def testComputationMethods():
import numpy as np
from utils import testSignal, sliceData, sliceY
irf = (0, 0, 0, 0, 0, 1, 0, 0, 0, 0)
offset = 0
# o->lag-by-5, -5=>no-lag, -9->lead-by-4
X, Y, st, A, B = testSignal(nTrl=1,
nSamp=500,
d=1,
nE=1,
nY=1,
isi=10,
tau=10,
offset=offset,
irf=irf,
noise2signal=0)
#plt.subplot(311);plt.plot(X[0,:,0],label='X');plt.plot(Y[0,:,0,0],label='Y');plt.title("offset={}, irf={}".format(offset,irf));plt.legend()
print("A={}\nB={}".format(A, B))
print("X{}={}".format(X.shape, X[:30, np.argmax(np.abs(A))]))
print("Y{}={}".format(Y.shape, Y[0, :30, 0]))
Y_true = Y[..., 0:1, :]
# other measures....
from updateSummaryStatistics import updateSummaryStatistics, autocov, cov, crossautocov, updateCyy, updateCxx, updateCxy
Cxdtxdt = autocov(X, tau=18) # (tau[0],dx,tau[1],dx)
Cxdtydt = crossautocov(X, Y[:, :, 0:1], tau=18) # (tau[0],dx,tau[1],dy)
Cydxdt = crossautocov(Y[:, :, 0:1], X, tau=[1, 18]) # (tauy,dy,taux,dx)
Cydtydt = np.stack([
crossautocov(Y[:, :, yi:yi + 1], Y[:, :, yi:yi + 1], tau=18)
for yi in range(Y.shape[-1])
], 0) # (nY,tau,d,tau,d)
Cxxdd = cov(X)
# compare
Cxx = updateCxx(None, X, None)
Cxy = updateCxy(None, X, Y[:, :, 0:1, np.newaxis], None,
tau=18) #(1,dy,tau,dx)
Cyy = updateCyy(None, Y[:, :, :, np.newaxis], None,
tau=18) #(nY,yd,tau,yd,tau)
print('Cxx-Cxxdd={}'.format(np.max(np.abs(Cxxdd - Cxx).ravel())))
print('Cxy-Cydxdt={}'.format(np.max(np.abs(Cxy - Cydxdt).ravel())))
print('Cyy-Cydtydt={}'.format(np.max(np.abs(Cyy.ravel() -
Cydtydt.ravel()))))
# test centering
cX = X - np.mean(X, (0, 1))
Cxx = updateCxx(None, X, None, center=True)
cCxx = updateCxx(None, cX, None, center=False)
print("Cxx(center) - cCxx={}".format(np.max(np.abs(Cxx - cCxx).ravel())))
Cxy = updateCxy(None, X, Y, None, tau=3, center=True)
cCxy = updateCxy(None, cX, Y, None, tau=3, center=False)
print("Cxy(center) - cCxy={}".format(np.max(np.abs(Cxy - cCxy).ravel())))
def testdataset(fn, **args):
''' a conforming toy dataset loader'''
fs=10
X,Y,st,A,B=testSignal(**args)
# make coords array for the meta-info about the dimensions of X
coords = [None]*X.ndim
coords[0] = {'name':'trial'}
coords[1] = {'name':'time','unit':'ms', \
'coords':[i*1000/fs for i in range(X.shape[1])], \
'fs':fs}
coords[2] = {'name':'channel','coords':None}
return (X, Y, coords)
def testcases():
# Fe = [nE x nEpoch x nTrl x nM ] similarity score for each event type for each stimulus
# Ye = [nE x nY x nEpoch x nTrl ] Indicator for which events occured for which outputs
nE = 2
nEpoch = 100
nTrl = 30
nY = 20
nM = 20
Fe = np.random.standard_normal((nM, nTrl, nEpoch, nE))
Ye = np.random.standard_normal((nTrl, nEpoch, nY, nE))
print("Fe={}".format(Fe.shape))
print("Ye={}".format(Ye.shape))
Fy = scoreOutput(Fe, Ye) # (nM,nTrl,nEp,nY)
print("Fy={}".format(Fy.shape))
import matplotlib.pyplot as plt
sFy = np.cumsum(Fy, axis=-2)
plt.clf()
plt.plot(sFy[0, 0, :, :])
plt.xlabel('epoch')
plt.ylabel('output')
plt.show()
# more complex example with actual signal/noise
from utils import testSignal
from scoreOutput import scoreOutput, plot_outputscore, convWX, convYR, convXYR
from scoreStimulus import scoreStimulus
from decodingSupervised import decodingSupervised
from normalizeOutputScores import normalizeOutputScores
import numpy as np
irf = (1, 1, -1, -1, 0, 0, 0, 0, 0, 0)
X, Y, st, W, R = testSignal(nTrl=1,
nSamp=1000,
d=1,
nE=1,
nY=10,
isi=2,
irf=irf,
noise2signal=0)
plot_outputscore(X[0, ...], Y[0, :, 0:3, :], W, R)
# add a correlated output
Y[:, :, 1, :] = Y[:, :, 0, :] * .5
plot_outputscore(X[0, ...], Y[0, :, 0:3, :], W, R)
def testcase():
from utils import testNoSignal, testSignal, sliceData, sliceY
#from multipleCCA import *
if False:
X, Y, st = testNoSignal()
else:
X, Y, st, A, B = testSignal(tau=10, noise2signal=10)
# TODO[]: summary stats directly without first slicing
Y_true = Y[:, :, 0:1, :] if Y.ndim > 3 else Y[:, 0:1, :] # N.B. hack with [0] to stop removal of dim...
from updateSummaryStatistics import updateSummaryStatistics, plot_summary_statistics
Cxx, Cxy, Cyy = updateSummaryStatistics(X, Y_true, tau=10)
#plot_summary_statistics(Cxx,Cxy,Cyy)
# single supervised training
from multipleCCA import multipleCCA, plot_multicca_solution
J, w, r = multipleCCA(Cxx, Cxy, Cyy, rcond=-.3, symetric=True) # 50% rank reduction
plot_multicca_solution(w, r)
# apply to the data
from scoreStimulus import scoreStimulus
Fe = scoreStimulus(X, w, r)
import matplotlib.pyplot as plt;
plt.clf();
plt.plot(np.einsum("Tsd,d->s",X,w.ravel()),'b',label="Xw");
plt.plot(Y_true.ravel(),'g',label="Y");
plt.plot(Fe.ravel()/np.max(Fe.ravel())/2,'r',label="Fe");
plt.legend()
from scoreOutput import scoreOutput
print("Fe={}".format(Fe.shape))
print("Y={}".format(Y.shape))
Fy = scoreOutput(Fe, Y, r) # (nM,nTrl,nEp,nY)
print("Fy={}".format(Fy.shape))
import matplotlib.pyplot as plt
sFy=np.cumsum(Fy, -2)
plt.clf();plt.plot(sFy[0, 0, :, :]);plt.xlabel('epoch');plt.ylabel('output');plt.show()
from decodingSupervised import decodingSupervised
decodingSupervised(Fy)
from decodingCurveSupervised import decodingCurveSupervised
decodingCurveSupervised(Fy)
def testCases():
import numpy as np
from utils import testSignal, sliceData, sliceY
from updateSummaryStatistics import updateSummaryStatistics, plot_summary_statistics
import matplotlib.pyplot as plt
irf = (0, 0, 0, 0, 0, 0, 0, 0, 0, 1)
offset = 0
# X->lag-by-10
X, Y, st, A, B = testSignal(nTrl=1,
nSamp=500,
d=1,
nE=1,
nY=1,
isi=10,
tau=10,
offset=offset,
irf=irf,
noise2signal=0)
print("A={}\nB={}".format(A, B))
print("X{}={}".format(X.shape, X[:30, np.argmax(np.abs(A))]))
print("Y{}={}".format(Y.shape, Y[0, :30, 0]))
tau = 10
uss_offset = 0
Cxx, Cxy, Cyy = updateSummaryStatistics(X,
Y,
None,
tau=tau,
offset=uss_offset)
print("Cxx={} Cxy={} Cyy={}".format(Cxx.shape, Cxy.shape, Cyy.shape))
plt.figure(1)
plot_summary_statistics(Cxx, Cxy, Cyy)
plt.show()
# leading X
irf = (1, 0, 0, 0, 0, 0, 0, 0, 0, 0)
offset = -9
# X leads by 9
X, Y, st, A, B = testSignal(nTrl=1,
nSamp=5000,
d=1,
nE=1,
nY=1,
isi=10,
tau=10,
offset=offset,
irf=irf,
noise2signal=0)
plt.figure(0)
plt.clf()
plt.plot(X[0, :, 0], label='X')
plt.plot(Y[0, :, 0, 0], label='Y')
plt.title("offset={}, irf={}".format(offset, irf))
plt.legend()
# no-shift in analysis window
tau = 10
uss_offset = 0
Cxx, Cxy, Cyy = updateSummaryStatistics(X,
Y,
None,
tau=tau,
offset=uss_offset)
print("Cxx={} Cxy={} Cyy={}".format(Cxx.shape, Cxy.shape, Cyy.shape))
plt.figure(1)
plt.clf()
plot_summary_statistics(Cxx, Cxy, Cyy)
plt.show()
# shifted analysis window
tau = 10
uss_offset = -9
Cxx, Cxy, Cyy = updateSummaryStatistics(X,
Y,
None,
tau=tau,
offset=uss_offset)
print("Cxx={} Cxy={} Cyy={}".format(Cxx.shape, Cxy.shape, Cyy.shape))
plt.figure(2)
plt.clf()
plot_summary_statistics(Cxx, Cxy, Cyy)
plt.show()
plt.figure(3)
plt.clf()
plot_erp(Cxy)
plt.show()
def testSlicedvsContinuous():
import numpy as np
from utils import testSignal, sliceData, sliceY
irf = (0, 0, 0, 0, 0, 1, 0, 0, 0, 0)
offset = 0
# o->lag-by-5, -5=>no-lag, -9->lead-by-4
X, Y, st, A, B = testSignal(nTrl=1,
nSamp=500,
d=1,
nE=1,
nY=1,
isi=10,
tau=10,
offset=offset,
irf=irf,
noise2signal=0)
#plt.subplot(311);plt.plot(X[0,:,0],label='X');plt.plot(Y[0,:,0,0],label='Y');plt.title("offset={}, irf={}".format(offset,irf));plt.legend()
print("A={}\nB={}".format(A, B))
print("X{}={}".format(X.shape, X[:30, np.argmax(np.abs(A))]))
print("Y{}={}".format(Y.shape, Y[0, :30, 0]))
Y_true = Y[..., 0:1, :]
# slice then compute
Xe = sliceData(X, stimTimes, tau=tau)
Ye = sliceY(Y, stimTimes)
Ye_true = Ye[..., 0:1, :]
print('Xe{}={}\nYe{}={}\nst{}={}'.format(Xe.shape, Xe[:5, :, 0], Ye.shape,
Ye[:5, 0, 0], stimTimes.shape,
stimTimes[:5]))
Cxx, Cxy, Cyy = updateSummaryStatistics(Xe,
Ye_true,
stimTimes,
tau=Xe.shape[-2])
print("Cxx={} Cxy={} Cyy={}".format(Cxx.shape, Cxy.shape, Cyy.shape))
plot_summary_statistics(Cxx, Cxy, Cyy)
plot_Cxy(Cxy)
dCxx = np.max((np.abs(Cxx / np.trace(Cxx) - Cxx1 / np.trace(Cxx1)) /
(np.abs(Cxx / np.trace(Cxx)) + 1e-5)).ravel())
dCxy = np.max((np.abs(Cxy - Cxy1) / (np.abs(Cxy) + 1e-5)).ravel())
dCyy = np.max((np.abs(Cyy - Cyy1) / (np.abs(Cyy) + 1e-5)).ravel())
print("dCxx= {} dCxy={} dCyy={}".format(dCxx, dCxy, dCyy))
# timing tests
from timeit import timeit
dur = timeit(
lambda: updateSummaryStatistics(Xe, Ye_true, stimTimes, tau=10),
number=1000,
globals=globals())
print("Sliced={}s".format(dur / 1000))
def slicenuss(X, Y, stimTimes, tau):
Xe = sliceData(X, stimTimes, tau)
Ye = sliceY(Y, stimTimes)
Cxx, Cxy, Cyy = updateSummaryStatistics(Xe, Ye, stimTimes, tau=tau)
dur = timeit(lambda: slicenuss(X, Y_true, stimTimes, tau=10),
number=1000,
globals=globals())
print("Slice+USS={}s".format(dur / 1000))
dur = timeit(lambda: updateSummaryStatistics(X, Y_true, None, tau=10),
number=1000,
globals=globals())
print("Raw={}s".format(dur / 1000))
def testLeadLag():
import numpy as np
from utils import testSignal
from model_fitting import MultiCCA
from decodingCurveSupervised import decodingCurveSupervised
import matplotlib.pyplot as plt
irf = (0, 0, 0, 0, 0, 0, 0, 0, 0, 1)
offset = 0
# X->lag-by-10
X, Y, st, A, B = testSignal(nTrl=1,
nSamp=500,
d=10,
nE=1,
nY=30,
isi=10,
tau=10,
offset=offset,
irf=irf,
noise2signal=0)
# reference case lagged response test
evtlabs = None
tau = 10
cca_offset = 0
cca = MultiCCA(evtlabs=evtlabs, tau=tau, offset=cca_offset)
scores = cca.cv_fit(X, Y)
Fy = scores['estimator']
(_) = decodingCurveSupervised(Fy)
# leading X
irf = (1, 0, 0, 0, 0, 0, 0, 0, 0, 0)
offset = -9
# X leads by 9
X, Y, st, A, B = testSignal(nTrl=1,
nSamp=5000,
d=10,
nE=1,
nY=30,
isi=10,
tau=10,
offset=offset,
irf=irf,
noise2signal=0)
plt.figure(0)
plt.clf()
plt.plot(X[0, :, 0], label='X')
plt.plot(Y[0, :, 0, 0], label='Y')
plt.title("offset={}, irf={}".format(offset, irf))
plt.legend()
# no-shift in analysis window
evtlabs = None
tau = 10
cca_offset = 0
cca = MultiCCA(evtlabs=evtlabs, tau=tau, offset=cca_offset)
scores = cca.cv_fit(X, Y)
Fy = scores['estimator']
(_) = decodingCurveSupervised(Fy)
# shifted analysis window
evtlabs = None
tau = 20
cca_offset = -9
cca = MultiCCA(evtlabs=evtlabs, tau=tau, offset=cca_offset)
scores = cca.cv_fit(X, Y)
Fy = scores['estimator']
(_) = decodingCurveSupervised(Fy)
plt.figure(1)
plt.clf()
plot_model_weights(cca)