simon-arc-mass-v23 / README.md
neoneye's picture
Update README.md
96cde5f verified
metadata
license: mit
task_categories:
  - image-to-text
  - text-to-image
language:
  - en
pretty_name: simons ARC (abstraction & reasoning corpus) object mass version 23
size_categories:
  - 10K<n<100K
configs:
  - config_name: default
    data_files:
      - split: train
        path: data.jsonl

Version 1

Measure the mass of objects for pixel connectivity 4 and pixel connectivity 8.

image size: 1-10.

max_mass: 4.

Version 2

image size: 1-20.

max_mass: 5.

This converged too slowly. I was too optimistic. I will have to proceed slower.

Version 3

image size: 1-12.

max_mass: 5.

Too big spikes in the training loss. I will have to lower the max_mass, and gradually increase it.

Version 4

image size: 1-15.

max_mass: 2.

The validation loss for this is fine. I'm increasing the max_mass.

Version 5

image size: 1-20.

max_mass: 3.

Version 6

image size: 1-25.

max_mass: 4.

Version 7

image size: 1-30.

max_mass: 2.

Version 8

image size: 1-30.

max_mass: 7.

Version 9

image size: 1-30.

max_mass: 8.

Version 10

image size: 1-30.

max_mass: 9.

Version 11

image size: 1-30.

max_mass: 10.

Version 12

image size: 1-30.

max_mass: 11.

Version 13

image size: 1-30.

max_mass: 15.

Version 14

image size: 1-30.

max_mass: 20.

Version 15

image size: 1-30.

max_mass: 25.

Version 16

Added PixelConnectivity.CORNER4, so it can identify diagonal structures.

image size: 1-10.

max_mass: 5.

Version 17

image size: 1-15.

max_mass: 10.

Version 18

Added mulple new PixelConnectivity modes: LR2, TB2, TLBR2, TRBL2, so it can identify line structures.

image size: 1-10.

max_mass: 5.

Version 19

image size: 1-15.

max_mass: 10.

Version 20

Disabled NEAREST4, ALL8, CORNER4. So the model can be trained only on line structures.

image size: 1-20.

max_mass: 15.

The model struggles with these settings. I have to lower the settings.

Version 21

Still with disabled NEAREST4, ALL8, CORNER4. So the model can be trained only on line structures.

Lowered the settings.

image size: 1-10.

max_mass: 6.

Version 22

Still with disabled NEAREST4, ALL8, CORNER4. So the model can be trained only on line structures.

Bigger images. image size: 1-15.

max_mass: 6.

Version 23

Enabled all the transformations.

Bigger images. image size: 1-20.

max_mass: 8.