LECOっぽいText Encoder only LoRAを学習する検証実装

sdxl_train_te_leco.py

# sdxl_train_network.py と同じ引数を指定してください
# --network_train_text_encoder_only オプションが必須です
#
# 260行目あたりの src_str, tgt_str に変換元と変換先のテキストを指定してください
# tagger の selected_tags.csv が必要ですので、適宜パスを変更してください
# この辺にあります : https://huggingface.co/SmilingWolf/wd-v1-4-swinv2-tagger-v2/tree/main
#
# "1girl" タグだけは必ず含まれる感じにしているので、必要なら 820 行目あたりを適宜変更してください
#
# 以下のオプションは指定できません（エラーになります）：
# --cache_latents
# --cache_latents_to_disk
# --cache_text_encoder_outputs
# --cache_text_encoder_outputs_to_disk
# --sample_every_n_epochs
# --sample_every_n_steps
# --gradient_checkpointing （エラーは出ませんが指定すると学習がうまく行かないかも）
#
# ほかにもエラーになるオプションがあるかもしれませんので、適宜外してください
#
# dataset 設定の .toml は無視されます
# VRAM 使用量を見ながら --train_batch_size を適宜指定してください（わりと大きくできます）

import argparse
import random
import sys
import time
import os
import math
import importlib
import json
import toml
import torch
from tqdm import tqdm
from library.device_utils import init_ipex, clean_memory_on_device

init_ipex()

from library import sdxl_model_util, sdxl_train_util, train_util
import train_network
from library.utils import setup_logging

setup_logging()
import logging

logger = logging.getLogger(__name__)

from accelerate.utils import set_seed
from diffusers import DDPMScheduler
from library import model_util

import library.train_util as train_util
from library.train_util import (
    DreamBoothDataset,
)
import library.config_util as config_util
from library.config_util import (
    ConfigSanitizer,
    BlueprintGenerator,
)
import library.huggingface_util as huggingface_util
import library.custom_train_functions as custom_train_functions
from library.custom_train_functions import (
    apply_snr_weight,
    get_weighted_text_embeddings,
    prepare_scheduler_for_custom_training,
    scale_v_prediction_loss_like_noise_prediction,
    add_v_prediction_like_loss,
    apply_debiased_estimation,
)


class SdxlNetworkTrainer(train_network.NetworkTrainer):
    def __init__(self):
        super().__init__()
        self.vae_scale_factor = sdxl_model_util.VAE_SCALE_FACTOR
        self.is_sdxl = True

    def assert_extra_args(self, args, train_dataset_group):
        super().assert_extra_args(args, train_dataset_group)
        sdxl_train_util.verify_sdxl_training_args(args)

        if args.cache_text_encoder_outputs:
            assert (
                train_dataset_group.is_text_encoder_output_cacheable()
            ), "when caching Text Encoder output, either caption_dropout_rate, shuffle_caption, token_warmup_step or caption_tag_dropout_rate cannot be used / Text Encoderの出力をキャッシュするときはcaption_dropout_rate, shuffle_caption, token_warmup_step, caption_tag_dropout_rateは使えません"

        assert (
            args.network_train_unet_only or not args.cache_text_encoder_outputs
        ), "network for Text Encoder cannot be trained with caching Text Encoder outputs / Text Encoderの出力をキャッシュしながらText Encoderのネットワークを学習することはできません"

        train_dataset_group.verify_bucket_reso_steps(32)

    def load_target_model(self, args, weight_dtype, accelerator):
        (
            load_stable_diffusion_format,
            text_encoder1,
            text_encoder2,
            vae,
            unet,
            logit_scale,
            ckpt_info,
        ) = sdxl_train_util.load_target_model(args, accelerator, sdxl_model_util.MODEL_VERSION_SDXL_BASE_V1_0, weight_dtype)

        self.load_stable_diffusion_format = load_stable_diffusion_format
        self.logit_scale = logit_scale
        self.ckpt_info = ckpt_info

        return sdxl_model_util.MODEL_VERSION_SDXL_BASE_V1_0, [text_encoder1, text_encoder2], vae, unet

    def load_tokenizer(self, args):
        tokenizer = sdxl_train_util.load_tokenizers(args)
        return tokenizer

    def is_text_encoder_outputs_cached(self, args):
        return args.cache_text_encoder_outputs

    def cache_text_encoder_outputs_if_needed(
        self, args, accelerator, unet, vae, tokenizers, text_encoders, dataset: train_util.DatasetGroup, weight_dtype
    ):
        if args.cache_text_encoder_outputs:
            if not args.lowram:
                # メモリ消費を減らす
                logger.info("move vae and unet to cpu to save memory")
                org_vae_device = vae.device
                org_unet_device = unet.device
                vae.to("cpu")
                unet.to("cpu")
                clean_memory_on_device(accelerator.device)

            # When TE is not be trained, it will not be prepared so we need to use explicit autocast
            with accelerator.autocast():
                dataset.cache_text_encoder_outputs(
                    tokenizers,
                    text_encoders,
                    accelerator.device,
                    weight_dtype,
                    args.cache_text_encoder_outputs_to_disk,
                    accelerator.is_main_process,
                )

            text_encoders[0].to("cpu", dtype=torch.float32)  # Text Encoder doesn't work with fp16 on CPU
            text_encoders[1].to("cpu", dtype=torch.float32)
            clean_memory_on_device(accelerator.device)

            if not args.lowram:
                logger.info("move vae and unet back to original device")
                vae.to(org_vae_device)
                unet.to(org_unet_device)
        else:
            # Text Encoderから毎回出力を取得するので、GPUに乗せておく
            text_encoders[0].to(accelerator.device, dtype=weight_dtype)
            text_encoders[1].to(accelerator.device, dtype=weight_dtype)

    def get_text_cond(self, args, accelerator, batch, tokenizers, text_encoders, weight_dtype):
        if "text_encoder_outputs1_list" not in batch or batch["text_encoder_outputs1_list"] is None:
            input_ids1 = batch["input_ids"]
            input_ids2 = batch["input_ids2"]
            # with torch.enable_grad():
            # Get the text embedding for conditioning
            # TODO support weighted captions
            # if args.weighted_captions:
            #     encoder_hidden_states = get_weighted_text_embeddings(
            #         tokenizer,
            #         text_encoder,
            #         batch["captions"],
            #         accelerator.device,
            #         args.max_token_length // 75 if args.max_token_length else 1,
            #         clip_skip=args.clip_skip,
            #     )
            # else:
            input_ids1 = input_ids1.to(accelerator.device)
            input_ids2 = input_ids2.to(accelerator.device)
            encoder_hidden_states1, encoder_hidden_states2, pool2 = train_util.get_hidden_states_sdxl(
                args.max_token_length,
                input_ids1,
                input_ids2,
                tokenizers[0],
                tokenizers[1],
                text_encoders[0],
                text_encoders[1],
                None if not args.full_fp16 else weight_dtype,
                accelerator=accelerator,
            )
        else:
            raise NotImplementedError("text_encoder_outputs1_list is not implemented yet")
            encoder_hidden_states1 = batch["text_encoder_outputs1_list"].to(accelerator.device).to(weight_dtype)
            encoder_hidden_states2 = batch["text_encoder_outputs2_list"].to(accelerator.device).to(weight_dtype)
            pool2 = batch["text_encoder_pool2_list"].to(accelerator.device).to(weight_dtype)

            # # verify that the text encoder outputs are correct
            # ehs1, ehs2, p2 = train_util.get_hidden_states_sdxl(
            #     args.max_token_length,
            #     batch["input_ids"].to(text_encoders[0].device),
            #     batch["input_ids2"].to(text_encoders[0].device),
            #     tokenizers[0],
            #     tokenizers[1],
            #     text_encoders[0],
            #     text_encoders[1],
            #     None if not args.full_fp16 else weight_dtype,
            # )
            # b_size = encoder_hidden_states1.shape[0]
            # assert ((encoder_hidden_states1.to("cpu") - ehs1.to(dtype=weight_dtype)).abs().max() > 1e-2).sum() <= b_size * 2
            # assert ((encoder_hidden_states2.to("cpu") - ehs2.to(dtype=weight_dtype)).abs().max() > 1e-2).sum() <= b_size * 2
            # assert ((pool2.to("cpu") - p2.to(dtype=weight_dtype)).abs().max() > 1e-2).sum() <= b_size * 2
            # logger.info("text encoder outputs verified")

        return encoder_hidden_states1, encoder_hidden_states2, pool2

    def call_unet(self, args, accelerator, unet, noisy_latents, timesteps, text_conds, batch, weight_dtype):
        noisy_latents = noisy_latents.to(weight_dtype)  # TODO check why noisy_latents is not weight_dtype

        # get size embeddings
        orig_size = batch["original_sizes_hw"]
        crop_size = batch["crop_top_lefts"]
        target_size = batch["target_sizes_hw"]
        embs = sdxl_train_util.get_size_embeddings(orig_size, crop_size, target_size, accelerator.device).to(weight_dtype)

        # concat embeddings
        encoder_hidden_states1, encoder_hidden_states2, pool2 = text_conds
        vector_embedding = torch.cat([pool2, embs], dim=1).to(weight_dtype)
        text_embedding = torch.cat([encoder_hidden_states1, encoder_hidden_states2], dim=2).to(weight_dtype)

        noise_pred = unet(noisy_latents, timesteps, text_embedding, vector_embedding)
        return noise_pred

    def sample_images(self, accelerator, args, epoch, global_step, device, vae, tokenizer, text_encoder, unet):
        sdxl_train_util.sample_images(accelerator, args, epoch, global_step, device, vae, tokenizer, text_encoder, unet)

    def train(self, args):
        assert args.network_train_text_encoder_only

        session_id = random.randint(0, 2**32)
        training_started_at = time.time()
        train_util.verify_training_args(args)
        train_util.prepare_dataset_args(args, True)
        setup_logging(args, reset=True)

        cache_latents = args.cache_latents
        use_dreambooth_method = args.in_json is None
        use_user_config = args.dataset_config is not None

        if args.seed is None:
            args.seed = random.randint(0, 2**32)
        train_network.set_seed(args.seed)

        # tokenizerは単体またはリスト、tokenizersは必ずリスト：既存のコードとの互換性のため
        tokenizer = self.load_tokenizer(args)
        tokenizers = tokenizer if isinstance(tokenizer, list) else [tokenizer]

        # datasetを準備する：単なるテキストなのでPyTorchのDatasetは使わないでOK

        # load tags from csv
        csv_file = "./wd14_tagger_model/selected_tags.csv"
        tags = []
        with open(csv_file, "r", encoding="utf-8") as f:
            lines = f.readlines()
        for line in lines[6:]:  # rating tag と 1girl を外す
            tags.append(line.split(",")[1].replace("_", " ").strip())
        print(f"tags loaded: {len(tags)}, {tags[:5]}")

        # srcをtgtに変換するように学習する：たぶんトークン長が同じでないと難しい
        src_str = "hatsune miku"
        tgt_str = "miki sayaka"
        num_steps_per_epoch = 100  # 適当に変えてね

        # acceleratorを準備する
        logger.info("preparing accelerator")
        accelerator = train_util.prepare_accelerator(args)
        is_main_process = accelerator.is_main_process

        # mixed precisionに対応した型を用意しておき適宜castする
        weight_dtype, save_dtype = train_util.prepare_dtype(args)
        vae_dtype = torch.float32 if args.no_half_vae else weight_dtype

        # モデルを読み込む
        model_version, text_encoder, vae, unet = self.load_target_model(args, weight_dtype, accelerator)

        # text_encoder is List[CLIPTextModel] or CLIPTextModel
        text_encoders = text_encoder if isinstance(text_encoder, list) else [text_encoder]

        # モデルに xformers とか memory efficient attention を組み込む
        train_util.replace_unet_modules(unet, args.mem_eff_attn, args.xformers, args.sdpa)
        if torch.__version__ >= "2.0.0":  # PyTorch 2.0.0 以上対応のxformersなら以下が使える
            vae.set_use_memory_efficient_attention_xformers(args.xformers)

        # 差分追加学習のためにモデルを読み込む
        sys.path.append(os.path.dirname(__file__))
        accelerator.print("import network module:", args.network_module)
        network_module = importlib.import_module(args.network_module)

        if args.base_weights is not None:
            # base_weights が指定されている場合は、指定された重みを読み込みマージする
            for i, weight_path in enumerate(args.base_weights):
                if args.base_weights_multiplier is None or len(args.base_weights_multiplier) <= i:
                    multiplier = 1.0
                else:
                    multiplier = args.base_weights_multiplier[i]

                accelerator.print(f"merging module: {weight_path} with multiplier {multiplier}")

                module, weights_sd = network_module.create_network_from_weights(
                    multiplier, weight_path, vae, text_encoder, unet, for_inference=True
                )
                module.merge_to(text_encoder, unet, weights_sd, weight_dtype, accelerator.device if args.lowram else "cpu")

            accelerator.print(f"all weights merged: {', '.join(args.base_weights)}")

        # 学習を準備する
        if cache_latents:
            vae.to(accelerator.device, dtype=vae_dtype)
            vae.requires_grad_(False)
            vae.eval()
            with torch.no_grad():
                train_dataset_group.cache_latents(vae, args.vae_batch_size, args.cache_latents_to_disk, accelerator.is_main_process)
            vae.to("cpu")
            clean_memory_on_device(accelerator.device)

            accelerator.wait_for_everyone()

        # 必要ならテキストエンコーダーの出力をキャッシュする: Text Encoderはcpuまたはgpuへ移される
        # cache text encoder outputs if needed: Text Encoder is moved to cpu or gpu
        # self.cache_text_encoder_outputs_if_needed(
        #     args, accelerator, unet, vae, tokenizers, text_encoders, train_dataset_group, weight_dtype
        # )

        # prepare network
        net_kwargs = {}
        if args.network_args is not None:
            for net_arg in args.network_args:
                key, value = net_arg.split("=")
                net_kwargs[key] = value

        # if a new network is added in future, add if ~ then blocks for each network (;'∀')
        if args.dim_from_weights:
            network, _ = network_module.create_network_from_weights(1, args.network_weights, vae, text_encoder, unet, **net_kwargs)
        else:
            if "dropout" not in net_kwargs:
                # workaround for LyCORIS (;^ω^)
                net_kwargs["dropout"] = args.network_dropout

            network = network_module.create_network(
                1.0,
                args.network_dim,
                args.network_alpha,
                vae,
                text_encoder,
                unet,
                neuron_dropout=args.network_dropout,
                **net_kwargs,
            )
        if network is None:
            return
        network_has_multiplier = hasattr(network, "set_multiplier")

        if hasattr(network, "prepare_network"):
            network.prepare_network(args)
        if args.scale_weight_norms and not hasattr(network, "apply_max_norm_regularization"):
            logger.warning(
                "warning: scale_weight_norms is specified but the network does not support it / scale_weight_normsが指定されていますが、ネットワークが対応していません"
            )
            args.scale_weight_norms = False

        train_unet = not args.network_train_text_encoder_only
        train_text_encoder = self.is_train_text_encoder(args)
        network.apply_to(text_encoder, unet, train_text_encoder, train_unet)

        if args.network_weights is not None:
            info = network.load_weights(args.network_weights)
            accelerator.print(f"load network weights from {args.network_weights}: {info}")

        if args.gradient_checkpointing:
            unet.enable_gradient_checkpointing()
            for t_enc in text_encoders:
                t_enc.gradient_checkpointing_enable()
            del t_enc
            network.enable_gradient_checkpointing()  # may have no effect

        # 学習に必要なクラスを準備する
        accelerator.print("prepare optimizer, data loader etc.")

        # 後方互換性を確保するよ
        try:
            trainable_params = network.prepare_optimizer_params(args.text_encoder_lr, args.unet_lr, args.learning_rate)
        except TypeError:
            accelerator.print(
                "Deprecated: use prepare_optimizer_params(text_encoder_lr, unet_lr, learning_rate) instead of prepare_optimizer_params(text_encoder_lr, unet_lr)"
            )
            trainable_params = network.prepare_optimizer_params(args.text_encoder_lr, args.unet_lr)

        optimizer_name, optimizer_args, optimizer = train_util.get_optimizer(args, trainable_params)

        # dataloaderを準備する
        # DataLoaderのプロセス数：0 は persistent_workers が使えないので注意
        n_workers = min(args.max_data_loader_n_workers, os.cpu_count())  # cpu_count or max_data_loader_n_workers

        # train_dataloader = torch.utils.data.DataLoader(
        #     train_dataset_group,
        #     batch_size=1,
        #     shuffle=True,
        #     collate_fn=collator,
        #     num_workers=n_workers,
        #     persistent_workers=args.persistent_data_loader_workers,
        # )

        # 学習ステップ数を計算する
        if args.max_train_epochs is not None:
            args.max_train_steps = args.max_train_epochs * math.ceil(
                num_steps_per_epoch / accelerator.num_processes / args.gradient_accumulation_steps
            )
            accelerator.print(
                f"override steps. steps for {args.max_train_epochs} epochs is / 指定エポックまでのステップ数: {args.max_train_steps}"
            )

        # データセット側にも学習ステップを送信
        # train_dataset_group.set_max_train_steps(args.max_train_steps)

        # lr schedulerを用意する
        lr_scheduler = train_util.get_scheduler_fix(args, optimizer, accelerator.num_processes)

        # 実験的機能：勾配も含めたfp16/bf16学習を行う　モデル全体をfp16/bf16にする
        if args.full_fp16:
            assert (
                args.mixed_precision == "fp16"
            ), "full_fp16 requires mixed precision='fp16' / full_fp16を使う場合はmixed_precision='fp16'を指定してください。"
            accelerator.print("enable full fp16 training.")
            network.to(weight_dtype)
        elif args.full_bf16:
            assert (
                args.mixed_precision == "bf16"
            ), "full_bf16 requires mixed precision='bf16' / full_bf16を使う場合はmixed_precision='bf16'を指定してください。"
            accelerator.print("enable full bf16 training.")
            network.to(weight_dtype)

        unet_weight_dtype = te_weight_dtype = weight_dtype
        # Experimental Feature: Put base model into fp8 to save vram
        if args.fp8_base:
            assert torch.__version__ >= "2.1.0", "fp8_base requires torch>=2.1.0 / fp8を使う場合はtorch>=2.1.0が必要です。"
            assert (
                args.mixed_precision != "no"
            ), "fp8_base requires mixed precision='fp16' or 'bf16' / fp8を使う場合はmixed_precision='fp16'または'bf16'が必要です。"
            accelerator.print("enable fp8 training.")
            unet_weight_dtype = torch.float8_e4m3fn
            te_weight_dtype = torch.float8_e4m3fn

        unet.requires_grad_(False)
        unet.to(dtype=unet_weight_dtype)
        for t_enc in text_encoders:
            t_enc.requires_grad_(False)

            # in case of cpu, dtype is already set to fp32 because cpu does not support fp8/fp16/bf16
            if t_enc.device.type != "cpu":
                t_enc.to(dtype=te_weight_dtype)
                # nn.Embedding not support FP8
                t_enc.text_model.embeddings.to(dtype=(weight_dtype if te_weight_dtype != weight_dtype else te_weight_dtype))

        # acceleratorがなんかよろしくやってくれるらしい / accelerator will do something good
        if train_unet:
            unet = accelerator.prepare(unet)
        else:
            unet.to(accelerator.device, dtype=unet_weight_dtype)  # move to device because unet is not prepared by accelerator
        if train_text_encoder:
            if len(text_encoders) > 1:
                text_encoder = text_encoders = [accelerator.prepare(t_enc) for t_enc in text_encoders]
            else:
                text_encoder = accelerator.prepare(text_encoder)
                text_encoders = [text_encoder]
        else:
            pass  # if text_encoder is not trained, no need to prepare. and device and dtype are already set

        # network, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(network, optimizer, train_dataloader, lr_scheduler)
        network, optimizer, lr_scheduler = accelerator.prepare(network, optimizer, lr_scheduler)

        if args.gradient_checkpointing:
            # according to TI example in Diffusers, train is required
            unet.train()
            for t_enc in text_encoders:
                t_enc.train()

                # set top parameter requires_grad = True for gradient checkpointing works
                if train_text_encoder:
                    t_enc.text_model.embeddings.requires_grad_(True)

        else:
            unet.eval()
            for t_enc in text_encoders:
                t_enc.eval()

        del t_enc

        accelerator.unwrap_model(network).prepare_grad_etc(text_encoder, unet)

        if not cache_latents:  # キャッシュしない場合はVAEを使うのでVAEを準備する
            vae.requires_grad_(False)
            vae.eval()
            vae.to(accelerator.device, dtype=vae_dtype)

        # 実験的機能：勾配も含めたfp16学習を行う　PyTorchにパッチを当ててfp16でのgrad scaleを有効にする
        if args.full_fp16:
            train_util.patch_accelerator_for_fp16_training(accelerator)

        # resumeする
        train_util.resume_from_local_or_hf_if_specified(accelerator, args)

        # epoch数を計算する
        num_update_steps_per_epoch = math.ceil(num_steps_per_epoch / args.gradient_accumulation_steps)
        num_train_epochs = math.ceil(args.max_train_steps / num_update_steps_per_epoch)
        if (args.save_n_epoch_ratio is not None) and (args.save_n_epoch_ratio > 0):
            args.save_every_n_epochs = math.floor(num_train_epochs / args.save_n_epoch_ratio) or 1

        # 学習する
        # TODO: find a way to handle total batch size when there are multiple datasets
        total_batch_size = args.train_batch_size * accelerator.num_processes * args.gradient_accumulation_steps

        accelerator.print("running training / 学習開始")
        # accelerator.print(f"  num train images * repeats / 学習画像の数×繰り返し回数: {train_dataset_group.num_train_images}")
        # accelerator.print(f"  num reg images / 正則化画像の数: {train_dataset_group.num_reg_images}")
        accelerator.print(f"  num batches per epoch / 1epochのバッチ数: {num_steps_per_epoch}")
        accelerator.print(f"  num epochs / epoch数: {num_train_epochs}")
        # accelerator.print(
        #     f"  batch size per device / バッチサイズ: {', '.join([str(d.batch_size) for d in train_dataset_group.datasets])}"
        # )
        # accelerator.print(f"  total train batch size (with parallel & distributed & accumulation) / 総バッチサイズ（並列学習、勾配合計含む）: {total_batch_size}")
        accelerator.print(f"  gradient accumulation steps / 勾配を合計するステップ数 = {args.gradient_accumulation_steps}")
        accelerator.print(f"  total optimization steps / 学習ステップ数: {args.max_train_steps}")

        # TODO refactor metadata creation and move to util
        metadata = {
            "ss_session_id": session_id,  # random integer indicating which group of epochs the model came from
            "ss_training_started_at": training_started_at,  # unix timestamp
            "ss_output_name": args.output_name,
            "ss_learning_rate": args.learning_rate,
            "ss_text_encoder_lr": args.text_encoder_lr,
            "ss_unet_lr": args.unet_lr,
            # "ss_num_train_images": train_dataset_group.num_train_images,
            # "ss_num_reg_images": train_dataset_group.num_reg_images,
            "ss_num_batches_per_epoch": num_steps_per_epoch,
            "ss_num_epochs": num_train_epochs,
            "ss_gradient_checkpointing": args.gradient_checkpointing,
            "ss_gradient_accumulation_steps": args.gradient_accumulation_steps,
            "ss_max_train_steps": args.max_train_steps,
            "ss_lr_warmup_steps": args.lr_warmup_steps,
            "ss_lr_scheduler": args.lr_scheduler,
            "ss_network_module": args.network_module,
            "ss_network_dim": args.network_dim,  # None means default because another network than LoRA may have another default dim
            "ss_network_alpha": args.network_alpha,  # some networks may not have alpha
            "ss_network_dropout": args.network_dropout,  # some networks may not have dropout
            "ss_mixed_precision": args.mixed_precision,
            "ss_full_fp16": bool(args.full_fp16),
            "ss_v2": bool(args.v2),
            "ss_base_model_version": model_version,
            "ss_clip_skip": args.clip_skip,
            "ss_max_token_length": args.max_token_length,
            "ss_cache_latents": bool(args.cache_latents),
            "ss_seed": args.seed,
            "ss_lowram": args.lowram,
            "ss_noise_offset": args.noise_offset,
            "ss_multires_noise_iterations": args.multires_noise_iterations,
            "ss_multires_noise_discount": args.multires_noise_discount,
            "ss_adaptive_noise_scale": args.adaptive_noise_scale,
            "ss_zero_terminal_snr": args.zero_terminal_snr,
            "ss_training_comment": args.training_comment,  # will not be updated after training
            "ss_sd_scripts_commit_hash": train_util.get_git_revision_hash(),
            "ss_optimizer": optimizer_name + (f"({optimizer_args})" if len(optimizer_args) > 0 else ""),
            "ss_max_grad_norm": args.max_grad_norm,
            "ss_caption_dropout_rate": args.caption_dropout_rate,
            "ss_caption_dropout_every_n_epochs": args.caption_dropout_every_n_epochs,
            "ss_caption_tag_dropout_rate": args.caption_tag_dropout_rate,
            "ss_face_crop_aug_range": args.face_crop_aug_range,
            "ss_prior_loss_weight": args.prior_loss_weight,
            "ss_min_snr_gamma": args.min_snr_gamma,
            "ss_scale_weight_norms": args.scale_weight_norms,
            "ss_ip_noise_gamma": args.ip_noise_gamma,
            "ss_debiased_estimation": bool(args.debiased_estimation_loss),
        }

        """
        if use_user_config:
            # save metadata of multiple datasets
            # NOTE: pack "ss_datasets" value as json one time
            #   or should also pack nested collections as json?
            datasets_metadata = []
            tag_frequency = {}  # merge tag frequency for metadata editor
            dataset_dirs_info = {}  # merge subset dirs for metadata editor

            for dataset in train_dataset_group.datasets:
                is_dreambooth_dataset = isinstance(dataset, DreamBoothDataset)
                dataset_metadata = {
                    "is_dreambooth": is_dreambooth_dataset,
                    "batch_size_per_device": dataset.batch_size,
                    "num_train_images": dataset.num_train_images,  # includes repeating
                    "num_reg_images": dataset.num_reg_images,
                    "resolution": (dataset.width, dataset.height),
                    "enable_bucket": bool(dataset.enable_bucket),
                    "min_bucket_reso": dataset.min_bucket_reso,
                    "max_bucket_reso": dataset.max_bucket_reso,
                    "tag_frequency": dataset.tag_frequency,
                    "bucket_info": dataset.bucket_info,
                }

                subsets_metadata = []
                for subset in dataset.subsets:
                    subset_metadata = {
                        "img_count": subset.img_count,
                        "num_repeats": subset.num_repeats,
                        "color_aug": bool(subset.color_aug),
                        "flip_aug": bool(subset.flip_aug),
                        "random_crop": bool(subset.random_crop),
                        "shuffle_caption": bool(subset.shuffle_caption),
                        "keep_tokens": subset.keep_tokens,
                        "keep_tokens_separator": subset.keep_tokens_separator,
                        "secondary_separator": subset.secondary_separator,
                        "enable_wildcard": bool(subset.enable_wildcard),
                        "caption_prefix": subset.caption_prefix,
                        "caption_suffix": subset.caption_suffix,
                    }

                    image_dir_or_metadata_file = None
                    if subset.image_dir:
                        image_dir = os.path.basename(subset.image_dir)
                        subset_metadata["image_dir"] = image_dir
                        image_dir_or_metadata_file = image_dir

                    if is_dreambooth_dataset:
                        subset_metadata["class_tokens"] = subset.class_tokens
                        subset_metadata["is_reg"] = subset.is_reg
                        if subset.is_reg:
                            image_dir_or_metadata_file = None  # not merging reg dataset
                    else:
                        metadata_file = os.path.basename(subset.metadata_file)
                        subset_metadata["metadata_file"] = metadata_file
                        image_dir_or_metadata_file = metadata_file  # may overwrite

                    subsets_metadata.append(subset_metadata)

                    # merge dataset dir: not reg subset only
                    # TODO update additional-network extension to show detailed dataset config from metadata
                    if image_dir_or_metadata_file is not None:
                        # datasets may have a certain dir multiple times
                        v = image_dir_or_metadata_file
                        i = 2
                        while v in dataset_dirs_info:
                            v = image_dir_or_metadata_file + f" ({i})"
                            i += 1
                        image_dir_or_metadata_file = v

                        dataset_dirs_info[image_dir_or_metadata_file] = {
                            "n_repeats": subset.num_repeats,
                            "img_count": subset.img_count,
                        }

                dataset_metadata["subsets"] = subsets_metadata
                datasets_metadata.append(dataset_metadata)

                # merge tag frequency:
                for ds_dir_name, ds_freq_for_dir in dataset.tag_frequency.items():
                    # あるディレクトリが複数のdatasetで使用されている場合、一度だけ数える
                    # もともと繰り返し回数を指定しているので、キャプション内でのタグの出現回数と、それが学習で何度使われるかは一致しない
                    # なので、ここで複数datasetの回数を合算してもあまり意味はない
                    if ds_dir_name in tag_frequency:
                        continue
                    tag_frequency[ds_dir_name] = ds_freq_for_dir

            metadata["ss_datasets"] = json.dumps(datasets_metadata)
            metadata["ss_tag_frequency"] = json.dumps(tag_frequency)
            metadata["ss_dataset_dirs"] = json.dumps(dataset_dirs_info)
        else:
            # conserving backward compatibility when using train_dataset_dir and reg_dataset_dir
            assert (
                len(train_dataset_group.datasets) == 1
            ), f"There should be a single dataset but {len(train_dataset_group.datasets)} found. This seems to be a bug. / データセットは1個だけ存在するはずですが、実際には{len(train_dataset_group.datasets)}個でした。プログラムのバグかもしれません。"

            dataset = train_dataset_group.datasets[0]

            dataset_dirs_info = {}
            reg_dataset_dirs_info = {}
            if use_dreambooth_method:
                for subset in dataset.subsets:
                    info = reg_dataset_dirs_info if subset.is_reg else dataset_dirs_info
                    info[os.path.basename(subset.image_dir)] = {"n_repeats": subset.num_repeats, "img_count": subset.img_count}
            else:
                for subset in dataset.subsets:
                    dataset_dirs_info[os.path.basename(subset.metadata_file)] = {
                        "n_repeats": subset.num_repeats,
                        "img_count": subset.img_count,
                    }

            metadata.update(
                {
                    "ss_batch_size_per_device": args.train_batch_size,
                    "ss_total_batch_size": total_batch_size,
                    "ss_resolution": args.resolution,
                    "ss_color_aug": bool(args.color_aug),
                    "ss_flip_aug": bool(args.flip_aug),
                    "ss_random_crop": bool(args.random_crop),
                    "ss_shuffle_caption": bool(args.shuffle_caption),
                    "ss_enable_bucket": bool(dataset.enable_bucket),
                    "ss_bucket_no_upscale": bool(dataset.bucket_no_upscale),
                    "ss_min_bucket_reso": dataset.min_bucket_reso,
                    "ss_max_bucket_reso": dataset.max_bucket_reso,
                    "ss_keep_tokens": args.keep_tokens,
                    "ss_dataset_dirs": json.dumps(dataset_dirs_info),
                    "ss_reg_dataset_dirs": json.dumps(reg_dataset_dirs_info),
                    "ss_tag_frequency": json.dumps(dataset.tag_frequency),
                    "ss_bucket_info": json.dumps(dataset.bucket_info),
                }
            )
        """

        # add extra args
        if args.network_args:
            metadata["ss_network_args"] = json.dumps(net_kwargs)

        # model name and hash
        if args.pretrained_model_name_or_path is not None:
            sd_model_name = args.pretrained_model_name_or_path
            if os.path.exists(sd_model_name):
                metadata["ss_sd_model_hash"] = train_util.model_hash(sd_model_name)
                metadata["ss_new_sd_model_hash"] = train_util.calculate_sha256(sd_model_name)
                sd_model_name = os.path.basename(sd_model_name)
            metadata["ss_sd_model_name"] = sd_model_name

        if args.vae is not None:
            vae_name = args.vae
            if os.path.exists(vae_name):
                metadata["ss_vae_hash"] = train_util.model_hash(vae_name)
                metadata["ss_new_vae_hash"] = train_util.calculate_sha256(vae_name)
                vae_name = os.path.basename(vae_name)
            metadata["ss_vae_name"] = vae_name

        metadata = {k: str(v) for k, v in metadata.items()}

        # make minimum metadata for filtering
        minimum_metadata = {}
        for key in train_util.SS_METADATA_MINIMUM_KEYS:
            if key in metadata:
                minimum_metadata[key] = metadata[key]

        progress_bar = tqdm(range(args.max_train_steps), smoothing=0, disable=not accelerator.is_local_main_process, desc="steps")
        global_step = 0

        noise_scheduler = DDPMScheduler(
            beta_start=0.00085, beta_end=0.012, beta_schedule="scaled_linear", num_train_timesteps=1000, clip_sample=False
        )
        prepare_scheduler_for_custom_training(noise_scheduler, accelerator.device)
        if args.zero_terminal_snr:
            custom_train_functions.fix_noise_scheduler_betas_for_zero_terminal_snr(noise_scheduler)

        if accelerator.is_main_process:
            init_kwargs = {}
            if args.wandb_run_name:
                init_kwargs["wandb"] = {"name": args.wandb_run_name}
            if args.log_tracker_config is not None:
                init_kwargs = toml.load(args.log_tracker_config)
            accelerator.init_trackers(
                "network_train" if args.log_tracker_name is None else args.log_tracker_name, init_kwargs=init_kwargs
            )

        loss_recorder = train_util.LossRecorder()
        # del train_dataset_group

        # callback for step start
        if hasattr(accelerator.unwrap_model(network), "on_step_start"):
            on_step_start = accelerator.unwrap_model(network).on_step_start
        else:
            on_step_start = lambda *args, **kwargs: None

        # function for saving/removing
        def save_model(ckpt_name, unwrapped_nw, steps, epoch_no, force_sync_upload=False):
            os.makedirs(args.output_dir, exist_ok=True)
            ckpt_file = os.path.join(args.output_dir, ckpt_name)

            accelerator.print(f"\nsaving checkpoint: {ckpt_file}")
            metadata["ss_training_finished_at"] = str(time.time())
            metadata["ss_steps"] = str(steps)
            metadata["ss_epoch"] = str(epoch_no)

            metadata_to_save = minimum_metadata if args.no_metadata else metadata
            sai_metadata = train_util.get_sai_model_spec(None, args, self.is_sdxl, True, False)
            metadata_to_save.update(sai_metadata)

            unwrapped_nw.save_weights(ckpt_file, save_dtype, metadata_to_save)
            if args.huggingface_repo_id is not None:
                huggingface_util.upload(args, ckpt_file, "/" + ckpt_name, force_sync_upload=force_sync_upload)

        def remove_model(old_ckpt_name):
            old_ckpt_file = os.path.join(args.output_dir, old_ckpt_name)
            if os.path.exists(old_ckpt_file):
                accelerator.print(f"removing old checkpoint: {old_ckpt_file}")
                os.remove(old_ckpt_file)

        # For --sample_at_first
        self.sample_images(accelerator, args, 0, global_step, accelerator.device, vae, tokenizer, text_encoder, unet)

        # training loop
        for epoch in range(num_train_epochs):
            accelerator.print(f"\nepoch {epoch+1}/{num_train_epochs}")
            # current_epoch.value = epoch + 1

            metadata["ss_epoch"] = str(epoch + 1)

            accelerator.unwrap_model(network).on_epoch_start(text_encoder, unet)

            # for step, batch in enumerate(train_dataloader):
            for step in range(num_steps_per_epoch):
                # current_step.value = global_step
                with accelerator.accumulate(network):
                    on_step_start(text_encoder, unet)

                    # make prompt
                    # Danbooru Tag と対象文字列を含むランダムなプロンプトを作成する
                    src_ids_list1 = []
                    src_ids_list2 = []
                    tgt_ids_list1 = []
                    tgt_ids_list2 = []
                    for bi in range(args.train_batch_size):
                        tag_count = random.randint(1, 20)
                        selected_tags = random.sample(tags, tag_count)
                        prompt_tags = ["1girl", src_str] + selected_tags  # 必要ならここも変える
                        random.shuffle(prompt_tags)  # selected_tags だけ shuffle する感じにすると実際のプロンプトに近くなるかも

                        src_prompt = ", ".join(prompt_tags)
                        tgt_prompt = ", ".join(prompt_tags).replace(src_str, tgt_str)
                        # print(f"src: {src_prompt}")
                        # print(f"tgt: {tgt_prompt}")

                        src_ids1 = (
                            tokenizers[0].encode(src_prompt, return_tensors="pt", truncation=True, padding="max_length").to("cpu")
                        )
                        src_ids2 = (
                            tokenizers[1].encode(src_prompt, return_tensors="pt", truncation=True, padding="max_length").to("cpu")
                        )
                        tgt_ids1 = (
                            tokenizers[0].encode(tgt_prompt, return_tensors="pt", truncation=True, padding="max_length").to("cpu")
                        )
                        tgt_ids2 = (
                            tokenizers[1].encode(tgt_prompt, return_tensors="pt", truncation=True, padding="max_length").to("cpu")
                        )

                        src_ids_list1.append(src_ids1)
                        src_ids_list2.append(src_ids2)
                        tgt_ids_list1.append(tgt_ids1)
                        tgt_ids_list2.append(tgt_ids2)

                    src_prompt_ids1 = torch.stack(src_ids_list1)
                    src_prompt_ids2 = torch.stack(src_ids_list2)
                    tgt_prompt_ids1 = torch.stack(tgt_ids_list1)
                    tgt_prompt_ids2 = torch.stack(tgt_ids_list2)

                    with torch.no_grad():
                        # Get the text embedding for conditioning
                        batch = {"input_ids": tgt_prompt_ids1, "input_ids2": tgt_prompt_ids2}
                        tgt_text_encoder_conds = self.get_text_cond(
                            args, accelerator, batch, tokenizers, text_encoders, weight_dtype
                        )
                    with torch.set_grad_enabled(True), accelerator.autocast():
                        batch = {"input_ids": src_prompt_ids1, "input_ids2": src_prompt_ids2}
                        src_text_encoder_conds = self.get_text_cond(
                            args, accelerator, batch, tokenizers, text_encoders, weight_dtype
                        )

                    src_emb1, src_emb2, src_pool = src_text_encoder_conds
                    tgt_emb1, tgt_emb2, tgt_pool = tgt_text_encoder_conds

                    loss1 = torch.nn.functional.mse_loss(src_emb1, tgt_emb1)
                    loss2 = torch.nn.functional.mse_loss(src_emb2, tgt_emb2)
                    loss_pool = torch.nn.functional.mse_loss(src_pool, tgt_pool)

                    # 必要ならここで重みづけする
                    loss = loss1 + loss2 + loss_pool

                    accelerator.backward(loss)
                    if accelerator.sync_gradients:
                        self.all_reduce_network(accelerator, network)  # sync DDP grad manually
                        if args.max_grad_norm != 0.0:
                            params_to_clip = accelerator.unwrap_model(network).get_trainable_params()
                            accelerator.clip_grad_norm_(params_to_clip, args.max_grad_norm)

                    optimizer.step()
                    lr_scheduler.step()
                    optimizer.zero_grad(set_to_none=True)

                if args.scale_weight_norms:
                    keys_scaled, mean_norm, maximum_norm = accelerator.unwrap_model(network).apply_max_norm_regularization(
                        args.scale_weight_norms, accelerator.device
                    )
                    max_mean_logs = {"Keys Scaled": keys_scaled, "Average key norm": mean_norm}
                else:
                    keys_scaled, mean_norm, maximum_norm = None, None, None

                # Checks if the accelerator has performed an optimization step behind the scenes
                if accelerator.sync_gradients:
                    progress_bar.update(1)
                    global_step += 1

                    self.sample_images(accelerator, args, None, global_step, accelerator.device, vae, tokenizer, text_encoder, unet)

                    # 指定ステップごとにモデルを保存
                    if args.save_every_n_steps is not None and global_step % args.save_every_n_steps == 0:
                        accelerator.wait_for_everyone()
                        if accelerator.is_main_process:
                            ckpt_name = train_util.get_step_ckpt_name(args, "." + args.save_model_as, global_step)
                            save_model(ckpt_name, accelerator.unwrap_model(network), global_step, epoch)

                            if args.save_state:
                                train_util.save_and_remove_state_stepwise(args, accelerator, global_step)

                            remove_step_no = train_util.get_remove_step_no(args, global_step)
                            if remove_step_no is not None:
                                remove_ckpt_name = train_util.get_step_ckpt_name(args, "." + args.save_model_as, remove_step_no)
                                remove_model(remove_ckpt_name)

                current_loss = loss.detach().item()
                current_loss1 = loss1.detach().item()
                current_loss2 = loss2.detach().item()
                current_loss_pool = loss_pool.detach().item()
                loss_recorder.add(epoch=epoch, step=step, loss=current_loss)
                avr_loss: float = loss_recorder.moving_average
                logs = {"avr_loss": avr_loss, "l1": current_loss1, "l2": current_loss2, "lp": current_loss_pool}
                progress_bar.set_postfix(**logs)

                if args.scale_weight_norms:
                    progress_bar.set_postfix(**{**max_mean_logs, **logs})

                if args.logging_dir is not None:
                    logs = self.generate_step_logs(args, current_loss, avr_loss, lr_scheduler, keys_scaled, mean_norm, maximum_norm)
                    accelerator.log(logs, step=global_step)

                if global_step >= args.max_train_steps:
                    break

            if args.logging_dir is not None:
                logs = {"loss/epoch": loss_recorder.moving_average}
                accelerator.log(logs, step=epoch + 1)

            accelerator.wait_for_everyone()

            # 指定エポックごとにモデルを保存
            if args.save_every_n_epochs is not None:
                saving = (epoch + 1) % args.save_every_n_epochs == 0 and (epoch + 1) < num_train_epochs
                if is_main_process and saving:
                    ckpt_name = train_util.get_epoch_ckpt_name(args, "." + args.save_model_as, epoch + 1)
                    save_model(ckpt_name, accelerator.unwrap_model(network), global_step, epoch + 1)

                    remove_epoch_no = train_util.get_remove_epoch_no(args, epoch + 1)
                    if remove_epoch_no is not None:
                        remove_ckpt_name = train_util.get_epoch_ckpt_name(args, "." + args.save_model_as, remove_epoch_no)
                        remove_model(remove_ckpt_name)

                    if args.save_state:
                        train_util.save_and_remove_state_on_epoch_end(args, accelerator, epoch + 1)

            self.sample_images(accelerator, args, epoch + 1, global_step, accelerator.device, vae, tokenizer, text_encoder, unet)

            # end of epoch

        # metadata["ss_epoch"] = str(num_train_epochs)
        metadata["ss_training_finished_at"] = str(time.time())

        if is_main_process:
            network = accelerator.unwrap_model(network)

        accelerator.end_training()

        if is_main_process and args.save_state:
            train_util.save_state_on_train_end(args, accelerator)

        if is_main_process:
            ckpt_name = train_util.get_last_ckpt_name(args, "." + args.save_model_as)
            save_model(ckpt_name, network, global_step, num_train_epochs, force_sync_upload=True)

            logger.info("model saved.")


def setup_parser() -> argparse.ArgumentParser:
    parser = train_network.setup_parser()
    sdxl_train_util.add_sdxl_training_arguments(parser)
    return parser


if __name__ == "__main__":
    parser = setup_parser()

    args = parser.parse_args()
    args = train_util.read_config_from_file(args, parser)

    trainer = SdxlNetworkTrainer()
    trainer.train(args)

容易に過学習するので、適当なepochで切り上げるか、適用時の重みで調整するかしてください。