def runThread(numTrain, numTest, agent, maxJobs, results, finished):
exp = None
try:
# pretrain
exp = SimpleSimulation(numDevices=numDevices,
maxJobs=maxJobs,
agentClass=agent,
tasks=[HARD],
systemStateClass=extendedSystemState,
jobInterval=1)
exp.setBatterySize(1e-1)
exp.setFpgaIdleSleep(1e-3)
exp.simulateEpisodes(int(numTrain))
except:
traceback.print_exc(file=sys.stdout)
if exp is not None:
print("Error in experiment:", maxJobs, exp.time)
exp.sharedAgent.setProductionMode()
for i in range(int(numTest)):
exp.simulateEpisode(int(numTrain) + i)
results.put([
"Agent %s" % exp.sharedAgent.__name__, maxJobs, exp.numFinishedJobs
])
# results.put(["", jobInterval, np.average([dev.numJobs for dev in exp.devices]) / exp.getCompletedJobs()])
finished.put(True)
def runThread(agent, numEpisodes, centralised, results, finished):
exp = SimpleSimulation(numDevices=2,
maxJobs=maxjobs,
agentClass=agent,
tasks=[HARD],
systemStateClass=minimalSystemState,
scenarioTemplate=REGULAR_SCENARIO_ROUND_ROBIN,
centralisedLearning=centralised)
# exp.scenario.setInterval(1)
exp.setBatterySize(1e-1)
exp.setFpgaIdleSleep(1e-3)
e = None
try:
for e in range(numEpisodes):
debug.infoEnabled = False
exp.simulateEpisode(e)
# results.put(["%s %s" % (exp.devices[0].agent.__name__, "Centralised" if centralised else "Decentralised"), e, exp.numFinishedJobs])
results.put([
"%s %s" % (exp.devices[0].agent.__name__,
"Centralised" if centralised else "Decentralised"),
e,
exp.getCurrentTime()
])
except:
debug.printCache()
traceback.print_exc(file=sys.stdout)
print(agent, e)
print("Error in experiment ̰:", exp.time)
sys.exit(0)
finished.put(True)
def threadRun(episodeNum, results, finished):
exp = SimpleSimulation(numDevices=numDevices,
maxJobs=100,
tasks=[HARD],
centralisedLearning=False)
exp.setBatterySize(1e-1)
exp.setFpgaIdleSleep(1e-3)
for agent, device in zip(
[lazyTableAgent, minimalTableAgent, minimalTableAgent], exp.devices):
device.agent = agent(reconsiderBatches=False,
systemState=exp.currentSystemState,
owner=device,
offPolicy=exp.offPolicy)
device.agent.setDevices(exp.devices)
print("set agent", agent, agent.__name__, device.agent,
device.agent.__name__)
print([device.agent.__name__ for device in exp.devices])
for e in range(int(episodeNum)):
exp.simulateEpisode(e)
for device in exp.devices:
# print("putting results", device.agent.__name__, device.numJobsDone)
# results.put(["Agent %s" % device.agent.__name__, e, device.currentTime.current])
results.put([
"Device %d Agent %s" % (device.index, device.agent.__name__),
e, device.numJobsDone
])
sys.stdout.write("\rProgress: %.2f%%" % ((e + 1) / episodeNum * 100.))
finished.put(True)
def runThread(jobInterval, fpgaSleepTime, numEpisodes, results, finished):
exp = SimpleSimulation(numDevices=4, maxJobs=maxJobs)
exp.setFpgaIdleSleep(fpgaSleepTime)
exp.scenario.setInterval(jobInterval)
print('interval', exp.scenario.timeInterval.mean)
exp.setBatterySize(1e1)
try:
for i in range(numEpisodes):
# exp.simulateTime(10)
exp.simulateEpisode()
results.put([
"FPGA Idle Sleep {} Interval {}".format(
fpgaSleepTime, jobInterval), i,
exp.getCurrentTime()
])
except:
traceback.print_exc(file=sys.stdout)
print(
jobInterval,
fpgaSleepTime,
)
print("Error in experiment:", jobInterval, fpgaSleepTime,
exp.getCurrentTime())
finished.put(True)
def runThread(agent, numEpisodes, numDevices, taskOptions, results, finished):
constants.CENTRALISED_LEARNING = False
exp = SimpleSimulation(numDevices=numDevices, maxJobs=10, agentClass=agent, tasks=taskOptions, systemStateClass=targetedSystemState, reconsiderBatches=False, scenarioTemplate=RANDOM_SCENARIO_ROUND_ROBIN, centralisedLearning=False)
exp.scenario.setInterval(1)
exp.setFpgaIdleSleep(60)
exp.setBatterySize(1e0)
e = None
try:
for e in range(numEpisodes):
debug.infoEnabled = False
exp.simulateEpisode()
times = np.zeros((len(taskOptions),))
for device in exp.devices:
for t in range(len(taskOptions)):
task = taskOptions[t]
times[t] = device.fpga.getConfigTime(task)
results.put(["Device %d %% time EASY" % device.index, e, times[0]/np.sum(times)])
# if times[1] == 0:
# ratio = inf
# else:
# ratio = times[0] / times[1]
# if ratio < 1 and ratio != 0: ratio = 1. / ratio
# results.put(["Device %d EASY/HARD" % (device.index), e, ratio])
# print(e, "Device %d Task %s" % (device.index, task), e, device.getNumTasksDone(task))
except:
debug.printCache()
traceback.print_exc(file=sys.stdout)
print(agent, e)
print("Error in experiment ̰:", exp.time)
sys.exit(0)
finished.put(True)
def runThread(agent, numEpisodes, taskOptions, results, finished):
exp = SimpleSimulation(numDevices=4,
maxJobs=6,
agentClass=agent,
tasks=taskOptions)
exp.scenario.setInterval(1)
exp.setFpgaIdleSleep(5)
exp.setBatterySize(1e0)
e = None
try:
for e in range(numEpisodes):
debug.infoEnabled = False
exp.simulateEpisode()
results.put([
"Agent %s (%s)" % (agent.__name__, len(taskOptions)), e,
exp.numFinishedJobs
])
except:
debug.printCache()
traceback.print_exc(file=sys.stdout)
print(agent, e)
print("Error in experiment ̰:", exp.time)
sys.exit(0)
finished.put(True)
# assert simulationResults.learningHistory is not None
# histories.put(simulationResults.learningHistory)
# print("\nsaving history", simulationResults.learningHistory, '\nr')
print("forward", counters.NUM_FORWARD, "backward", counters.NUM_BACKWARD)
def runThread(agent, numEpisodes, results, finished, histories):
exp = SimpleSimulation(numDevices=2, maxJobs=6, agentClass=agent)
exp.setFpgaIdleSleep(5)
exp.setBatterySize(1e0)
try:
for e in range(numEpisodes):
debug.infoEnabled = False
exp.simulateEpisode()
results.put(["Duration %s" % exp.sharedAgent, e, exp.getCurrentTime()])
# results.put(["Episode reward %s" % exp.sharedAgent, e, exp.sharedAgent.episodeReward])
except:
debug.printCache(200)
traceback.print_exc(file=sys.stdout)
print(agent, e)
print("Error in experiment ̰:", exp.time)
sys.exit(0)
finished.put(True)
assert simulationResults.learningHistory is not None
histories.put(simulationResults.learningHistory)
print("\nsaving history", simulationResults.learningHistory, '\nr')
print("forward", counters.NUM_FORWARD, "backward", counters.NUM_BACKWARD)
exp.sharedAgent.printModel()
def runThread(agent, numEpisodes, numDevices, taskOptions, results, finished):
constants.CENTRALISED_LEARNING = False
exp = SimpleSimulation(numDevices=numDevices, maxJobs=10, agentClass=agent, tasks=taskOptions, systemStateClass=targetedSystemState, reconsiderBatches=False, scenarioTemplate=RANDOM_SCENARIO_ROUND_ROBIN, centralisedLearning=False)
exp.scenario.setInterval(1)
exp.setFpgaIdleSleep(60)
exp.setBatterySize(1e0)
e = None
try:
for e in range(numEpisodes):
debug.infoEnabled = False
exp.simulateEpisode()
for device in exp.devices:
for task in taskOptions:
results.put(["Device %d Task %d" % (device.index, task.identifier), e, device.getNumTasksDone(task)])
# print(e, "Device %d Task %s" % (device.index, task), e, device.getNumTasksDone(task))
except:
debug.printCache()
traceback.print_exc(file=sys.stdout)
print(agent, e)
print("Error in experiment ̰:", exp.time)
sys.exit(0)
finished.put(True)
def runThread(agent, numEpisodes, results, finished, histories):
exp = SimpleSimulation(numDevices=1, maxJobs=3, agentClass=agent)
exp.setFpgaIdleSleep(5)
exp.setBatterySize(1e0)
# sim.simulations.constants.FPGA_IDLE_SLEEP = 5
# sim.simulations.constants.OFFLOADING_POLICY = REINFORCEMENT_LEARNING
# sim.simulations.constants.TOTAL_TIME = 1e3
try:
for e in range(numEpisodes):
debug.infoEnabled = False
exp.simulateEpisode()
results.put([
"Episode reward %s" % exp.sharedAgent, e,
exp.sharedAgent.episodeReward
])
except:
debug.printCache(200)
traceback.print_exc(file=sys.stdout)
print(agent, e)
print("Error in experiment ̰:", exp.time)
sys.exit(0)
finished.put(True)
assert simulationResults.learningHistory is not None
histories.put(simulationResults.learningHistory)
print("\nsaving history", simulationResults.learningHistory, '\nr')
print("forward", counters.NUM_FORWARD, "backward", counters.NUM_BACKWARD)
exp.sharedAgent.printModel()
def runThread(agent, numEpisodes, results, finished):
exp = SimpleSimulation(numDevices=2,
maxJobs=maxjobs,
agentClass=agent,
tasks=[HARD],
systemStateClass=minimalSystemState,
scenarioTemplate=REGULAR_SCENARIO_ROUND_ROBIN,
centralisedLearning=True,
numEnergyLevels=numEnergyStates,
trainClassification=True)
# exp.scenario.setInterval(1)
exp.sharedAgent.loadModel()
exp.sharedAgent.setProductionMode()
exp.setBatterySize(1e-1)
exp.setFpgaIdleSleep(1e-3)
e = None
try:
for e in range(numEpisodes):
debug.infoEnabled = False
exp.simulateEpisode(e)
agentName = exp.devices[0].agent.__name__
result = [f"{agentName}", e, exp.numFinishedJobs]
print(result)
results.put(result)
# results.put([f"{agentName}", e, exp.getCurrentTime()])
except:
debug.printCache()
traceback.print_exc(file=sys.stdout)
print(agent, e)
print("Error in experiment :", exp.time)
sys.exit(0)
finished.put(True)
def runThread(agent, numEpisodes, numDevices, taskOptions, interval, results,
finished):
constants.CENTRALISED_LEARNING = False
exp = SimpleSimulation(numDevices=numDevices,
maxJobs=50,
agentClass=agent,
tasks=taskOptions,
systemStateClass=targetedSystemState,
reconsiderBatches=False,
scenarioTemplate=RANDOM_SCENARIO_ROUND_ROBIN,
centralisedLearning=False)
exp.scenario.setInterval(interval)
exp.setFpgaIdleSleep(1e-3)
exp.setBatterySize(1e-1)
assert numEpisodes % 2 == 0
offset = 0
e = 0
reduced = False
try:
#pretrain
# for e in range(0):
# exp.simulateEpisode()
debug.infoEnabled = False
for i in range(2):
for e in range(int(numEpisodes / 2)):
exp.simulateEpisode(e)
dol_ind_task, dol_task_ind = DOL(exp.devices, taskOptions)
results.put(
["DOL %d devices" % numDevices, offset + e, dol_ind_task])
results.put([
"Jobs Completed %d devices" % numDevices, offset + e,
exp.numFinishedJobs
])
# results.put(["Interval %.2f" % interval, e, dol_ind_task])
if not reduced:
print()
# remove half
for i in range(int(numDevices / 2)):
exp.removeDevice()
reduced = True
print("reduce to", exp.devices)
offset = e
finished.put("")
except:
debug.printCache()
traceback.print_exc(file=sys.stdout)
print(agent, e)
print("Error in experiment ̰:", exp.time)
print(agent, numEpisodes, numDevices, taskOptions, interval, e, offset,
reduced)
sys.exit(0)
finished.put(True)
def runThread(agent, numEpisodes, numDevices, taskOptions, interval, results, finished):
constants.CENTRALISED_LEARNING = False
exp = SimpleSimulation(numDevices=numDevices, maxJobs=10, agentClass=agent, tasks=taskOptions, systemStateClass=targetedSystemState, reconsiderBatches=False, scenarioTemplate=RANDOM_SCENARIO_ROUND_ROBIN, centralisedLearning=False)
exp.scenario.setInterval(interval)
exp.setFpgaIdleSleep(5)
exp.setBatterySize(1e0)
e = None
try:
#pretrain
for e in range(0):
exp.simulateEpisode()
for e in range(numEpisodes):
debug.infoEnabled = False
exp.simulateEpisode()
times = np.zeros((len(taskOptions),))
percentages = np.zeros((numDevices,))
for device in exp.devices:
for t in range(len(taskOptions)):
task = taskOptions[t]
times[t] = device.fpga.getConfigTime(task)
sum = np.sum(times)
if sum > 0:
perc = times[0] / sum
else:
perc = 0.5
if perc < 0.5: perc = 1. - perc
percentages[device.index] = perc
# results.put(["", np.average(percentages), exp.numFinishedJobs])
results.put(["Interval %.2f" % interval, np.average(percentages), exp.numFinishedJobs])
# for device in exp.devices:
# device.agent.setProductionMode()
# exp.simulateEpisode()
finished.put("")
# if times[1] == 0:
# ratio = inf
# else:
# ratio = times[0] / times[1]
# if ratio < 1 and ratio != 0: ratio = 1. / ratio
# results.put(["Device %d EASY/HARD" % (device.index), e, ratio])
# print(e, "Device %d Task %s" % (device.index, task), e, device.getNumTasksDone(task))
except:
debug.printCache()
traceback.print_exc(file=sys.stdout)
print(agent, e)
print("Error in experiment ̰:", exp.time)
sys.exit(0)
finished.put(True)
def runThread(id, agent, systemState, productionMode, offPolicy, numPhases,
numEpisodes, results, finished):
exp = SimpleSimulation(numDevices=4,
maxJobs=maxjobs,
agentClass=agent,
tasks=[HARD],
systemStateClass=systemState,
scenarioTemplate=REGULAR_SCENARIO_ROUND_ROBIN,
centralisedLearning=True,
numEnergyLevels=numEnergyStates,
trainClassification=True,
offPolicy=offPolicy,
allowExpansion=True)
exp.setBatterySize(1e-1)
exp.setFpgaIdleSleep(1e-3)
exp.sharedAgent.loadModel()
exp.sharedAgent.setProductionMode(productionMode)
e = None
overallEpisode = 0
try:
for phase in range(numPhases):
for e in range(numEpisodes):
debug.infoEnabled = False
exp.simulateEpisode(e)
agentName = exp.devices[0].agent.__name__
result = [
f"{agentName} Production: {productionMode} OffPolicy: {offPolicy}",
overallEpisode + e, exp.numFinishedJobs
]
# print("result", result)
results.put(result)
# results.put([f"{agentName}", e, exp.getCurrentTime()])
# check if not the last one
if phase < numPhases - 1:
beforeStates = exp.currentSystemState.getUniqueStates()
for i in range(5):
exp.expandState("jobsInQueue")
# print("\nexpand:", beforeStates, exp.currentSystemState.getUniqueStates())
# print()
overallEpisode += numEpisodes
except:
debug.printCache()
traceback.print_exc(file=sys.stdout)
print(agent, e, offPolicy, productionMode)
print("Error in experiment :", exp.time)
sys.exit(0)
finished.put(True)
def runThread(id, agent, productionMode, offPolicy, numEpisodes, results,
finished):
startTime = datetime.now()
exp = SimpleSimulation(numDevices=4,
maxJobs=maxjobs,
agentClass=agent,
tasks=[HARD],
systemStateClass=extendedSystemState,
scenarioTemplate=REGULAR_SCENARIO_ROUND_ROBIN,
centralisedLearning=True,
numEnergyLevels=numEnergyStates,
trainClassification=True,
offPolicy=offPolicy)
exp.sharedAgent.precache = True
# exp.scenario.setInterval(1)
exp.sharedAgent.loadModel()
if productionMode:
exp.sharedAgent.setProductionMode()
exp.setBatterySize(1e-1)
exp.setFpgaIdleSleep(1e-3)
e = None
try:
for e in range(numEpisodes):
debug.infoEnabled = False
exp.simulateEpisode(e)
agentName = exp.devices[0].agent.__name__
result = [
f"{agentName} PM: {productionMode} OP: {offPolicy}", e,
exp.numFinishedJobs
]
results.put(result)
# result = [f"{agentName} PM: {productionMode} OP: {offPolicy} JOBS", e, exp.jobCounter]
# results.put(result)
except:
debug.printCache()
traceback.print_exc(file=sys.stdout)
print(agent, e)
print("Error in experiment :", exp.time)
sys.exit(0)
finished.put(True)
print(
f"duration: {agent} PM {productionMode} OP {offPolicy}: {datetime.now() - startTime}"
)
def runThread(agent, numEpisodes, results, finished):
# constants.CENTRALISED_LEARNING = False
exp = SimpleSimulation(numDevices=2,
maxJobs=maxjobs,
agentClass=agent,
tasks=[HARD],
systemStateClass=minimalSystemState,
scenarioTemplate=REGULAR_SCENARIO_ROUND_ROBIN,
reconsiderBatches=False,
centralisedLearning=False)
# exp.scenario.setInterval(1)
exp.setFpgaIdleSleep(1e-3)
exp.setBatterySize(1e-1)
e = None
try:
for e in range(numEpisodes):
debug.infoEnabled = False
exp.simulateEpisode(e)
results.put([
"Job %s" % exp.devices[0].agent.__name__, e,
exp.numFinishedJobs
])
results.put([
"Duration %s" % exp.devices[0].agent.__name__, e,
exp.getCurrentTime()
])
except:
debug.printCache()
traceback.print_exc(file=sys.stdout)
print(agent, e)
print("Error in experiment ̰:", exp.time)
sys.exit(0)
finished.put(True)
# assert simulationResults.learningHistory is not None
# histories.put(simulationResults.learningHistory)
# print("\nsaving history", simulationResults.learningHistory, '\nr')
print("forward", counters.NUM_FORWARD, "backward", counters.NUM_BACKWARD)
if exp.sharedAgent.__class__ == minimalTableAgent:
plotting.plotModel(exp.sharedAgent, drawLabels=False)
def runThread(agent, numEpisodes, numDevices, taskOptions, interval, results,
finished):
constants.CENTRALISED_LEARNING = False
exp = SimpleSimulation(numDevices=numDevices,
maxJobs=10,
agentClass=agent,
tasks=taskOptions,
systemStateClass=targetedSystemState,
reconsiderBatches=False,
scenarioTemplate=RANDOM_SCENARIO_ROUND_ROBIN,
centralisedLearning=False)
exp.scenario.setInterval(interval)
exp.setFpgaIdleSleep(5)
exp.setBatterySize(1e0)
e = None
try:
#pretrain
# for e in range(0):
# exp.simulateEpisode()
for e in range(numEpisodes):
debug.infoEnabled = False
exp.simulateEpisode()
dol_ind_task, dol_task_ind = DOL(exp.devices, taskOptions)
results.put(["DOL Devices %d" % numDevices, e, dol_ind_task])
results.put([
"Jobs Devices %d" % numDevices, e, exp.numFinishedJobs / 1000
])
# results.put(["Interval %.2f" % interval, e, dol_ind_task])
finished.put("")
except:
debug.printCache()
traceback.print_exc(file=sys.stdout)
print(agent, e)
print("Error in experiment ̰:", exp.time)
sys.exit(0)
finished.put(True)
def threadRun(percentageMinimal, episodeNum, results, finished):
exp = SimpleSimulation(numDevices=numDevices,
maxJobs=50,
tasks=[HARD],
centralisedLearning=False,
systemStateClass=minimalSystemState,
scenarioTemplate=REGULAR_SCENARIO_ROUND_ROBIN)
exp.scenario.setInterval(interval / numDevices)
exp.setBatterySize(1e-1)
exp.setFpgaIdleSleep(1e-3)
numMinimal = int(numDevices * percentageMinimal)
numRandom = numDevices - numMinimal
agents = [minimalTableAgent] * numMinimal + [randomAgent] * numRandom
for agent, device in zip(agents, exp.devices):
device.agent = agent(reconsiderBatches=False,
systemState=exp.currentSystemState,
owner=device,
offPolicy=exp.offPolicy)
device.agent.setDevices(exp.devices)
# print("set agent", agent, agent.__name__, device.agent, device.agent.__name__)
for e in range(int(episodeNum)):
exp.simulateEpisode(e)
totalMinimalAgentJobs = 0
for i in range(numMinimal):
dev = exp.devices[i]
assert dev.agent.__class__ == minimalTableAgent
totalMinimalAgentJobs += dev.numJobsDone
results.put([
"%d %% Basic Agents" % (int(percentageMinimal * 100.)), e,
totalMinimalAgentJobs / numMinimal
])
# sys.stdout.write("\rProgress: %.2f%%" % ((e+1)/episodeNum*100.))
finished.put(True)
def runThread(results, finished):
exp = SimpleSimulation(numDevices=4,
maxJobs=100,
reconsiderBatches=False,
systemStateClass=minimalSystemState,
agentClass=minimalTableAgent)
exp.setFpgaIdleSleep(1e-2)
exp.setBatterySize(1e6)
try:
for i in range(numJobs):
exp.simulateUntilJobDone()
results.put(
["Loss", exp.completedJobs, exp.sharedAgent.latestLoss, False])
# results.put(["Reward", exp.completedJobs, exp.sharedAgent.latestReward, True])
# results.put(["Action", exp.completedJobs, exp.sharedAgent.latestAction, True])
# results.put(["MAE", exp.completedJobs, exp.sharedAgent.latestMAE, True])
# results.put(["MeanQ", exp.completedJobs, exp.sharedAgent.latestMeanQ, True])
except:
traceback.print_exc(file=sys.stdout)
sys.exit(0)
print("Error in experiment:", exp.time)
finished.put(True)
def runThread(jobInterval, fpgaSleepTime, numEpisodes, results, finished):
exp = SimpleSimulation(numDevices=4)
exp.scenario.setInterval(1)
exp.setFpgaIdleSleep(fpgaSleepTime)
exp.setBatterySize(1e1)
try:
for i in range(numEpisodes):
# exp.simulateTime(10)
exp.simulateEpisode()
except:
traceback.print_exc(file=sys.stdout)
print(
jobInterval,
fpgaSleepTime,
)
print("Error in experiment:", jobInterval, fpgaSleepTime,
exp.getCurrentTime())
results.put([
"FPGA Idle Sleep {}".format(fpgaSleepTime), jobInterval,
np.sum(exp.totalDevicesEnergy()) / exp.getNumDevices()
])
finished.put(True)
def singleThread(ID, depth, width, loss, epochs, results):
# import tensorflow as tf
# import tensorflow.config.gpu
# tf.config.gpu.set_per_process_memory_growth(True)
# tf.debugging.set_log_device_placement(True)
# create deep agent for learning
dqnExp = SimpleSimulation(numDevices=4,
maxJobs=MAX_JOBS,
agentClass=dqnAgent,
tasks=[HARD],
systemStateClass=systemStateClass,
scenarioTemplate=REGULAR_SCENARIO_ROUND_ROBIN,
centralisedLearning=True,
numEnergyLevels=NUM_ENERGY_STATES,
trainClassification=CLASSIFICATION)
dqnExp.setBatterySize(1e-1)
dqnExp.setFpgaIdleSleep(1e-3)
# %%
# train deep agent
deepAgent = dqnExp.sharedAgent
deepAgent.loss = loss
deepAgent.createModel(depth, width)
deepAgent.fit(trainingData,
trainingTargetOneHot,
epochs=epochs,
split=0.0,
class_weights=class_weights_dict)
if CLASSIFICATION:
output = deepAgent.model.evaluate(trainingData,
trainingTargetOneHot,
verbose=0)
else:
output = deepAgent.fullModel.evaluate(
trainingData, trainingTargetOneHot.argmax(axis=1), verbose=0)
predictions = deepAgent.model.predict_on_batch(trainingData)
# print("evaluate: ", ID, output)
if CLASSIFICATION:
confusion = metrics.confusion_matrix(trainingTarget,
predictions.argmax(axis=1))
else:
# print(trainingTargetOneHot.argmax(axis=1))
onehot = np.eye(
np.shape(trainingTargetOneHot)[1])[trainingTargetOneHot.argmax(
axis=1)]
# print(onehot)
# print(deepAgent.fullModel.predict_on_batch(trainingData).argmax(axis=1))
confusion = metrics.confusion_matrix(
trainingTargetOneHot.argmax(axis=1),
deepAgent.fullModel.predict_on_batch(trainingData).argmax(axis=1))
# full estimate
# print(deepAgent.model.predict_on_batch(trainingData))
# print(deepAgent.fullModel.predict_on_batch(trainingData))
# deepAgent.model.summary()
# deepAgent.fullModel.summary()
# print('done')
deepAgent.saveModel(ID)
result = (depth, width, loss, output[1], confusion, epochs, ID)
# print(result)
results.put(result)
def runThread(numEpisodes, results, finished):
dqnExp = SimpleSimulation(numDevices=4,
maxJobs=maxjobs,
agentClass=dqnAgent,
tasks=[HARD],
systemStateClass=minimalSystemState,
scenarioTemplate=REGULAR_SCENARIO_ROUND_ROBIN,
centralisedLearning=True,
numEnergyLevels=numEnergyStates,
trainClassification=False)
# exp.scenario.setInterval(1)
dqnExp.sharedAgent.loadModel()
dqnExp.sharedAgent.setProductionMode()
dqnExp.setBatterySize(1e-1)
dqnExp.setFpgaIdleSleep(1e-3)
tableExp = SimpleSimulation(numDevices=4,
maxJobs=maxjobs,
agentClass=minimalTableAgent,
tasks=[HARD],
systemStateClass=minimalSystemState,
scenarioTemplate=REGULAR_SCENARIO_ROUND_ROBIN,
centralisedLearning=True,
numEnergyLevels=numEnergyStates,
trainClassification=False)
# exp.scenario.setInterval(1)
tableExp.sharedAgent.loadModel()
tableExp.sharedAgent.setProductionMode()
tableExp.setBatterySize(1e-1)
tableExp.setFpgaIdleSleep(1e-3)
print("states:", tableExp.sharedAgent.systemState.uniqueStates)
# iterate through possible states
for i in range(tableExp.sharedAgent.systemState.uniqueStates):
# set state in both experiments
state = tableExp.sharedAgent.systemState.fromIndex(i)
# table
agentName = tableExp.sharedAgent.__name__
action = tableExp.sharedAgent.selectAction(state.currentState)
result = [f"{agentName}", i, action]
results.put(result)
# dqn
dqnState = dqnExp.sharedAgent.systemState.fromIndex(i)
agentName = dqnExp.sharedAgent.__name__
# action = dqnExp.sharedAgent.selectAction(state.currentState)
action = dqnExp.sharedAgent.policy.select_action(
q_values=tableExp.sharedAgent.predict(state))
result = [f"{agentName}", i, action]
results.put(result)
# results.put([f"{agentName}", e, exp.getCurrentTime()])
# except:
# debug.printCache()
# traceback.print_exc(file=sys.stdout)
# print(agent, e)
# print("Error in experiment :", exp.time)
# sys.exit(0)
finished.put(True)
# %%
# create table agent for training data
tableExp = SimpleSimulation(numDevices=4,
maxJobs=MAX_JOBS,
agentClass=minimalTableAgent,
tasks=[HARD],
systemStateClass=systemStateClass,
scenarioTemplate=REGULAR_SCENARIO_ROUND_ROBIN,
centralisedLearning=True,
numEnergyLevels=NUM_ENERGY_STATES,
trainClassification=CLASSIFICATION)
tableExp.sharedAgent.loadModel()
tableExp.setBatterySize(1e-1)
tableExp.setFpgaIdleSleep(1e-3)
# %%
def chooseState(state, chosenJobs, chosenBatteryLevel):
state = tableExp.sharedAgent.systemState
maxJobs = state.maxJobs
numBatteryLevels = state.singlesDiscrete['energyRemaining']
state.setField('jobsInQueue', chosenJobs / maxJobs)
state.setField('energyRemaining', chosenBatteryLevel / numBatteryLevels)
# %%
# prepare training and testing data
state = tableExp.sharedAgent.systemState
def runThread(numEpisodes, results, finished):
dqnExp = SimpleSimulation(numDevices=2,
maxJobs=maxjobs,
agentClass=dqnAgent,
tasks=[HARD],
systemStateClass=minimalSystemState,
scenarioTemplate=REGULAR_SCENARIO_ROUND_ROBIN,
centralisedLearning=True,
numEnergyLevels=numEnergyStates,
trainClassification=False)
# exp.scenario.setInterval(1)
dqnExp.sharedAgent.loadModel()
# dqnExp.sharedAgent.fullModel.summary()
# dqnExp.sharedAgent.model.summary()
dqnExp.sharedAgent.setProductionMode()
dqnExp.setBatterySize(1e-1)
dqnExp.setFpgaIdleSleep(1e-3)
tableExp = SimpleSimulation(numDevices=2,
maxJobs=maxjobs,
agentClass=minimalTableAgent,
tasks=[HARD],
systemStateClass=minimalSystemState,
scenarioTemplate=REGULAR_SCENARIO_ROUND_ROBIN,
centralisedLearning=True,
numEnergyLevels=numEnergyStates,
trainClassification=True)
# exp.scenario.setInterval(1)
tableExp.sharedAgent.loadModel()
tableExp.sharedAgent.setProductionMode()
tableExp.setBatterySize(1e-1)
tableExp.setFpgaIdleSleep(1e-3)
print("states:", tableExp.sharedAgent.systemState.uniqueStates)
duplicates = []
# iterate through possible states
for i in range(tableExp.sharedAgent.systemState.uniqueStates):
# set state in both experiments
state = tableExp.sharedAgent.systemState.fromIndex(i)
# table
agentName = tableExp.sharedAgent.__name__
qvalues = tableExp.sharedAgent.predict(state)
print(np.where(qvalues == np.max(qvalues))[0])
if np.where(qvalues == np.max(qvalues))[0].shape[0] > 1:
duplicates.append(i)
for j in range(tableExp.sharedAgent.numActions):
result = [f"{agentName} {j}", i, qvalues[j]]
results.put(result)
# dqn
agentName = dqnExp.sharedAgent.__name__
qvalues = dqnExp.sharedAgent.predict(dqnExp.sharedAgent.model,
state.currentState)
for j in range(dqnExp.sharedAgent.numActions):
result = [f"{agentName} {j}", i, qvalues[j]]
results.put(result)
# results.put([f"{agentName}", e, exp.getCurrentTime()])
print("DUPLICATES", duplicates)
# except:
# debug.printCache()
# traceback.print_exc(file=sys.stdout)
# print(agent, e)
# print("Error in experiment :", exp.time)
# sys.exit(0)
finished.put(True)
processes = list()
results = multiprocessing.Queue()
finished = multiprocessing.Queue()
localConstants.REPEATS = 1
# localConstants.REPEATS = 8
numEpisodes = int(1e3)
# agentsToTest = [minimalTableAgent]
minimalTableAgent # , localAgent]
agent = minimalTableAgent # [minimalAgent, lazyAgent]:
centralised = True
exp = SimpleSimulation(numDevices=4, maxJobs=maxjobs, agentClass=agent, tasks=[HARD], numEnergyLevels=numEnergyLevels, systemStateClass=extendedSystemState, scenarioTemplate=REGULAR_SCENARIO_ROUND_ROBIN, centralisedLearning=centralised, trainClassification=CLASSIFICATION)
print(exp.currentSystemState.uniqueStates)
# exp.scenario.setInterval(1)
exp.setBatterySize(1e-1)
exp.setFpgaIdleSleep(1e-3)
results = []
e = None
latest = None
try:
print("running episodes")
for e in range(numEpisodes):
if e % 100 == 0: print ('.')
debug.infoEnabled = False
exp.simulateEpisode(e)
latest = exp.numFinishedJobs
results.append(latest)
# results.put(["%s %s" % (exp.devices[0].agent.__name__, "Centralised" if centralised else "Decentralised"), e, exp.numFinishedJobs])
