neoneye/simon-arc-image-v39

Version 1

Have dataset items that are somewhat evenly of each type. The LLM learned some of the types fine. However rotated images are causing problems. The image sizes are between 1 and 10 pixels.

Version 2

Here the majority of dataset items are rotated images. Since this is what my LLM is struggling with. Smaller images. Here the image sizes are between 1 and 5 pixels. This helped a lot on the validation loss.

Version 3

Main focus is now on count_same_color_as_center_with_8neighbors_nowrap and image size 1-6. Which the LLM has struggeld with in the past, maybe due to too big image sizes. Struggles somewhat with the count_same_color_as_center_with_8neighbors_nowrap.

Version 4

I'm trying smaller images again. Here the image sizes are between 1 and 5 pixels. Added same_color_inside_3x3_area_nowrap that checks if all surrounding pixels agree on the same color, maybe that have some synergy with the count_same_color_as_center_with_8neighbors_nowrap. It helped a little, but it's still not as good at counting neighbors as I would like.

Version 5

I have added a pixels_with_k_matching_neighbors with a k parameter between 1-8. This may help improve on counting the number of neighboring pixels. The image size 1-6. This did indeed help on counting the number of surrounding pixels.

Version 6

Same weight to all the transformations. Image size 1-11.

Version 7

Focus on histogram and k-nearest neighbors. image size 1-12. It seems like the LLM has gotten the hang of it.

Version 8

Focus on histogram and k-nearest neighbors. image size 5-20.

Version 9

Focus on histogram and k-nearest neighbors. image size 10-30.

Version 10

Same weight to all the transformations. image width 10-30. image height 2-5.

Version 11

Same weight to all the transformations. image width 2-5. image height 10-30.

Version 12

Focus on k-nearest neighbors. image width 2-5. image height 10-30.

Version 13

Focus on compres_x, compres_y, compres_xy. image size is 1-10.

Version 14

Focus on histograms and k-nearest-neighbors. image size 5-20.

Version 15

Focus on histograms and k-nearest-neighbors. image size 10-30.

Version 16

Focus on k-nearest-neighbors. image size 10-25.

Version 17

Disabled k-nearest-neighbors, I suspect this is the reason why it converges so slowly. image size 15-30.

Version 18

Disabled k-nearest-neighbors, and compression. image size 15-25.

Version 19

Translate x/y by plus/minus 1. Disabled rotation and transpose. image size 22-30.

Version 20

Focus on k-nearest-neighbors. image size 5-15.

Version 21

Focus on k-nearest-neighbors. image size 8-18.

Version 22

Same weight to all the transformations. image size 8-20.

Version 23

Same weight to all the transformations. image size 5-30.

The LLM is struggling learning this. I'm going to try with small images.

Version 24

Focus on rotate cw, rotate ccw, transpose. image size 2-10.

The LLM is struggling learning this. Despite being small images. I'm going to try with even small images.

Version 25

Focus on rotate cw, rotate ccw, transpose. image size 2-5.

The LLM is struggling learning this. Despite being small images. I'm going to try with even small images.

Version 26

Focus on rotate cw, rotate ccw, transpose, k-nearest-neighbors. image size 1-3.

The LLM is struggling learning this. Despite being small images.

Version 27

Focus on rotate cw, rotate ccw, transpose. image size 1-4.

The LLM is struggling learning this. Despite being small images.

Version 28

Focus on rotate cw, rotate ccw, transpose. image size 1-5.

Version 29

Focus on rotate cw, rotate ccw, transpose. image size 1-6.

Version 30

Focus on rotate cw, rotate ccw, transpose. image size 1-8.

Version 31

Focus on rotate cw, rotate ccw, transpose. image size 1-10.

Version 32

Focus on rotate cw, rotate ccw, transpose. image size 1-12.

Version 33

Focus on rotate cw, rotate ccw, transpose. image size 1-14. The serialize items were using fewer names to identify the dataset, now uses the same names as deserialize.

Version 34

Focus only on rotate cw. All other operations have been disabled. image size 1-30.

Version 35

Focus only on rotate ccw. All other operations have been disabled. image size 1-30.

Version 36

Focus only on rotate cw and rotate ccw. All other operations have been disabled. image size 1-30.

Version 37

Focus only on rotate cw and rotate ccw. All other operations have been disabled. image size 1-30. Using the same image and apply both rotate cw and rotate ccw. My hypothesis is that it will learn better to distinguish between the two rotation types.

Version 38

Argh, the validation loss was seriously bad on this one.

I guess what happened is that rotate cw always was followed by rotate ccw, causing the model to be biased, always expecting the opposite transformation. Now I have suffled the entire dataset. So there are still 50% of each operation, in random order.

Using the same image and apply both rotate cw and rotate ccw. I still think my hypothesis is sound, that it will learn better to distinguish between the two rotation types when it's the same image.

Focus only on rotate cw and rotate ccw. All other operations have been disabled. image size 1-30.

Version 39

The validation loss is not improving.

I'm going back to small image sizes.

Focus only on rotate cw and rotate ccw. All other operations have been disabled. image size 1-10.