dcao / babble Goto Github PK

On 9cf4759, the following is totally uncompressed:

(list
  (@ foo (@ (@ + (@ f x)) (@ f x)))
  (@ foo (@ (@ + (@ f x)) (@ f x)))
)

Changing basically anything about one of the inputs, (foo -> bar, x -> y) makes it work again.

Running smiley on the following example (without DSRs):

(+ 
   (move 1 0 circle)
   (move 2 0 circle)
   (move 3 0 circle)
   (move 4 0 circle)
   (move 5 0 circle)
   (move 6 0 circle)
)

produces the expected result:

lib l0 =
  λx1 -> move <x1, 0> circle
in
  [l0 1,
    l0 2,
    l0 3,
    l0 4,
    l0 5,
    l0 6]

Adding a DSR "circle" => "(scale 1 circle)" works well, and even allows to obtain the above result if some of the circles in the input are scaled by 1.

However, adding a more general rewrite ?x -> (scale 1 ?x) results in not learning any libraries, and moreover returning an expression that is larger than the original:

scale 1 [move <1, 0> circle,
  move <2, 0> circle,
  move <3, 0> circle,
  move <4, 0> circle,
  move <5, 0> circle,
  move <6, 0> circle]

There might be two bugs here:

1. AU always returns empty for all pairs of states except the diagonal (might have to do with the e-graph being too loopy?)
2. Extraction also picks the wrong expression (not the original)

Running smiley on the following input:

(+ (move -20 0 (+ (scale 5 circle)
                  (move -2 -1.5 (rotate 90 (scale 2 line)))
                  (move 2 -1.5 (rotate 90 (scale 2 line)))
                  (move 0 2 (scale 4 line))))
   (move -7.5 0 (+ (scale 5 circle)
                   (move -2 -1.5 (rotate 90 (scale 2 line)))
                   (move 2 -1.5 (rotate 90 (scale 2 line)))
                   (move 0 2 (scale 1 circle))))
   (move 5 0 (+ (scale 5 circle)
                (move -2 -1.5 (rotate 90 (scale 1 circle)))
                (move 2 -1.5 (rotate 90 (scale 1 circle)))
                (move 0 2 (scale 1 circle))))
   (move 17.5 0 (+ (scale 5 circle)
                   (move -2 -1.5 (rotate 90 (scale 1 circle)))
                   (move 2 -1.5 (rotate 90 (scale 1 circle)))
                   (move 0 2 (scale 4 line)))))

produces the following output:

lib l15 =
  λx1 x2 x3 -> [scale 5 circle,
    move <-2, -1.5> (rotate 90 (scale x3 x2)),
    move <2, -1.5> (rotate 90 (scale x3 x2)),
    x1]
in
  [move <-20, 0> (l15 (move <0, 2> (scale 4 line)) line 2),
    move <-7.5, 0> (l15 (move <0, 2> (scale 1 circle)) line 2),
    move <5, 0> (l15 (move <0, 2> (scale 1 circle)) circle 1),
    move <17.5, 0> (l15 (move <0, 2> (scale 4 line)) circle 1)]

The issue here is that the four expressions of the form (move <0, 2> (scale ? ?)) are all abstracted into x1, which is incorrect AU, since they share top-level structure. The suspicious rewrite shows up in the log as "anti-unify 6".

Hello,
using the code from compression, we can read and compress inputs in the DreamCoder format. This seems to work, but am I overlooking any functionality for exporting compression results in the same format (e.g. as CompressionOutput as defined in dreamcoder/json.rs? This seems like basic functionality, but I'm confused by all the different experiment classes. My current understanding is that this is not implemented, and I'll have to manually post-process the ExperimentResult. This is of course possible, but appears a bit indirect - is there some simple way to access the invented library functions (and rewritten frontiers) such that we can parse them into the correct format?

Thanks for your help!