experiments/exp_base.py

"""
This repo is forked from [Boyuan Chen](https://boyuan.space/)'s research 
template [repo](https://github.com/buoyancy99/research-template). 
By its MIT license, you must keep the above sentence in `README.md` 
and the `LICENSE` file to credit the author.
"""

from abc import ABC, abstractmethod
from typing import Optional, Union, Literal, List, Dict
import pathlib
import os
import re
from datetime import timedelta

import hydra
import torch
from lightning.pytorch.strategies.ddp import DDPStrategy

import lightning.pytorch as pl
from lightning.pytorch.loggers.wandb import WandbLogger
from lightning.pytorch.utilities.types import TRAIN_DATALOADERS
from lightning.pytorch.callbacks import LearningRateMonitor, ModelCheckpoint
from pytorch_lightning.utilities import rank_zero_info

from omegaconf import DictConfig, open_dict

from utils.print_utils import cyan
from utils.distributed_utils import is_rank_zero
from safetensors.torch import load_file
from pathlib import Path
from huggingface_hub import hf_hub_download
from huggingface_hub import model_info

torch.set_float32_matmul_precision("high")


class LastKStepModelCheckpoint(ModelCheckpoint):
    """ModelCheckpoint variant that keeps only the latest k step checkpoints."""

    def __init__(self, *args, save_last_k: Optional[int] = None, **kwargs):
        self._save_last_k = None
        if save_last_k is not None:
            self._save_last_k = int(save_last_k)
            if self._save_last_k < 1:
                raise ValueError("save_last_k must be a positive integer")
            # Lightning needs save-all mode here; this subclass prunes by step after each save.
            kwargs["save_top_k"] = -1
        super().__init__(*args, **kwargs)

    def _save_checkpoint(self, trainer: "pl.Trainer", filepath: str) -> None:
        super()._save_checkpoint(trainer, filepath)
        self._prune_old_step_checkpoints(trainer)

    def _prune_old_step_checkpoints(self, trainer: "pl.Trainer") -> None:
        if self._save_last_k is None or self.dirpath is None:
            return

        checkpoint_dir = pathlib.Path(self.dirpath)
        if not checkpoint_dir.exists():
            return

        step_checkpoints = []
        for checkpoint_path in checkpoint_dir.glob("*.ckpt"):
            match = re.search(r"(?:^|[_=-])step[=_]?(\d+)", checkpoint_path.stem)
            if match:
                step_checkpoints.append((int(match.group(1)), checkpoint_path.name, str(checkpoint_path)))

        for _, _, checkpoint_path in sorted(step_checkpoints)[:-self._save_last_k]:
            self._remove_checkpoint(trainer, checkpoint_path)

def is_huggingface_model(path: str) -> bool:
    hf_ckpt = str(path).split('/')
    repo_id = '/'.join(hf_ckpt[:2])
    try:
        model_info(repo_id)
        return True
    except:
        return False
    
_IGNORED_STATE_KEYS = {"data_mean", "data_std"}
_WRAPPER_PREFIXES = ("module.", "_orig_mod.")
_SOURCE_PREFIXES = (
    "",
    "algo.",
    "diffusion_model.model.",
    "diffusion_model.",
    "model.",
    "vae.",
    "pose_prediction_model.",
)
_TARGET_PREFIXES = (
    "",
    "_orig_mod.",
    "model.",
    "diffusion_model.",
    "diffusion_model.model.",
    "vae.",
    "pose_prediction_model.",
)
_DEMEMWM_FRAME_MEMORY_MARKERS = (
    "r_attn_anchor",
    "r_attn_dynamic",
    "r_attn_revisit",
    "r_adaLN_modulation",
    "r_mlp",
    "frame_memory",
    "memory_adapter",
)


def _checkpoint_step_from_path(ckpt_path) -> Optional[int]:
    if ckpt_path is None:
        return None
    step_match = re.search(r"step[=_-]?(\d+)", pathlib.Path(str(ckpt_path)).stem)
    return int(step_match.group(1)) if step_match else None


def _short_list(items, limit=8):
    items = list(items)
    suffix = "" if len(items) <= limit else f", ... (+{len(items) - limit} more)"
    return ", ".join(str(item) for item in items[:limit]) + suffix


def _extract_state_dict(ckpt, checkpoint_path):
    if isinstance(ckpt, dict):
        for key in ("state_dict", "model_state_dict", "model"):
            value = ckpt.get(key)
            if isinstance(value, dict):
                return value
        return ckpt
    raise ValueError(f"Checkpoint {checkpoint_path} does not contain a state dict.")


def _read_checkpoint_state_dict(checkpoint_path):
    if checkpoint_path.suffix == ".safetensors":
        return load_file(checkpoint_path)
    if checkpoint_path.suffix == ".pt":
        return _extract_state_dict(
            torch.load(checkpoint_path, map_location=torch.device("cpu"), weights_only=True),
            checkpoint_path,
        )
    if checkpoint_path.suffix == ".ckpt":
        return _extract_state_dict(
            # Lightning .ckpt envelopes can contain callback/loop metadata that
            # weights_only cannot unpickle; tensors are filtered before loading.
            torch.load(checkpoint_path, map_location=torch.device("cpu"), weights_only=False),
            checkpoint_path,
        )
    raise ValueError(f"Unsupported checkpoint: {checkpoint_path}")


def _load_checkpoint_state_dict(checkpoint_path):
    if not checkpoint_path.is_dir():
        return _read_checkpoint_state_dict(checkpoint_path), checkpoint_path

    ckpt_files = sorted(checkpoint_path.glob("*.ckpt"), key=lambda p: p.stat().st_mtime, reverse=True)
    if not ckpt_files:
        raise FileNotFoundError("No .ckpt files found in the specified directory!")

    errors = []
    for ckpt_file in ckpt_files:
        try:
            state_dict = _read_checkpoint_state_dict(ckpt_file)
            if any(torch.is_tensor(value) for value in state_dict.values()):
                rank_zero_info(f"Checkpoint file selected for loading: {ckpt_file}")
                return state_dict, ckpt_file
            errors.append(f"{ckpt_file.name}: no tensor entries")
        except Exception as exc:
            errors.append(f"{ckpt_file.name}: {exc}")
    raise RuntimeError(f"No readable .ckpt files found in {checkpoint_path}: {_short_list(errors, limit=4)}")


def _clean_state_dict(state_dict):
    clean_state_dict = {}
    ignored_keys, non_tensor_keys = [], []
    for key, value in state_dict.items():
        key = str(key)
        if key.split(".")[-1] in _IGNORED_STATE_KEYS:
            ignored_keys.append(key)
        elif not torch.is_tensor(value):
            non_tensor_keys.append(key)
        else:
            clean_state_dict[key] = value

    if ignored_keys:
        rank_zero_info(f"Ignored checkpoint statistics keys: {_short_list(ignored_keys)}")
    if non_tensor_keys:
        rank_zero_info(f"Ignored non-tensor checkpoint entries: {_short_list(non_tensor_keys)}")
    if not clean_state_dict:
        raise ValueError("Checkpoint contains no tensor weights after filtering.")
    return clean_state_dict


def _strip_wrapper_prefixes(key):
    changed = True
    while changed:
        changed = False
        for prefix in _WRAPPER_PREFIXES:
            if key.startswith(prefix):
                key = key[len(prefix):]
                changed = True
    return key


def _best_prefix_match(state_dict, target_state):
    source_prefixes = [
        prefix for prefix in _SOURCE_PREFIXES
        if prefix == "" or any(_strip_wrapper_prefixes(key).startswith(prefix) for key in state_dict)
    ]
    target_prefixes = [
        prefix for prefix in _TARGET_PREFIXES
        if prefix == "" or any(key.startswith(prefix) for key in target_state)
    ]

    best = None
    for source_prefix in source_prefixes:
        for target_prefix in target_prefixes:
            mapped, collisions = {}, 0
            for source_key, value in state_dict.items():
                key = _strip_wrapper_prefixes(source_key)
                if source_prefix and key.startswith(source_prefix):
                    key = key[len(source_prefix):]
                target_key = target_prefix + key
                if target_key in mapped:
                    collisions += 1
                    continue
                mapped[target_key] = (source_key, value)
            matched = [key for key in mapped if key in target_state]
            shape_matched = [
                key for key in matched
                if tuple(mapped[key][1].shape) == tuple(target_state[key].shape)
            ]
            score = (len(shape_matched), len(matched), -collisions, source_prefix == "", target_prefix == "")
            if best is None or score > best[0]:
                best = (score, source_prefix, target_prefix, mapped)
    return best[1], best[2], best[3]


def _is_dememwm_frame_memory_key(key):
    return any(marker in key for marker in _DEMEMWM_FRAME_MEMORY_MARKERS)


def _zero_frame_memory_gate_slices(algo):
    zeroed = []
    with torch.no_grad():
        for name, param in algo.named_parameters():
            if "r_adaLN_modulation" not in name or param.shape[0] % 6 != 0:
                continue
            chunk = param.shape[0] // 6
            param[2 * chunk:3 * chunk].zero_()
            param[5 * chunk:6 * chunk].zero_()
            zeroed.append(name)
    return zeroed


def load_custom_checkpoint(algo, checkpoint_path):
    if not checkpoint_path:
        rank_zero_info("No checkpoint path provided, skipping checkpoint loading.")
        return None

    if not isinstance(checkpoint_path, Path):
        checkpoint_path = Path(checkpoint_path)

    if not checkpoint_path.exists() and is_huggingface_model(str(checkpoint_path)):
        # Resolve non-local Hub checkpoint paths before normal state-dict filtering.
        hf_ckpt = str(checkpoint_path).split('/')
        repo_id = '/'.join(hf_ckpt[:2])
        file_name = '/'.join(hf_ckpt[2:])
        model_path = hf_hub_download(repo_id=repo_id, filename=file_name)
        checkpoint_path = Path(model_path)

    raw_state_dict, loaded_path = _load_checkpoint_state_dict(checkpoint_path)
    state_dict = _clean_state_dict(raw_state_dict)
    target_state = algo.state_dict()
    source_prefix, target_prefix, mapped_state = _best_prefix_match(state_dict, target_state)

    compatible_state_dict, unexpected_keys, shape_mismatches = {}, [], []
    for target_key, (source_key, value) in mapped_state.items():
        if target_key not in target_state:
            unexpected_keys.append(source_key)
            continue
        if tuple(value.shape) != tuple(target_state[target_key].shape):
            shape_mismatches.append(
                f"{source_key}->{target_key}: {tuple(value.shape)} != {tuple(target_state[target_key].shape)}"
            )
            continue
        compatible_state_dict[target_key] = value

    missing_keys = [key for key in target_state if key not in compatible_state_dict]
    checkpoint_is_dememwm = any(_is_dememwm_frame_memory_key(key) for key in state_dict)
    dememwm_target_keys = [key for key in target_state if _is_dememwm_frame_memory_key(key)]
    dememwm_missing = [key for key in dememwm_target_keys if key not in compatible_state_dict]

    rank_zero_info(
        "Checkpoint prefix match: "
        f"strip '{source_prefix or '<none>'}', add '{target_prefix or '<none>'}'."
    )
    if shape_mismatches:
        rank_zero_info(f"Skipped shape-mismatched checkpoint keys: {_short_list(shape_mismatches)}")
    if unexpected_keys:
        rank_zero_info(f"Skipped unexpected checkpoint keys: {_short_list(unexpected_keys)}")
    if missing_keys:
        rank_zero_info(f"Missing checkpoint keys: {_short_list(missing_keys)}")
    if not compatible_state_dict:
        raise RuntimeError(f"No compatible tensor weights found in checkpoint {loaded_path}.")
    if checkpoint_is_dememwm and dememwm_missing:
        raise RuntimeError(
            "Incomplete DeMemWM checkpoint: missing frame-memory keys "
            f"{_short_list(dememwm_missing)}"
        )

    algo.load_state_dict(compatible_state_dict, strict=False)
    if dememwm_target_keys and not checkpoint_is_dememwm:
        zeroed = _zero_frame_memory_gate_slices(algo)
        if zeroed:
            rank_zero_info(f"Zeroed frame-memory gate slices for base checkpoint load: {_short_list(zeroed)}")
    rank_zero_info(f"Model weights loaded from {loaded_path}: {len(compatible_state_dict)} tensors.")

class BaseExperiment(ABC):
    """
    Abstract class for an experiment. This generalizes the pytorch lightning Trainer & lightning Module to more
    flexible experiments that doesn't fit in the typical ml loop, e.g. multi-stage reinforcement learning benchmarks.
    """

    # each key has to be a yaml file under '[project_root]/configurations/algorithm' without .yaml suffix
    compatible_algorithms: Dict = NotImplementedError

    def __init__(
        self,
        root_cfg: DictConfig,
        logger: Optional[WandbLogger] = None,
        ckpt_path: Optional[Union[str, pathlib.Path]] = None,
    ) -> None:
        """
        Constructor

        Args:
            cfg: configuration file that contains everything about the experiment
            logger: a pytorch-lightning WandbLogger instance
            ckpt_path: an optional path to saved checkpoint
        """
        super().__init__()
        self.root_cfg = root_cfg
        self.cfg = root_cfg.experiment
        self.debug = root_cfg.debug
        self.logger = logger
        self.ckpt_path = ckpt_path
        self.algo = None
        self.customized_load = getattr(root_cfg, "customized_load", False)
        self.seperate_load = getattr(root_cfg, "seperate_load", False)
        self.zero_init_gate= getattr(root_cfg, "zero_init_gate", False)
        self.only_tune_memory = getattr(root_cfg, "only_tune_memory", False)
        if self.root_cfg.algorithm._name == "dememwm_base":
            if self.zero_init_gate:
                raise ValueError("zero_init_gate is not supported for algorithm._name=dememwm_base.")
            if self.only_tune_memory:
                raise ValueError("only_tune_memory is not supported for algorithm._name=dememwm_base.")
        self.diffusion_model_path = getattr(root_cfg, "diffusion_model_path", None)
        self.vae_path = getattr(root_cfg, "vae_path", None)
        self.pose_predictor_path = getattr(root_cfg, "pose_predictor_path", None)
        self.auto_resuming = getattr(root_cfg, "_auto_resuming", False)

    def _load_custom_checkpoint_for_task(self) -> None:
        if self.ckpt_path:
            load_custom_checkpoint(algo=self.algo, checkpoint_path=self.ckpt_path)
        elif self.seperate_load:
            if 'oasis500m' in self.diffusion_model_path:
                load_custom_checkpoint(algo=self.algo.diffusion_model.model, checkpoint_path=self.diffusion_model_path)
            else:
                load_custom_checkpoint(algo=self.algo.diffusion_model, checkpoint_path=self.diffusion_model_path)
            load_custom_checkpoint(algo=self.algo.vae, checkpoint_path=self.vae_path)
        else:
            load_custom_checkpoint(algo=self.algo, checkpoint_path=self.ckpt_path)

    def _build_algo(self):
        """
        Build the lightning module
        :return:  a pytorch-lightning module to be launched
        """
        algo_name = self.root_cfg.algorithm._name
        if algo_name not in self.compatible_algorithms:
            raise ValueError(
                f"Algorithm {algo_name} not found in compatible_algorithms for this Experiment class. "
                "Make sure you define compatible_algorithms correctly and make sure that each key has "
                "same name as yaml file under '[project_root]/configurations/algorithm' without .yaml suffix"
            )
        return self.compatible_algorithms[algo_name](self.root_cfg.algorithm)

    def exec_task(self, task: str) -> None:
        """
        Executing a certain task specified by string. Each task should be a stage of experiment.
        In most computer vision / nlp applications, tasks should be just train and test.
        In reinforcement learning, you might have more stages such as collecting dataset etc

        Args:
            task: a string specifying a task implemented for this experiment
        """
        if hasattr(self, task) and callable(getattr(self, task)):
            if is_rank_zero:
                print(cyan("Executing task:"), f"{task} out of {self.cfg.tasks}")
            getattr(self, task)()
        else:
            raise ValueError(
                f"Specified task '{task}' not defined for class {self.__class__.__name__} or is not callable."
            )

    def exec_interactive(self, task: str) -> None:
        """
        Executing a certain task specified by string. Each task should be a stage of experiment.
        In most computer vision / nlp applications, tasks should be just train and test.
        In reinforcement learning, you might have more stages such as collecting dataset etc

        Args:
            task: a string specifying a task implemented for this experiment
        """
        if hasattr(self, task) and callable(getattr(self, task)):
            if is_rank_zero:
                print(cyan("Executing task:"), f"{task} out of {self.cfg.tasks}")
            return getattr(self, task)()
        else:
            raise ValueError(
                f"Specified task '{task}' not defined for class {self.__class__.__name__} or is not callable."
            )

class BaseLightningExperiment(BaseExperiment):
    """
    Abstract class for pytorch lightning experiments. Useful for computer vision & nlp where main components are
    simply models, datasets and train loop.
    """

    # each key has to be a yaml file under '[project_root]/configurations/algorithm' without .yaml suffix
    compatible_algorithms: Dict = NotImplementedError

    # each key has to be a yaml file under '[project_root]/configurations/dataset' without .yaml suffix
    compatible_datasets: Dict = NotImplementedError

    def _build_trainer_callbacks(self):
        callbacks = []
        if self.logger:
            callbacks.append(LearningRateMonitor("step", True))
        if "checkpointing" in self.cfg.training:
            checkpointing_cfg = dict(self.cfg.training.checkpointing)
            save_last_k = checkpointing_cfg.pop("save_last_k", None)
            checkpoint_cls = LastKStepModelCheckpoint if save_last_k is not None else ModelCheckpoint
            if save_last_k is not None:
                checkpointing_cfg["save_last_k"] = save_last_k
            callbacks.append(
                checkpoint_cls(
                    pathlib.Path(hydra.core.hydra_config.HydraConfig.get()["runtime"]["output_dir"]) / "checkpoints",
                    filename="epoch{epoch}_step{step}",
                    auto_insert_metric_name=False,
                    **checkpointing_cfg,
                )
            )
        return callbacks

    def _build_training_trainer(self, max_steps: int):
        return pl.Trainer(
            accelerator="auto",
            devices="auto",
            strategy=DDPStrategy(find_unused_parameters=True) if torch.cuda.device_count() > 1 else "auto",
            logger=self.logger or False,
            callbacks=self._build_trainer_callbacks(),
            gradient_clip_val=self.cfg.training.optim.gradient_clip_val or 0.0,
            val_check_interval=self.cfg.validation.val_every_n_step if self.cfg.validation.val_every_n_step else None,
            limit_val_batches=self.cfg.validation.limit_batch,
            check_val_every_n_epoch=self.cfg.validation.val_every_n_epoch if not self.cfg.validation.val_every_n_step else None,
            accumulate_grad_batches=self.cfg.training.optim.accumulate_grad_batches or 1,
            precision=self.cfg.training.precision or 32,
            detect_anomaly=False,
            num_sanity_val_steps=int(self.cfg.debug) if self.cfg.debug else 0,
            max_epochs=self.cfg.training.max_epochs,
            max_steps=max_steps,
            max_time=self.cfg.training.max_time
        )

    def _curriculum_stages(self):
        curriculum = self.cfg.training.get("curriculum", None)
        if curriculum is None or not bool(curriculum.get("enabled", True)):
            return None
        stages = curriculum.get("stages", None)
        return stages if stages is not None and len(stages) > 0 else None

    def _merge_curriculum_updates(self, target_cfg: DictConfig, updates: DictConfig) -> None:
        with open_dict(target_cfg):
            for key, value in updates.items():
                if key in target_cfg and isinstance(target_cfg[key], DictConfig) and isinstance(value, DictConfig):
                    self._merge_curriculum_updates(target_cfg[key], value)
                else:
                    target_cfg[key] = value

    def _apply_curriculum_stage(self, stage: DictConfig, stage_idx: int, until_step: int) -> None:
        for cfg_name in ("dataset", "algorithm"):
            updates = stage.get(cfg_name, None)
            if updates is not None:
                self._merge_curriculum_updates(self.root_cfg[cfg_name], updates)
        if is_rank_zero:
            print(cyan("Curriculum stage:"), f"{stage.get('name', stage_idx + 1)} until step {until_step}")

    def _curriculum_until_step(self, stage: DictConfig, stage_idx: int) -> int:
        if "until_step" not in stage:
            raise ValueError(f"experiment.training.curriculum.stages[{stage_idx}] must define until_step")
        return int(stage.until_step)

    def _curriculum_checkpoint_path(self, stage: DictConfig, stage_idx: int, until_step: int) -> pathlib.Path:
        checkpoint_dir = pathlib.Path(hydra.core.hydra_config.HydraConfig.get()["runtime"]["output_dir"]) / "checkpoints"
        stage_name = str(stage.get("name", f"stage_{stage_idx + 1}"))
        safe_name = "".join(ch if ch.isalnum() or ch in "._-" else "_" for ch in stage_name)
        return checkpoint_dir / f"curriculum_{stage_idx + 1:02d}_{safe_name}_step{until_step}.ckpt"

    def _prepare_training_weights(self):
        if self.auto_resuming:
            return self.ckpt_path
        if self.customized_load:
            self._load_custom_checkpoint_for_task()

        if self.zero_init_gate:
            for name, para in self.algo.diffusion_model.named_parameters():
                if 'r_adaLN_modulation' in name:
                    para.requires_grad_(False)
                    para[2*1024:3*1024] = 0
                    para[5*1024:6*1024] = 0
                    para.requires_grad_(True)

        if self.only_tune_memory:
            for name, para in self.algo.diffusion_model.named_parameters():
                para.requires_grad_(False)
                if 'r_' in name or 'pose_embedder' in name or 'pose_cond_mlp' in name or 'lora_' in name:
                    para.requires_grad_(True)

        return None if self.customized_load else self.ckpt_path

    def _fit_training_curriculum(self, stages, ckpt_path) -> None:
        resume_step = _checkpoint_step_from_path(ckpt_path) if self.auto_resuming else None
        for stage_idx, stage in enumerate(stages):
            until_step = self._curriculum_until_step(stage, stage_idx)
            self._apply_curriculum_stage(stage, stage_idx, until_step)
            if resume_step is not None and resume_step >= until_step:
                if is_rank_zero:
                    print(
                        cyan("Skipping curriculum stage:"),
                        f"{stage.get('name', stage_idx + 1)} until step {until_step}; "
                        f"resume checkpoint is already at step {resume_step}",
                    )
                continue
            trainer = self._build_training_trainer(max_steps=until_step)
            trainer.fit(
                self.algo,
                train_dataloaders=self._build_training_loader(),
                val_dataloaders=self._build_validation_loader(),
                ckpt_path=ckpt_path,
            )
            if stage_idx + 1 < len(stages):
                ckpt_path = self._curriculum_checkpoint_path(stage, stage_idx, until_step)
                ckpt_path.parent.mkdir(parents=True, exist_ok=True)
                trainer.save_checkpoint(ckpt_path)

    def _build_training_loader(self) -> Optional[Union[TRAIN_DATALOADERS, pl.LightningDataModule]]:
        train_dataset = self._build_dataset("training")
        shuffle = (
            False if isinstance(train_dataset, torch.utils.data.IterableDataset) else self.cfg.training.data.shuffle
        )
        if train_dataset:
            return torch.utils.data.DataLoader(
                train_dataset,
                batch_size=self.cfg.training.batch_size,
                num_workers=min(os.cpu_count(), self.cfg.training.data.num_workers),
                shuffle=shuffle,
                persistent_workers=True,
            )
        else:
            return None

    def _build_validation_loader(self) -> Optional[Union[TRAIN_DATALOADERS, pl.LightningDataModule]]:
        validation_dataset = self._build_dataset("validation")
        shuffle = (
            False
            if isinstance(validation_dataset, torch.utils.data.IterableDataset)
            else self.cfg.validation.data.shuffle
        )
        if validation_dataset:
            return torch.utils.data.DataLoader(
                validation_dataset,
                batch_size=self.cfg.validation.batch_size,
                num_workers=min(os.cpu_count(), self.cfg.validation.data.num_workers),
                shuffle=shuffle,
                persistent_workers=True,
            )
        else:
            return None

    def _build_test_loader(self) -> Optional[Union[TRAIN_DATALOADERS, pl.LightningDataModule]]:
        test_dataset = self._build_dataset("test")
        shuffle = False if isinstance(test_dataset, torch.utils.data.IterableDataset) else self.cfg.test.data.shuffle
        if test_dataset:
            return torch.utils.data.DataLoader(
                test_dataset,
                batch_size=self.cfg.test.batch_size,
                num_workers=min(os.cpu_count(), self.cfg.test.data.num_workers),
                shuffle=shuffle,
                persistent_workers=True,
            )
        else:
            return None

    def training(self) -> None:
        """
        All training happens here
        """
        if not self.algo:
            self.algo = self._build_algo()
        if self.cfg.training.compile:
            self.algo = torch.compile(self.algo)

        ckpt_path = self._prepare_training_weights()
        curriculum_stages = self._curriculum_stages()
        if curriculum_stages is not None:
            self._fit_training_curriculum(curriculum_stages, ckpt_path)
            return

        trainer = self._build_training_trainer(max_steps=self.cfg.training.max_steps)
        trainer.fit(
            self.algo,
            train_dataloaders=self._build_training_loader(),
            val_dataloaders=self._build_validation_loader(),
            ckpt_path=ckpt_path,
        )

    def validation(self) -> None:
        """
        All validation happens here
        """
        if not self.algo:
            self.algo = self._build_algo()
        if self.cfg.validation.compile:
            self.algo = torch.compile(self.algo)

        callbacks = []

        trainer = pl.Trainer(
            accelerator="auto",
            logger=self.logger,
            devices="auto",
            num_nodes=self.cfg.num_nodes,
            strategy=DDPStrategy(find_unused_parameters=False, timeout=timedelta(hours=1)) if torch.cuda.device_count() > 1 else "auto",
            callbacks=callbacks,
            limit_val_batches=self.cfg.validation.limit_batch,
            precision=self.cfg.validation.precision,
            detect_anomaly=False,  # self.cfg.debug,
            inference_mode=self.cfg.validation.inference_mode,
        )

        if self.customized_load:
            self._load_custom_checkpoint_for_task()

            if self.zero_init_gate:
                for name, para in self.algo.diffusion_model.named_parameters():
                    if 'r_adaLN_modulation' in name:
                        para.requires_grad_(False)
                        para[2*1024:3*1024] = 0
                        para[5*1024:6*1024] = 0
                        para.requires_grad_(True)
            
            trainer.validate(
                self.algo,
                dataloaders=self._build_validation_loader(),
                ckpt_path=None,
            )
        else:
            trainer.validate(
                self.algo,
                dataloaders=self._build_validation_loader(),
                ckpt_path=self.ckpt_path,
            )

    def test(self) -> None:
        """
        All testing happens here
        """
        if not self.algo:
            self.algo = self._build_algo()
        if self.cfg.test.compile:
            self.algo = torch.compile(self.algo)

        callbacks = []

        trainer = pl.Trainer(
            accelerator="auto",
            logger=self.logger,
            devices="auto",
            num_nodes=self.cfg.num_nodes,
            strategy=DDPStrategy(find_unused_parameters=False, timeout=timedelta(hours=1)) if torch.cuda.device_count() > 1 else "auto",
            callbacks=callbacks,
            limit_test_batches=self.cfg.test.limit_batch,
            precision=self.cfg.test.precision,
            inference_mode=self.cfg.test.inference_mode,
            detect_anomaly=False,  # self.cfg.debug,
        )

        if self.customized_load:
            self._load_custom_checkpoint_for_task()

            if self.zero_init_gate:
                for name, para in self.algo.diffusion_model.named_parameters():
                    if 'r_adaLN_modulation' in name:
                        para.requires_grad_(False)
                        para[2*1024:3*1024] = 0
                        para[5*1024:6*1024] = 0
                        para.requires_grad_(True)
            
            trainer.test(
                self.algo,
                dataloaders=self._build_test_loader(),
                ckpt_path=None,
            )
        else:
            trainer.test(
                self.algo,
                dataloaders=self._build_test_loader(),
                ckpt_path=self.ckpt_path,
            )


        if not self.algo:
            self.algo = self._build_algo()
        if self.cfg.validation.compile:
            self.algo = torch.compile(self.algo)

    def _build_dataset(self, split: str) -> Optional[torch.utils.data.Dataset]:
        if split in ["training", "test", "validation"]:
            return self.compatible_datasets[self.root_cfg.dataset._name](self.root_cfg.dataset, split=split)
        else:
            raise NotImplementedError(f"split '{split}' is not implemented")