def test_ModelFitandFreez1():
typs = ["SL", "SL2"]
for typ in typs:
NNx = NN.clsNN(1,1,2, Type=typs[0], FreezeModel=[1])
NN.ModelFit(NNx.model[0],x1,f20(x1), Epochs=10, ReTry=100, StopAt=[1.0e-2])
oldw = list(NNx.model[0].layers[1].get_weights()[0][0])
NN.ModelFit(NNx.model[1],x1,f21(x1), Epochs=10, ReTry=100, StopAt=[1.0e-2])
assert list(NNx.model[0].layers[1].get_weights()[0][0]) == \
list(NNx.model[1].layers[1].get_weights()[0][0]), "shared layer weights not identical"
assert not list(NNx.model[0].layers[2].get_weights()[0][0]) == \
list(NNx.model[1].layers[2].get_weights()[0][0]), "last /parallel layer weights identical"
assert oldw == list(NNx.model[0].layers[1].get_weights()[0][0]), "frozen weights changed"
def review_ModelFitMultiOut():
typ = "STD"
NNx = NN.clsNN(1,1,2,Type=typ)
Y = np.swapaxes([f20(x1), f21(x1)],0,1)
NN.ModelFit(NNx.model[0],x1,Y, Epochs=10, StopAt=[1.0e-2])
Ynn = np.swapaxes(NN.ModelPredict(NNx.model[0],x1),0,1)
plot(x1,f20(x1),Ynn[0])
plot(x1,f21(x1),Ynn[1])
typs = ["SI", "SL", "SL2"]
for typ in typs:
NNx = NN.clsNN(1,1,2,Type=typ)
NN.ModelFit(NNx.model[0],x1,f20(x1), Epochs=10, StopAt=[1.0e-2])
NN.ModelFit(NNx.model[1],x1,f21(x1), Epochs=10, StopAt=[1.0e-2])
Ynn = [NN.ModelPredict(NNx.model[0],x1), NN.ModelPredict(NNx.model[1],x1)]
plot(x1,f20(x1),Ynn[0])
plot(x1,f21(x1),Ynn[1])
def xTrainQToReward(self, FromSequence, StopAt=[1.0e-03, 0]):
SequenceSample = FromSequence.ReturnSample(self.batch)
Xs, _, As, Rs, _ = SequenceSample.AsList()
X_01 = self._RetXActAs01(Xs, As)
R = Rs
if FromSequence.Nterminal > 0:
SequenceTerminal = FromSequence.ReturnSample(
idxs=FromSequence.ReturnIdx(1))
_, Xt, _, _, Rt = SequenceTerminal.AsList(multi=4)
At = [
j for _ in range(SequenceTerminal.len)
for j in range(len(self.actions))
]
X_01 = self._RetXActAs01(Xs + Xt, As + At)
R = Rs + Rt
# Fit
NN.ModelFit(self.QModel[0],
np.array(X_01),
np.array(R),
Epochs=100,
StopAt=StopAt)
def xTrain(self, FromSequence, BellmanIterations, StopAt=[1.0e-03, 0]):
# Terminal States
if FromSequence.Nterminal > 0:
SequenceTerminal = FromSequence.ReturnSample(
idxs=FromSequence.ReturnIdx(1))
_, Xt, _, _, Rt = SequenceTerminal.AsList(multi=4)
At = [
j for _ in range(SequenceTerminal.len)
for j in range(len(self.actions))
]
for k in range(BellmanIterations):
# Non Terminal States
SequenceSample = FromSequence.ReturnSample(self.batch)
X0s, X1s, As, Rs, _ = SequenceSample.AsList()
# Q under current policy
Qpolicy_X0 = self.Predict(X0s)
Qpolicy_X1 = self.Predict(X1s)
QpAcn_X0 = [Qpolicy_X0[i][As[i]] for i in range(len(Qpolicy_X0))]
QpMax_X1 = [max(Qpolicy_X1[i]) for i in range(len(Qpolicy_X1))]
# taget Q
QtAcn_X0 = [
QpAcn_X0[i] + self.alpha * (Rs[i] + QpMax_X1[i] - QpAcn_X0[i])
for i in range(len(QpAcn_X0))
]
if FromSequence.Nterminal > 0:
X0s = X0s + Xt
As = As + At
QtAcn_X0 = QtAcn_X0 + Rt
# Generate X As Combination of State and Action
XA = self._RetXActAs01(X0s, As) # MOHI
print("Iteration" + str(k))
# Fit
NN.ModelFit(self.QModel[0],
np.array(XA),
np.array(QtAcn_X0),
Epochs=100,
StopAt=StopAt)
return
def review_ModelFit():
for typ in ModelTypes:
NNx = NN.clsNN(1,1,1,Type=typ)
NN.ModelFit(NNx.model[0],x1,f20(x1), Epochs=10, StopAt=[1.0e-3])
Ynn = NN.ModelPredict(NNx.model[0],x1)
plot(x1,f20(x1),Ynn)
