def __init__(self, memory, shared, semaphore):
multiprocessing.Process.__init__(self)
# hyperparameters
self.TRAIN_MAX = 10
self.TRANSFER = 10
self.BATCH_SIZE = 128
#self.BATCH_SIZE = 5
self.GAMMA = 0.99
#self.SAMPLE_ALPHA = 0.5
#self.SAMPLE_EPISLON = 0.
#self.SAMPLE_BETA = 0.
#self.SAMPLE_S = 44.8
self.SAMPLE_S = 5.0
self.SAMPLE_Q = 1.0
LEARNING_RATE = 0.00025
MOMENTUM = 0.95
SQUARED_MOMENTUM = 0.95
MIN_SQUARED_GRAD = 0.01
self.net = DQN() # Deep Net
self.targetNet = DQN()
self.copy_weights()
self.net.setOptimizer(
optim.RMSprop(self.net.parameters(),
lr=LEARNING_RATE,
momentum=MOMENTUM,
alpha=SQUARED_MOMENTUM,
eps=MIN_SQUARED_GRAD))
self.memory = memory
self.shared = shared # shared resources, {'memory', 'SENT_FLAG'}
self.semaphore = semaphore
def __init__(self, inputs):
mp.Process.__init__(self)
self.BATCH_SIZE = 32
self.TRAIN_MAX = 500
self.TRANSFER = 100
self.GAMMA = 1.0
LEARNING_RATE = 0.00025
MOMENTUM = 0.95
SQUARED_MOMENTUM = 0.95
MIN_SQUARED_GRAD = 0.01
self.demonet = DQN()
self.targetnet = DQN()
self.copy_weights()
self.demonet.setOptimizer(
optim.RMSprop(self.demonet.parameters(),
lr=LEARNING_RATE,
momentum=MOMENTUM,
alpha=SQUARED_MOMENTUM,
eps=MIN_SQUARED_GRAD))
self.inputs = inputs
def __init__(self, shared):
multiprocessing.Process.__init__(self)
# hyperparameters
self.TRAIN_MAX = 500
self.TRANSFER = 100
self.BATCH_SIZE = 32
self.GAMMA = 1.0
LEARNING_RATE = 0.00025
MOMENTUM = 0.95
SQUARED_MOMENTUM = 0.95
MIN_SQUARED_GRAD = 0.01
self.net = DQN() # Deep Net
self.targetNet = DQN()
self.copy_weights()
self.net.setOptimizer(
optim.RMSprop(self.net.parameters(),
lr=LEARNING_RATE,
momentum=MOMENTUM,
alpha=SQUARED_MOMENTUM,
eps=MIN_SQUARED_GRAD))
self.shared = shared # shared resources, {'memory', 'SENT_FLAG'}
#from Env.Environment import Environment
from Env.gymEnv_V2 import myGym
#from Env.gymEnv import myGym
from DQN.Improver_Q_Learning import Improver
from DQN.Evaluator_Dense_Q_Learning import Evaluator
from DQN.ReplayMemory import ReplayMemory
import os
os.system(
"taskset -p 0xff %d" % os.getpid()
) #https://stackoverflow.com/questions/15639779/why-does-multiprocessing-use-only-a-single-core-after-i-import-numpy
if __name__ == '__main__':
# hyperparameters
MEMORY_SIZE = 5000
#MEMORY_SIZE = 5
imp_net = DQN()
# populate memory
# let improver populate first
manager = SyncManager()
manager.start()
memory = ReplayMemory(MEMORY_SIZE)
s = multiprocessing.Semaphore(1)
#memory = multiprocessing.Queue(MEMORY_SIZE)
memory = manager.list()
shared = manager.dict({'SENT_FLAG': True, 'weights': None})
#shared = manager.dict({'memory':memory, 'SENT_FLAG':True, 'weights':None})
#improver = Improver(imp_net, shared, myGym(), s)
improver = Improver(imp_net, MEMORY_SIZE, memory, shared, myGym(), s)
# improver is executed by the main process
evaluator = Evaluator(memory, shared, s)
from multiprocessing.managers import SyncManager
import torch.nn as nn
import time
from DQN.DQNcartpole import DQN
from Env.gymEnv import myGym
from DQN.CartPoleDQN import CartPoleDQN
from DQN.ReplayMemory import ReplayMemory
if __name__ == '__main__':
demonet = DQN()
#manager = SyncManager()
#manager.start()
memory = ReplayMemory(10000)
#for i in range(memory.capacity):
# memory.push(torch.FloatTensor(1, 3, 40, 80))
shared = dict({'memory': memory, 'SENT_FLAG': True, 'weights': None})
p = CartPoleDQN(DQN(), shared, myGym())
p.run()