########################################################### # MAIN trainsize = 4000 washout = 2000 testsize = 8000 # choose ESN: compare STD-ESN with BP-ESN net, trainnoise, testnoise = setup_STD_ESN() # net, trainnoise, testnoise = setup_ESN_BP() net.init() # generate signals slsine = generate_slow_sine(trainsize + testsize) trainin, trainout, testin, testout = get_esn_data(slsine, trainsize, testsize) # ESN training net.setNoise(trainnoise) net.train(trainin, trainout, washout) print "output weights:" print "\tmean: ", net.getWout().mean(), "\tmax: ", abs(net.getWout()).max() # ESN simulation esnout = N.empty(testout.shape) net.setNoise(testnoise) net.simulate(testin, esnout) print "\nNRMSE: ", errorcalc.nrmse(esnout, testout, 50) plot(esnout, testout)
washout = 200 testsize = 2200 # choose ESN: compare performance between square and STD-ESN #net, trainnoise, testnoise = setup_STD_ESN() net, trainnoise, testnoise = setup_SQUARE_ESN() # generate train/test signals size = trainsize+testsize x = random.rand(size)*0.5 y = narma10(x) # create in/outs with bias input trainin, trainout, testin, testout = get_esn_data(x,y,trainsize,testsize) # ESN training net.setNoise(trainnoise) print "training ..." net.train(trainin,trainout,washout) print "output weights:" print "\tmean: ", net.getWout().mean(), "\tmax: ", abs(net.getWout()).max() # ESN simulation esnout = empty(testout.shape) net.setNoise(testnoise) net.simulate(testin,esnout) nrmse = errorcalc.nrmse( esnout, testout, washout ) print "\nNRMSE: ", nrmse print "\nNMSE: ", errorcalc.nmse( esnout, testout, washout ) plot(esnout,testout)
x = random.rand(size) * 0.5
# choose system: for very large stepsize increase trainsize !!
#y = narma10sparse(x,10)
y = sparseSystem2(x, step=50)
# create in/out data
trainin, trainout, testin, testout = get_esn_data(x, y, trainsize, testsize)
# ESN training
net.setNoise(net.trainnoise)
net.train(trainin, trainout, washout)
if (net.ds == 1):
delays = zeros((1, 102))
net.getDelays(delays)
print "trained delays:"
print delays
print "output weights:"
print "\tmean: ", net.getWout().mean(), "\tmax: ", abs(net.getWout()).max()
# ESN simulation
esnout = empty(testout.shape)
net.setNoise(net.testnoise)
net.simulate(testin, esnout)
nrmse = errorcalc.nrmse(esnout, testout, washout)
print "\nNRMSE: ", nrmse
print "\nNMSE: ", errorcalc.nmse(esnout, testout, washout)
# final plotting
plot(esnout, testout)
testsize = 300 # make sine n = N.arange(float(trainsize+testsize)) signal = N.sin(2*N.pi*n* 0.05) # generate train/test data trainout = signal[0:trainsize] trainout.shape = 1, -1 trainin = N.zeros(trainout.shape) testout = signal[trainsize:trainsize+testsize] testout.shape = 1, -1 testin = N.zeros(testout.shape) # ESN training net.train(trainin, trainout, washout) delays = N.zeros((1,2)) net.getDelays(delays) print "output weights:", net.getWout() print "calculated delays:", delays # ESN simulation esnout = N.empty(testout.shape) net.simulate(testin, esnout) print "\nNRMSE: ", errorcalc.nrmse(esnout, testout, 50) # some plotting P.plot(testout.flatten()) P.plot(esnout.flatten(), 'r') P.show()
