Improvements: debug lcu and add new metrics (#38)

README.md

@@ -9,6 +9,15 @@ It's a Python implementation of a image segmentation algorithm that
 combines superpixel with complex networks dynamics. In this setup, we
 classify the algorithm as transductive as well.
 
+# Showcase on grabcut dataset
+
+![showcase](pics/egsis-showcase.png)
+
+First segmentation mask is the result of EGSIS segmentation over lasso
+annotation from GrabCut dataset, second segmentation mask is the
+ground truth.
+
+
 # What is transductive segmentation?
 
 Transductive segmentation is a concept in machine learning and

egsis/lcu.py

@@ -6,35 +6,14 @@
 
 from loguru import logger
 
-"""lerax - sex 09 dez 2022 13:55:34
-Alguma coisa não está certa... as subnetworks não são disjuntas, a
-função g que gera novas particulas quase sempre está zerada... a
-matriz delta só fica com elementos na diagonal... tem algo errado aqui
-
-lerax - sáb 17 dez 2022 02:47:16
-
-óh céus!!! A função de evolução não tá mudando em nada depois das
-primeiras iterações! Será que existe mais algum bug na implementação
-das equações!? Será que eu deveria fazer o que o verri falou e usar o
-paper mais novo?! Deixei em research/lcu_simplified_improved.pdf
-
-esse documento deve ajudar
-
-lerax - qua 21 dez 2022 10:10:58
-
-Parece que a evolução de nc está incorreta, inverti a multiplicação de
-matrizes, pois geralmente é matrix x vetor, não vetor x matrix...
-Os resultados começaram a ter maior variação, mas parecem estar invertidos..
-"""
-
 
 class LabeledComponentUnfolding:
 
     """
     Collective Dynamic Labeled Component Unfolding
 
 
-    It can be used to solve Semi-Supervised problems.
+    It can be used to solve Transductive Semi-Supervised problems.
 
 
     Parameters
@@ -184,11 +163,13 @@ def p(self, G: nx.graph, i: int, j: int, c: int) -> float:
         edge_weight = G.edges[i, j]["weight"]
         walk = edge_weight / G.degree[i]
         survival = 1 - self.competition_level * self.sigma(G, i, j, c)
-
+        # FIXME: weirdly, sigma function only outputs 1, 0.5  or 0
+        # it doesn't makes sense
+        # logger.trace(f"survival factor = {survival}")
         return walk * survival
 
     def probability(self, G: nx.Graph) -> np.ndarray:
-        """Matrix with probabilities of particle survival"""
+        """Matrix with probabilities of particle surviving"""
         P = np.zeros(shape=self.N.shape)
         C, nodes, _ = P.shape
         for c in range(C):
@@ -198,6 +179,8 @@ def probability(self, G: nx.Graph) -> np.ndarray:
         return P
 
     def probability_of_new_particles(self, G: nx.Graph, c: int) -> np.ndarray:
+        # FIXME: review this code based on the paper
+        # equation on page 4
         node_degrees = [
             G.degree[node] for node in G.nodes
         ]
@@ -222,7 +205,7 @@ def n0(self, G: nx.Graph) -> np.ndarray:
             if label != 0:
                 labels[cls, idx] = label
         particles = labels * population * self.scale_particles
-        logger.debug(f"n0: {particles}")
+        logger.debug(f"n0: \n{particles}")
         return particles
 
     def N0(self, G: nx.Graph):
@@ -241,17 +224,18 @@ def delta0(self, G: nx.Graph):
         return np.zeros(shape=(self.n_classes, nodes, nodes))
 
     def sigma(self, G: nx.Graph, i: int, j: int, c: int) -> float:
-        """Matrix with current relative domination sigma_ij[c]
+        """Current relative subordination sigma_ij[c]
 
-        Number of particles with label=c which moved from v_i to
-        v_j in the current time.
+        The fraction of particles that do not belong to class c and
+        have sucessfuly passed thorugh edge (i, j) in any direction at
+        the current time.
         """
-        S = np.sum((self.N[:, i, j] + self.N[:, j, i].T).flatten())
+        S = np.sum((self.N[:, i, j] + self.N[:, j, i]).flatten())
         result: float
-        if S <= 0:
-            result = 1 - (1 / self.n_classes)
-        else:
+        if S > 0:
             result = 1 - ((self.N[c][i][j] + self.N[c][j][i]) / S)
+        else:
+            result = 1 - (1 / self.n_classes)
 
         return result
 

egsis/metrics.py

@@ -30,7 +30,42 @@ def f1(y_true: np.ndarray, y_pred: np.ndarray) -> float:
 
 
 def err(y_true: np.ndarray, y_pred: np.ndarray) -> float:
-    """Error rate (mis-segmentation rate)"""
-    roi_pixels = y_true
-    mis_segmentation = (y_true | y_pred) - y_true
-    return mis_segmentation.sum() / roi_pixels.sum()
+    """This function calculates the error rate, also known as the
+     mis-segmentation rate, between the true and predicted values.
+
+    Parameters
+    ---------
+    y_true : np.ndarray
+        The ground truth binary labels. The binary label indicates
+        whether each pixel is within the region of interest (ROI) or
+        not.
+    y_pred : np.ndarray:
+        The predicted binary labels. The binary label indicates
+        whether each pixel is predicted to be within the ROI or not.
+
+    Returns
+    -------
+    err : float
+
+    The mis-segmentation rate. This is calculated as the number of
+    mis-segmented pixels (false-positives and false-negatives) divided
+    by the total number of pixels in the ROI.
+    """
+    union = (y_true | y_pred)
+    intersection = (y_true & y_pred)
+    mis_segmentation = (union - intersection).sum()
+    w, h = y_true.shape
+    roi = w * h
+    return mis_segmentation / roi
+
+
+def recall(y_true: np.ndarray, y_pred: np.ndarray) -> float:
+    tp = (y_true & y_pred).sum()
+    fn = ((y_true | y_pred) - y_pred).sum()
+    return tp / (tp + fn)
+
+
+def precision(y_true: np.ndarray, y_pred: np.ndarray) -> float:
+    tp = (y_true & y_pred).sum()
+    fp = ((y_true | y_pred) - y_true).sum()
+    return tp / (tp + fp)

egsis/model.py

@@ -107,16 +107,19 @@ def build_complex_network(
             segments=segments,
             labels=y
         )
+        logger.info("Complex networks: compute node labels finished.")
         complex_networks.compute_node_features(
             graph=G,
             img=X,
             segments=segments,
             feature_method=self.feature_extraction
         )
+        logger.info("Complex networks: feature extraction finished.")
         complex_networks.compute_edge_weights(
             graph=G,
             similarity_function=self.feature_similarity
         )
+        logger.info("Complex networks: compute node weights finished.")
         return G
 
     def sub_networks_to_matrix(self, sub_networks: List[networkx.Graph]):
@@ -134,15 +137,17 @@ def fit_predict(self, X: numpy.ndarray, y: numpy.ndarray):
              luminosity of each color channel of RGB
         y : numpy.ndarray (shape=(n, m))
              it's the label matrix with partial annotation, to be full
-             filled Every non-zero value it's a label, and zero it's
+             filled every non-zero value it's a label, and zero it's
              an unlabeled pixel.
         Returns
         -------
         new y matrix with full filled labels.
         """
         logger.info("Run!")
         self.segments = self.build_superpixels(X)
+        logger.info("Superpixels: finished.")
         self.G = self.build_complex_network(X, y, self.segments)
+        logger.info("Complex networks: finished.")
         n_classes = len(numpy.unique(y)) - 1
         collective_dynamic = lcu.LabeledComponentUnfolding(
             competition_level=self.lcu_competition_level,
@@ -152,6 +157,7 @@ def fit_predict(self, X: numpy.ndarray, y: numpy.ndarray):
 
         self.sub_networks = collective_dynamic.fit_predict(self.G)
         self.G_pred = collective_dynamic.classify_vertexes(self.sub_networks)
+        logger.info("Dynamic collective LCU: finished.")
 
         # FIXME: should return a matrix y with new labels
         return self.G_pred

pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "egsis"
-version = "0.1.0a0"
+version = "0.1.0"
 description = "Exploratory Graph-based Semi-Supervised Image Segmentation"
 authors = [
     {name = "Manoel Vilela", email = "[email protected]"},