Merge pull request #46 from bonsai-rx/fix/get-gaussian-observation-st…

…atistics

Fix get gaussian observation statistics

src/Bonsai.ML.HiddenMarkovModels.Design/GaussianObservationsStatisticsClustersVisualizer.cs

@@ -149,7 +149,6 @@ public override void Show(object value)
         {
             if (value is Observations.GaussianObservationsStatistics gaussianObservationsStatistics)
             {
-
                 var statesCount = gaussianObservationsStatistics.Means.GetLength(0);
                 var observationDimensions = gaussianObservationsStatistics.Means.GetLength(1);
 
@@ -227,12 +226,14 @@ public override void Show(object value)
 
                 var batchObservationsCount = gaussianObservationsStatistics.BatchObservations.GetLength(0);
                 var offset = BufferData && batchObservationsCount > BufferCount ? batchObservationsCount - BufferCount : 0;
+                var predictedStatesCount = gaussianObservationsStatistics.PredictedStates.Length;
+
                 for (int i = offset; i < batchObservationsCount; i++)
                 {
                     var dim1 = gaussianObservationsStatistics.BatchObservations[i, dimension1SelectedIndex];
                     var dim2 = gaussianObservationsStatistics.BatchObservations[i, dimension2SelectedIndex];
-                    var state = gaussianObservationsStatistics.InferredMostProbableStates[i];
-                    allScatterSeries[(int)state].Points.Add(new ScatterPoint(dim1, dim2, value: state, tag: state));
+                    var state = gaussianObservationsStatistics.PredictedStates[i];
+                    allScatterSeries[Convert.ToInt32(state)].Points.Add(new ScatterPoint(dim1, dim2, value: state, tag: state));
                 }
 
                 for (int i = 0; i < statesCount; i++)

src/Bonsai.ML.HiddenMarkovModels/InferState.bonsai

@@ -52,7 +52,7 @@
             </Expression>
             <Expression xsi:type="Combinator">
               <Combinator xsi:type="py:Exec">
-                <py:Script>hmm.most_likely_states([59.7382107943162,3.99285183724331])</py:Script>
+                <py:Script>hmm.infer_state([59.7382107943162,3.99285183724331])</py:Script>
               </Combinator>
             </Expression>
             <Expression xsi:type="WorkflowOutput" />

src/Bonsai.ML.HiddenMarkovModels/Observations/GaussianObservationsStatistics.cs

@@ -45,11 +45,11 @@ public class GaussianObservationsStatistics
         public double[,] BatchObservations { get; set; }
 
         /// <summary>
-        /// The sequence of inferred most probable states.
+        /// The predicted state for each observation in the batch of observations.
         /// </summary>
-        [Description("The sequence of inferred most probable states.")]
+        [Description("The predicted state for each observation in the batch of observations.")]
         [XmlIgnore]
-        public int[] InferredMostProbableStates { get; set; }
+        public long[] PredictedStates { get; set; }
 
         /// <summary>
         /// Transforms an observable sequence of <see cref="PyObject"/> into an observable sequence 
@@ -64,15 +64,15 @@ public IObservable<GaussianObservationsStatistics> Process(IObservable<PyObject>
                 var covarianceMatricesPyObj = (double[,,])observationsPyObj.GetArrayAttr("Sigmas");
                 var stdDevsPyObj = DiagonalSqrt(covarianceMatricesPyObj);
                 var batchObservationsPyObj = (double[,])pyObject.GetArrayAttr("batch_observations");
-                var inferredMostProbableStatesPyObj = (int[])pyObject.GetArrayAttr("inferred_most_probable_states");
+                var predictedStatesPyObj = (long[])pyObject.GetArrayAttr("predicted_states");
 
                 return new GaussianObservationsStatistics
                 {
                     Means = meansPyObj,
                     StdDevs = stdDevsPyObj,
                     CovarianceMatrices = covarianceMatricesPyObj,
                     BatchObservations = batchObservationsPyObj,
-                    InferredMostProbableStates = inferredMostProbableStatesPyObj
+                    PredictedStates = predictedStatesPyObj
                 };
             });
         }

src/Bonsai.ML.HiddenMarkovModels/main.py

@@ -78,14 +78,15 @@ def get_nonlinearity_type(func):
         self.state_probabilities = None
 
         self.batch = None
-        self.batch_observations = np.array([[]], dtype=float)
+        self.batch_observations = np.array([[]], dtype=float).reshape((0, dimensions))
         self.is_running = False
         self._fit_finished = False
         self.loop = None
         self.thread = None
         self.curr_batch_size = 0
         self.flush_data_between_batches = True
-        self.inferred_most_probable_states = np.array([], dtype=int)
+        self.predicted_states = np.array([], dtype=int)
+        self.buffer_count = 250
 
     def update_params(self, initial_state_distribution, transitions_params, observations_params):
         hmm_params = self.params
@@ -124,10 +125,17 @@ def update_params(self, initial_state_distribution, transitions_params, observat
 
     def infer_state(self, observation: list[float]):
 
-        self.log_alpha = self.compute_log_alpha(
-            np.expand_dims(np.array(observation), 0), self.log_alpha)
+        observation = np.expand_dims(np.array(observation), 0)
+        self.log_alpha = self.compute_log_alpha(observation, self.log_alpha)
         self.state_probabilities = np.exp(self.log_alpha).astype(np.double)
-        return self.state_probabilities.argmax()
+        prediction = self.state_probabilities.argmax()
+        self.predicted_states = np.append(self.predicted_states, prediction)
+        if self.predicted_states.shape[0] > self.buffer_count:
+            self.predicted_states = self.predicted_states[1:]
+        self.batch_observations = np.vstack([self.batch_observations, observation])
+        if self.batch_observations.shape[0] == self.buffer_count:
+            self.batch_observations = self.batch_observations[1:]
+        return prediction
 
     def compute_log_alpha(self, obs, log_alpha=None):
 
@@ -171,8 +179,6 @@ def fit_async(self,
             self.batch = np.vstack(
                 [self.batch[1:], np.expand_dims(np.array(observation), 0)])
 
-        self.batch_observations = self.batch
-
         if not self.is_running and self.loop is None and self.thread is None:
 
             if self.curr_batch_size >= batch_size:
@@ -221,8 +227,6 @@ def on_completion(future):
                     if self.flush_data_between_batches:
                         self.batch = None
 
-                    self.inferred_most_probable_states = np.array([self.infer_state(obs) for obs in self.batch_observations]).astype(int)
-
                 self.is_running = True
 
                 if self.loop is None or self.loop.is_closed():