Merge pull request #41 from NiklasMelton/test-bayesian-art

Test bayesian art

common/utils.py

@@ -1,4 +1,5 @@
 import numpy as np
+from matplotlib.axes import Axes
 
 def normalize(data: np.ndarray) -> np.ndarray:
     normalized = (data-np.min(data))/(np.max(data)-np.min(data))
@@ -16,3 +17,95 @@ def l2norm2(data: np.ndarray) -> float:
 
 def fuzzy_and(x: np.ndarray, y: np.ndarray) -> np.ndarray:
     return np.minimum(x, y)
+
+def plot_gaussian_contours_fading(
+        ax: Axes,
+        mean: np.ndarray,
+        std_dev: np.ndarray,
+        color: np.ndarray,
+        max_std: int = 2,
+        sigma_steps: float = 0.25,
+        linewidth: int = 1
+):
+    """
+    Plots concentric ellipses to represent the contours of a 2D Gaussian distribution, with fading colors.
+
+    Parameters:
+        - ax: Matplotlib axis object. If None, creates a new figure and axis.
+    - mean: A numpy array representing the mean (μ) of the distribution.
+    - std_dev: A numpy array representing the standard deviation (σ) of the distribution.
+    - color: A 4D numpy array including RGB and alpha channels to specify the color and initial opacity.
+    - max_std: Max standard deviations to draw contours to. Default is 2.
+    - sigma_steps: Step size in standard deviations for each contour. Default is 0.25.
+
+    """
+    from matplotlib.patches import Ellipse
+
+    # Calculate the number of steps
+    steps = int(max_std / sigma_steps)
+    alphas = np.linspace(1, 0.1, steps)
+
+    if len(color) != 4:
+        color = np.concatenate([color, [1.]])
+
+    for i, alpha in zip(range(1, steps + 1), alphas):
+        # Adjust the alpha value of the color
+        current_color = np.copy(color)
+        current_color[-1] = alpha  # Update the alpha channel
+
+        # Width and height of the ellipse are 2*i*sigma_steps times the std_dev values
+        width, height = 2 * i * sigma_steps * std_dev[0], 2 * i * sigma_steps * std_dev[1]
+        ellipse = Ellipse(xy=(mean[0], mean[1]), width=width, height=height, edgecolor=current_color, facecolor='none', linewidth=linewidth,
+                          linestyle='dashed', label=f'{i * sigma_steps}σ')
+        ax.add_patch(ellipse)
+
+
+def plot_gaussian_contours_covariance(
+        ax: Axes,
+        mean: np.ndarray,
+        covariance: np.ndarray,
+        color: np.ndarray,
+        max_std: int = 2,
+        sigma_steps: float = 0.25,
+        linewidth: int = 1
+):
+    """
+    Plots concentric ellipses to represent the contours of a 2D Gaussian distribution, with fading colors.
+    Accepts a covariance matrix to properly represent the distribution's orientation and shape.
+
+
+    Parameters:
+        - ax: Matplotlib axis object. If None, creates a new figure and axis.
+    - mean: A numpy array representing the mean (μ) of the distribution.
+    - covariance: A 2x2 numpy array representing the covariance matrix of the distribution.
+    - color: A 4D numpy array including RGB and alpha channels to specify the color and initial opacity.
+    - max_std: Max standard deviations to draw contours to. Default is 2.
+    - sigma_steps: Step size in standard deviations for each contour. Default is 0.25.
+
+    """
+    from matplotlib.patches import Ellipse
+
+    # Calculate the eigenvalues and eigenvectors of the covariance matrix
+    eigenvalues, eigenvectors = np.linalg.eig(covariance)
+    major_axis = np.sqrt(eigenvalues[0])  # The major axis length (sqrt of larger eigenvalue)
+    minor_axis = np.sqrt(eigenvalues[1])  # The minor axis length (sqrt of smaller eigenvalue)
+    angle = np.arctan2(
+        *eigenvectors[:, 0][::-1])  # Angle in radians between the x-axis and the major axis of the ellipse
+
+    # Calculate the number of steps
+    steps = int(max_std / sigma_steps)
+    alphas = np.linspace(1, 0.1, steps)
+
+    for i, alpha in zip(range(1, steps + 1), alphas):
+        # Adjust the alpha value of the color
+        current_color = np.copy(color)
+        current_color[-1] = alpha  # Update the alpha channel
+
+        # Width and height of the ellipse based on the covariance
+        width, height = 2 * i * sigma_steps * major_axis * 2, 2 * i * sigma_steps * minor_axis * 2
+        ellipse = Ellipse(xy=(mean[0], mean[1]), width=width, height=height, angle=float(np.degrees(angle)),
+                          edgecolor=current_color, facecolor='None', linewidth=linewidth,
+                          linestyle='dashed', label=f'{i * sigma_steps}σ')
+        ax.add_patch(ellipse)
+
+

elementary/BayesianART.py

@@ -4,9 +4,11 @@
 IEEE Transactions on Neural Networks, 18, 1628–1644. doi:10.1109/TNN.2007.900234.
 """
 import numpy as np
-from typing import Optional
+from typing import Optional, Iterable
+from matplotlib.axes import Axes
 from common.BaseART import BaseART
 from common.utils import normalize
+from common.utils import plot_gaussian_contours_covariance
 
 def prepare_data(data: np.ndarray) -> np.ndarray:
     normalized = normalize(data)
@@ -24,7 +26,7 @@ def validate_params(params: dict):
         assert params["rho"] > 0
 
     def check_dimensions(self, X: np.ndarray):
-        if not self.dim_:
+        if not hasattr(self, "dim_"):
             self.dim_ = X.shape[1]
             assert self.params["cov_init"].shape[0] == self.dim_
             assert self.params["cov_init"].shape[1] == self.dim_
@@ -33,43 +35,56 @@ def check_dimensions(self, X: np.ndarray):
 
     def category_choice(self, i: np.ndarray, w: np.ndarray, params: dict) -> tuple[float, Optional[dict]]:
         mean = w[:self.dim_]
-        cov = w[self.dim_:self.dim_*self.dim_].reshape((self.dim_, self.dim_))
+        cov = w[self.dim_:-1].reshape((self.dim_, self.dim_))
         n = w[-1]
         dist = mean - i
-        exp_dist_cov_dist = np.exp(-0.5 * np.matmul(dist.T, np.matmul((1 / cov), dist)))
+        exp_dist_cov_dist = np.exp(-0.5 * np.matmul(dist.T, np.matmul(np.linalg.inv(cov), dist)))
+        det_cov = np.linalg.det(cov)
         cache = {
-            "exp_dist_cov_dist": exp_dist_cov_dist,
-            "cov": cov
+            "cov": cov,
+            "det_cov": det_cov
         }
-        p_i_cj = exp_dist_cov_dist / np.sqrt((self.pi2 ** self.dim_) * np.linalg.det(cov))
+        p_i_cj = exp_dist_cov_dist / np.sqrt((self.pi2 ** self.dim_) * det_cov)
         p_cj = n / np.sum(w_[-1] for w_ in self.W)
 
-        return p_i_cj * p_cj, cache
+        activation = p_i_cj * p_cj
+
+        return activation, cache
 
     def match_criterion(self, i: np.ndarray, w: np.ndarray, params: dict, cache: Optional[dict] = None) -> float:
         if cache is None:
             raise ValueError("No cache provided")
-        return cache["cov"]
+        # return cache["det_cov"]
+        return np.prod(np.diag(cache["cov"]))
 
     def match_criterion_bin(self, i: np.ndarray, w: np.ndarray, params: dict, cache: Optional[dict] = None) -> bool:
-        return self.match_criterion(i, w, params=params, cache=cache) >= params["rho"]
+        return self.match_criterion(i, w, params=params, cache=cache) <= params["rho"]
 
     def update(self, i: np.ndarray, w: np.ndarray, params, cache: Optional[dict] = None) -> np.ndarray:
         if cache is None:
             raise ValueError("No cache provided")
 
         mean = w[:self.dim_]
-        cov = cache["cov"]
+        cov = w[self.dim_:-1].reshape((self.dim_, self.dim_))
         n = w[-1]
 
         n_new = n+1
         mean_new = (1-(1/n_new))*mean + (1/n_new)*i
-        cov_new = (n/n_new)*cov + (1/n_new)*np.multiply(
-            ((i-mean_new).reshape((-1, 1))*(i-mean_new).reshape((1, -1))).T,
-            np.identity(self.dim_)
-        )
+
+        i_mean_dist = i-mean_new
+        i_mean_dist_2 = i_mean_dist.reshape((-1, 1))*i_mean_dist.reshape((1, -1))
+
+        cov_new = (n / n_new) * cov + (1 / n_new) * i_mean_dist_2
 
         return np.concatenate([mean_new, cov_new.flatten(), [n_new]])
 
     def new_weight(self, i: np.ndarray, params: dict) -> np.ndarray:
-        return np.concatenate[i, params["cov_init"].flatten(), [1]]
+        return np.concatenate([i, params["cov_init"].flatten(), [1]])
+
+
+    def plot_cluster_bounds(self, ax: Axes, colors: Iterable, linewidth: int = 1):
+        for w, col in zip(self.W, colors):
+            mean = w[:self.dim_]
+            cov = w[self.dim_:-1].reshape((self.dim_, self.dim_))
+            # sigma = np.sqrt(np.diag(cov))
+            plot_gaussian_contours_covariance(ax, mean, cov, col, linewidth=linewidth)

elementary/GaussianART.py

@@ -8,48 +8,7 @@
 from typing import Optional, Iterable
 from matplotlib.axes import Axes
 from common.BaseART import BaseART
-
-
-def plot_gaussian_contours_fading(
-        ax: Axes,
-        mean: np.ndarray,
-        std_dev: np.ndarray,
-        color: np.ndarray,
-        max_std: int = 2,
-        sigma_steps: float = 0.25,
-        linewidth: int = 1
-):
-    """
-    Plots concentric ellipses to represent the contours of a 2D Gaussian distribution, with fading colors.
-
-    Parameters:
-        - ax: Matplotlib axis object. If None, creates a new figure and axis.
-    - mean: A numpy array representing the mean (μ) of the distribution.
-    - std_dev: A numpy array representing the standard deviation (σ) of the distribution.
-    - color: A 4D numpy array including RGB and alpha channels to specify the color and initial opacity.
-    - max_std: Max standard deviations to draw contours to. Default is 2.
-    - sigma_steps: Step size in standard deviations for each contour. Default is 0.25.
-
-    """
-    from matplotlib.patches import Ellipse
-
-    # Calculate the number of steps
-    steps = int(max_std / sigma_steps)
-    alphas = np.linspace(1, 0.1, steps)
-
-    if len(color) != 4:
-        color = np.concatenate([color, [1.]])
-
-    for i, alpha in zip(range(1, steps + 1), alphas):
-        # Adjust the alpha value of the color
-        current_color = np.copy(color)
-        current_color[-1] = alpha  # Update the alpha channel
-
-        # Width and height of the ellipse are 2*i*sigma_steps times the std_dev values
-        width, height = 2 * i * sigma_steps * std_dev[0], 2 * i * sigma_steps * std_dev[1]
-        ellipse = Ellipse(xy=(mean[0], mean[1]), width=width, height=height, edgecolor=current_color, facecolor='none', linewidth=linewidth,
-                          linestyle='dashed', label=f'{i * sigma_steps}σ')
-        ax.add_patch(ellipse)
+from common.utils import plot_gaussian_contours_fading
 
 
 class GaussianART(BaseART):

examples/test_bayesian_art.py

@@ -0,0 +1,39 @@
+from sklearn.datasets import make_blobs
+import matplotlib.pyplot as plt
+import path
+import sys
+
+# directory reach
+directory = path.Path(__file__).abspath()
+
+print(directory.parent)
+# setting path
+sys.path.append(directory.parent.parent)
+
+from elementary.BayesianART import BayesianART
+from common.utils import normalize
+import numpy as np
+
+
+def cluster_blobs():
+    data, target = make_blobs(n_samples=150, centers=3, cluster_std=0.50, random_state=0, shuffle=False)
+    print("Data has shape:", data.shape)
+
+    X = normalize(data)
+    print("Prepared data has shape:", X.shape)
+
+    params = {
+        "rho": 0.00002,
+        "cov_init": np.array([[0.0001, 0.0], [0.0, 0.0001]]),
+    }
+    cls = BayesianART(params)
+    y = cls.fit_predict(X)
+
+    print(f"{cls.n_clusters} clusters found")
+
+    cls.visualize(X, y)
+    plt.show()
+
+
+if __name__ == "__main__":
+    cluster_blobs()