@Override
public void moveToFinetune() {
log.info("Moving to finetune");
isPretrain.set(false);
// new phase: resets the counter
numBatches.set(0);
}
@Override
public boolean addJobToCurrent(Job j) throws Exception {
IAtomicReference<Job> r = h.getAtomicReference("job-" + j.getWorkerId());
if (r.get() != null || !r.isNull()) {
boolean sent = false;
while (!sent) {
// always update
for (String s : workers()) {
if (jobFor(s) == null) {
log.info(
"Redirecting worker "
+ j.getWorkerId()
+ " to "
+ s
+ " due to work already being allocated");
r = h.getAtomicReference("job-" + s);
j.setWorkerId(s);
sent = true;
}
}
}
}
r.set(j);
// iterate over jobs without the work/data
j.setWork(null);
jobs.add(j);
return true;
}
@Test
public void testHazelcastInstances() {
assertNotNull(map1);
assertNotNull(map2);
assertNotNull(multiMap);
assertNotNull(queue);
assertNotNull(topic);
assertNotNull(set);
assertNotNull(list);
assertNotNull(executorService);
assertNotNull(idGenerator);
assertNotNull(atomicLong);
assertNotNull(atomicReference);
assertNotNull(countDownLatch);
assertNotNull(semaphore);
assertNotNull(lock);
assertEquals("map1", map1.getName());
assertEquals("map2", map2.getName());
assertEquals("testMultimap", multiMap.getName());
assertEquals("testQ", queue.getName());
assertEquals("testTopic", topic.getName());
assertEquals("set", set.getName());
assertEquals("list", list.getName());
assertEquals("idGenerator", idGenerator.getName());
assertEquals("atomicLong", atomicLong.getName());
assertEquals("atomicReference", atomicReference.getName());
assertEquals("countDownLatch", countDownLatch.getName());
assertEquals("semaphore", semaphore.getName());
}
@Override
public void clearJob(String id) throws Exception {
if (id == null) {
log.warn("No job to clear; was null, returning");
return;
}
IAtomicReference<Job> jRef = h.getAtomicReference("job-" + id);
if (jRef.isNull()) return;
jRef.clear();
log.info("Destroyed job ref " + id);
Job remove = null;
for (Job j : jobs) {
if (j.getWorkerId().equals(id)) {
remove = j;
break;
}
}
jobs.remove(remove);
}
@Override
public boolean isDone() {
// reason being that isDone() may getFromOrigin called and throw errors
// this ensures a safe method call happens and just silently
// returns true in case hazelcast is shutdown
try {
return done.get();
} catch (Exception e) {
log.warn("Hazelcast already shutdown...returning true on isDone()");
return true;
}
}
@Override
public void finish() {
// reason being that isDone() may getFromOrigin called and throw errors
// this ensures a safe method call happens and just silently
// returns true in case hazelcast is shutdown
try {
done.set(true);
updateSaver().cleanup();
} catch (Exception e) {
log.warn(
"Hazelcast already shutdown...done() being applyTransformToDestination is pointless");
}
}
/**
* Creates a training evaluator using the given neural network
*
* @param network the neural network to use
* @return a training evaluator based on the configuration of the state tracker and the given
* network
*/
@Override
public TrainingEvaluator create(BaseMultiLayerNetwork network) {
OutputLayerTrainingEvaluator eval =
new OutputLayerTrainingEvaluator.Builder()
.bestLoss(bestLoss())
.improvementThreshold(improvementThreshold())
.patience(patience())
.testSet(testSet())
.withNetwork(network)
.validationEpochs(validationEpochs())
.patienceIncrease(patienceIncrease.get())
.build();
return eval;
}
@Override
public E getCurrent() throws Exception {
E u = (E) master.get();
if (u == null) return null;
return u;
}
/**
* Whether the cluster has begun training
*
* @return whether the cluster has begun training
*/
@Override
public boolean hasBegun() {
return begunTraining.get();
}
@Override
public void incrementNumTimesPreTrainRan() {
numTimesPretrainRan.set(numTimesPreTrainRun() + 1);
}
@Override
public int runPreTrainIterations() {
return numTimesPretrain.get();
}
@Override
public Job jobFor(String id) {
IAtomicReference<Job> j = h.getAtomicReference("job-" + id);
if (j.isNull() || isCurrentlyJob(id)) return null;
return j.get();
}
/**
* The number of epochs to test on
*
* @return the number of epochs to test on
*/
@Override
public int validationEpochs() {
return validationEpochs.get();
}
/**
* The best validation loss so far
*
* @return the best validation loss so far
*/
@Override
public double bestLoss() {
return bestLoss.get();
}
/**
* Improvement threshold for early stopping, aka the minimum
*
* @return
*/
@Override
public double improvmentThreshold() {
return improvementThreshold.get();
}
/**
* Patience is what controls early stopping
*
* @return the patience for the trainer
*/
@Override
public double patience() {
return patience.get();
}
/**
* The patience improvement to use
*
* @param improvmentThreshold the patience improvement to applyTransformToDestination
*/
@Override
public void setImprovmentThreshold(double improvmentThreshold) {
improvementThreshold.set(improvmentThreshold);
}
@Override
public boolean isPretrain() {
return isPretrain.get();
}
/**
* Current mini batch size
*
* @return
*/
@Override
public int miniBatchSize() {
return miniBatchSize.get();
}
/**
* Whether to validate against a held out test applyTransformToDestination and test for validation
* error.
*
* @return whether to validate against a held out test applyTransformToDestination and test for
* validation error.
*/
@Override
public boolean isEarlyStopTesting() {
return earlyStop.get();
}
@Override
public void runPreTrainIterations(int numTimes) {
numTimesPretrain.set(numTimes);
}
/**
* The input split to use. This means that each data applyTransformToDestination that is trained
* on and loaded will be this batch size or lower per worker
*
* @return the input split to use
*/
@Override
public int inputSplit() {
return (miniBatchSize.get() * numWorkers()) / numWorkers();
}
@Override
public int numTimesPreTrainRun() {
return numTimesPretrainRan.get();
}
public BaseHazelCastStateTracker(String connectionString, String type, int stateTrackerPort)
throws Exception {
log.info(
"Setting up hazelcast with type "
+ type
+ " connection string "
+ connectionString
+ " and port "
+ stateTrackerPort);
if (type.equals("master") && !PortTaken.portTaken(stateTrackerPort)) {
// sets up a proper connection string for reference wrt external actors needing a reference
if (connectionString.equals("master")) {
String host = InetAddress.getLocalHost().getHostName();
this.connectionString = host + ":" + stateTrackerPort;
}
this.hazelCastPort = stateTrackerPort;
config = hazelcast();
h = Hazelcast.newHazelcastInstance(config);
h.getCluster()
.addMembershipListener(
new MembershipListener() {
@Override
public void memberAdded(MembershipEvent membershipEvent) {
log.info("Member added " + membershipEvent.toString());
}
@Override
public void memberRemoved(MembershipEvent membershipEvent) {
log.info("Member removed " + membershipEvent.toString());
}
@Override
public void memberAttributeChanged(MemberAttributeEvent memberAttributeEvent) {
log.info("Member changed " + memberAttributeEvent.toString());
}
});
} else if (type.equals("master") && PortTaken.portTaken(stateTrackerPort))
throw new IllegalStateException(
"Specified type was master and the port specified was taken, please specify a different port");
else {
setConnectionString(connectionString);
log.info("Connecting to hazelcast on " + connectionString);
ClientConfig client = new ClientConfig();
client.getNetworkConfig().addAddress(connectionString);
h = HazelcastClient.newHazelcastClient(client);
}
this.type = type;
jobs = h.getList(JOBS);
workers = h.getList(WORKERS);
// we can make the assumption workers isn't empty because
// the master node by default comes with a applyTransformToDestination of workers
if (!this.type.equals("master")) {
while (workers.isEmpty()) {
log.warn("Waiting for data sync...");
Thread.sleep(1000);
}
log.info("Workers is " + workers.size());
}
begunTraining = h.getAtomicReference(BEGUN);
miniBatchSize = h.getAtomicReference(INPUT_SPLIT);
workerEnabled = h.getMap(WORKER_ENABLED);
replicate = h.getList(REPLICATE_WEIGHTS);
topics = h.getList(TOPICS);
updates = h.getList(UPDATES);
heartbeat = h.getMap(HEART_BEAT);
master = h.getAtomicReference(RESULT);
isPretrain = h.getAtomicReference(IS_PRETRAIN);
numTimesPretrain = h.getAtomicReference(NUM_TIMES_RUN_PRETRAIN);
numTimesPretrainRan = h.getAtomicReference(NUM_TIMES_PRETRAIN_RAN);
done = h.getAtomicReference(DONE);
validationEpochs = h.getAtomicReference(VALIDATION_EPOCHS);
improvementThreshold = h.getAtomicReference(IMPROVEMENT_THRESHOLD);
bestLoss = h.getAtomicReference(BEST_LOSS);
earlyStop = h.getAtomicReference(EARLY_STOP);
patience = h.getAtomicReference(PATIENCE);
patienceIncrease = h.getAtomicReference(PATIENCE_INCREASE);
numBatches = h.getAtomicReference(NUM_BATCHES_SO_FAR_RAN);
// applyTransformToDestination defaults only when master, otherwise, overrides previous values
if (type.equals("master")) {
begunTraining.set(false);
saver = createUpdateSaver();
numTimesPretrainRan.set(0);
numTimesPretrain.set(1);
isPretrain.set(true);
done.set(false);
resource = new StateTrackerDropWizardResource(this);
bestLoss.set(Double.POSITIVE_INFINITY);
earlyStop.set(true);
patience.set(40.0);
patienceIncrease.set(2.0);
improvementThreshold.set(0.995);
validationEpochs.set((int) Math.min(10, patience() / 2));
numBatches.set(0);
}
}
/**
* Increments the number of batches ran. This is purely a count and does not necessarily mean
* progress.
*
* @param numBatchesRan the number of batches ran to increment by
*/
@Override
public void incrementBatchesRan(int numBatchesRan) {
numBatches.set(numBatchesRan + numBatches.get());
}
/**
* Assuming a job already exists, updates the job
*
* @param j the job to update
*/
@Override
public void updateJob(Job j) {
IAtomicReference<Job> jRef = h.getAtomicReference("job-" + j.getWorkerId());
jRef.set(j);
}
/**
* Number of batches ran so far
*
* @return the number of batches ran so far
*/
@Override
public int numBatchesRan() {
return numBatches.get();
}
/** Begin training */
@Override
public void beginTraining() {
begunTraining.set(true);
}