So some explanation.
ECJ's new Evolve.java code (which I personally wrote a few months
back :-) has the additional wrinkle of a jobs facility because so
many people asked for one. The problem is how to start up from
checkpoint and continue right at the job number you had done before.
That's where it's a little complicated, but not too bad.
A simpler, job-free main, would go something like this (I'm doing
this at 12:30 PM so pardon me if I screw up somewhere!)
main(...)
{
state = possiblyRestoreFromCheckpoint(args);
if (state!=null) // meaning we loaded from checkpoint
state.run(EvolutionState.C_STARTED_FROM_CHECKPOINT);
else
{
parameters = loadParameterDatabase(args);
state = initialize(parameters, 0); // no job offset
startFresh(state, null); // no job prefix
}
System.exit(0);
}
All the job stuff proceeds from there.
The way we're doing jobs is that ECJ maintains an Object[array] in
EvolutionState.java called "job". You can use this array for
whatever you want if you're writing your own main(). Or you can, in
the example above, not use it at all. The purpose of the array is to
enable you to store your job information and have it survive
checkpoints and restores. There is another variable in
EvolutionState also called "runtimeArguments", which is the array of
strings passed in as args[] originally to main(). Because jobs might
need that info to survive too.
The basic main() we have is very simple -- it just increments a
single job number, starting at 0. We store the number as an Integer
in jobs[0]. So here's some explanation of the code. Keep in mind
this is the DEFAULT main(). ECJ has been carefully arranged so that
you can do your own main loop in any way you like.
1. We first try to load the state from a checkpoint if there was
one. If so, we finish that run.
state = possiblyRestoreFromCheckpoint(args);
if (state!=null)
state.run(EvolutionState.C_STARTED_FROM_CHECKPOINT);
2. Okay, that job is finished if it existed. Now, we make the job
number 0 if we're starting from the very beginning (not from
checkpoint), but if we had loaded from checkpoint, we want to set the
new job number to the old job number (stored away in job[0]) plus 1.
int jobCurrent = 0;
if (state != null)
{
try
{
if (state.runtimeArguments == null)
Output.initialError("...");
args = state.runtimeArguments;
jobCurrent = ((Integer)(state.job[0])).intValue() + 1;
}
catch (Exception e)
{
Output.initialError("...");
}
}
3. Now we load the parameter database. Here's where the confusion
comes in. We're loading the parameter database every job in our
simple main(). Why do we do this? Mostly because parameter
databases can get modified at runtime by the experimenter. Otherwise
there's not much reason I think. The ParameterDatabase has a File
called "directory" which might be problematic if it's restored from
checkpoint, and a Vector called listeners, but I don't think either
is probably a problem in most cases. So sure, you could try just
loading the parameter database only if we're starting fresh, and
reusing the database in other cases. I was throwing away the old
database and reloading it because it was cleaner and less likely to
generate experimenter-tickled bugs.
parameters = loadParameterDatabase(args);
int numJobs = parameters.getIntWithDefault(new Parameter
("jobs"), null, 1);
if (numJobs < 1)
Output.initialError("...");
4. Now we just loop through the jobs. Each time we reload the
parameter database if necessary, then initialize the EvolutionState
from the database (this calls all the setup() methods, and the random
number generators, more or less. The RNG's seeds are incremented
using the job# perhaps). Then we stash away the job number in the
EvolutionState in case it's going to get checkpointed out. We also
figure out the prefix that the job should use to make filenames
specific to that job so jobs don't write on each other's output
files. If only one job, we don't use a prefix (for backwards
compatibility). And then we start the run with startFresh! Then
trash the parameters:
for(int job = jobCurrent ; job < numJobs; job++)
{
if (parameters == null)
parameters = loadParameterDatabase(args);
state = initialize(parameters, job);
state.output.systemMessage("Job: " + job);
state.job = new Object[1];
state.job[0] = new Integer(job);
state.runtimeArguments = args;
String jobFilePrefix = null;
if (numJobs > 1)
jobFilePrefix = "job." + job + ".";
startFresh(state, jobFilePrefix);
parameters = null;
}
5. And then we quit.
System.exit(0);
In truth, you don't have to follow this scheme at all; you can setup
and pulse EvolutionState any way you like. Here's the general
procedure:
A. If starting up from checkpoint, call
Checkpoint.restoreFromCheckpoint(filename),
which does a bunch of stuff for you.
B. Else you need to make a parameter database and set up various things:
1. The Output and logs for stdout, stderr, etc.
2. The random number generators
3. The evolutionState, and set some variables in it
[All this is done by Evolve.initialize if you want to use that instead]
4. if (jobFilePrefix != null)
state.output.setFilePrefix(jobFilePrefix);
5. state.startFresh()
C. Now you have a functioning EvolutionState. Repeatedly call
result = state.evolve();
... until result = EvolutionState.R_NOTDONE
D. state.finish(result)
E. state.output.flush()
F. print out the used parameters etc. if you wish (See
Evolve.startFresh(...) for code)
G. If you're all done, flush various streams and close the Output,
and exit(0)
And you don't even have to load a parameter database from a file --
you can construct one by hand if you like.
But in explaining this on this email I see an ugliness I didn't clean
up. I use state.run(...) to finish out the checkpointed run, but I
use startFresh(...) to do a noncheckpointed run. startFresh(...)
does a bunch of items, including the same basic code as state.run
(...). But they're [1] not orthogonal and [2] startFresh cleans up
afterwards and dumps parameters but it looks like my main(...) code
does *not* do that if restoring from checkpoint. Not a big bug, but
I should make that prettier. Thanks guys, even if inadvertently!
Sean
|
