save weights to "toy.weights" doesn't work.

[don.tu...@gmail.com]

Dear all,

I'm trying to save weights after learning in the non-binary format. However, i get an error message:

# save weights to "toy.weights";

Error on line 0: save weights to !ERROR!"toy.weights";

However, I can save to the binary format with

# save weights to dump "toy.weights";

335 weights saved.

Has anyone experienced this problem?

Thanks & best

Don

The long story: 

I'm using thebeast as an instance classifier in a python environment using the subprocess modul and write commands to it through PIPE from python. 

thebeast=subprocess.Popen(['thebeast'], stdin=subprocess.PIPE, stdout=subprocess.PIPE)      #staring thebeast

thebeast.stdin.write('set solver.model.solver = "gurobi";include "types.pml";include "coref.pml";set solver.model.initIntegers = true;load weights from dump "toy.weights";\n')    #loading the model

I then write a file "test.atoms" (in python) and pass this file to the running thebeast:

thebeast.stdin.write('load corpus from "test.atoms";save corpus to ram;next; solve;print atoms.coref;\n')   #testing an instance

and read back in thebeast output, and so forth.

What I noticed is that loading the weights from a dump can be problematic when the training instances are not present. I.e. if I start thebeast and do

include "types.pml";include "coref.pml";set solver.model.initIntegers = true;load weights from dump "toy.weights";print weights;

I get the error:

>w_animacy

java.lang.IndexOutOfBoundsException: Index: 1, Size: 0

at java.util.ArrayList.rangeCheck(ArrayList.java:571)

at java.util.ArrayList.get(ArrayList.java:349)

at thebeast.nodmem.type.MemCategoricalType.representation(MemCategoricalType.java:45)

at thebeast.nodmem.value.MemCategorical.representation(MemCategorical.java:29)

at thebeast.nod.util.TabularValuePrinter.visitCategorical(TabularValuePrinter.java:59)

at thebeast.nodmem.value.MemCategorical.acceptValueVisitor(MemCategorical.java:25)

at thebeast.nod.util.TabularValuePrinter.visitTuple(TabularValuePrinter.java:32)

at thebeast.nodmem.value.MemTuple.acceptValueVisitor(MemTuple.java:34)

at thebeast.nod.util.TabularValuePrinter.visitRelation(TabularValuePrinter.java:52)

at thebeast.nodmem.value.MemRelation.acceptValueVisitor(MemRelation.java:64)

at thebeast.nodmem.value.MemRelation.writeTo(MemRelation.java:81)

at thebeast.pml.Weights.save(Weights.java:185)

at thebeast.pml.Weights.save(Weights.java:161)

at thebeast.pml.parser.Shell.visitPrint(Shell.java:541)

at thebeast.pml.parser.ParserPrint.acceptParserStatementVisitor(ParserPrint.java:15)

at thebeast.pml.parser.Shell.interactive(Shell.java:186)

at thebeast.pml.TheBeast.main(TheBeast.java:107)

It seems that the constants aren't instantiated properly.. I also noticed that formulas with different constant values can display different weights for the same constant combination given different test instances. I assume there is a problem with loading the weighs from dumps when the instances from which the weights were inferred aren't in memory..