def __init__(self,
stimuli,
fixations,
saliency_map_provider,
remove_doublicates=False,
antonio_gaussian=False,
empirical_maps=None,
cache_empirical_maps=True):
self.stimuli = stimuli
if remove_doublicates:
fixations = remove_doublicate_fixations(fixations)
self.fixations = fixations
self.saliency_map_provider = saliency_map_provider
if empirical_maps is not None:
self.empirical_maps = empirical_maps
elif not antonio_gaussian:
self.empirical_maps = pysaliency.FixationMap(
stimuli,
fixations,
kernel_size=23.99,
caching=cache_empirical_maps)
else:
self.empirical_maps = MITFixationMap(stimuli,
fixations,
fc=8,
caching=cache_empirical_maps)
def test_fixation_map_model(stimuli, fixation_trains):
fixation_map = pysaliency.FixationMap(stimuli, fixation_trains)
smap1 = fixation_map.saliency_map(stimuli[0])
assert smap1.min() == 0
assert smap1.max() == 3
assert smap1.sum() == (fixation_trains.n == 0).sum()
def test_fixation_map_model_ignore_doublicates(stimuli, fixation_trains):
fixation_map = pysaliency.FixationMap(stimuli,
fixation_trains,
ignore_doublicates=True)
smap1 = fixation_map.saliency_map(stimuli[0])
assert smap1.min() == 0
assert smap1.max() == 1
assert smap1.sum() == (fixation_trains.n == 0).sum() - 2
def test_cc_gauss(stimuli, fixation_trains):
gsmm = GaussianSaliencyMapModel()
gold = pysaliency.FixationMap(stimuli,
fixation_trains,
kernel_size=10,
ignore_doublicates=True)
cc = gsmm.CC(stimuli, gold)
np.testing.assert_allclose(cc, -0.1542654)
def test_cc_constant(stimuli, fixation_trains):
model = ConstantSaliencyMapModel()
gold = pysaliency.FixationMap(stimuli,
fixation_trains,
kernel_size=10,
ignore_doublicates=True)
cc = model.CC(stimuli, gold)
np.testing.assert_allclose(cc, 0)
def test_conditional_saliency_maps(stimuli, fixation_trains):
model = pysaliency.FixationMap(stimuli, fixation_trains, kernel_size=10)
saliency_maps_1 = model.conditional_saliency_maps(stimuli, fixation_trains)
saliency_maps_2 = [
model.conditional_saliency_map_for_fixation(stimuli, fixation_trains,
i)
for i in range(len(fixation_trains))
]
np.testing.assert_allclose(saliency_maps_1, saliency_maps_2)
def test_image_based_kldiv_gauss_functions(stimuli, fixation_trains):
gsmm = GaussianSaliencyMapModel()
gold = pysaliency.FixationMap(stimuli,
fixation_trains,
kernel_size=10,
ignore_doublicates=True)
stimuli, fixation_trains = pysaliency.create_subset(
stimuli, fixation_trains, [0])
saliency_map_1 = gsmm.saliency_map(stimuli[0])
saliency_map_2 = gold.saliency_map(stimuli[0])
value1 = gsmm.image_based_kl_divergence(stimuli, gsmm)
value2 = pysaliency.metrics.image_based_kl_divergence(
saliency_map_1, saliency_map_1)
np.testing.assert_allclose(value1, value2)
constant_gold = ConstantSaliencyMapModel()
saliency_map_1 = gsmm.saliency_map(stimuli[0])
saliency_map_2 = constant_gold.saliency_map(stimuli[0])
value1 = gsmm.image_based_kl_divergence(stimuli, constant_gold)
value2 = pysaliency.metrics.image_based_kl_divergence(
saliency_map_1, saliency_map_2)
np.testing.assert_allclose(value1, value2)
constant_gold = ConstantSaliencyMapModel(value=0.0)
saliency_map_2 = constant_gold.saliency_map(stimuli[0])
value1 = gsmm.image_based_kl_divergence(stimuli, constant_gold)
value2 = pysaliency.metrics.image_based_kl_divergence(
saliency_map_1, saliency_map_2)
np.testing.assert_allclose(value1, value2)
# test MIT Benchmarking settings
# (minimum_value=0 can be problematic for constant models)
value1 = gsmm.image_based_kl_divergence(stimuli,
constant_gold,
minimum_value=0,
log_regularization=2.2204e-16,
quotient_regularization=2.2204e-16)
value2 = pysaliency.metrics.MIT_KLDiv(saliency_map_1, saliency_map_2)
np.testing.assert_allclose(value1, value2)
saliency_map_2 = gold.saliency_map(stimuli[0])
value1 = gsmm.image_based_kl_divergence(stimuli, gold)
value2 = pysaliency.metrics.image_based_kl_divergence(
saliency_map_1, saliency_map_2)
np.testing.assert_allclose(value1, value2)
def test_SIM_gauss(stimuli, fixation_trains):
gsmm = GaussianSaliencyMapModel()
constant_gold = ConstantSaliencyMapModel()
sim = gsmm.SIM(stimuli, constant_gold)
np.testing.assert_allclose(sim, 0.54392, rtol=1e-6)
constant_gold = ConstantSaliencyMapModel(value=0.0)
sim = gsmm.SIM(stimuli, constant_gold)
np.testing.assert_allclose(sim, 0.54392, rtol=1e-6)
gold = pysaliency.FixationMap(stimuli,
fixation_trains,
kernel_size=10,
ignore_doublicates=True)
sim = gsmm.SIM(stimuli, gold)
np.testing.assert_allclose(sim, 0.315899, rtol=1e-6)
def test_gauss(self):
stimuli = pysaliency.Stimuli([np.random.randn(40, 40, 3),
np.random.randn(40, 40, 3)])
gsmm = GaussianSaliencyMapModel()
constant_gold = ConstantSaliencyMapModel()
gold = pysaliency.FixationMap(stimuli, self.f, kernel_size = 10)
ib_kl = gsmm.image_based_kl_divergence(stimuli, gsmm)
np.testing.assert_allclose(ib_kl, 0.0)
ib_kl = gsmm.image_based_kl_divergence(stimuli, constant_gold)
np.testing.assert_allclose(ib_kl, 0.8152272380729648)
ib_kl = gsmm.image_based_kl_divergence(stimuli, gold)
np.testing.assert_allclose(ib_kl, 1.961124862592289)
ib_kl = gold.image_based_kl_divergence(stimuli, gold)
np.testing.assert_allclose(ib_kl, 0.0)
def test_image_based_kldiv_gauss(stimuli, fixation_trains):
gsmm = GaussianSaliencyMapModel()
gold = pysaliency.FixationMap(stimuli,
fixation_trains,
kernel_size=10,
ignore_doublicates=True)
ib_kl = gsmm.image_based_kl_divergence(stimuli, gsmm)
np.testing.assert_allclose(ib_kl, 0.0)
ib_kl = gsmm.KLDiv(stimuli, gsmm)
np.testing.assert_allclose(ib_kl, 0.0)
constant_gold = ConstantSaliencyMapModel()
ib_kl = gsmm.image_based_kl_divergence(stimuli, constant_gold)
np.testing.assert_allclose(ib_kl, 0.8396272788909165)
ib_kl = gsmm.KLDiv(stimuli, constant_gold)
np.testing.assert_allclose(ib_kl, 0.8396272788909165)
constant_gold = ConstantSaliencyMapModel(value=0.0)
ib_kl = gsmm.image_based_kl_divergence(stimuli, constant_gold)
np.testing.assert_allclose(ib_kl, 0.8396272788909165)
ib_kl = gsmm.KLDiv(stimuli, constant_gold)
np.testing.assert_allclose(ib_kl, 0.8396272788909165)
# test MIT Benchmarking settings
# (minimum_value=0 can be problematic for constant models)
ib_kl = gsmm.image_based_kl_divergence(stimuli,
constant_gold,
minimum_value=0,
log_regularization=2.2204e-16,
quotient_regularization=2.2204e-16)
np.testing.assert_allclose(ib_kl, 0.8396272788909165)
ib_kl = gsmm.image_based_kl_divergence(stimuli, gold)
np.testing.assert_allclose(ib_kl, 2.029283206727416)
ib_kl = gold.image_based_kl_divergence(stimuli, gold)
np.testing.assert_allclose(ib_kl, 0.0)
def test_image_based_kldiv_gauss(stimuli, fixation_trains):
gsmm = GaussianSaliencyMapModel()
constant_gold = ConstantSaliencyMapModel()
gold = pysaliency.FixationMap(stimuli,
fixation_trains,
kernel_size=10,
ignore_doublicates=True)
ib_kl = gsmm.image_based_kl_divergence(stimuli, gsmm)
np.testing.assert_allclose(ib_kl, 0.0)
ib_kl = gsmm.KLDiv(stimuli, gsmm)
np.testing.assert_allclose(ib_kl, 0.0)
ib_kl = gsmm.image_based_kl_divergence(stimuli, constant_gold)
np.testing.assert_allclose(ib_kl, 0.8396272788909165)
ib_kl = gsmm.KLDiv(stimuli, constant_gold)
np.testing.assert_allclose(ib_kl, 0.8396272788909165)
ib_kl = gsmm.image_based_kl_divergence(stimuli, gold)
np.testing.assert_allclose(ib_kl, 2.029283206727416)
ib_kl = gold.image_based_kl_divergence(stimuli, gold)
np.testing.assert_allclose(ib_kl, 0.0)
def __init__(self, stimuli, fixations, fc=8):
fixation_map = pysaliency.FixationMap(stimuli, fixations)
super(MITFixationMap, self).__init__(fixation_map, fc=fc)