ec
ec copied to clipboard
ellisk42
Demo example with python solver not working
I've been trying the demo script given in this section that works fine with the ocaml solver. This works perfectly and is able to solve the tasks
python bin/incr.py
Failure to load pregex. This is only acceptable if using pypy
Failure to load pregex. This is only acceptable if using pypy
DEPRECATION NOTICE: this module (dreamcoder.ec) will be deleted soon, please update your code to import from dreamcoder.dreamcoder instead
CUDA is available?: False
using cuda?: False
Warning: Recognition model needs feature extractor. Ignoring recognition model.
Running EC on grok-animate @ 2021-04-28 05:34:10.237338 with 8 CPUs and parameters:
noConsolidation = False
iterations = 10
enumerationTimeout = 10
useRecognitionModel = False
topk_use_only_likelihood = False
pseudoCounts = 30.0
aic = 1.0
structurePenalty = 1.0
arity = 3
taskReranker = default
storeTaskMetrics = True
rewriteTaskMetrics = False
maximumFrontier = 10
solver = ocaml
topK = 2
evaluationTimeout = 1.0
cuda = False
Currently using this much memory: 202817536
Currently using this much memory: 202817536
Using a waking task batch of size: 3
Disabling parallelism on the Python side because we only have one job.
If you are using ocaml, there could still be parallelism.
(python) Launching int -> int (3 tasks) w/ 8 CPUs. 0.000000 <= MDL < 1.500000. Timeout 10.000000.
(ocaml: 8 CPUs. shatter: 80. |fringe| = 0. |finished| = 1.)
(python) Launching int -> int (3 tasks) w/ 8 CPUs. 1.500000 <= MDL < 3.000000. Timeout 9.705448.
(ocaml: 8 CPUs. shatter: 80. |fringe| = 0. |finished| = 2.)
(python) Launching int -> int (3 tasks) w/ 8 CPUs. 3.000000 <= MDL < 4.500000. Timeout 9.429657.
(ocaml: 8 CPUs. shatter: 80. |fringe| = 0. |finished| = 12.)
(python) Launching int -> int (3 tasks) w/ 8 CPUs. 4.500000 <= MDL < 6.000000. Timeout 9.119417.
(ocaml: 8 CPUs. shatter: 80. |fringe| = 0. |finished| = 16.)
(python) Launching int -> int (3 tasks) w/ 8 CPUs. 6.000000 <= MDL < 7.500000. Timeout 8.869352.
(ocaml: 8 CPUs. shatter: 80. |fringe| = 0. |finished| = 32.)
(python) Launching int -> int (3 tasks) w/ 8 CPUs. 7.500000 <= MDL < 9.000000. Timeout 8.613590.
(ocaml: 8 CPUs. shatter: 80. |fringe| = 80. |finished| = 16.)
(python) Launching int -> int (3 tasks) w/ 8 CPUs. 9.000000 <= MDL < 10.500000. Timeout 8.307677.
(ocaml: 8 CPUs. shatter: 80. |fringe| = 80. |finished| = 0.)
(python) Launching int -> int (3 tasks) w/ 8 CPUs. 10.500000 <= MDL < 12.000000. Timeout 8.014529.
(ocaml: 8 CPUs. shatter: 80. |fringe| = 80. |finished| = 0.)
(python) Launching int -> int (3 tasks) w/ 8 CPUs. 12.000000 <= MDL < 13.500000. Timeout 7.743182.
(ocaml: 8 CPUs. shatter: 80. |fringe| = 80. |finished| = 0.)
(python) Launching int -> int (3 tasks) w/ 8 CPUs. 13.500000 <= MDL < 15.000000. Timeout 7.429715.
(ocaml: 8 CPUs. shatter: 80. |fringe| = 80. |finished| = 0.)
(python) Launching int -> int (3 tasks) w/ 8 CPUs. 15.000000 <= MDL < 16.500000. Timeout 7.124215.
(ocaml: 8 CPUs. shatter: 80. |fringe| = 80. |finished| = 0.)
(python) Launching int -> int (3 tasks) w/ 8 CPUs. 16.500000 <= MDL < 18.000000. Timeout 6.732337.
(ocaml: 8 CPUs. shatter: 80. |fringe| = 80. |finished| = 0.)
(python) Launching int -> int (3 tasks) w/ 8 CPUs. 18.000000 <= MDL < 19.500000. Timeout 6.307096.
(ocaml: 8 CPUs. shatter: 80. |fringe| = 80. |finished| = 0.)
(python) Launching int -> int (3 tasks) w/ 8 CPUs. 19.500000 <= MDL < 21.000000. Timeout 5.796163.
(ocaml: 8 CPUs. shatter: 80. |fringe| = 80. |finished| = 0.)
(python) Launching int -> int (3 tasks) w/ 8 CPUs. 21.000000 <= MDL < 22.500000. Timeout 4.227132.
(ocaml: 8 CPUs. shatter: 80. |fringe| = 80. |finished| = 0.)
(python) Launching int -> int (3 tasks) w/ 8 CPUs. 22.500000 <= MDL < 24.000000. Timeout 2.094563.
(ocaml: 8 CPUs. shatter: 80. |fringe| = 80. |finished| = 0.)
We enumerated this many programs, for each task:
[1824168, 1824168, 1824168]
Generative model enumeration results:
HIT add1 w/ (lambda (incr $0)) ; log prior = -2.197225 ; log likelihood = 0.000000
HIT add2 w/ (lambda (incr2 $0)) ; log prior = -2.197225 ; log likelihood = 0.000000
HIT add3 w/ (lambda (incr (incr2 $0))) ; log prior = -3.295837 ; log likelihood = 0.000000
Hits 3/3 tasks
Average description length of a program solving a task: 2.563429 nats
Generative model average: 0 sec. median: 0 max: 1 standard deviation 0
Currently using this much memory: 214368256
Frontiers discovered top down: 3
Total frontiers: 3
Currently using this much memory: 214368256
Showing the top 5 programs in each frontier being sent to the compressor:
add1
0.00 (lambda (incr $0))
add2
-0.29 (lambda (incr2 $0))
-1.39 (lambda (incr (incr $0)))
add3
-0.85 (lambda (incr (incr2 $0)))
-0.85 (lambda (incr2 (incr $0)))
-1.95 (lambda (incr (incr (incr $0))))
But the same script with python solver fails. (No change in script, just the solver is python one) Is the support for python solver limited?
python bin/incr.py --solver python
Failure to load pregex. This is only acceptable if using pypy
Failure to load pregex. This is only acceptable if using pypy
DEPRECATION NOTICE: this module (dreamcoder.ec) will be deleted soon, please update your code to import from dreamcoder.dreamcoder instead
CUDA is available?: False
using cuda?: False
Warning: Recognition model needs feature extractor. Ignoring recognition model.
Running EC on grok-animate @ 2021-04-28 05:27:37.193696 with 8 CPUs and parameters:
noConsolidation = False
iterations = 10
enumerationTimeout = 10
useRecognitionModel = False
topk_use_only_likelihood = False
pseudoCounts = 30.0
aic = 1.0
structurePenalty = 1.0
arity = 3
taskReranker = default
storeTaskMetrics = True
rewriteTaskMetrics = False
maximumFrontier = 10
solver = python
topK = 2
evaluationTimeout = 1.0
cuda = False
Currently using this much memory: 203104256
Currently using this much memory: 203104256
Using a waking task batch of size: 3
Disabling parallelism on the Python side because we only have one job.
If you are using ocaml, there could still be parallelism.
(python) Launching int -> int (3 tasks) w/ 8 CPUs. 0.000000 <= MDL < 1.500000. Timeout 10.000000.
(python) Launching int -> int (3 tasks) w/ 8 CPUs. 1.500000 <= MDL < 3.000000. Timeout 9.840024.
(python) Launching int -> int (3 tasks) w/ 8 CPUs. 3.000000 <= MDL < 4.500000. Timeout 9.683462.
(python) Launching int -> int (3 tasks) w/ 8 CPUs. 4.500000 <= MDL < 6.000000. Timeout 9.505921.
(python) Launching int -> int (3 tasks) w/ 8 CPUs. 6.000000 <= MDL < 7.500000. Timeout 9.346369.
(python) Launching int -> int (3 tasks) w/ 8 CPUs. 7.500000 <= MDL < 9.000000. Timeout 9.182197.
(python) Launching int -> int (3 tasks) w/ 8 CPUs. 9.000000 <= MDL < 10.500000. Timeout 8.985159.
(python) Launching int -> int (3 tasks) w/ 8 CPUs. 10.500000 <= MDL < 12.000000. Timeout 8.759223.
(python) Launching int -> int (3 tasks) w/ 8 CPUs. 12.000000 <= MDL < 13.500000. Timeout 8.463651.
(python) Launching int -> int (3 tasks) w/ 8 CPUs. 13.500000 <= MDL < 15.000000. Timeout 7.592893.
(python) Launching int -> int (3 tasks) w/ 8 CPUs. 15.000000 <= MDL < 16.500000. Timeout 6.155116.
We enumerated this many programs, for each task:
[31972, 31972, 31972]
Generative model enumeration results:
MISS add1
MISS add2
MISS add3
Hits 0/3 tasks
Average description length of a program solving a task: 0.000000 nats
Generative model no successful times to report statistics on!
Currently using this much memory: 214769664
Frontiers discovered top down: 0
Total frontiers: 0
Currently using this much memory: 214769664
Showing the top 5 programs in each frontier being sent to the compressor:
No compression frontiers; not inducing a grammar this iteration.
Grammar after iteration 3:
0.000000 t0 $_
0.000000 int -> int incr
0.000000 int -> int incr2
Currently using this much memory: 214990848
Am using the singularity container for all my work.
Huh, you're right, I'm getting exactly the same problem. Not sure what the deal is. :/
I am also interested in the differences between the ocaml and python solver
I don't know if this is still relevant for someone but I think I found the problem with the python solver.
So I also did some debugging why the python solver is not working for the demo example and actually it was only a false definition of the python primitives. The OCaml primitives are correct and therefore it could solve the problem. If you change the python primitives from:
def _incr(x): return lambda x: x + 1
def _incr2(x): return lambda x: x + 2
to:
def _incr(x): return x + 1
def _incr2(x): return x + 2
it will work. When a lambda function is returned, it waits for a second parameter, but we only need one parameter for this primitive. After changing this, the example works with the Python and pypy solver.
