Skip to content

Latest commit

 

History

History
36 lines (27 loc) · 1.28 KB

README.md

File metadata and controls

36 lines (27 loc) · 1.28 KB

SHAPES-SyGeT

SHAPES-SyGeT (SHAPES Systematic Generalization Test) is a split of the SHAPES dataset [1] that can be used to evaluate systematic generalization.

Templates

Train templates:

  1. is a COLOR shape TRANSFORM a COLOR shape
  2. is a SHAPE TRANSFORM a COLOR shape
  3. is a SHAPE TRANSFORM TRANSFORM a COLOR shape
  4. is a SHAPE TRANSFORM TRANSFORM a SHAPE
  5. is a COLOR shape a SHAPE
  6. is a SHAPE COLOR
  7. is a SHAPE a SHAPE

Evaluation templates:

  1. is a COLOR shape TRANSFORM TRANSFORM a COLOR shape
  2. is a COLOR shape TRANSFORM a SHAPE
  3. is a COLOR shape TRANSFORM TRANSFORM a SHAPE
  4. is a SHAPE TRANSFORM a SHAPE
  5. is a COLOR shape COLOR

COLOR can take values in 'red', 'green', 'blue'
SHAPE can take values in 'circle', 'triangle', 'square'
TRANSFORM can take values in 'above', 'below', 'left of', 'right of'

Splits

Train and Val-IID use train templates. Val-OOD uses evaluation templates.

Train size: 7560
Val-IID size: 1080
Val-OOD size: 6976

References

[1] Jacob Andreas, Marcus Rohrbach, Trevor Darrell, and Dan Klein. “Neural module networks.” In: Proceedings of the IEEE conference on computer vision and pattern recognition. 2016, pp. 39–48.