interpolation_scale[args.image_size],
use_additional_conditions=args.micro_condition,
)
transformer.load_state_dict(converted_state_dict, strict=True)
assert transformer.pos_embed.pos_embed is not None
try:
state_dict.pop("y_embedder.y_embedding")
state_dict.pop("pos_embed")
except Exception as e:
print(f"Skipping {str(e)}")
pass
assert len(state_dict) == 0, f"State dict is not empty, {state_dict.keys()}"
num_model_params = sum(p.numel() for p in transformer.parameters())
print(f"Total number of transformer parameters: {num_model_params}")
if args.only_transformer:
transformer.save_pretrained(os.path.join(args.dump_path, "transformer"))
else:
# pixart-Sigma vae link: https://huggingface.co/PixArt-alpha/pixart_sigma_sdxlvae_T5_diffusers/tree/main/vae
vae = AutoencoderKL.from_pretrained(f"{ckpt_id}/pixart_sigma_sdxlvae_T5_diffusers", subfolder="vae")
scheduler = DPMSolverMultistepScheduler()
tokenizer = T5Tokenizer.from_pretrained(f"{ckpt_id}/pixart_sigma_sdxlvae_T5_diffusers", subfolder="tokenizer")
text_encoder = T5EncoderModel.from_pretrained(
f"{ckpt_id}/pixart_sigma_sdxlvae_T5_diffusers", subfolder="text_encoder"
)
pipeline = PixArtSigmaPipeline(
tokenizer=tokenizer, text_encoder=text_encoder, transformer=transformer, vae=vae, scheduler=scheduler
)
pipeline.save_pretrained(args.dump_path)
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument(
"--micro_condition", action="store_true", help="If use Micro-condition in PixArtMS structure during training."
)
parser.add_argument("--qk_norm", action="store_true", help="If use qk norm during training.")
parser.add_argument(
"--orig_ckpt_path", default=None, type=str, required=False, help="Path to the checkpoint to convert."
)
parser.add_argument(
"--image_size",
default=1024,
type=int,
choices=[256, 512, 1024, 2048],
required=False,
help="Image size of pretrained model, 256, 512, 1024, or 2048.",
)
parser.add_argument("--dump_path", default=None, type=str, required=True, help="Path to the output pipeline.")
parser.add_argument("--only_transformer", default=True, type=bool, required=True)
args = parser.parse_args()
main(args)
|