from __future__ import annotations from typing import Any from psi.config.config import DataConfig class SeedVideoDataConfig(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 video_sample_fps: float = 1.0 video_size_pixels: int = 50176 video_cache_size: int = 32 video_embed_dir: str = "" use_video_instruction: bool = False def __call__(self, split: str = "train", transform_kwargs={}, **kwargs) -> Any: from psi.data.dataset import Dataset as MapStyleDataset from psi.data.seed.seed_video_motion_dataset import SeedVideoMotionDataset 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 = SeedVideoMotionDataset( 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, video_sample_fps=self.video_sample_fps, video_size_pixels=self.video_size_pixels, video_cache_size=self.video_cache_size, video_embed_dir=self.video_embed_dir, use_video_instruction=self.use_video_instruction, ) return MapStyleDataset(self, dataset, transform_kwargs=transform_kwargs, **kwargs)