Publish 0.2.2

Author: wq2012

README.md

@@ -185,6 +185,15 @@ clusterer = spectral_clusterer.SpectralClusterer(
 labels = clusterer.predict(matrix, constraint_matrix)
 ```
 
+The constraint matrix can be constructed from a `speaker_turn_scores` list:
+
+```python
+from spectralcluster import constraint
+
+constraint_matrix = constraint.ConstraintMatrix(
+    spk_turn_entries, threshold=1).compute_diagonals()
+```
+
 ## Citations
 
 Our paper is cited as:

docs/configs.html

@@ -29,16 +29,22 @@ <h1 class="title">Module <code>spectralcluster.configs</code></h1>
 </summary>
 <pre><code class="python">&#34;&#34;&#34;Example configurations.&#34;&#34;&#34;
 
+from spectralcluster import autotune
+from spectralcluster import constraint
+from spectralcluster import laplacian
 from spectralcluster import refinement
 from spectralcluster import spectral_clusterer
 
+AutoTune = autotune.AutoTune
+ConstraintName = constraint.ConstraintName
+ConstraintOptions = constraint.ConstraintOptions
 RefinementName = refinement.RefinementName
 RefinementOptions = refinement.RefinementOptions
 ThresholdType = refinement.ThresholdType
 SymmetrizeType = refinement.SymmetrizeType
+LaplacianType = laplacian.LaplacianType
 SpectralClusterer = spectral_clusterer.SpectralClusterer
 
-
 # Configurations that are closest to the ICASSP2018 paper
 # &#34;Speaker Diarization with LSTM&#34;.
 ICASSP2018_REFINEMENT_SEQUENCE = [
@@ -63,6 +69,41 @@ <h1 class="title">Module <code>spectralcluster.configs</code></h1>
     autotune=None,
     laplacian_type=None,
     refinement_options=icassp2018_refinement_options,
+    custom_dist=&#34;cosine&#34;)
+
+# Configurations of Turn-To-Diarize system using the
+# Turn + Constraint Propagation + AutoTune method
+TURNTODIARIZE_REFINEMENT_SEQUENCE = [
+    RefinementName.RowWiseThreshold, RefinementName.Symmetrize
+]
+
+turntodiarize_refinement_options = RefinementOptions(
+    thresholding_soft_multiplier=0.01,
+    thresholding_type=ThresholdType.Percentile,
+    thresholding_with_binarization=True,
+    thresholding_preserve_diagonal=True,
+    symmetrize_type=SymmetrizeType.Average,
+    refinement_sequence=TURNTODIARIZE_REFINEMENT_SEQUENCE)
+
+turntodiarize_constraint_options = ConstraintOptions(
+    constraint_name=ConstraintName.ConstraintPropagation,
+    apply_before_refinement=True,
+    constraint_propagation_alpha=0.4)
+
+turntodiarize_auto_tune = AutoTune(
+    p_percentile_min=0.40,
+    p_percentile_max=0.95,
+    init_search_step=0.05,
+    search_level=1)
+
+turntodiarize_clusterer = SpectralClusterer(
+    min_clusters=2,
+    max_clusters=7,
+    refinement_options=turntodiarize_refinement_options,
+    constraint_options=turntodiarize_constraint_options,
+    autotune=turntodiarize_auto_tune,
+    laplacian_type=LaplacianType.GraphCut,
+    row_wise_renorm=True,
     custom_dist=&#34;cosine&#34;)</code></pre>
 </details>
 </section>

docs/index.html

@@ -35,7 +35,7 @@ <h1 class="title">Package <code>spectralcluster</code></h1>
 from . import laplacian
 from . import refinement
 from . import spectral_clusterer
-
+from . import utils
 
 AutoTune = autotune.AutoTune
 
@@ -52,6 +52,8 @@ <h1 class="title">Package <code>spectralcluster</code></h1>
 
 SpectralClusterer = spectral_clusterer.SpectralClusterer
 
+EigenGapType = utils.EigenGapType
+
 ICASSP2018_REFINEMENT_SEQUENCE = configs.ICASSP2018_REFINEMENT_SEQUENCE</code></pre>
 </details>
 </section>

docs/laplacian.html

@@ -43,10 +43,10 @@ <h1 class="title">Module <code>spectralcluster.laplacian</code></h1>
   # The unnormalied Laplacian: L = D - W
   Unnormalized = 1
 
-  # The random walk view normalized Laplacian:  D^{-1}L
+  # The random walk view normalized Laplacian:  D^{-1} * L
   RandomWalk = 2
 
-  # The graph cut view normalized Laplacian: D^{-1/2}LD^{-1/2}
+  # The graph cut view normalized Laplacian: D^{-1/2} * L * D^{-1/2}
   GraphCut = 3
 
 
@@ -181,10 +181,10 @@ <h2 class="section-title" id="header-classes">Classes</h2>
   # The unnormalied Laplacian: L = D - W
   Unnormalized = 1
 
-  # The random walk view normalized Laplacian:  D^{-1}L
+  # The random walk view normalized Laplacian:  D^{-1} * L
   RandomWalk = 2
 
-  # The graph cut view normalized Laplacian: D^{-1/2}LD^{-1/2}
+  # The graph cut view normalized Laplacian: D^{-1/2} * L * D^{-1/2}
   GraphCut = 3</code></pre>
 </details>
 <h3>Ancestors</h3>

docs/spectral_clusterer.html

@@ -39,6 +39,7 @@ <h1 class="title">Module <code>spectralcluster.spectral_clusterer</code></h1>
 RefinementName = refinement.RefinementName
 LaplacianType = laplacian.LaplacianType
 ConstraintName = constraint.ConstraintName
+EigenGapType = utils.EigenGapType
 
 
 class SpectralClusterer:
@@ -54,7 +55,8 @@ <h1 class="title">Module <code>spectralcluster.spectral_clusterer</code></h1>
                row_wise_renorm=False,
                custom_dist=&#34;cosine&#34;,
                max_iter=300,
-               constraint_options=None):
+               constraint_options=None,
+               eigengap_type=EigenGapType.Ratio):
     &#34;&#34;&#34;Constructor of the clusterer.
 
     Args:
@@ -77,6 +79,7 @@ <h1 class="title">Module <code>spectralcluster.spectral_clusterer</code></h1>
       max_iter: the maximum number of iterations for the custom k-means
       constraint_options: a ConstraintOptions object that contains constraint
         arguments
+      eigengap_type: the type of the eigengap computation
     &#34;&#34;&#34;
     self.min_clusters = min_clusters
     self.max_clusters = max_clusters
@@ -91,6 +94,7 @@ <h1 class="title">Module <code>spectralcluster.spectral_clusterer</code></h1>
     self.custom_dist = custom_dist
     self.max_iter = max_iter
     self.constraint_options = constraint_options
+    self.eigengap_type = eigengap_type
 
   def _compute_eigenvectors_ncluster(self, affinity, constraint_matrix=None):
     &#34;&#34;&#34;Perform eigen decomposition and estiamte the number of clusters.
@@ -133,7 +137,11 @@ <h1 class="title">Module <code>spectralcluster.spectral_clusterer</code></h1>
       (eigenvalues, eigenvectors) = utils.compute_sorted_eigenvectors(affinity)
       # Get number of clusters.
       n_clusters, max_delta_norm = utils.compute_number_of_clusters(
-          eigenvalues, self.max_clusters, self.stop_eigenvalue, descend=True)
+          eigenvalues,
+          self.max_clusters,
+          self.stop_eigenvalue,
+          self.eigengap_type,
+          descend=True)
     else:
       # Compute Laplacian matrix
       laplacian_norm = laplacian.compute_laplacian(
@@ -144,7 +152,7 @@ <h1 class="title">Module <code>spectralcluster.spectral_clusterer</code></h1>
           laplacian_norm, descend=False)
       # Get number of clusters. Eigen values are sorted in an ascending order
       n_clusters, max_delta_norm = utils.compute_number_of_clusters(
-          eigenvalues, self.max_clusters, descend=False)
+          eigenvalues, self.max_clusters, self.eigengap_type, descend=False)
     return eigenvectors, n_clusters, max_delta_norm
 
   def predict(self, embeddings, constraint_matrix=None):
@@ -187,7 +195,7 @@ <h1 class="title">Module <code>spectralcluster.spectral_clusterer</code></h1>
         (eigenvectors, n_clusters,
          max_delta_norm) = self._compute_eigenvectors_ncluster(
              affinity, constraint_matrix)
-        ratio = (1 - p_percentile) / max_delta_norm
+        ratio = np.sqrt(1 - p_percentile) / max_delta_norm
         return ratio, eigenvectors, n_clusters
 
       eigenvectors, n_clusters, _ = self.autotune.tune(p_percentile_to_ratio)
@@ -228,7 +236,7 @@ <h2 class="section-title" id="header-classes">Classes</h2>
 <dl>
 <dt id="spectralcluster.spectral_clusterer.SpectralClusterer"><code class="flex name class">
 <span>class <span class="ident">SpectralClusterer</span></span>
-<span>(</span><span>min_clusters=None, max_clusters=None, refinement_options=None, autotune=None, laplacian_type=None, stop_eigenvalue=0.01, row_wise_renorm=False, custom_dist='cosine', max_iter=300, constraint_options=None)</span>
+<span>(</span><span>min_clusters=None, max_clusters=None, refinement_options=None, autotune=None, laplacian_type=None, stop_eigenvalue=0.01, row_wise_renorm=False, custom_dist='cosine', max_iter=300, constraint_options=None, eigengap_type=EigenGapType.Ratio)</span>
 </code></dt>
 <dd>
 <div class="desc"><p>Spectral clustering class.</p>
@@ -264,6 +272,8 @@ <h2 id="args">Args</h2>
 <dt><strong><code>constraint_options</code></strong></dt>
 <dd>a ConstraintOptions object that contains constraint
 arguments</dd>
+<dt><strong><code>eigengap_type</code></strong></dt>
+<dd>the type of the eigengap computation</dd>
 </dl></div>
 <details class="source">
 <summary>
@@ -282,7 +292,8 @@ <h2 id="args">Args</h2>
                row_wise_renorm=False,
                custom_dist=&#34;cosine&#34;,
                max_iter=300,
-               constraint_options=None):
+               constraint_options=None,
+               eigengap_type=EigenGapType.Ratio):
     &#34;&#34;&#34;Constructor of the clusterer.
 
     Args:
@@ -305,6 +316,7 @@ <h2 id="args">Args</h2>
       max_iter: the maximum number of iterations for the custom k-means
       constraint_options: a ConstraintOptions object that contains constraint
         arguments
+      eigengap_type: the type of the eigengap computation
     &#34;&#34;&#34;
     self.min_clusters = min_clusters
     self.max_clusters = max_clusters
@@ -319,6 +331,7 @@ <h2 id="args">Args</h2>
     self.custom_dist = custom_dist
     self.max_iter = max_iter
     self.constraint_options = constraint_options
+    self.eigengap_type = eigengap_type
 
   def _compute_eigenvectors_ncluster(self, affinity, constraint_matrix=None):
     &#34;&#34;&#34;Perform eigen decomposition and estiamte the number of clusters.
@@ -361,7 +374,11 @@ <h2 id="args">Args</h2>
       (eigenvalues, eigenvectors) = utils.compute_sorted_eigenvectors(affinity)
       # Get number of clusters.
       n_clusters, max_delta_norm = utils.compute_number_of_clusters(
-          eigenvalues, self.max_clusters, self.stop_eigenvalue, descend=True)
+          eigenvalues,
+          self.max_clusters,
+          self.stop_eigenvalue,
+          self.eigengap_type,
+          descend=True)
     else:
       # Compute Laplacian matrix
       laplacian_norm = laplacian.compute_laplacian(
@@ -372,7 +389,7 @@ <h2 id="args">Args</h2>
           laplacian_norm, descend=False)
       # Get number of clusters. Eigen values are sorted in an ascending order
       n_clusters, max_delta_norm = utils.compute_number_of_clusters(
-          eigenvalues, self.max_clusters, descend=False)
+          eigenvalues, self.max_clusters, self.eigengap_type, descend=False)
     return eigenvectors, n_clusters, max_delta_norm
 
   def predict(self, embeddings, constraint_matrix=None):
@@ -415,7 +432,7 @@ <h2 id="args">Args</h2>
         (eigenvectors, n_clusters,
          max_delta_norm) = self._compute_eigenvectors_ncluster(
              affinity, constraint_matrix)
-        ratio = (1 - p_percentile) / max_delta_norm
+        ratio = np.sqrt(1 - p_percentile) / max_delta_norm
         return ratio, eigenvectors, n_clusters
 
       eigenvectors, n_clusters, _ = self.autotune.tune(p_percentile_to_ratio)
@@ -516,7 +533,7 @@ <h2 id="raises">Raises</h2>
       (eigenvectors, n_clusters,
        max_delta_norm) = self._compute_eigenvectors_ncluster(
            affinity, constraint_matrix)
-      ratio = (1 - p_percentile) / max_delta_norm
+      ratio = np.sqrt(1 - p_percentile) / max_delta_norm
       return ratio, eigenvectors, n_clusters
 
     eigenvectors, n_clusters, _ = self.autotune.tune(p_percentile_to_ratio)