opentau.datasets.vqa.clevr

CLEVR dataset for visual reasoning and vqa tasks.

This module provides the CLEVR (Compositional Language and Elementary Visual Reasoning) dataset implementation for training vision-language models on compositional visual reasoning tasks. The dataset contains synthetic scenes with geometric objects and questions requiring compositional reasoning.

The dataset is loaded from HuggingFace and formatted for vqa tasks, providing images, questions, and answers for visual reasoning.

Classes:

CLEVRDataset: Dataset class that loads and formats CLEVR data from

MMInstruction/Clevr_CoGenT_TrainA_70K_Complex on HuggingFace.

Functions:

_img_to_normalized_tensor: Convert PIL Image to normalized torch tensor

with channel-first format and [0, 1] normalization.

Example

Use CLEVR dataset in training:

>>> from opentau.configs.default import DatasetConfig
>>> cfg = DatasetConfig(vqa="clevr")
>>> dataset = make_dataset(cfg, train_cfg)

Classes

CLEVRDataset(cfg[, consecutive_bad_tolerance])

CLEVR dataset for visual reasoning and vqa tasks.

class opentau.datasets.vqa.clevr.CLEVRDataset(cfg: TrainPipelineConfig, consecutive_bad_tolerance=100)

Bases: VQADataset

CLEVR dataset for visual reasoning and vqa tasks.

Loads the MMInstruction/Clevr_CoGenT_TrainA_70K_Complex dataset from HuggingFace and formats it for vqa tasks.

__init__(cfg: TrainPipelineConfig, consecutive_bad_tolerance=100)