def plss(X, y, cv, n_components=1):
"""
"""
pls = PLSRegression(n_components=n_components)
sse = np.zeros(y.shape[1])
for train, test in cv:
X_train, X_test = X[train], X[test]
y_train, y_test = y[train], y[test]
y0 = y_train.mean(0)
X0 = X_train.mean(0)
pls.fit(X_train - X0, y_train - y0)
sse += np.sum((y_test - y0 - pls.predict(X_test - X0)) ** 2, 0)
return sse
def plss(X, y, cv, n_components=1):
"""
"""
pls = PLSRegression(n_components=n_components)
sse = np.zeros(y.shape[1])
for train, test in cv:
X_train, X_test = X[train], X[test]
y_train, y_test = y[train], y[test]
y0 = y_train.mean(0)
X0 = X_train.mean(0)
pls.fit(X_train - X0, y_train - y0)
sse += np.sum((y_test - y0 - pls.predict(X_test - X0))**2, 0)
return sse
def pls_kfold( sample_set, kfold_group_count, max_components, preprocess ):
print "load...";
l = AttrDict(load('linre_big'+sample_set+'.npz'))
disa = l.disa
expa = l.expa
Y = disa[:,None]
X = l.flum.T
X, Y, expa = shuffle(X, Y, expa, random_state=1)
print "fix...";
X_err, X = find_peaks(X,l.exa)
pls = PLSRegression( scale=False, algorithm='svd' )
pls.fit(X=X,Y=Y)
PC = pls.transform(X.copy())
PC1 = PC[:,0]
good = PC1 > -PC1.std()*2
X, Y, expa = X[good,:], Y[good,:], expa[good]
if preprocess:
X[X<0.5]=0.5
X = X**0.25
#save?
print "cross-validation...";
group_count = kfold_group_count(len(disa))
Ypred4n_components = empty((len(Y),max_components))
for n_components in arange(max_components)+1:
Ypred = empty_like(Y)
loo = KFold( n=len(Y), k=group_count, indices=False )
for fit, test in loo:
pls = PLSRegression(
scale=False,
algorithm='svd',
n_components=n_components
)
pls.fit( X=X[fit].copy(), Y=Y[fit].copy() )
Ypred[test] = pls.predict(X[test].copy())
Ypred4n_components[:,n_components-1] = Ypred[:,0]
print "done for "+str(n_components)+" components"
savez('out23/'+preprocess+'pred.npz',
X=X, Y=Y, expa=expa, Ypred4n_components=Ypred4n_components
)
def dict2mean(X, dict):
plsca = PLSRegression(n_components=np.shape(dict['coefs'])[0])
plsca.x_mean_ = dict['x_mean']
plsca.y_mean_ = dict['y_mean']
plsca.coefs = dict['coefs']
return plsca.predict(X)
if preprocess:
X[X<0.5]=0.5
X = X**0.25
print "fit..."
a4fit = arange(len(X)) >= samples_in_testing_set
a4test = logical_not(a4fit)
X4fit, Y4fit, expa4fit = X[a4fit ,:], Y[a4fit ,:], expa[a4fit ]
X4test, Y4test, expa4test = X[a4test,:], Y[a4test,:], expa[a4test]
pls = PLSRegression(n_components=n_components,algorithm='svd',scale=False)
pls.fit(X=X4fit,Y=Y4fit)
print "predict..."
Y_pred = pls.predict(X4test.copy())
dis4test = Y4test[:,0]
dis_pred = Y_pred[:,0]
dis_max = max(disa)
#dis_pred = where(dis_pred<dis_max+1,where(dis_pred<-1,-1,dis_pred),dis_max+1)
persons = Persons(expa4test)
#print ia4test, ia4fit, logical_not(a4fit & good_std)
#print expa4test.shape, dis4test.shape, dis_pred.shape, Y.shape, Y4test.shape, Y_pred.shape
plt.plot([0,dis_max],[0,dis_max],'g-')
persons.plot(plt,dis4test,dis_pred)
plt.savefig(out_pre+"pred.png")
plt.cla()
n = 1000
q = 3
p = 10
X = np.random.normal(size=n * p).reshape((n, p))
B = np.array([[1, 2] + [0] * (p - 2)] * q).T
# each Yj = 1*X1 + 2*X2 + noize
Y = np.dot(X, B) + np.random.normal(size=n * q).reshape((n, q)) + 5
pls2 = PLSRegression(n_components=3)
pls2.fit(X, Y)
print ("True B (such that: Y = XB + Err)")
print (B)
# compare pls2.coefs with B
print ("Estimated B")
print (np.round(pls2.coefs, 1))
pls2.predict(X)
###############################################################################
# PLS regression, with univariate response, a.k.a. PLS1
n = 1000
p = 10
X = np.random.normal(size=n * p).reshape((n, p))
y = X[:, 0] + 2 * X[:, 1] + np.random.normal(size=n * 1) + 5
pls1 = PLSRegression(n_components=3)
pls1.fit(X, y)
# note that the number of compements exceeds 1 (the dimension of y)
print ("Estimated betas")
print (np.round(pls1.coefs, 1))
###############################################################################
def dict2mean(X, dict):
plsca = PLSRegression(n_components=np.shape(dict['coefs'])[0])
plsca.x_mean_ = dict['x_mean']
plsca.y_mean_ = dict['y_mean']
plsca.coefs = dict['coefs']
return plsca.predict(X)
n = 1000 q = 3 p = 10 X = np.random.normal(size=n * p).reshape((n, p)) B = np.array([[1, 2] + [0] * (p - 2)] * q).T # each Yj = 1*X1 + 2*X2 + noize Y = np.dot(X, B) + np.random.normal(size=n * q).reshape((n, q)) + 5 pls2 = PLSRegression(n_components=3) pls2.fit(X, Y) print "True B (such that: Y = XB + Err)" print B # compare pls2.coefs with B print "Estimated B" print np.round(pls2.coefs, 1) pls2.predict(X) ############################################################################### # PLS regression, with univariate response, a.k.a. PLS1 n = 1000 p = 10 X = np.random.normal(size=n * p).reshape((n, p)) y = X[:, 0] + 2 * X[:, 1] + np.random.normal(size=n * 1) + 5 pls1 = PLSRegression(n_components=3) pls1.fit(X, y) # note that the number of compements exceeds 1 (the dimension of y) print "Estimated betas" print np.round(pls1.coefs, 1) ###############################################################################
dumb_scores = []
for ncomp in max_comps:
print 'Trying %d components' % ncomp
pls = PLSRegression(n_components=ncomp)
dumb = DummyRegressor(strategy='mean')
mae = 0
dumb_mae = 0
for oidx, (train, test) in enumerate(cv):
X_fmri_train = X_fmri[train]
X_fmri_test = X_fmri[test]
X_meg_train = X_meg[train]
X_meg_test = X_meg[test]
pls.fit(X_fmri_train, X_meg_train)
pred = pls.predict(X_fmri_test)
mae += mean_absolute_error(X_meg_test, pred)
dumb.fit(X_fmri_train, X_meg_train)
dumb_pred = dumb.predict(X_fmri_test)
dumb_mae += mean_absolute_error(X_meg_test, dumb_pred)
comp_scores.append(mae / nfolds)
dumb_scores.append(dumb_mae / nfolds)
import matplotlib
matplotlib.use('Agg')
import matplotlib.pyplot as plt
plt.plot(max_comps, comp_scores, max_comps, dumb_scores)
t_str = seed + str(band)
c = DatasetCreator(dtk_params=params, encoder_params=[4096, 3])
D = c.get_d()
n = len(D[0])
print(D[1])
train_X, train_Y = D[0][:n / 2], D[1][:n / 2]
test_X, test_Y = D[0][n / 2:], D[1][n / 2:]
pls2 = PLSRegression()
pls2.fit(train_X, train_Y)
#print(pls2.coefs)
pred = pls2.predict(test_X)
mean_err = np.mean((pred - test_Y)**2)
print(mean_err)
mean_cos = 0
mean_cos_original = 0
for i, j in zip(pred, test_Y):
mean_cos = mean_cos + np.dot(i, j) / np.sqrt(np.dot(i, i) * np.dot(j, j))
print(mean_cos / n)
mae = 0
dumb_mae = 0
meg_mae, fmri_mae = 0, 0
for oidx, (train, test) in enumerate(cv):
X_fmri_train = X_fmri[train]
X_fmri_test = X_fmri[test]
X_meg_train = X_meg[train]
X_meg_test = X_meg[test]
y_train = y[train]
y_test = y[test]
X_train = np.hstack([X_fmri_train,X_meg_train])
X_test = np.hstack([X_fmri_test,X_meg_test])
pls.fit(X_train, y_train)
pred = pls.predict(X_test)
mae += mean_absolute_error(y_test, pred)
dumb.fit(X_train, y_train)
dumb_pred = dumb.predict(X_test)
dumb_mae += mean_absolute_error(y_test,dumb_pred)
if within:
pls.fit(X_fmri_train, y_train)
pred = pls.predict(X_fmri_test)
fmri_mae += mean_absolute_error(y_test, pred)
pls.fit(X_meg_train, y_train)
pred = pls.predict(X_meg_test)
meg_mae += mean_absolute_error(y_test, pred)
dumb_scores = []
for ncomp in max_comps:
print 'Trying %d components'%ncomp
pls = PLSRegression(n_components=ncomp)
dumb = DummyRegressor(strategy='mean')
mae = 0
dumb_mae = 0
for oidx, (train, test) in enumerate(cv):
X_fmri_train = X_fmri[train]
X_fmri_test = X_fmri[test]
X_meg_train = X_meg[train]
X_meg_test = X_meg[test]
pls.fit(X_fmri_train, X_meg_train)
pred = pls.predict(X_fmri_test)
mae += mean_absolute_error(X_meg_test, pred)
dumb.fit(X_fmri_train, X_meg_train)
dumb_pred = dumb.predict(X_fmri_test)
dumb_mae += mean_absolute_error(X_meg_test,dumb_pred)
comp_scores.append(mae/nfolds)
dumb_scores.append(dumb_mae/nfolds)
import matplotlib
matplotlib.use('Agg')
import matplotlib.pyplot as plt
plt.plot(max_comps,comp_scores,max_comps,dumb_scores)
t_str = seed + str(band)
X = X**0.25
group_count = 11
n_components_list = range(9,20)
for n_components in n_components_list:
Ypred = empty_like(Y)
loo = KFold( n=len(Y), k=group_count, indices=False )
for fit, test in loo:
pls = PLSRegression(
scale=False,
algorithm='svd',
n_components=n_components
)
pls.fit( X=X[fit].copy(), Y=Y[fit].copy() )
Ypred[test] = pls.predict(X[test].copy())
print n_components, RMSEP(Y[:,0],Ypred[:,0])
#n, bins, patches = plt.hist(X.flatten(),40,range=(0,2))
#plt.show()
"""
stuff:
print [v for v in X.flatten() if v<-0.4] #only 3 numbers from X <-0.4
X = (1-X)**2/X*2 #Kubelka-Munk function
"""
"""-
5 0.407984567727
6 0.354843551016
7 0.340217332243
8 0.328329231934
def test_predictions():
d = load_linnerud()
X = d.data
Y = d.target
tol = 5e-12
miter = 1000
num_comp = 2
Xorig = X.copy()
Yorig = Y.copy()
# SSY = np.sum(Yorig**2)
# center = True
scale = False
pls1 = PLSRegression(n_components = num_comp, scale = scale,
tol = tol, max_iter = miter, copy = True)
pls1.fit(Xorig, Yorig)
Yhat1 = pls1.predict(Xorig)
SSYdiff1 = np.sum((Yorig-Yhat1)**2)
# print "PLSRegression: R2Yhat = %.4f" % (1 - (SSYdiff1 / SSY))
# Compare PLSR and sklearn.PLSRegression
pls3 = PLSR(num_comp = num_comp, center = True, scale = scale,
tolerance = tol, max_iter = miter)
pls3.fit(X, Y)
Yhat3 = pls3.predict(X)
assert_array_almost_equal(Yhat1, Yhat3, decimal = 5,
err_msg = "PLSR gives wrong prediction")
SSYdiff3 = np.sum((Yorig-Yhat3)**2)
# print "PLSR : R2Yhat = %.4f" % (1 - (SSYdiff3 / SSY))
assert abs(SSYdiff1 - SSYdiff3) < 0.00005
pls2 = PLSCanonical(n_components = num_comp, scale = scale,
tol = tol, max_iter = miter, copy = True)
pls2.fit(Xorig, Yorig)
Yhat2 = pls2.predict(Xorig)
SSYdiff2 = np.sum((Yorig-Yhat2)**2)
# print "PLSCanonical : R2Yhat = %.4f" % (1 - (SSYdiff2 / SSY))
# Compare PLSC and sklearn.PLSCanonical
pls4 = PLSC(num_comp = num_comp, center = True, scale = scale,
tolerance = tol, max_iter = miter)
pls4.fit(X, Y)
Yhat4 = pls4.predict(X)
SSYdiff4 = np.sum((Yorig-Yhat4)**2)
# print "PLSC : R2Yhat = %.4f" % (1 - (SSYdiff4 / SSY))
# Compare O2PLS and sklearn.PLSCanonical
pls5 = O2PLS(num_comp = [num_comp, 1, 0], center = True, scale = scale,
tolerance = tol, max_iter = miter)
pls5.fit(X, Y)
Yhat5 = pls5.predict(X)
SSYdiff5 = np.sum((Yorig-Yhat5)**2)
# print "O2PLS : R2Yhat = %.4f" % (1 - (SSYdiff5 / SSY))
assert abs(SSYdiff2 - SSYdiff4) < 0.00005
assert SSYdiff2 > SSYdiff5
