Updated notebook with empirically derived radius for latent neighborhood

deepof/model_utils.py

@@ -211,10 +211,20 @@ class neighbor_cluster_purity(tf.keras.callbacks.Callback):
     """
 
     def __init__(
-        self, r, variational=True, validation_data=None, samples=10000, log_dir="."
+        self,
+        encoding_dim,
+        variational=True,
+        validation_data=None,
+        samples=10000,
+        log_dir=".",
     ):
         super().__init__()
-        self.r = r
+        self.enc = encoding_dim
+        self.r = (
+            -0.14220132706202965 * np.log2(validation_data.shape[0])
+            + 0.17189696892334544 * self.enc
+            + 1.6940295848037952
+        )  # Empirically derived from data. See examples/set_default_entropy_radius.ipynb for details
         self.variational = variational
         self.validation_data = validation_data
         self.samples = samples
@@ -249,6 +259,7 @@ class neighbor_cluster_purity(tf.keras.callbacks.Callback):
             encoding = encoder.predict(self.validation_data)
             groups = grouper.predict(self.validation_data)
             hard_groups = groups.argmax(axis=1)
+            max_groups = groups.max(axis=1)
 
             # compute pairwise distances on latent space
             pdist = pairwise_distances(encoding)
@@ -259,7 +270,7 @@ class neighbor_cluster_purity(tf.keras.callbacks.Callback):
                 range(encoding.shape[0]), self.samples, replace=False
             )
             purity_vector = np.zeros(self.samples)
-            purity_weights = np.zeros(self.samples)
+            neighbor_number = np.zeros(self.samples)
 
             for i, sample in enumerate(random_idxs):
 
@@ -270,7 +281,10 @@ class neighbor_cluster_purity(tf.keras.callbacks.Callback):
                 neigh_entropy = entropy(z)
 
                 purity_vector[i] = neigh_entropy
-                purity_weights[i] = np.sum(neighborhood)
+                neighbor_number[i] = np.sum(neighborhood)
+
+            # Compute weights multiplying neighbor number and target confidence
+            purity_weights = neighbor_number * max_groups
 
             writer = tf.summary.create_file_writer(self.log_dir)
             with writer.as_default():
@@ -279,6 +293,16 @@ class neighbor_cluster_purity(tf.keras.callbacks.Callback):
                     data=np.average(purity_vector, weights=purity_weights),
                     step=epoch,
                 )
+                tf.summary.scalar(
+                    "average_neighbors_in_radius",
+                    data=np.average(neighbor_number),
+                    step=epoch,
+                )
+                tf.summary.scalar(
+                    "average_confidence_in_selected_cluster",
+                    data=np.average(max_groups),
+                    step=epoch,
+                )
 
 
 class uncorrelated_features_constraint(Constraint):

deepof/train_model.py

@@ -399,7 +399,6 @@ else:
         cp=False,
         variational=variational,
         entropy_samples=entropy_samples,
-        entropy_radius=entropy_radius,
         phenotype_class=pheno_class,
         predictor=predictor,
         loss=loss,

deepof/train_utils.py

@@ -76,7 +76,6 @@ def get_callbacks(
     reg_cat_clusters: bool = False,
     reg_cluster_variance: bool = False,
     entropy_samples: int = 15000,
-    entropy_radius: float = None,
     logparam: dict = None,
     outpath: str = ".",
 ) -> List[Union[Any]]:
@@ -113,13 +112,7 @@ def get_callbacks(
     )
 
     entropy = deepof.model_utils.neighbor_cluster_purity(
-        r=(
-            entropy_radius
-            if entropy_radius is not None
-            else 0.15 * logparam["encoding"]
-            - 0.18  # equation derived empirically to keep neighbor number constant.
-            # See examples/set_default_entropy_radius.ipynb for details
-        ),
+        encoding_dim=logparam["encoding"],
         samples=entropy_samples,
         validation_data=X_val,
         log_dir=os.path.join(outpath, "metrics", run_ID),
@@ -270,7 +263,6 @@ def autoencoder_fitting(
     variational: bool,
     reg_cat_clusters: bool,
     reg_cluster_variance: bool,
-    entropy_radius: float,
     entropy_samples: int,
 ):
     """Implementation function for deepof.data.coordinates.deep_unsupervised_embedding"""
@@ -300,7 +292,6 @@ def autoencoder_fitting(
         phenotype_class=phenotype_class,
         predictor=predictor,
         loss=loss,
-        entropy_radius=entropy_radius,
         entropy_samples=entropy_samples,
         reg_cat_clusters=reg_cat_clusters,
         reg_cluster_variance=reg_cluster_variance,

deepof/utils.py

@@ -572,9 +572,9 @@ def circular_arena_recognition(frame: np.array) -> np.array:
     ellipse_params = cv2.fitEllipse(cnts[0])
     center_coordinates = tuple([int(i) for i in ellipse_params[0]])
     axes_length = tuple([int(i) // 2 for i in ellipse_params[1]])
-    angle = ellipse_params[2]
+    ellipse_angle = ellipse_params[2]
 
-    return center_coordinates, axes_length, angle
+    return center_coordinates, axes_length, ellipse_angle
 
 
 def rolling_speed(

tests/test_model_utils.py

@@ -270,7 +270,7 @@ def test_find_learning_rate():
 
 
 def test_neighbor_cluster_purity():
-    X = np.random.uniform(0, 10, [1500, 5, 6])
+    X = np.random.normal(0, 1, [1500, 25, 6])
 
     test_model = deepof.models.SEQ_2_SEQ_GMVAE()
     gmvaep = test_model.build(X.shape)[3]
@@ -279,6 +279,6 @@ def test_neighbor_cluster_purity():
         X,
         X,
         callbacks=deepof.model_utils.neighbor_cluster_purity(
-            r=0.5, validation_data=X, variational=True
+            encoding_dim=6, validation_data=X, variational=True
         ),
     )

tests/test_train_utils.py

@@ -67,7 +67,6 @@ def test_get_callbacks(
         True,
         True,
         None,
-        entropy_radius=0.5,
     )
     assert type(runID) == str
     assert type(tbc) == tf.keras.callbacks.TensorBoard
@@ -171,7 +170,6 @@ def test_tune_search(
             cp=False,
             reg_cat_clusters=True,
             reg_cluster_variance=True,
-            entropy_radius=0.75,
             entropy_samples=10,
             logparam=None,
         )