Merge pull request #92 from NiklasMelton/match-tracking-update

Match tracking update; improve Gaussian ART; add verbose fit

artlib/common/BaseART.py

@@ -340,7 +340,7 @@ def step_fit(self, x: np.ndarray, match_reset_func: Optional[Callable] = None, m
             return 0
         else:
 
-            if match_reset_method == "MY~" and match_reset_func is not None:
+            if match_reset_method in ["MT~"] and match_reset_func is not None:
                 T_values, T_cache = zip(*[
                     self.category_choice(x, w, params=self.params)
                     if match_reset_func(x, w, c_, params=self.params, cache=None)
@@ -355,17 +355,20 @@ def step_fit(self, x: np.ndarray, match_reset_func: Optional[Callable] = None, m
                 w = self.W[c_]
                 cache = T_cache[c_]
                 m, cache = self.match_criterion_bin(x, w, params=self.params, cache=cache, op=mt_operator)
-                no_match_reset = (
-                        match_reset_func is None or
-                        match_reset_func(x, w, c_, params=self.params, cache=cache)
-                )
+                if match_reset_method in ["MT~"] and match_reset_func is not None:
+                    no_match_reset = True
+                else:
+                    no_match_reset = (
+                            match_reset_func is None or
+                            match_reset_func(x, w, c_, params=self.params, cache=cache)
+                    )
                 if m and no_match_reset:
                     self.set_weight(c_, self.update(x, w, self.params, cache=cache))
                     self._set_params(base_params)
                     return c_
                 else:
                     T[c_] = np.nan
-                    if not no_match_reset:
+                    if m and not no_match_reset:
                         keep_searching = self._match_tracking(cache, epsilon, self.params, match_reset_method)
                         if not keep_searching:
                             T[:] = np.nan
@@ -427,7 +430,7 @@ def post_fit(self, X: np.ndarray):
         pass
 
 
-    def fit(self, X: np.ndarray, y: Optional[np.ndarray] = None, match_reset_func: Optional[Callable] = None, max_iter=1, match_reset_method:Literal["MT+", "MT-", "MT0", "MT1", "MT~"] = "MT+", epsilon: float = 0.0):
+    def fit(self, X: np.ndarray, y: Optional[np.ndarray] = None, match_reset_func: Optional[Callable] = None, max_iter=1, match_reset_method:Literal["MT+", "MT-", "MT0", "MT1", "MT~"] = "MT+", epsilon: float = 0.0, verbose: bool = False):
         """
         Fit the model to the data
 
@@ -453,7 +456,12 @@ def fit(self, X: np.ndarray, y: Optional[np.ndarray] = None, match_reset_func: O
         self.W: list[np.ndarray] = []
         self.labels_ = np.zeros((X.shape[0], ), dtype=int)
         for _ in range(max_iter):
-            for i, x in enumerate(X):
+            if verbose:
+                from tqdm import tqdm
+                x_iter = tqdm(enumerate(X), total=int(X.shape[0]))
+            else:
+                x_iter = enumerate(X)
+            for i, x in x_iter:
                 self.pre_step_fit(X)
                 c = self.step_fit(x, match_reset_func=match_reset_func, match_reset_method=match_reset_method, epsilon=epsilon)
                 self.labels_[i] = c

artlib/elementary/GaussianART.py

@@ -5,6 +5,7 @@
 """
 
 import numpy as np
+from decimal import Decimal
 from typing import Optional, Iterable, List
 from matplotlib.axes import Axes
 from artlib.common.BaseART import BaseART
@@ -13,20 +14,22 @@
 
 class GaussianART(BaseART):
     # implementation of GaussianART
-    pi2 = np.pi*2
-    def __init__(self, rho: float, sigma_init: np.ndarray):
+    def __init__(self, rho: float, sigma_init: np.ndarray, alpha: float = 1e-10):
         """
         Parameters:
         - rho: vigilance parameter
         - sigma_init: initial estimate of the diagonal std
+        - alpha: used to prevent division by zero errors
 
         """
         params = {
             "rho": rho,
             "sigma_init": sigma_init,
+            "alpha": alpha
         }
         super().__init__(params)
 
+
     @staticmethod
     def validate_params(params: dict):
         """
@@ -38,7 +41,9 @@ def validate_params(params: dict):
         """
         assert "rho" in params
         assert "sigma_init" in params
-        assert 1.0 >= params["rho"] > 0.
+        assert "alpha" in params
+        assert 1.0 >= params["rho"] >= 0.
+        assert params["alpha"] > 0.
         assert isinstance(params["rho"], float)
         assert isinstance(params["sigma_init"], np.ndarray)
 
@@ -57,15 +62,19 @@ def category_choice(self, i: np.ndarray, w: np.ndarray, params: dict) -> tuple[f
 
         """
         mean = w[:self.dim_]
-        sigma = w[self.dim_:-1]
+        # sigma = w[self.dim_:2*self.dim]
+        inv_sig = w[2*self.dim_:3*self.dim_]
+        sqrt_det_sig = w[-2]
         n = w[-1]
-        sig = np.diag(np.multiply(sigma,sigma))
+
         dist = mean-i
-        exp_dist_sig_dist = np.exp(-0.5*np.matmul(dist.T, np.matmul(np.linalg.inv(sig), dist)))
+        exp_dist_sig_dist = np.exp(-0.5 * np.dot(dist, np.multiply(inv_sig, dist)))
+
         cache = {
             "exp_dist_sig_dist": exp_dist_sig_dist
         }
-        p_i_cj = exp_dist_sig_dist/np.sqrt((self.pi2**self.dim_)*np.linalg.det(sig))
+        # ignore the (2*pi)^d term as that is constant
+        p_i_cj = exp_dist_sig_dist/(params["alpha"]+sqrt_det_sig)
         p_cj = n/np.sum(w_[-1] for w_ in self.W)
 
         activation = p_i_cj*p_cj
@@ -109,14 +118,18 @@ def update(self, i: np.ndarray, w: np.ndarray, params: dict, cache: Optional[dic
 
         """
         mean = w[:self.dim_]
-        sigma = w[self.dim_:-1]
+        sigma = w[self.dim_:2*self.dim_]
         n = w[-1]
 
         n_new = n+1
         mean_new = (1-(1/n_new))*mean + (1/n_new)*i
         sigma_new = np.sqrt((1-(1/n_new))*np.multiply(sigma, sigma) + (1/n_new)*((mean_new - i)**2))
 
-        return np.concatenate([mean_new, sigma_new, [n_new]])
+        sigma2 = np.multiply(sigma_new, sigma_new)
+        inv_sig = 1 / sigma2
+        det_sig = np.sqrt(np.prod(sigma2))
+
+        return np.concatenate([mean_new, sigma_new, inv_sig, [det_sig], [n_new]])
 
 
     def new_weight(self, i: np.ndarray, params: dict) -> np.ndarray:
@@ -132,7 +145,10 @@ def new_weight(self, i: np.ndarray, params: dict) -> np.ndarray:
             updated cluster weight
 
         """
-        return np.concatenate([i, params["sigma_init"], [1.]])
+        sigma2 = np.multiply(params["sigma_init"], params["sigma_init"])
+        inv_sig_init = 1 / sigma2
+        det_sig_init = np.sqrt(np.prod(sigma2))
+        return np.concatenate([i, params["sigma_init"], inv_sig_init, [det_sig_init], [1.]])
 
     def get_cluster_centers(self) -> List[np.ndarray]:
         """

artlib/supervised/SimpleARTMAP.py

@@ -129,7 +129,7 @@ def step_fit(self, x: np.ndarray, c_b: int, match_reset_method: Literal["MT+", "
             assert self.map[c_a] == c_b
         return c_a
 
-    def fit(self, X: np.ndarray, y: np.ndarray, max_iter=1, match_reset_method: Literal["MT+", "MT-", "MT0", "MT1", "MT~"] = "MT+", epsilon: float = 1e-10):
+    def fit(self, X: np.ndarray, y: np.ndarray, max_iter=1, match_reset_method: Literal["MT+", "MT-", "MT0", "MT1", "MT~"] = "MT+", epsilon: float = 1e-10, verbose: bool = False):
         """
         Fit the model to the data
 
@@ -155,7 +155,12 @@ def fit(self, X: np.ndarray, y: np.ndarray, max_iter=1, match_reset_method: Lite
         self.module_a.labels_ = np.zeros((X.shape[0],), dtype=int)
 
         for _ in range(max_iter):
-            for i, (x, c_b) in enumerate(zip(X, y)):
+            if verbose:
+                from tqdm import tqdm
+                x_y_iter = tqdm(enumerate(zip(X, y)), total=int(X.shape[0]))
+            else:
+                x_y_iter = enumerate(zip(X, y))
+            for i, (x, c_b) in x_y_iter:
                 self.module_a.pre_step_fit(X)
                 c_a = self.step_fit(x, c_b, match_reset_method=match_reset_method, epsilon=epsilon)
                 self.module_a.labels_[i] = c_a