StimulusDependentSaliencyMapModel

CHANGELOG.md

@@ -1,6 +1,7 @@
 # Changelog
 
 * 0.2.22 (dev):
+  * Feature: `StimulusDependentSaliencyMapModel`
   * Bugfix: The NUSEF dataset scaled some fixations not correctly to image coordinates. Also, we now account for some typos in the
     dataset source data.
   * Feature: CrossvalMultipleRegularizations and GeneralMixtureKernelDensityEstimator in baseline utils (names might change!)

pysaliency/__init__.py

@@ -28,6 +28,7 @@
     ExpSaliencyMapModel,
     DisjointUnionSaliencyMapModel,
     SubjectDependentSaliencyMapModel,
+    StimulusDependentSaliencyMapModel,
     ResizingSaliencyMapModel,
     BluringSaliencyMapModel,
     DigitizeMapModel,

pysaliency/saliency_map_models.py

@@ -1,4 +1,5 @@
 from __future__ import absolute_import, print_function, division, unicode_literals
+from itertools import combinations
 
 import os
 from abc import ABCMeta, abstractmethod
@@ -15,7 +16,7 @@
 from .numba_utils import fill_fixation_map, auc_for_one_positive
 
 from .utils import TemporaryDirectory, run_matlab_cmd, Cache, average_values, deprecated_class, remove_trailing_nans
-from .datasets import Stimulus, Fixations
+from .datasets import Stimulus, Fixations, get_image_hash
 from .metrics import CC, NSS, SIM
 from .sampling_models import SamplingModelMixin
 
@@ -934,6 +935,31 @@ def conditional_saliency_map(self, stimulus, x_hist, y_hist, t_hist, attributes=
             stimulus, x_hist, y_hist, t_hist, attributes=attributes, **kwargs)
 
 
+class StimulusDependentSaliencyMapModel(SaliencyMapModel):
+    def __init__(self, stimuli_models, check_stimuli=True, fallback_model=None, **kwargs):
+        super(StimulusDependentSaliencyMapModel, self).__init__(**kwargs)
+        self.stimuli_models = stimuli_models
+        self.fallback_model = fallback_model
+        if check_stimuli:
+            self.check_stimuli()
+
+    def check_stimuli(self):
+        for s1, s2 in tqdm(list(combinations(self.stimuli_models, 2))):
+            if not set(s1.stimulus_ids).isdisjoint(s2.stimulus_ids):
+                raise ValueError('Stimuli not disjoint')
+
+    def _saliency_map(self, stimulus):
+        stimulus_hash = get_image_hash(stimulus)
+        for stimuli, model in self.stimuli_models.items():
+            if stimulus_hash in stimuli.stimulus_ids:
+                return model.saliency_map(stimulus)
+        else:
+            if self.fallback_model is not None:
+                return self.fallback_model.saliency_map(stimulus)
+            else:
+                raise ValueError('stimulus not provided by these models')
+
+
 class ExpSaliencyMapModel(SaliencyMapModel):
     def __init__(self, parent_model):
         super(ExpSaliencyMapModel, self).__init__(caching=False)

tests/test_saliency_map_models.py

@@ -493,3 +493,36 @@ def test_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_stimulus_dependent_saliency_map_model(stimuli, fixation_trains):
+    # Create stimulus models
+    stimulus_model_1 = ConstantSaliencyMapModel(value=0.5)
+    stimulus_model_2 = GaussianSaliencyMapModel()
+
+    # Create the stimulus-dependent saliency map model
+    stimuli_models = {stimuli[[0]]: stimulus_model_1, stimuli[[1]]: stimulus_model_2}
+    fallback_model = ConstantSaliencyMapModel(value=0.2)
+    sdsmm = pysaliency.saliency_map_models.StimulusDependentSaliencyMapModel(stimuli_models, fallback_model=fallback_model)
+
+    # Test saliency map for stimulus 1
+    saliency_map_1 = sdsmm.saliency_map(stimuli[0])
+    np.testing.assert_allclose(saliency_map_1, np.ones((40, 40)) * 0.5)
+
+    # Test saliency map for stimulus 2
+    saliency_map_2 = sdsmm.saliency_map(stimuli[1])
+    height = stimuli[1].shape[0]
+    width = stimuli[1].shape[1]
+    expected_saliency_map_2 = np.exp(-0.5 * ((np.mgrid[:height, :width][1] - 0.5 * width) ** 2 +
+                                             (np.mgrid[:height, :width][0] - 0.5 * height) ** 2) /
+                                     np.sqrt(width ** 2 + height ** 2))
+    np.testing.assert_allclose(saliency_map_2, expected_saliency_map_2)
+
+    # Test fallback model
+    fallback_saliency_map = fallback_model.saliency_map(np.random.randn(50, 50, 3))
+    np.testing.assert_allclose(fallback_saliency_map, np.ones((50, 50)) * 0.2)
+
+    # Test saliency map for stimulus not provided by the models if there is no fallback model
+    sdsmm.fallback_model = None
+    with pytest.raises(ValueError):
+        sdsmm.saliency_map(np.random.randn(50, 50, 3))