from __future__ import annotations from typing import Any from psi.config.config import DataConfig class SeedAudioDataConfig(DataConfig): root_dir: str video_dir: str val_ratio: float = 0.1 split_seed: int = 42 chunk_size: int = 30 sample_stride: int = 1 use_delta_actions: bool = True motion_fps: int = 50 audio_sampling_rate: int = 16000 audio_window_seconds: float = 1.0 audio_left_context_seconds: float = 0.0 # True时每个样本直接使用整段视频音频,audio_window_seconds仅保留为旧窗口实验配置。 full_audio: bool = False align_start_hold: bool = True use_audio_instruction: bool = False audio_task: str = "dance" def __call__(self, split: str = "train", transform_kwargs={}, **kwargs) -> Any: from psi.data.dataset import Dataset as MapStyleDataset from psi.data.seed.seed_audio_motion_dataset import SeedAudioMotionDataset repack = self.transform.repack action_chunk_size = int(getattr(repack, "action_chunk_size", self.chunk_size)) use_delta_actions = bool(getattr(repack, "use_delta_actions", self.use_delta_actions)) num_past_frames = int(getattr(repack, "num_past_frames", 0)) self.chunk_size = action_chunk_size self.use_delta_actions = use_delta_actions dataset = SeedAudioMotionDataset( root_dir=self.root_dir, video_dir=self.video_dir, split=split, val_ratio=self.val_ratio, split_seed=self.split_seed, action_chunk_size=action_chunk_size, sample_stride=self.sample_stride, use_delta_actions=use_delta_actions, num_past_frames=num_past_frames, motion_fps=self.motion_fps, audio_sampling_rate=self.audio_sampling_rate, audio_window_seconds=self.audio_window_seconds, audio_left_context_seconds=self.audio_left_context_seconds, full_audio=self.full_audio, align_start_hold=self.align_start_hold, use_audio_instruction=self.use_audio_instruction, audio_task=self.audio_task, ) return MapStyleDataset(self, dataset, transform_kwargs=transform_kwargs, **kwargs)