def runTester(testWeightsQueue, resultQueue, abort):
try:
env = getEnv()
env.reset()
tester = Tester(env)
tester.build()
while not abort.value:
weights = testWeightsQueue.get()
tester.model.set_weights(weights)
log = tester.runOneEpisode(abort)
resultQueue.put(log)
except KeyboardInterrupt:
pass
def runExplorer(ID, weightsQueue, expQueue, abort, weightNoiseProb=0.):
try:
env = getEnv()
env.reset()
nc = noiseConfig[ID % len(noiseConfig)]
noise = getNoise(nc, env.action_space.shape[0])
exp = Explorer(ID=ID,
env=env,
weightNoiseProb=weightNoiseProb,
noiseProcess=noise)
exp.build()
while not abort.value:
weights = weightsQueue.get()
exp.model.set_weights(weights)
results = exp.runOneEpisode(abort)
expQueue.put(results)
except KeyboardInterrupt:
pass
parser.add_argument('--interval',
dest='interval',
action='store',
default=100,
type=int) #checkpoint interval
parser.add_argument('--obs', dest='obs', action='store_true', default=False)
parser.add_argument('--vis',
dest='visualize',
action='store_true',
default=False)
args = parser.parse_args()
from source.agents.testActor import TestActor
from config import getActor, getEnv, ENV_TAG
env = getEnv(visualize=args.visualize)
env.reset()
if args.visualize:
if ENV_TAG == 'OSIM':
vis = env.osim_model.model.updVisualizer().updSimbodyVisualizer()
vis.setBackgroundType(vis.GroundAndSky)
actor = getActor(env)
print(actor.summary())
tester = TestActor(actor)
epoch = args.start
log = []
while True:
try:
wQueues.append(weightsQueue)
from source.agents.tester import runTester
p = Process(target=runTester, args=(
testWeights,
testResults,
abort,
))
p.daemon = True
p.start()
#### Initiate main process agent for training ####
from source.customKerasLayers import LayerNormDense
from source.agents.customDDPG import CollectiveDDPG
env = getEnv()
env.reset()
numActions = env.action_space.shape[0]
memory_ = memory
optimizer_ = optimizer
actor, critic, metrics_, action_input = getActorCritic(env)
print(actor.summary())
print(critic.summary())
agent = CollectiveDDPG(weightNoiseProb=wnprob,
nb_actions=numActions,
actor=actor,
critic=critic,
critic_action_input=action_input,