Update to eventseg

brainiak/eventseg/event.py

@@ -33,6 +33,7 @@
 import copy
 from sklearn.base import BaseEstimator
 from sklearn.utils.validation import check_is_fitted, check_array
+from sklearn.exceptions import NotFittedError
 
 from . import _utils as utils  # type: ignore
 
@@ -86,8 +87,25 @@ def __init__(self, n_events=2,
                  step_var=_default_var_schedule,
                  n_iter=500):
         self.n_events = n_events
+        self.classes_ = np.arange(self.n_events)
         self.step_var = step_var
         self.n_iter = n_iter
+        self.event_var_ = None
+
+        # Set up transition matrix, with final sink state
+        # For transition matrix of this form, the transition probability has
+        # no impact on the final solution, since all valid paths must take
+        # the same number of transitions
+        self.p_start = np.zeros((1, self.n_events + 1))
+        self.p_start[0, 0] = 1
+        self.P = np.vstack((np.hstack((
+            0.5 * np.diag(np.ones(self.n_events))
+            + 0.5 * np.diag(np.ones(self.n_events - 1), 1),
+            np.append(np.zeros((self.n_events - 1, 1)), [[0.5]], axis=0))),
+                            np.append(np.zeros((1, self.n_events)), [[1]],
+                                      axis=1)))
+        self.p_end = np.zeros((1, self.n_events + 1))
+        self.p_end[0, -2] = 1
 
     def fit(self, X, y=None):
         """Learn a segmentation on training data
@@ -126,22 +144,6 @@ def fit(self, X, y=None):
         for i in range(n_train):
             X[i] = stats.zscore(X[i], axis=1, ddof=1)
 
-        # Set up transition matrix, with final sink state
-        # For transition matrix of this form, the transition probability has
-        # no impact on the final solution, since all valid paths must take
-        # the same number of transitions
-        self.p_start = np.zeros((1, self.n_events + 1))
-        self.p_start[0, 0] = 1
-        self.P = np.vstack((np.hstack((
-            0.5 * np.diag(np.ones(self.n_events))
-            + 0.5 * np.diag(np.ones(self.n_events - 1), 1),
-            np.append(np.zeros((self.n_events - 1, 1)), [[0.5]], axis=0))),
-            np.append(np.zeros((1, self.n_events)), [[1]], axis=1)))
-        self.p_end = np.zeros((1, self.n_events + 1))
-        self.p_end[0, -2] = 1
-
-        self.classes_ = np.arange(self.n_events)
-
         # Initialize variables for fitting
         log_gamma = []
         for i in range(n_train):
@@ -313,12 +315,28 @@ def _log(self, x):
         y = utils.masked_log(_x)
         return y.reshape(xshape)
 
+    def set_event_patterns(self, event_pat):
+        """Set HMM event patterns manually
+
+        Rather than fitting the event patterns automatically using fit(), this
+        function allows them to be set explicitly. They can then be used to
+        find corresponding events in a new dataset, using find_events().
+
+        Parameters
+        ----------
+        event_pat: voxel by event ndarray
+        """
+        if event_pat.shape[1] != self.n_events:
+            raise ValueError(("Number of columns of event_pat must match "
+                              "number of events"))
+        self.event_pat_ = event_pat.copy()
+
     def find_events(self, testing_data, var=None, scramble=False):
         """Applies learned event segmentation to new testing dataset
 
-        After fitting an event segmentation using fit(),
-        this function finds the same sequence of event patterns in a new
-        testing dataset.
+        After fitting an event segmentation using fit() or setting event
+        patterns directly using set_event_patterns(), this function finds the
+        same sequence of event patterns in a new testing dataset.
 
         Parameters
         ----------
@@ -328,7 +346,8 @@ def find_events(self, testing_data, var=None, scramble=False):
         var: float or 1D ndarray of length equal to the number of events
             default: uses variance that maximized training log-likelihood
             Variance of the event Gaussians. If scalar, all events are
-            assumed to have the same variance.
+            assumed to have the same variance. If fit() has not previously
+            been run, this must be specifed (cannot be None).
 
         scramble: bool : default False
             If true, the order of the learned events are shuffled before
@@ -345,7 +364,11 @@ def find_events(self, testing_data, var=None, scramble=False):
         """
 
         if var is None:
-            var = self.event_var_
+            if self.event_var_ is None:
+                raise NotFittedError(("The event patterns must first be set "
+                                      "by fit() or set_event_patterns()"))
+            else:
+                var = self.event_var_
 
         if scramble:
             mean_pat = self.event_pat_[:, np.random.permutation(self.n_events)]

docs/newsfragments/266.feature

@@ -0,0 +1 @@
+eventseg: add set_event_patterns() and notebook example

examples/eventseg/requirements.txt

@@ -1 +1,6 @@
 matplotlib
+notebook
+sklearn
+deepdish
+scipy
+

examples/isfc/download_data.sh

@@ -1,4 +1,4 @@
 #!/bin/sh
-curl --location -o isfc.zip https://api.onedrive.com/v1.0/shares/s!Aobi2ryypFQCgqQOm2Zhs-TSLu9leQ/root/content
+wget https://ndownloader.figshare.com/files/9342556 -O isfc.zip
 unzip -qo isfc.zip
 rm -f isfc.zip

tests/eventseg/test_event.py

@@ -38,3 +38,14 @@ def test_simple_boundary():
     events = np.argmax(es.segments_[0], axis=1)
     assert np.array_equal(events, [0, 0, 0, 1, 1, 1, 1]),\
         "Failed to correctly segment two events"
+
+
+def test_event_transfer():
+    es = brainiak.eventseg.event.EventSegment(2)
+    es.set_event_patterns(np.asarray([[1, 0], [0, 1]]))
+    sample_data = np.asarray([[1, 1, 1, 0, 0, 0, 0], [0, 0, 0, 1, 1, 1, 1]])
+    seg = es.find_events(sample_data.T, np.asarray([1, 1]))[0]
+
+    events = np.argmax(seg, axis=1)
+    assert np.array_equal(events, [0, 0, 0, 1, 1, 1, 1]),\
+        "Failed to correctly transfer two events to new data"