def kldiv(p, q, distp = None, distq = None, scale_factor = 1):
"""
Computes the Kullback-Leibler divergence between two distributions.
Parameters
p : Matrix
The first probability distribution
q : Matrix
The second probability distribution
distp : fixmat
If p is None, distp is used to compute a FDM which
is then taken as 1st probability distribution.
distq : fixmat
If q is None, distq is used to compute a FDM which is
then taken as 2dn probability distribution.
scale_factor : double
Determines the size of FDM computed from distq or distp.
"""
assert q != None or distq != None, "Either q or distq have to be given"
assert p != None or distp != None, "Either p or distp have to be given"
try:
if p == None:
p = compute_fdm(distp, scale_factor = scale_factor)
if q == None:
q = compute_fdm(distq, scale_factor = scale_factor)
except NoFixations:
return np.NaN
q += np.finfo(q.dtype).eps
p += np.finfo(p.dtype).eps
kl = np.sum( p * (np.log2(p / q)))
return kl
def skip_emd(self):
fm1 = fixmat.TestFactory(params={'image_size': [93, 128]})
fm2 = fixmat.TestFactory(points=list(
zip(list(range(10, 50)), list(range(10, 50)))),
params={'image_size': [93, 128]})
self.assertEqual(measures.emd_model(fixmat.compute_fdm(fm1), fm1), 0)
self.assertTrue(not (
measures.emd_model(fixmat.compute_fdm(fm1), fm2) == 0))
def test_corners(self):
""" ``test_corners(self)``
Tests whether handling of fixations in the corners is correct.
It manually generates an fdm with four fixations in the corners of a
922x1272 array and compares it to the map generated by compute_fdm.
The difference between the maps must not be larger than the machine
precision for floats.
"""
yvec = [922, 922, 0, 0]
xvec = [1272, 0, 1272, 0]
self.fm = fixmat.TestFixmatFactory(points=(xvec, yvec))
fdm = fixmat.compute_fdm(self.fm)
# manually calculate the fdm
fdm_man = np.zeros((922, 1272))
fdm_man[0][0] = 1
fdm_man[921][1271] = 1
fdm_man[0][1271] = 1
fdm_man[921][0] = 1
# use default settings for fwhm, pixels_per_degree and scale_factor
kernel_sigma = 2 * 36 * 1
kernel_sigma = kernel_sigma / (2 * (2 * np.log(2))**.5)
fdm_man = gaussian_filter(np.array(fdm_man),
kernel_sigma,
order=0,
mode='constant')
fdm_man = fdm_man / fdm_man.sum()
diff = fdm - fdm_man
self.assertFalse((diff > np.finfo('float').eps).any())
def test_kldiv(self):
arr = scipy.random.random((21, 13))
fm = fixmat.TestFixmatFactory(categories=[1, 2, 3],
filenumbers=[1, 2, 3, 4, 5, 6],
subjectindices=[1, 2, 3, 4, 5, 6],
params={
'pixels_per_degree': 10,
'image_size': [100, 500]
})
kl = measures.kldiv(arr, arr)
self.assertEqual(
kl, 0, "KL Divergence between same distribution should be 0")
kl = measures.kldiv(None, None, distp=fm, distq=fm, scale_factor=0.25)
self.assertEqual(
kl, 0, "KL Divergence between same distribution should be 0")
fdm = fixmat.compute_fdm(fm)
kl = measures.kldiv_model(fdm, fm)
self.assertTrue(
kl < 10**-13,
"KL Divergence between same distribution should be almost 0")
fm.x = np.array([])
fm.y = np.array([])
kl = measures.kldiv(None, None, distp=fm, distq=fm, scale_factor=0.25)
self.assertTrue(np.isnan(kl))
def test_corners(self):
""" ``test_corners(self)``
Tests whether handling of fixations in the corners is correct.
It manually generates an fdm with four fixations in the corners of a
922x1272 array and compares it to the map generated by compute_fdm.
The difference between the maps must not be larger than the machine
precision for floats.
"""
yvec = [922, 922,0 ,0]
xvec = [1272,0 ,1272,0]
self.fm = fixmat.TestFixmatFactory(points = (xvec, yvec))
fdm = fixmat.compute_fdm(self.fm)
# manually calculate the fdm
fdm_man = np.zeros((922,1272))
fdm_man[0][0] = 1
fdm_man[921][1271] = 1
fdm_man[0][1271] = 1
fdm_man[921][0] = 1
# use default settings for fwhm, pixels_per_degree and scale_factor
kernel_sigma = 2 * 36 * 1
kernel_sigma = kernel_sigma / (2 * (2 * np.log(2)) ** .5)
fdm_man = gaussian_filter(np.array(fdm_man), kernel_sigma, order=0,
mode='constant')
fdm_man = fdm_man / fdm_man.sum()
diff = fdm - fdm_man
self.assertFalse((diff > np.finfo('float').eps).any())
def test_nss(self):
fm = fixmat.TestFixmatFactory(points=zip([0,500,1000],[1,10,10]),
params = {'image_size':[100,10]})
fm.SUBJECTINDEX = np.array([1,1,1])
fm.filenumber = np.array([1,1,1])
fm.category = np.array([1,1,1])
fm.x = np.array([0,50,1000])
fm.y = np.array([1,10,10])
fm.fix = np.array([1,2,3])
fm._num_fix = 3
fdm = fixmat.compute_fdm(fm[(fm.x<10) & (fm.y<10)])
self.assertRaises(IndexError, lambda: measures.nss(fdm, (fm.y, fm.x)))
def upper_bound(fm, nr_subs=None, scale_factor=1):
"""
compute the inter-subject consistency upper bound for a fixmat.
Input:
fm : a fixmat instance
nr_subs : the number of subjects used for the prediction. Defaults
to the total number of subjects in the fixmat minus 1
scale_factor : the scale factor of the FDMs. Default is 1.
Returns:
A list of scores; the list contains one dictionary for each measure.
Each dictionary contains one key for each category and corresponding
values is an array with scores for each subject.
"""
nr_subs_total = len(np.unique(fm.SUBJECTINDEX))
if not nr_subs:
nr_subs = nr_subs_total - 1
assert (nr_subs < nr_subs_total)
# initialize output structure; every measure gets one dict with
# category numbers as keys and numpy-arrays as values
intersub_scores = []
for measure in range(len(measures.scores)):
res_dict = {}
result_vectors = [
np.empty(nr_subs_total) + np.nan for _ in np.unique(fm.category)
]
res_dict.update(zip(np.unique(fm.category), result_vectors))
intersub_scores.append(res_dict)
#compute inter-subject scores for every stimulus, with leave-one-out
#over subjects
for fm_cat in fm.by_field('category'):
cat = fm_cat.category[0]
for (sub_counter, sub) in enumerate(np.unique(fm_cat.SUBJECTINDEX)):
image_scores = []
for fm_single in fm_cat.by_field('filenumber'):
predicting_subs = (np.setdiff1d(
np.unique(fm_single.SUBJECTINDEX), [sub]))
np.random.shuffle(predicting_subs)
predicting_subs = predicting_subs[0:nr_subs]
predicting_fm = fm_single[(ismember(fm_single.SUBJECTINDEX,
predicting_subs))]
predicted_fm = fm_single[fm_single.SUBJECTINDEX == sub]
try:
predicting_fdm = compute_fdm(predicting_fm,
scale_factor=scale_factor)
except RuntimeError:
predicting_fdm = None
image_scores.append(
measures.prediction_scores(predicting_fdm, predicted_fm))
for (measure, score) in enumerate(nanmean(image_scores, 0)):
intersub_scores[measure][cat][sub_counter] = score
return intersub_scores
def correlation_model(prediction, fm):
"""
wraps numpy.corrcoef functionality for model evaluation
input:
prediction: 2D Matrix
the model salience map
fm: fixmat
Used to compute a FDM to which the prediction is compared.
"""
(_, r_x) = calc_resize_factor(prediction, fm.image_size)
fdm = compute_fdm(fm, scale_factor = r_x)
return np.corrcoef(fdm.flatten(), prediction.flatten())[0,1]
def test_scaling(self):
"""``test_scaling(self)``
Tests that the size of the resulting fdm corresponds to
image_size*scale_factor. Tests with different image sizes.
"""
size = self.fm.image_size
sf = [0.6, 1.6, 1.0 / 2.0]
# if the scaling factor is not a float, this does not work but seems to
# be a problem of the inbuilt function.
for i in range(len(sf)):
fdm = fixmat.compute_fdm(self.fm, scale_factor=sf[i])
self.assertEqual((int(size[0] * sf[i]), int(size[1] * sf[i])),
np.shape(fdm))
def test_scaling(self):
"""``test_scaling(self)``
Tests that the size of the resulting fdm corresponds to
image_size*scale_factor. Tests with different image sizes.
"""
size = self.fm.image_size
sf = [0.6, 1.6, 1.0/2.0]
# if the scaling factor is not a float, this does not work but seems to
# be a problem of the inbuilt function.
for i in range(len(sf)):
fdm = fixmat.compute_fdm(self.fm,scale_factor = sf[i])
self.assertEqual((int(size[0]*sf[i]),int(size[1]*sf[i])),
np.shape(fdm))
def upper_bound(fm, nr_subs = None, scale_factor = 1):
"""
compute the inter-subject consistency upper bound for a fixmat.
Input:
fm : a fixmat instance
nr_subs : the number of subjects used for the prediction. Defaults
to the total number of subjects in the fixmat minus 1
scale_factor : the scale factor of the FDMs. Default is 1.
Returns:
A list of scores; the list contains one dictionary for each measure.
Each dictionary contains one key for each category and corresponding
values is an array with scores for each subject.
"""
nr_subs_total = len(np.unique(fm.SUBJECTINDEX))
if not nr_subs:
nr_subs = nr_subs_total - 1
assert (nr_subs < nr_subs_total)
# initialize output structure; every measure gets one dict with
# category numbers as keys and numpy-arrays as values
intersub_scores = []
for measure in range(len(measures.scores)):
res_dict = {}
result_vectors = [np.empty(nr_subs_total) + np.nan
for _ in np.unique(fm.category)]
res_dict.update(zip(np.unique(fm.category), result_vectors))
intersub_scores.append(res_dict)
#compute inter-subject scores for every stimulus, with leave-one-out
#over subjects
for fm_cat in fm.by_field('category'):
cat = fm_cat.category[0]
for (sub_counter, sub) in enumerate(np.unique(fm_cat.SUBJECTINDEX)):
image_scores = []
for fm_single in fm_cat.by_field('filenumber'):
predicting_subs = (np.setdiff1d(np.unique(
fm_single.SUBJECTINDEX),[sub]))
np.random.shuffle(predicting_subs)
predicting_subs = predicting_subs[0:nr_subs]
predicting_fm = fm_single[
(ismember(fm_single.SUBJECTINDEX, predicting_subs))]
predicted_fm = fm_single[fm_single.SUBJECTINDEX == sub]
try:
predicting_fdm = compute_fdm(predicting_fm,
scale_factor = scale_factor)
except NoFixations:
predicting_fdm = None
image_scores.append(measures.prediction_scores(
predicting_fdm, predicted_fm))
for (measure, score) in enumerate(nanmean(image_scores, 0)):
intersub_scores[measure][cat][sub_counter] = score
return intersub_scores
def emd_model(prediction, fm):
"""
wraps emd functionality for model evaluation
requires:
OpenCV python bindings
input:
prediction: the model salience map
fm : fixmat filtered for the image corresponding to the prediction
"""
(_, r_x) = calc_resize_factor(prediction, fm.image_size)
gt = fixmat.compute_fdm(fm, scale_factor = r_x)
return emd(prediction, gt)
def test_correlation(self):
fm = fixmat.TestFixmatFactory(categories = [1,2,3],
filenumbers = [1,2,3,4,5,6],
subjectindices = [1, 2, 3, 4, 5, 6],
params = {'pixels_per_degree':1, 'image_size':[100,500]})
# Arr has zero variance, should return nan
arr = np.ones(fm.image_size)
corr = measures.correlation_model(arr, fm)
self.assertTrue(np.isnan(corr))
# With itself should give 1
fdm = fixmat.compute_fdm(fm)
corr = measures.correlation_model(fdm,fm)
self.assertEquals(corr,1)
# Anti-correlation should give -1
corr = measures.correlation_model(-1*fdm,fm)
self.assertEquals(corr,-1)
def test_nss_values(self):
fm = fixmat.TestFixmatFactory(categories = [1,2,3],
filenumbers = [1,2,3,4,5,6],
subjectindices = [1, 2, 3, 4, 5, 6],
params = {'pixels_per_degree':0.1, 'image_size':[200,500]})
# Arr has zero variance, should return nan
arr = np.ones(fm.image_size)
nss = measures.nss_model(arr, fm)
self.assertTrue(np.isnan(nss))
# With itself should yield a high value
fdm = fixmat.compute_fdm(fm)
nss = measures.nss_model(fdm, fm)
self.assertTrue(nss>15)
# Fixations at these locations should give nss < 0
nss = measures.nss(fdm, [[100, 101, 102, 103, 104, 105],[0, 0, 0, 0, 0, 0]])
self.assertTrue(nss < 0)
def test_emd(self):
try:
import opencv
except ImportError:
print "Skipping EMD test - no opencv available"
return
opencv # pyflakes
fm = fixmat.TestFixmatFactory(categories = [1,2,3],
filenumbers = [1,2,3,4,5,6],
subjectindices = [1, 2, 3, 4, 5, 6],
params = {'pixels_per_degree':1, 'image_size':[20,50]})
arr = np.ones(fm.image_size)
fdm = fixmat.compute_fdm(fm)
e = measures.emd_model(arr, fm)
self.assertTrue(e > 0)
e = measures.emd(fdm, fdm)
self.assertEquals(e, 0)
def test_prediction_scores(self):
measures.set_scores([measures.roc_model,
measures.kldiv_model,
measures.nss_model])
fm = fixmat.TestFixmatFactory(categories = [1,2,3],
filenumbers = [1,2,3,4,5,6],
subjectindices = [1, 2, 3, 4, 5, 6],
params = {'pixels_per_degree':10, 'image_size':[100,500]})
fdm = fixmat.compute_fdm(fm)
measures.set_scores([measures.roc_model,
measures.kldiv_model,
measures.nss_model])
scores = measures.prediction_scores(fdm, fm)
self.assertEquals(len(scores), 3)
measures.set_scores([measures.roc_model])
scores = measures.prediction_scores(fdm, fm)
self.assertEquals(len(scores), 1)
measures.set_scores([measures.roc_model,
measures.kldiv_model,
measures.nss_model])
scores = measures.prediction_scores(fdm, fm)
self.assertEquals(len(scores), 3)
def test_kldiv(self):
arr = scipy.random.random((21,13))
fm = fixmat.TestFixmatFactory(categories = [1,2,3],
filenumbers = [1,2,3,4,5,6],
subjectindices = [1, 2, 3, 4, 5, 6],
params = {'pixels_per_degree':10, 'image_size':[100,500]})
kl = measures.kldiv(arr, arr)
self.assertEqual(kl, 0,
"KL Divergence between same distribution should be 0")
kl = measures.kldiv(None, None, distp = fm, distq = fm, scale_factor = 0.25)
self.assertEqual(kl, 0,
"KL Divergence between same distribution should be 0")
fdm = fixmat.compute_fdm(fm)
kl = measures.kldiv_model(fdm, fm)
self.assertTrue(kl < 10**-13,
"KL Divergence between same distribution should be almost 0")
fm.x = np.array([])
fm.y = np.array([])
kl = measures.kldiv(None, None, distp = fm, distq = fm, scale_factor = 0.25)
self.assertTrue(np.isnan(kl))
def lower_bound(fm, nr_subs=None, nr_imgs=None, scale_factor=1):
"""
Compute the spatial bias lower bound for a fixmat.
Input:
fm : a fixmat instance
nr_subs : the number of subjects used for the prediction. Defaults
to the total number of subjects in the fixmat minus 1
nr_imgs : the number of images used for prediction. If given, the
same number will be used for every category. If not given,
leave-one-out will be used in all categories.
scale_factor : the scale factor of the FDMs. Default is 1.
Returns:
A list of spatial bias scores; the list contains one dictionary for each
measure. Each dictionary contains one key for each category and
corresponding values is an array with scores for each subject.
"""
nr_subs_total = len(np.unique(fm.SUBJECTINDEX))
if nr_subs is None:
nr_subs = nr_subs_total - 1
assert (nr_subs < nr_subs_total)
# initialize output structure; every measure gets one dict with
# category numbers as keys and numpy-arrays as values
sb_scores = []
for measure in range(len(measures.scores)):
res_dict = {}
result_vectors = [
np.empty(nr_subs_total) + np.nan for _ in np.unique(fm.category)
]
res_dict.update(zip(np.unique(fm.category), result_vectors))
sb_scores.append(res_dict)
# compute mean spatial bias predictive power for all subjects in all
# categories
for fm_cat in fm.by_field('category'):
cat = fm_cat.category[0]
nr_imgs_cat = len(np.unique(fm_cat.filenumber))
if not nr_imgs:
nr_imgs_current = nr_imgs_cat - 1
else:
nr_imgs_current = nr_imgs
assert (nr_imgs_current < nr_imgs_cat)
for (sub_counter, sub) in enumerate(np.unique(fm.SUBJECTINDEX)):
image_scores = []
for fm_single in fm_cat.by_field('filenumber'):
# Iterating by field filenumber makes filenumbers
# in fm_single unique: Just take the first one to get the
# filenumber for this fixmat
fn = fm_single.filenumber[0]
predicting_subs = (np.setdiff1d(np.unique(fm_cat.SUBJECTINDEX),
[sub]))
np.random.shuffle(predicting_subs)
predicting_subs = predicting_subs[0:nr_subs]
predicting_fns = (np.setdiff1d(np.unique(fm_cat.filenumber),
[fn]))
np.random.shuffle(predicting_fns)
predicting_fns = predicting_fns[0:nr_imgs_current]
predicting_fm = fm_cat[
(ismember(fm_cat.SUBJECTINDEX, predicting_subs))
& (ismember(fm_cat.filenumber, predicting_fns))]
predicted_fm = fm_single[fm_single.SUBJECTINDEX == sub]
try:
predicting_fdm = compute_fdm(predicting_fm,
scale_factor=scale_factor)
except RuntimeError:
predicting_fdm = None
image_scores.append(
measures.prediction_scores(predicting_fdm, predicted_fm))
for (measure, score) in enumerate(nanmean(image_scores, 0)):
sb_scores[measure][cat][sub_counter] = score
return sb_scores
def intersubject_scores(fm,
category,
predicting_filenumbers,
predicting_subjects,
predicted_filenumbers,
predicted_subjects,
controls=True,
scale_factor=1):
"""
Calculates how well the fixations from a set of subjects on a set of
images can be predicted with the fixations from another set of subjects
on another set of images.
The prediction is carried out by computing a fixation density map from
fixations of predicting_subjects subjects on predicting_images images.
Prediction accuracy is assessed by measures.prediction_scores.
Parameters
fm : fixmat instance
category : int
Category from which the fixations are taken.
predicting_filenumbers : list
List of filenumbers used for prediction, i.e. images where fixations
for the prediction are taken from.
predicting_subjects : list
List of subjects whose fixations on images in predicting_filenumbers
are used for the prediction.
predicted_filenumnbers : list
List of images from which the to be predicted fixations are taken.
predicted_subjects : list
List of subjects used for evaluation, i.e subjects whose fixations
on images in predicted_filenumbers are taken for evaluation.
controls : bool, optional
If True (default), n_predict subjects are chosen from the fixmat.
If False, 1000 fixations are randomly generated and used for
testing.
scale_factor : int, optional
specifies the scaling of the fdm. Default is 1.
Returns
auc : area under the roc curve for sets of actuals and controls
true_pos_rate : ndarray
Rate of true positives for every given threshold value.
All values appearing in actuals are taken as thresholds. Uses lower
sum interpolation.
false_pos_rate : ndarray
See true_pos_rate but for false positives.
"""
predicting_fm = fm[(ismember(fm.SUBJECTINDEX, predicting_subjects))
& (ismember(fm.filenumber, predicting_filenumbers)) &
(fm.category == category)]
predicted_fm = fm[(ismember(fm.SUBJECTINDEX, predicted_subjects))
& (ismember(fm.filenumber, predicted_filenumbers)) &
(fm.category == category)]
try:
predicting_fdm = compute_fdm(predicting_fm, scale_factor=scale_factor)
except RuntimeError:
predicting_fdm = None
if controls == True:
fm_controls = fm[(ismember(fm.SUBJECTINDEX, predicted_subjects)) & (
(ismember(fm.filenumber, predicted_filenumbers)) != True) &
(fm.category == category)]
return measures.prediction_scores(predicting_fdm,
predicted_fm,
controls=(fm_controls.y,
fm_controls.x))
return measures.prediction_scores(predicting_fdm,
predicted_fm,
controls=None)
def lower_bound(fm, nr_subs = None, nr_imgs = None, scale_factor = 1):
"""
Compute the spatial bias lower bound for a fixmat.
Input:
fm : a fixmat instance
nr_subs : the number of subjects used for the prediction. Defaults
to the total number of subjects in the fixmat minus 1
nr_imgs : the number of images used for prediction. If given, the
same number will be used for every category. If not given,
leave-one-out will be used in all categories.
scale_factor : the scale factor of the FDMs. Default is 1.
Returns:
A list of spatial bias scores; the list contains one dictionary for each
measure. Each dictionary contains one key for each category and
corresponding values is an array with scores for each subject.
"""
nr_subs_total = len(np.unique(fm.SUBJECTINDEX))
if nr_subs is None:
nr_subs = nr_subs_total - 1
assert (nr_subs < nr_subs_total)
# initialize output structure; every measure gets one dict with
# category numbers as keys and numpy-arrays as values
sb_scores = []
for measure in range(len(measures.scores)):
res_dict = {}
result_vectors = [np.empty(nr_subs_total) + np.nan
for _ in np.unique(fm.category)]
res_dict.update(zip(np.unique(fm.category),result_vectors))
sb_scores.append(res_dict)
# compute mean spatial bias predictive power for all subjects in all
# categories
for fm_cat in fm.by_field('category'):
cat = fm_cat.category[0]
nr_imgs_cat = len(np.unique(fm_cat.filenumber))
if not nr_imgs:
nr_imgs_current = nr_imgs_cat - 1
else:
nr_imgs_current = nr_imgs
assert(nr_imgs_current < nr_imgs_cat)
for (sub_counter, sub) in enumerate(np.unique(fm.SUBJECTINDEX)):
image_scores = []
for fm_single in fm_cat.by_field('filenumber'):
# Iterating by field filenumber makes filenumbers
# in fm_single unique: Just take the first one to get the
# filenumber for this fixmat
fn = fm_single.filenumber[0]
predicting_subs = (np.setdiff1d(np.unique(
fm_cat.SUBJECTINDEX), [sub]))
np.random.shuffle(predicting_subs)
predicting_subs = predicting_subs[0:nr_subs]
predicting_fns = (np.setdiff1d(np.unique(
fm_cat.filenumber), [fn]))
np.random.shuffle(predicting_fns)
predicting_fns = predicting_fns[0:nr_imgs_current]
predicting_fm = fm_cat[
(ismember(fm_cat.SUBJECTINDEX, predicting_subs)) &
(ismember(fm_cat.filenumber, predicting_fns))]
predicted_fm = fm_single[fm_single.SUBJECTINDEX == sub]
try:
predicting_fdm = compute_fdm(predicting_fm,
scale_factor = scale_factor)
except NoFixations:
predicting_fdm = None
image_scores.append(measures.prediction_scores(predicting_fdm,
predicted_fm))
for (measure, score) in enumerate(nanmean(image_scores, 0)):
sb_scores[measure][cat][sub_counter] = score
return sb_scores
def skip_emd(self):
fm1 = fixmat.TestFactory(params = {'image_size':[93,128]})
fm2 = fixmat.TestFactory(points=zip(range(10,50),range(10,50)),params = {'image_size':[93,128]})
self.assertEquals(measures.emd_model(fixmat.compute_fdm(fm1), fm1), 0)
self.assertTrue(not (measures.emd_model(fixmat.compute_fdm(fm1), fm2) == 0))
def test_pdf(self):
fdm = fixmat.compute_fdm(self.fm)
# ensure that we have a probability density function
if fdm.min() < 0:
fdm = fdm - fdm.min()
self.assertFalse(abs(fdm.sum() - 1) > np.finfo('single').eps)
def test_pdf(self):
fdm = fixmat.compute_fdm(self.fm)
# ensure that we have a probability density function
if fdm.min()< 0:
fdm = fdm - fdm.min()
self.assertFalse(abs(fdm.sum()-1) > np.finfo('single').eps)
def intersubject_scores(fm, category, predicting_filenumbers,
predicting_subjects, predicted_filenumbers,
predicted_subjects, controls = True, scale_factor = 1):
"""
Calculates how well the fixations from a set of subjects on a set of
images can be predicted with the fixations from another set of subjects
on another set of images.
The prediction is carried out by computing a fixation density map from
fixations of predicting_subjects subjects on predicting_images images.
Prediction accuracy is assessed by measures.prediction_scores.
Parameters
fm : fixmat instance
category : int
Category from which the fixations are taken.
predicting_filenumbers : list
List of filenumbers used for prediction, i.e. images where fixations
for the prediction are taken from.
predicting_subjects : list
List of subjects whose fixations on images in predicting_filenumbers
are used for the prediction.
predicted_filenumnbers : list
List of images from which the to be predicted fixations are taken.
predicted_subjects : list
List of subjects used for evaluation, i.e subjects whose fixations
on images in predicted_filenumbers are taken for evaluation.
controls : bool, optional
If True (default), n_predict subjects are chosen from the fixmat.
If False, 1000 fixations are randomly generated and used for
testing.
scale_factor : int, optional
specifies the scaling of the fdm. Default is 1.
Returns
auc : area under the roc curve for sets of actuals and controls
true_pos_rate : ndarray
Rate of true positives for every given threshold value.
All values appearing in actuals are taken as thresholds. Uses lower
sum interpolation.
false_pos_rate : ndarray
See true_pos_rate but for false positives.
"""
predicting_fm = fm[
(ismember(fm.SUBJECTINDEX, predicting_subjects)) &
(ismember(fm.filenumber, predicting_filenumbers)) &
(fm.category == category)]
predicted_fm = fm[
(ismember(fm.SUBJECTINDEX,predicted_subjects)) &
(ismember(fm.filenumber,predicted_filenumbers))&
(fm.category == category)]
try:
predicting_fdm = compute_fdm(predicting_fm, scale_factor = scale_factor)
except NoFixations:
predicting_fdm = None
if controls == True:
fm_controls = fm[
(ismember(fm.SUBJECTINDEX, predicted_subjects)) &
((ismember(fm.filenumber, predicted_filenumbers)) != True) &
(fm.category == category)]
return measures.prediction_scores(predicting_fdm, predicted_fm,
controls = (fm_controls.y, fm_controls.x))
return measures.prediction_scores(predicting_fdm, predicted_fm, controls = None)
