miscellaneous improvements

ec.py

@@ -276,6 +276,9 @@ def checkpointPath(iteration, extra=""):
 
     if addFullTaskMetrics:
         assert resume is not None, "--addFullTaskMetrics requires --resume"
+
+    def reportMemory():
+        eprint(f"Currently using this much memory: {getThisMemoryUsage()}")
 
     # Restore checkpoint
     if resume is not None:
@@ -368,6 +371,8 @@ def checkpointPath(iteration, extra=""):
             eprint("Resetting task metrics for next iteration.")
             result.recognitionTaskMetrics = {}
 
+        reportMemory()
+
         # Evaluate on held out tasks if we have them
         if testingTimeout > 0 and ((j % testEvery == 0) or (j == iterations - 1)):
             eprint("Evaluating on held out testing tasks for iteration: %d" % (j))
@@ -388,6 +393,8 @@ def checkpointPath(iteration, extra=""):
         else:
             helmholtzFrontiers = lambda: []
 
+        reportMemory()
+
         # Get waking task batch.
         wakingTaskBatch = taskBatcher.getTaskBatch(result, tasks, taskBatchSize, j)
         eprint("Using a waking task batch of size: " + str(len(wakingTaskBatch)))
@@ -407,6 +414,8 @@ def checkpointPath(iteration, extra=""):
         tasksHitTopDown = {f.task for f in topDownFrontiers if not f.empty}
         result.hitsAtEachWake.append(len(tasksHitTopDown))
 
+        reportMemory()
+
         # Combine topDownFrontiers from this task batch with all frontiers.
         for f in topDownFrontiers:
             if f.task not in result.allFrontiers: continue # backward compatibility with old checkpoints
@@ -452,9 +461,11 @@ def checkpointPath(iteration, extra=""):
 
         # Sleep-G
         if useDSL:
+            eprint(f"Currently using this much memory: {getThisMemoryUsage()}")
             grammar = consolidate(result, grammar, topK=topK, pseudoCounts=pseudoCounts, arity=arity, aic=aic,
                                   structurePenalty=structurePenalty, compressor=compressor, CPUs=CPUs,
                                   iteration=j)
+            eprint(f"Currently using this much memory: {getThisMemoryUsage()}")
         else:
             eprint("Skipping consolidation.")
             result.grammars.append(grammar)
@@ -561,6 +572,7 @@ def sleep_recognition(result, grammar, taskBatch, tasks, testingTasks, allFronti
                                     contextual=contextual,
                                     previousRecognitionModel=previousRecognitionModel,
                                     id=i) for i in range(ensembleSize)]
+    eprint(f"Currently using this much memory: {getThisMemoryUsage()}")
     trainedRecognizers = parallelMap(min(CPUs,len(recognizers)),
                                      lambda recognizer: recognizer.train(allFrontiers,
                                                                          biasOptimal=biasOptimal,
@@ -574,6 +586,7 @@ def sleep_recognition(result, grammar, taskBatch, tasks, testingTasks, allFronti
                                                                          vectorized=True),
                                      recognizers,
                                      seedRandom=True)
+    eprint(f"Currently using this much memory: {getThisMemoryUsage()}")
     # Enumerate frontiers for each of the recognizers.
     eprint("Trained an ensemble of %d recognition models, now enumerating." % len(trainedRecognizers))
     ensembleFrontiers, ensembleTimes, ensembleRecognitionTimes = [], [], []
@@ -614,6 +627,7 @@ def sleep_recognition(result, grammar, taskBatch, tasks, testingTasks, allFronti
                                  'startProductions')
 
     result.hitsAtEachWake.append(len(totalTasksHitBottomUp))
+    eprint(f"Currently using this much memory: {getThisMemoryUsage()}")
 
     """ Rescore and combine the frontiers across the ensemble of recognition models."""
     eprint("Recognition model enumeration results for the best recognizer.")

examineFrontier.py

@@ -97,4 +97,4 @@ def examineProgram(entry):
 
 print(f"Best posteriorc hits task {posteriorHits}/{totalTasks} = {posteriorHits/totalTasks}")
 print(f"Best likelihood hits task {likelihoodHits}/{totalTasks} = {likelihoodHits/totalTasks}")
->>>>>>> 56832f1fe9d3731ce0d8dfc524025c95f1b428d7
+

logo.py

@@ -26,11 +26,14 @@
 global prefix_dreams
 
 def dreamFromGrammar(g, directory, N=100):
-    programs = [ p
-                 for _ in range(N)
-                 for p in [g.sample(arrow(turtle,turtle),
-                                    maximumDepth=20)]
-                 if p is not None]
+    if isinstance(g,Grammar):
+        programs = [ p
+                     for _ in range(N)
+                     for p in [g.sample(arrow(turtle,turtle),
+                                        maximumDepth=20)]
+                     if p is not None]
+    else:
+        programs = g
     drawLogo(*programs,
              pretty=False, smoothPretty=False,
              resolution=512,
@@ -178,6 +181,27 @@ def outputDreams(checkpoint, directory):
     os.system("mkdir  -p %s"%directory)
     dreamFromGrammar(g, directory)
 
+def enumerateDreams(checkpoint, directory):
+    from recognition import backgroundHelmholtzEnumeration
+    from utilities import loadPickle,standardDeviation,mean
+    result = loadPickle(checkpoint)
+    eprint(" [+] Loaded checkpoint",checkpoint)
+    g = result.grammars[-1]
+    if directory is None: assert False, "please specify a directory"
+    eprint(" Dreaming into",directory)
+    os.system("mkdir  -p %s"%directory)
+    frontiers = backgroundHelmholtzEnumeration(makeTasks(None,None), g, 500,
+                                               evaluationTimeout=0.01,
+                                               special=LogoFeatureCNN.special)()
+    random.shuffle(frontiers)
+    frontiers = frontiers[:500]
+    md = [list(f.entries)[0].logPrior for f in frontiers]
+    eprint("MDLs",md)
+    eprint(f"average MDL {mean(md)} +/- {standardDeviation(md)}")
+
+    dreamFromGrammar([list(f.entries)[0].program for f in frontiers],
+                     directory)
+
 def visualizePrimitives(primitives, export='/tmp/logo_primitives.png'):
     from itertools import product
     from pylab import imshow,show
@@ -300,7 +324,8 @@ def argumentChoices(t):
     proto = args.pop("proto")
 
     if dreamCheckpoint is not None:
-        outputDreams(dreamCheckpoint, dreamDirectory)
+        #outputDreams(dreamCheckpoint, dreamDirectory)
+        enumerateDreams(dreamCheckpoint, dreamDirectory)
         sys.exit(0)        
 
     target = args.pop("target")

official_experiments

@@ -43,6 +43,7 @@ Scientific laws:
 python launch.py -k  -c -z n1-highmem-64 scientific_1h "python scientificLaws.py  -t 3600 --topK 5 --arity 3 --maximumFrontier 5 -i 10 -R 3600 -RS 5000 --biasOptimal --contextual --mask  -r 0."
 python launch.py -k  -c -z n1-highmem-64 scientific_12m_10b "python scientificLaws.py  -t 720 --taskReranker randomShuffle --taskBatchSize 10 --topK 5 --arity 3 --maximumFrontier 5 -i 10 -R 3600 -RS 5000 --biasOptimal --contextual --mask  -r 0."
 python launch.py -k  -c -z n1-highmem-64 scientific_60m_10b "python scientificLaws.py  -t 3600 --taskReranker randomShuffle --taskBatchSize 10 --topK 5 --arity 3 --maximumFrontier 5 -i 10 -R 3600 -RS 5000 --biasOptimal --contextual --mask  -r 0."
+python launch.py -k  -c -z n1-highmem-64 scientific_30m_40b "python scientificLaws.py  -t 3600 --taskReranker randomShuffle --taskBatchSize 40 --topK 5 --arity 3 --maximumFrontier 5 -i 10 -R 3600 -RS 5000 --biasOptimal --contextual --mask  -r 0."
 python launch.py -k  -c -z n1-highmem-64 scientific_4h "python scientificLaws.py  -t 3600 --topK 5 --arity 3 --maximumFrontier 5 -i 10 -R 3600 -RS 5000 --biasOptimal --contextual --mask  -r 0."
 
 

physics.py

@@ -250,14 +250,14 @@ def physics_options(parser):
 
     tasks = physicsTasks()
 
-    generator = ecIterator(g, tasks,
-                           outputPrefix="experimentOutputs/physics",
-                           evaluationTimeout=0.01,
-                           **arguments)
-    for result in generator:
-        pass
+    # generator = ecIterator(g, tasks,
+    #                        outputPrefix="experimentOutputs/physics",
+    #                        evaluationTimeout=0.01,
+    #                        **arguments)
+    # for result in generator:
+    #     pass
 
-    assert False
+    # assert False
 
 
     def showLikelihood(e):
@@ -286,7 +286,7 @@ def showLikelihood(e):
         dt = "$%d"%(Program.parse(dv).numberOfFreeVariables)
         dv = "(*v %s %s)"%(dt,dv)
         showLikelihood(dv)
-        dv = "(+v (velocity $0) %s)"%dv
+        dv = "(+v (get-velocity $0) %s)"%dv
         showLikelihood(dv)
 
         eprint()

solvers/logoLib/logoInterpreter.ml

@@ -86,6 +86,20 @@ let eval_turtle ?sequence (t2t : turtle -> turtle) =
   List.iter eval_instruction p ;
   !c
 
+let animate_turtle ?sequence (t2t : turtle -> turtle) =
+  let p,_ = (t2t logo_NOP) (init_state ()) in
+  let p = center_logo_list p in
+  let c = ref (new_canvas ()) in
+  let lineto x y = (c := (lineto !c x y))
+  and moveto x y = (c := (moveto !c x y)) in
+  let t = init_state () in
+  moveto t.x t.y ;
+  let rec eval_instruction i = match i with
+    | SEGMENT(x1,y1,x2,y2) ->
+      (moveto x1 y1; lineto x2 y2)  in
+  List.iter eval_instruction p ;
+  !c
+
 let logo_PU : turtle =
   fun s -> ([], {s with p = false})
 

tower.py

@@ -313,6 +313,7 @@ def visualizeSolutions(solutions, export, tasks=None):
     dreamOfTowers(g0, "%s/random_0"%outputDirectory)
 
     for result in generator:
+        continue
         iteration = len(result.learningCurve)
         newTowers = [tuple(centerTower(executeTower(frontier.sample().program)))
                      for frontier in result.taskSolutions.values() if not frontier.empty]