In [32]:

    

from nilearn.datasets import fetch_haxby
haxby = fetch_haxby(n_subjects=2)

import numpy as np

%matplotlib inline


    

In [19]:

    

func_file = haxby.func[0]
target_file = haxby.session_target[0]

target = np.recfromcsv(target_file, delimiter=" ")  # check recfromtxt and pandas, make them find delimiter


    

In [21]:

    

labels = target['labels']
sessions = target['chunks']


    

In [26]:

    

face_place_mask = (labels == 'face') | (labels == 'house')


    

In [28]:

    

from sklearn.feature_selection import SelectKBest, f_classif


    

In [30]:

    

from nilearn.input_data import NiftiMasker
masker = NiftiMasker(mask_strategy='epi')
masker.fit(func)


    

    
Out[30]:





NiftiMasker(detrend=False, high_pass=None, low_pass=None, mask_args=None,
      mask_img=None, mask_strategy='epi', memory=Memory(cachedir=None),
      memory_level=1, sample_mask=None, sessions=None, smoothing_fwhm=None,
      standardize=False, t_r=None, target_affine=None, target_shape=None,
      verbose=0)

In [34]:

    

mask = masker.mask_img_.get_data()
import matplotlib.pyplot as plt
plt.matshow(mask[20])


    

    
Out[34]:





<matplotlib.image.AxesImage at 0x7fc0f95ed450>






    

In [35]:

    

X = masker.transform(func)


    

In [36]:

    

X.shape


    

    
Out[36]:





(1452, 23707)

In [37]:

    

selector = SelectKBest(f_classif, 1)


    

In [38]:

    

x = selector.fit_transform(X[face_place_mask], labels[face_place_mask])


    

In [39]:

    

plt.plot(x)


    

    
Out[39]:





[<matplotlib.lines.Line2D at 0x7fc0f94f2590>]






    

In [44]:

    

fp_labels = labels[face_place_mask]
plt.plot(x[fp_labels == 'face'], np.zeros_like(x[fp_labels=='face']), 'rx')
plt.plot(x[fp_labels == 'house'], np.ones_like(x[fp_labels=='house']), 'bo')
plt.axis([1500, 1800, -10, 10])


    

    
Out[44]:





[1500, 1800, -10, 10]






    

In [62]:

    

from sklearn.linear_model import LogisticRegression

estimator = LogisticRegression(C=100000)

estimator.fit(x, fp_labels)


    

    
Out[62]:





LogisticRegression(C=100000, class_weight=None, dual=False,
          fit_intercept=True, intercept_scaling=1, max_iter=100,
          multi_class='ovr', n_jobs=1, penalty='l2', random_state=None,
          solver='liblinear', tol=0.0001, verbose=0, warm_start=False)

In [63]:

    

sigmoid = 1. / (1 + np.exp(-(estimator.coef_[0] * x + estimator.intercept_[0]))).ravel()


    

In [72]:

    

plt.plot(x[fp_labels == 'face'], np.zeros_like(x[fp_labels=='face']), 'rx')
plt.plot(x[fp_labels == 'house'], np.ones_like(x[fp_labels=='house']), 'bo')
plt.axis([1500, 1800, -2, 2])
order = x.ravel().argsort()
plt.plot(x.ravel()[order], sigmoid[order], 'g')
boundary = -estimator.intercept_[0] / estimator.coef_[0]
plt.vlines([boundary], -2, 2)


    

    
Out[72]:





<matplotlib.collections.LineCollection at 0x7fc0f7675250>






    

In [80]:

    

from scipy.stats import kde

d_face = kde.gaussian_kde(x[fp_labels == 'face'].ravel())
d_house = kde.gaussian_kde(x[fp_labels == 'house'].ravel())

p = np.linspace(1500, 1800, 1000)
p_face = d_face(p)
p_house = d_house(p)

decision = p_house / (p_face + p_house)

plt.plot(x[fp_labels == 'face'], np.zeros_like(x[fp_labels=='face']), 'rx')
plt.plot(x[fp_labels == 'house'], np.ones_like(x[fp_labels=='house']), 'bo')
plt.axis([1500, 1800, -2, 2])
order = x.ravel().argsort()
plt.plot(x.ravel()[order], sigmoid[order], 'g')
boundary = -estimator.intercept_[0] / estimator.coef_[0]
plt.vlines([boundary], -2, 2)

plt.plot(p, p_face / p_face.max() * .5, 'r')
plt.plot(p, p_house / p_house.max() * .5, 'b')
plt.plot(p, decision, 'm')


    

    
Out[80]:





[<matplotlib.lines.Line2D at 0x7fc0f7744210>]






    

In [78]:

    

plt.plot(p_face)
plt.plot(p_house)

p_face.sum()


    

    
Out[78]:





3.3297524877522573






    

In [84]:

    

selector2 = SelectKBest(f_classif, 2)
x2 = selector2.fit_transform(X[face_place_mask], labels[face_place_mask])
colors = np.array(['r', 'b'])[(labels[face_place_mask] == 'house').astype('int')]
plt.scatter(x2.T[0], x2.T[1], c=colors)


    

    
Out[84]:





<matplotlib.collections.PathCollection at 0x7fc0f928b290>






    

In [82]:

    

plt.scatter?


    

In [ ]: