Implementation:Facebookresearch Habitat lab EqaDataset
| Knowledge Sources | |
|---|---|
| Domains | Embodied_AI, Embodied_Question_Answering |
| Last Updated | 2026-02-15 00:00 GMT |
Overview
EQADataset is a PyTorch-compatible WebDataset for Embodied Question Answering (EQA) that caches rendered episode frames to disk as tar archives and streams them during training for both VQA and PACMAN models.
Description
EQADataset extends webdataset.Dataset to support the Embodied Q&A task. On initialization, it creates a Habitat environment, loads the episode dataset, sorts episodes by ID, computes maximum question and action lengths for padding, and restructures the answer vocabulary for contiguous indexing. If a cached tar archive of rendered frames does not exist, the dataset iterates over all scenes and episodes, renders frames along the shortest path, saves them as JPEG images, and archives them into a tar file. A custom group_by_keys_ pipeline function groups the tar entries by episode and attaches question tokens and answer labels. The dataset supports both VQA-only mode (using only the last N frames) and full PACMAN mode (using all shortest path frames).
Usage
Use EQADataset when training EQA models (VQA or PACMAN). Provide a configuration specifying the Habitat environment, dataset split, and frame storage paths. The dataset handles caching automatically.
Code Reference
Source Location
- Repository: Facebookresearch_Habitat_lab
- File: habitat-baselines/habitat_baselines/il/data/data.py
- Lines: 28-262
Signature
class EQADataset(wds.Dataset):
def __init__(
self,
config,
input_type,
num_frames=5,
max_controller_actions=5,
):
Import
from habitat_baselines.il.data.data import EQADataset
I/O Contract
Inputs
| Name | Type | Required | Description |
|---|---|---|---|
| config | DictConfig | Yes | Configuration object containing habitat environment and baselines settings |
| input_type | str | Yes | Type of model being trained: "vqa" or "pacman" |
| num_frames | int | No | Number of frames used as input to VQA model (default: 5) |
| max_controller_actions | int | No | Maximum number of controller actions (default: 5) |
Outputs
| Name | Type | Description |
|---|---|---|
| (iterable samples) | dict | Each sample contains episode_id, question (LongTensor), answer (int), and frame images keyed by index suffix |
Key Methods
get_vocab_dicts
def get_vocab_dicts(self) -> Tuple[VocabDict, VocabDict]
Returns the question and answer VocabDict objects.
calc_max_length
def calc_max_length(self) -> None
Computes maximum question token length and maximum action sequence length across all episodes.
save_frame_queue
def save_frame_queue(self, pos_queue: List[ShortestPathPoint], episode_id) -> None
Renders and saves JPEG frames for an episode's position queue.
Usage Examples
Basic Usage
from habitat_baselines.il.data.data import EQADataset
from torch.utils.data import DataLoader
dataset = EQADataset(
config=my_config,
input_type="vqa",
num_frames=5,
)
q_vocab, ans_vocab = dataset.get_vocab_dicts()
dataloader = DataLoader(
dataset,
batch_size=32,
shuffle=False,
)
for batch in dataloader:
questions = batch["question"]
answers = batch["answer"]
# Process frames and questions through VQA model