convert_dit_to_diffusers.py

import argparse
import os

import torch
from torchvision.datasets.utils import download_url

from diffusers import AutoencoderKL, DDIMScheduler, DiTPipeline, Transformer2DModel


pretrained_models = {512: "DiT-XL-2-512x512.pt", 256: "DiT-XL-2-256x256.pt"}


def download_model(model_name):
    """
    Downloads a pre-trained DiT model from the web.
    """
    local_path = f"pretrained_models/{model_name}"
    if not os.path.isfile(local_path):
        os.makedirs("pretrained_models", exist_ok=True)
        web_path = f"https://dl.fbaipublicfiles.com/DiT/models/{model_name}"
        download_url(web_path, "pretrained_models")
    model = torch.load(local_path, map_location=lambda storage, loc: storage)
    return model


def main(args):
    state_dict = download_model(pretrained_models[args.image_size])

    state_dict["pos_embed.proj.weight"] = state_dict["x_embedder.proj.weight"]
    state_dict["pos_embed.proj.bias"] = state_dict["x_embedder.proj.bias"]
    state_dict.pop("x_embedder.proj.weight")
    state_dict.pop("x_embedder.proj.bias")

    for depth in range(28):
        state_dict[f"transformer_blocks.{depth}.norm1.emb.timestep_embedder.linear_1.weight"] = state_dict[
            "t_embedder.mlp.0.weight"
        ]
        state_dict[f"transformer_blocks.{depth}.norm1.emb.timestep_embedder.linear_1.bias"] = state_dict[
            "t_embedder.mlp.0.bias"
        ]
        state_dict[f"transformer_blocks.{depth}.norm1.emb.timestep_embedder.linear_2.weight"] = state_dict[
            "t_embedder.mlp.2.weight"
        ]
        state_dict[f"transformer_blocks.{depth}.norm1.emb.timestep_embedder.linear_2.bias"] = state_dict[
            "t_embedder.mlp.2.bias"
        ]
        state_dict[f"transformer_blocks.{depth}.norm1.emb.class_embedder.embedding_table.weight"] = state_dict[
            "y_embedder.embedding_table.weight"
        ]

        state_dict[f"transformer_blocks.{depth}.norm1.linear.weight"] = state_dict[
            f"blocks.{depth}.adaLN_modulation.1.weight"
        ]
        state_dict[f"transformer_blocks.{depth}.norm1.linear.bias"] = state_dict[
            f"blocks.{depth}.adaLN_modulation.1.bias"
        ]

        q, k, v = torch.chunk(state_dict[f"blocks.{depth}.attn.qkv.weight"], 3, dim=0)
        q_bias, k_bias, v_bias = torch.chunk(state_dict[f"blocks.{depth}.attn.qkv.bias"], 3, dim=0)

        state_dict[f"transformer_blocks.{depth}.attn1.to_q.weight"] = q
        state_dict[f"transformer_blocks.{depth}.attn1.to_q.bias"] = q_bias
        state_dict[f"transformer_blocks.{depth}.attn1.to_k.weight"] = k
        state_dict[f"transformer_blocks.{depth}.attn1.to_k.bias"] = k_bias
        state_dict[f"transformer_blocks.{depth}.attn1.to_v.weight"] = v
        state_dict[f"transformer_blocks.{depth}.attn1.to_v.bias"] = v_bias

        state_dict[f"transformer_blocks.{depth}.attn1.to_out.0.weight"] = state_dict[
            f"blocks.{depth}.attn.proj.weight"
        ]
        state_dict[f"transformer_blocks.{depth}.attn1.to_out.0.bias"] = state_dict[f"blocks.{depth}.attn.proj.bias"]

        state_dict[f"transformer_blocks.{depth}.ff.net.0.proj.weight"] = state_dict[f"blocks.{depth}.mlp.fc1.weight"]
        state_dict[f"transformer_blocks.{depth}.ff.net.0.proj.bias"] = state_dict[f"blocks.{depth}.mlp.fc1.bias"]
        state_dict[f"transformer_blocks.{depth}.ff.net.2.weight"] = state_dict[f"blocks.{depth}.mlp.fc2.weight"]
        state_dict[f"transformer_blocks.{depth}.ff.net.2.bias"] = state_dict[f"blocks.{depth}.mlp.fc2.bias"]

        state_dict.pop(f"blocks.{depth}.attn.qkv.weight")
        state_dict.pop(f"blocks.{depth}.attn.qkv.bias")
        state_dict.pop(f"blocks.{depth}.attn.proj.weight")
        state_dict.pop(f"blocks.{depth}.attn.proj.bias")
        state_dict.pop(f"blocks.{depth}.mlp.fc1.weight")
        state_dict.pop(f"blocks.{depth}.mlp.fc1.bias")
        state_dict.pop(f"blocks.{depth}.mlp.fc2.weight")
        state_dict.pop(f"blocks.{depth}.mlp.fc2.bias")
        state_dict.pop(f"blocks.{depth}.adaLN_modulation.1.weight")
        state_dict.pop(f"blocks.{depth}.adaLN_modulation.1.bias")

    state_dict.pop("t_embedder.mlp.0.weight")
    state_dict.pop("t_embedder.mlp.0.bias")
    state_dict.pop("t_embedder.mlp.2.weight")
    state_dict.pop("t_embedder.mlp.2.bias")
    state_dict.pop("y_embedder.embedding_table.weight")

    state_dict["proj_out_1.weight"] = state_dict["final_layer.adaLN_modulation.1.weight"]
    state_dict["proj_out_1.bias"] = state_dict["final_layer.adaLN_modulation.1.bias"]
    state_dict["proj_out_2.weight"] = state_dict["final_layer.linear.weight"]
    state_dict["proj_out_2.bias"] = state_dict["final_layer.linear.bias"]

    state_dict.pop("final_layer.linear.weight")
    state_dict.pop("final_layer.linear.bias")
    state_dict.pop("final_layer.adaLN_modulation.1.weight")
    state_dict.pop("final_layer.adaLN_modulation.1.bias")

    # DiT 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,
        activation_fn="gelu-approximate",
        num_embeds_ada_norm=1000,
        norm_type="ada_norm_zero",
        norm_elementwise_affine=False,
    )
    transformer.load_state_dict(state_dict, strict=True)

    scheduler = DDIMScheduler(
        num_train_timesteps=1000,
        beta_schedule="linear",
        prediction_type="epsilon",
        clip_sample=False,
    )

    vae = AutoencoderKL.from_pretrained(args.vae_model)

    pipeline = DiTPipeline(transformer=transformer, vae=vae, scheduler=scheduler)

    if args.save:
        pipeline.save_pretrained(args.checkpoint_path)


if __name__ == "__main__":
    parser = argparse.ArgumentParser()

    parser.add_argument(
        "--image_size",
        default=256,
        type=int,
        required=False,
        help="Image size of pretrained model, either 256 or 512.",
    )
    parser.add_argument(
        "--vae_model",
        default="stabilityai/sd-vae-ft-ema",
        type=str,
        required=False,
        help="Path to pretrained VAE model, either stabilityai/sd-vae-ft-mse or stabilityai/sd-vae-ft-ema.",
    )
    parser.add_argument(
        "--save", default=True, type=bool, required=False, help="Whether to save the converted pipeline or not."
    )
    parser.add_argument(
        "--checkpoint_path", default=None, type=str, required=True, help="Path to the output pipeline."
    )

    args = parser.parse_args()
    main(args)