Have a question? Connect with the community at the TensorFlow Forum Visit Forum


CLEVR is a diagnostic dataset that tests a range of visual reasoning abilities. It contains minimal biases and has detailed annotations describing the kind of reasoning each question requires.

Split Examples
'test' 15,000
'train' 70,000
'validation' 15,000
  • Features:
    'file_name': Text(shape=(), dtype=tf.string),
    'image': Image(shape=(None, None, 3), dtype=tf.uint8),
    'objects': Sequence({
        '3d_coords': Tensor(shape=(3,), dtype=tf.float32),
        'color': ClassLabel(shape=(), dtype=tf.int64, num_classes=8),
        'material': ClassLabel(shape=(), dtype=tf.int64, num_classes=2),
        'pixel_coords': Tensor(shape=(3,), dtype=tf.float32),
        'rotation': tf.float32,
        'shape': ClassLabel(shape=(), dtype=tf.int64, num_classes=3),
        'size': ClassLabel(shape=(), dtype=tf.int64, num_classes=2),
    'question_answer': Sequence({
        'answer': Text(shape=(), dtype=tf.string),
        'question': Text(shape=(), dtype=tf.string),


  • Citation:
  title={ {CLEVR}: A diagnostic dataset for compositional language and elementary visual reasoning},
  author={Johnson, Justin and Hariharan, Bharath and van der Maaten, Laurens and Fei-Fei, Li and Lawrence Zitnick, C and Girshick, Ross},
  booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition},