def pongMaker():
return pong_env(seqLength=1, nBallsEpisode=20, useSeqLength=False)
def netMaker():
net = mxT.a3cHybridSequential(useInitStates= True)
net.add(mx.gluon.nn.Conv2D(channels = 32, kernel_size = (3,3), strides = (2,2), padding = (1,1), activation = None, prefix = "c1"))
net.add(mx.gluon.nn.ELU())
net.add(mx.gluon.nn.Conv2D(channels = 32, kernel_size = (3,3), strides = (2,2), padding = (1,1), activation = None, prefix = "c2"))
net.add(mx.gluon.nn.ELU())
net.add(mx.gluon.nn.Conv2D(channels = 32, kernel_size = (3,3), strides = (2,2), padding = (1,1), activation = None, prefix = "c3"))
net.add(mx.gluon.nn.ELU())
net.add(mx.gluon.nn.Conv2D(channels = 32, kernel_size = (3,3), strides = (2,2), padding = (1,1), activation = None, prefix = "c4"))
net.add(mx.gluon.nn.ELU())
net.add(mx.gluon.nn.Flatten())
net.add(mxT.a3cLSTM(mx.gluon.rnn.LSTMCell(256, prefix = 'lstm_')))
net.add(mx.gluon.nn.ELU())
net.add(mxT.a3cOutput(n_policy = 3, prefix = ""))
net.initialize(init = mx.initializer.Xavier(magnitude = 0.1), ctx= mx.cpu())
## set inital parameters from per-trained model
params = mx.gluon.nn.SymbolBlock.imports(symbol_file = "/home/markus/Documents/Nerding/python/a3c/test/pong/array/1200/net-symbol.json",
param_file = "/home/markus/Documents/Nerding/python/a3c/test/pong/array/1200/net-0001.params",
input_names = ['data'])
net.copyParams(fromNet=params)
return(net)
mainThread = mT(netMaker = netMaker ,
envMaker = pongMaker,
configFile = 'a3c/test/pong/pong.cfg')
mainThread.run()
import mxnetTools as mxT
import mxnet as mx
from mainThread import mainThread as mT
def cartpoleMaker():
return cartpole_env()
def netMaker():
net = mxT.a3cHybridSequential()
net.add(mx.gluon.rnn.LSTM(hidden_size=32, input_size=4, prefix="lstm_"))
net.add(mx.gluon.nn.Activation("relu"))
net.add(mxT.a3cOutput(n_policy=2, prefix=""))
net.add()
net.initialize(init=mx.initializer.Xavier())
return (net)
mainThread = mT(netMaker=netMaker,
envMaker=cartpoleMaker,
configFile='a3c/test/cartpole/cartpole.cfg',
verbose=False)
mainThread.run()
mainThread.save("a3c/test/cartpole/smallModel", overwrite=True, savePlots=True)
#after = mainThread.module.get_params()[0]['fullyconnected0_weight'].asnumpy()
mainThread.getPerformancePlots("test", overwrite=True)
def netMaker():
net = mxT.a3cHybridSequential(useInitStates=True)
net.add(
mx.gluon.nn.Conv1D(channels=64,
kernel_size=270,
strides=270,
activation=None,
prefix="c1"))
net.add(mx.gluon.nn.ELU())
net.add(mx.gluon.nn.Flatten())
net.add(mxT.a3cLSTM(mx.gluon.rnn.LSTMCell(128, prefix='lstm1')))
net.add(mx.gluon.nn.ELU())
net.add(mxT.a3cOutput(n_policy=50, prefix=""))
net.initialize(init=mx.initializer.Xavier(magnitude=0.1), ctx=mx.cpu())
## set inital parameters from per-trained model
params = mx.gluon.nn.SymbolBlock.imports(
symbol_file=
"/home/markus/Documents/Nerding/python/a3c/test/dinner/test50noPadnoIntoleranceNoValidRestriction_continued/final/net-symbol.json",
param_file=
"/home/markus/Documents/Nerding/python/a3c/test/dinner/test50noPadnoIntoleranceNoValidRestriction_continued/final/net-0001.params",
input_names=['data'])
net.copyParams(fromNet=params)
return (net)
mainThread = mT(netMaker=netMaker,
envMaker=dinnerMaker,
configFile='a3c/test/dinner/dinner.cfg')
mainThread.run()
def run():
mainThread = mT(netMaker=netMaker,
envMaker=dinnerMaker,
configFile='a3c/test/dinner_simple/dinner_simple.cfg')
mainThread.run()
dinnerAssigned = assigner.assignDinnerCourses()
dinnerTime = datetime(2018, 07, 01, 20)
environment = state(data=dinnerAssigned,
dinnerTime=dinnerTime,
travelMode='simple')
tmp = mx.sym.Variable('data')
tmp = mx.sym.FullyConnected(data=tmp, num_hidden=100)
tmp = mx.sym.Dropout(data=tmp, p=0.2)
tmp = mx.sym.Activation(data=tmp, act_type='relu')
tmp = mx.sym.FullyConnected(data=tmp, num_hidden=100)
tmp = mx.sym.Dropout(data=tmp, p=0.1)
tmp = mx.sym.Activation(data=tmp, act_type='relu')
mainThread = mT(tmp, environment, 'a3c/test/a3c.cfg', verbose=True)
mainThread.run()
tmp = mxT.a3cOutput(tmp, environment.getRewards().size)
mod = a3cModule(tmp, 13500)
data = mxT.state2a3cInput(environment.state)
mod.forward(data)
mod = mx.mod.Module(tmp)
mod.bind(data_shapes=[('data', (1, 10))],
label_shapes=[('valueLabel', (1, 1)),
('advantageLabel', (1, environment.getRewards().size))],
grad_req='add')