""" Diffusion Policy image dataset for Open-Drawer demos. Follows the PushTImageDataset pattern. """ from typing import Dict import copy import numpy as np import torch from diffusion_policy.common.pytorch_util import dict_apply from diffusion_policy.common.replay_buffer import ReplayBuffer from diffusion_policy.common.sampler import ( SequenceSampler, get_val_mask, downsample_mask, ) from diffusion_policy.common.normalize_util import get_image_range_normalizer from diffusion_policy.model.common.normalizer import LinearNormalizer from diffusion_policy.dataset.base_dataset import BaseImageDataset class OpenDrawerImageDataset(BaseImageDataset): def __init__( self, zarr_path: str, horizon: int = 16, pad_before: int = 0, pad_after: int = 0, seed: int = 42, val_ratio: float = 0.0, max_train_episodes=None, ): super().__init__() self.replay_buffer = ReplayBuffer.copy_from_path( zarr_path, keys=["image", "proprio", "action"] ) val_mask = get_val_mask( n_episodes=self.replay_buffer.n_episodes, val_ratio=val_ratio, seed=seed, ) train_mask = ~val_mask train_mask = downsample_mask( mask=train_mask, max_n=max_train_episodes, seed=seed, ) self.sampler = SequenceSampler( replay_buffer=self.replay_buffer, sequence_length=horizon, pad_before=pad_before, pad_after=pad_after, episode_mask=train_mask, ) self.train_mask = train_mask self.horizon = horizon self.pad_before = pad_before self.pad_after = pad_after def get_validation_dataset(self): val_set = copy.copy(self) val_set.sampler = SequenceSampler( replay_buffer=self.replay_buffer, sequence_length=self.horizon, pad_before=self.pad_before, pad_after=self.pad_after, episode_mask=~self.train_mask, ) val_set.train_mask = ~self.train_mask return val_set def get_normalizer(self, mode="limits", **kwargs): data = { "action": self.replay_buffer["action"], "proprio": self.replay_buffer["proprio"], } normalizer = LinearNormalizer() normalizer.fit(data=data, last_n_dims=1, mode=mode, **kwargs) normalizer["image"] = get_image_range_normalizer() return normalizer def __len__(self) -> int: return len(self.sampler) def _sample_to_data(self, sample): image = sample["image"].astype(np.float32) / 255.0 # (T, 3, H, W) proprio = sample["proprio"].astype(np.float32) # (T, 50) action = sample["action"].astype(np.float32) # (T, 38) return { "obs": { "image": image, "proprio": proprio, }, "action": action, } def __getitem__(self, idx: int) -> Dict[str, torch.Tensor]: sample = self.sampler.sample_sequence(idx) data = self._sample_to_data(sample) return dict_apply(data, torch.from_numpy)