oikan 0.0.3.7__py3-none-any.whl → 0.0.3.9__py3-none-any.whl

oikan/__init__.py

@@ -9,6 +9,8 @@ Docs: https://silvermete0r.github.io/oikan/
 '''
 
 from .model import OIKAN, OIKANClassifier, OIKANRegressor
+from .neural import TabularNet
+from .elasticnet import ElasticNet
 
-__all__ = ['OIKAN', 'OIKANClassifier', 'OIKANRegressor']
+__all__ = ['OIKAN', 'OIKANClassifier', 'OIKANRegressor', 'TabularNet', 'ElasticNet']
 __version__ = '0.0.3'

oikan/elasticnet.py

@@ -0,0 +1,71 @@
+import torch.nn as nn
+import torch
+import numpy as np
+
+class ElasticNet(nn.Module):
+    def __init__(self, alpha=1.0, l1_ratio=0.5, fit_intercept=False, max_iter=1000, tol=1e-4, random_state=None):
+        super().__init__()
+        self.alpha = alpha
+        self.l1_ratio = l1_ratio
+        self.fit_intercept = fit_intercept
+        self.max_iter = max_iter
+        self.tol = tol
+        self.random_state = random_state
+        self.coef_ = None
+        self.intercept_ = None
+
+    def fit(self, X, y):
+        X = np.asarray(X, dtype=np.float32)
+        y = np.asarray(y, dtype=np.float32)
+        n_samples, n_features = X.shape
+        if y.ndim == 1:
+            y = y.reshape(-1, 1)
+        n_targets = y.shape[1]
+
+        if self.random_state is not None:
+            torch.manual_seed(self.random_state)
+            np.random.seed(self.random_state)
+
+        X_tensor = torch.tensor(X, dtype=torch.float32)
+        y_tensor = torch.tensor(y, dtype=torch.float32)
+
+        W = torch.zeros((n_features, n_targets), requires_grad=True, dtype=torch.float32)
+        if self.fit_intercept:
+            b = torch.zeros(n_targets, requires_grad=True, dtype=torch.float32)
+        else:
+            b = None
+
+        optimizer = torch.optim.Adam([W] + ([b] if b is not None else []), lr=0.05)
+
+        prev_loss = None
+        for _ in range(self.max_iter):
+            optimizer.zero_grad()
+            pred = X_tensor @ W
+            if b is not None:
+                pred = pred + b
+            mse = torch.mean((pred - y_tensor) ** 2)
+            l1 = torch.sum(torch.abs(W))
+            l2 = torch.sum(W ** 2)
+            loss = mse + self.alpha * (self.l1_ratio * l1 + (1 - self.l1_ratio) * l2)
+            loss.backward()
+            optimizer.step()
+            if prev_loss is not None and abs(prev_loss - loss.item()) < self.tol:
+                break
+            prev_loss = loss.item()
+
+        self.coef_ = W.detach().cpu().numpy().T if n_targets > 1 else W.detach().cpu().numpy().flatten()
+        if b is not None:
+            self.intercept_ = b.detach().cpu().numpy()
+        else:
+            self.intercept_ = np.zeros(n_targets) if n_targets > 1 else 0.0
+        return self
+
+    def predict(self, X):
+        X = np.asarray(X, dtype=np.float32)
+        if self.coef_ is None:
+            raise RuntimeError("Model not fitted yet.")
+        W = self.coef_.T if self.coef_.ndim == 2 else self.coef_
+        y_pred = X @ W
+        if self.intercept_ is not None:
+            y_pred += self.intercept_
+        return y_pred

oikan/model.py

@@ -3,15 +3,16 @@ import torch
 import torch.nn as nn
 import torch.optim as optim
 from sklearn.preprocessing import PolynomialFeatures
-from sklearn.linear_model import ElasticNet
 from abc import ABC, abstractmethod
 import json
+from .elasticnet import ElasticNet
 from .neural import TabularNet
 from .utils import evaluate_basis_functions, get_features_involved, sympify_formula, get_latex_formula
 from sklearn.model_selection import train_test_split
 from sklearn.metrics import r2_score, accuracy_score
 from .exceptions import *
 import sys
+from tqdm import tqdm
 
 class OIKAN(ABC):
     """
@@ -25,9 +26,12 @@ class OIKAN(ABC):
         Activation function for the neural network ('relu', 'tanh', 'leaky_relu', 'elu', 'swish', 'gelu').
     augmentation_factor : int, optional (default=10)
         Number of augmented samples per original sample.
-    alpha : float, optional (default=0.1)
-        L1 regularization strength for Lasso in symbolic regression.
-    sigma : float, optional (default=0.1)
+    alpha : float, optional (default=1.0)
+        ElasticNet regularization strength.
+    l1_ratio: float, optional (default=0.5)
+        ElasticNet mixing parameter (0 <= l1_ratio <= 1).
+        0 is equivalent to Ridge regression, 1 is equivalent to Lasso.
+    sigma : float, optional (default=5.0)
         Standard deviation of Gaussian noise for data augmentation.
     top_k : int, optional (default=5)
         Number of top features to select in hierarchical symbolic regression.
@@ -45,7 +49,7 @@ class OIKAN(ABC):
         Random seed for reproducibility.
     """
     def __init__(self, hidden_sizes=[64, 64], activation='relu', augmentation_factor=10, 
-                 alpha=0.1, sigma=0.1, epochs=100, lr=0.001, batch_size=32, 
+                 alpha=1.0, l1_ratio=0.5, sigma=5.0, epochs=100, lr=0.001, batch_size=32, 
                  verbose=False, evaluate_nn=False, top_k=5, random_state=None):
         if not isinstance(hidden_sizes, list) or not all(isinstance(x, int) and x > 0 for x in hidden_sizes):
             raise InvalidParameterError("hidden_sizes must be a list of positive integers")
@@ -63,6 +67,8 @@ class OIKAN(ABC):
             raise InvalidParameterError("epochs must be a positive integer")
         if not 0 <= alpha <= 1:
             raise InvalidParameterError("alpha must be between 0 and 1")
+        if not 0 <= l1_ratio <= 1:
+            raise InvalidParameterError("l1_ratio must be between 0 and 1")
         if sigma <= 0:
             raise InvalidParameterError("sigma must be positive")
         
@@ -70,6 +76,7 @@ class OIKAN(ABC):
         self.activation = activation
         self.augmentation_factor = augmentation_factor
         self.alpha = alpha
+        self.l1_ratio = l1_ratio
         self.sigma = sigma
         self.epochs = epochs
         self.lr = lr
@@ -353,14 +360,23 @@ class OIKAN(ABC):
         if np.any(np.isinf(X)) or np.any(np.isinf(y)):
             raise NumericalInstabilityError("Input data contains infinite values")
 
-        # Stage 1: Coarse Model
+        if self.verbose:
+            print("\nStage 1: Coarse Model Fitting")
+            
         coarse_degree = 2  # Fixed low degree for coarse model
         poly_coarse = PolynomialFeatures(degree=coarse_degree, include_bias=True)
+        
+        if self.verbose:
+            print("Generating polynomial features...")
         X_poly_coarse = poly_coarse.fit_transform(X)
-        model_coarse = ElasticNet(alpha=self.alpha, fit_intercept=False)
+        
+        if self.verbose:
+            print("Fitting coarse elastic net model...")
+        model_coarse = ElasticNet(alpha=self.alpha, l1_ratio=self.l1_ratio, fit_intercept=False, random_state=self.random_state)
         model_coarse.fit(X_poly_coarse, y)
 
-        # Compute feature importances for original features
+        if self.verbose:
+            print("Computing feature importances...")
         basis_functions_coarse = poly_coarse.get_feature_names_out()
         if len(y.shape) == 1 or y.shape[1] == 1:
             coef_coarse = model_coarse.coef_.flatten()
@@ -368,7 +384,7 @@ class OIKAN(ABC):
             coef_coarse = np.sum(np.abs(model_coarse.coef_), axis=0)
 
         importances = np.zeros(X.shape[1])
-        for i, func in enumerate(basis_functions_coarse):
+        for i, func in enumerate(tqdm(basis_functions_coarse, disable=not self.verbose, desc="Analyzing features")):
             features_involved = get_features_involved(func)
             for idx in features_involved:
                 importances[idx] += np.abs(coef_coarse[i])
@@ -379,11 +395,13 @@ class OIKAN(ABC):
         # Select top K features
         top_k_indices = np.argsort(importances)[::-1][:self.top_k]
 
-        # Stage 2: Refined Model
-        # ~ generate additional non-linear features for top K features
+        if self.verbose:
+            print(f"\nStage 2: Refined Model with top {self.top_k} features")
+            print("Generating additional non-linear features...")
+
         additional_features = []
         additional_names = []
-        for i in top_k_indices:
+        for i in tqdm(top_k_indices, disable=not self.verbose, desc="Generating features"):
             # Higher-degree polynomial
             additional_features.append(X[:, i]**3)
             additional_names.append(f'x{i}^3')
@@ -395,15 +413,18 @@ class OIKAN(ABC):
             additional_features.append(np.sin(X[:, i]))
             additional_names.append(f'sin_x{i}')
         
-        # Combine features
+        if self.verbose:
+            print("Combining features and fitting final model...")
         X_additional = np.column_stack(additional_features)
         X_refined = np.hstack([X_poly_coarse, X_additional])
         basis_functions_refined = list(basis_functions_coarse) + additional_names
 
-        # Fit refined model
-        model_refined = ElasticNet(alpha=self.alpha, fit_intercept=False)
+        model_refined = ElasticNet(alpha=self.alpha, l1_ratio=self.l1_ratio, fit_intercept=False, random_state=self.random_state)
         model_refined.fit(X_refined, y)
 
+        if self.verbose:
+            print("Building final symbolic model...")
+
         # Store symbolic model
         if len(y.shape) == 1 or y.shape[1] == 1:
             # Regression
@@ -418,7 +439,7 @@ class OIKAN(ABC):
             # Classification
             coefficients_list = []
             selected_indices = set()
-            for c in range(y.shape[1]):
+            for c in tqdm(range(y.shape[1]), disable=not self.verbose, desc="Processing classes"):
                 coef = model_refined.coef_[c]
                 indices = np.where(np.abs(coef) > 1e-6)[0]
                 selected_indices.update(indices)
@@ -454,7 +475,7 @@ class OIKANRegressor(OIKAN):
             self._train_neural_net(X, y, output_size=1, loss_fn=nn.MSELoss())
             
             if self.verbose:
-                print(f"Original data: features shape: {X.shape} | target shape: {y.shape}")
+                print(f"Original data: features shape: {X.shape} | target shape: {y.shape} | size: {X.nbytes / (1024 * 1024):.2f} MB")
             
             X_aug = self._generate_augmented_data(X)
             
@@ -463,13 +484,14 @@ class OIKANRegressor(OIKAN):
                 y_aug = self.neural_net(torch.tensor(X_aug, dtype=torch.float32)).detach().numpy()
             
             if self.verbose:
-                print(f"Augmented data: features shape: {X_aug.shape} | target shape: {y_aug.shape}")
+                print(f"Augmented data: features shape: {X_aug.shape} | target shape: {y_aug.shape} | size: {X_aug.nbytes / (1024 * 1024):.2f} MB")
             
             X_combined = np.vstack([X, X_aug])
             y_combined = np.vstack([y, y_aug])
         else:
             if self.verbose:
                 print("Skipping neural network training (augmentation_factor=1)")
+                print(f"Data: features shape: {X.shape} | target shape: {y.shape} | size: {X.nbytes / (1024 * 1024):.2f} MB")
             X_combined = X
             y_combined = y
             
@@ -523,7 +545,7 @@ class OIKANClassifier(OIKAN):
             self._train_neural_net(X, y_onehot, output_size=n_classes, loss_fn=nn.CrossEntropyLoss())
             
             if self.verbose:
-                print(f"Original data: features shape: {X.shape} | target shape: {y.shape}")
+                print(f"Original data: features shape: {X.shape} | target shape: {y.shape} | size: {X.nbytes / (1024 * 1024):.2f} MB")
             
             X_aug = self._generate_augmented_data(X)
             
@@ -532,13 +554,14 @@ class OIKANClassifier(OIKAN):
                 logits_aug = self.neural_net(torch.tensor(X_aug, dtype=torch.float32)).detach().numpy()
             
             if self.verbose:
-                print(f"Augmented data: features shape: {X_aug.shape} | target shape: {logits_aug.shape}")
+                print(f"Augmented data: features shape: {X_aug.shape} | target shape: {logits_aug.shape} | size: {X_aug.nbytes / (1024 * 1024):.2f} MB")
             
             X_combined = np.vstack([X, X_aug])
             y_combined = np.vstack([y_onehot.numpy(), logits_aug])
         else:
             if self.verbose:
                 print("Skipping neural network training (augmentation_factor=1)")
+                print(f"Data: features shape: {X.shape} | target shape: {y.shape} | size: {X.nbytes / (1024 * 1024):.2f} MB")
             X_combined = X
             y_combined = y_onehot.numpy()
             

{oikan-0.0.3.7.dist-info → oikan-0.0.3.9.dist-info}/METADATA

@@ -1,9 +1,11 @@
 Metadata-Version: 2.4
 Name: oikan
-Version: 0.0.3.7
+Version: 0.0.3.9
 Summary: OIKAN: Neuro-Symbolic ML for Scientific Discovery
 Author: Arman Zhalgasbayev
 License: MIT
+Project-URL: Homepage, https://github.com/silvermete0r/oikan
+Project-URL: Bug Tracker, https://github.com/silvermete0r/oikan/issues
 Classifier: Programming Language :: Python :: 3
 Classifier: License :: OSI Approved :: MIT License
 Classifier: Operating System :: OS Independent
@@ -49,19 +51,9 @@ OIKAN is a neuro-symbolic machine learning framework inspired by Kolmogorov-Arno
 - 🔬 **Research-Focused**: Designed for academic exploration and experimentation
 - 📈 **Multi-Task**: Supports both regression and classification problems
 
-## Scientific Foundation
+## Key Aspects
 
-OIKAN implements a modern interpretation of the Kolmogorov-Arnold Representation Theorem through a hybrid neural architecture:
-
-1. **Theoretical Foundation**: The Kolmogorov-Arnold theorem states that any continuous n-dimensional function can be decomposed into a combination of single-variable functions:
-
-   ```
-   f(x₁,...,xₙ) = ∑(j=0 to 2n){ φⱼ( ∑(i=1 to n) ψᵢⱼ(xᵢ) ) }
-   ```
-
-   where φⱼ and ψᵢⱼ are continuous univariate functions.
-
-2. **Neural Implementation**: OIKAN uses a specialized architecture combining:
+1. **Neural Implementation**: OIKAN uses a specialized architecture combining:
    - Feature transformation layers with interpretable basis functions
    - Symbolic regression for formula extraction (ElasticNet-based)
    - Automatic pruning of insignificant terms
@@ -78,7 +70,7 @@ OIKAN implements a modern interpretation of the Kolmogorov-Arnold Representation
             self.symbolic_regression = SymbolicRegression(alpha=alpha)
    ```
 
-3. **Basis Functions**: Core set of interpretable transformations:
+2. **Basis Functions**: Core set of interpretable transformations:
    ```python
    SYMBOLIC_FUNCTIONS = {
        'linear': 'x',           # Direct relationships
@@ -92,10 +84,10 @@ OIKAN implements a modern interpretation of the Kolmogorov-Arnold Representation
    }
    ```
 
-4. **Formula Extraction Process**:
+3. **Formula Extraction Process**:
    - Train neural network on raw data
    - Generate augmented samples for better coverage
-   - Perform L1-regularized symbolic regression (alpha)
+   - Perform ElasticNet-regularization
    - Prune terms with coefficients below threshold
    - Export human-readable mathematical expressions
 
@@ -127,6 +119,9 @@ pip install -e .  # Install in development mode
 | Dependencies      | torch, numpy, scikit-learn, sympy, tqdm   |
 
 ### Regression Example
+
+> **Suggestion:** Please ensure that the data is normalized using standard scaling (or another suitable normalization method), as Elastic Net assumes that the model intercept has already been accounted for.
+
 ```python
 from oikan import OIKANRegressor
 from sklearn.metrics import mean_squared_error
@@ -136,8 +131,9 @@ model = OIKANRegressor(
     hidden_sizes=[32, 32], # Hidden layer sizes
     activation='relu', # Activation function (other options: 'tanh', 'leaky_relu', 'elu', 'swish', 'gelu')
     augmentation_factor=5, # Augmentation factor for data generation
-    alpha=0.1, # L1 regularization strength (Symbolic regression)
-    sigma=0.1, # Standard deviation of Gaussian noise for data augmentation
+    alpha=1.0, # ElasticNet regularization strength (Symbolic regression)
+    l1_rate=0.5, # ElasticNet mixing parameter (0 <= l1_ratio <= 1). 0 is equivalent to Ridge regression, 1 is equivalent to Lasso (Symbolic regression)
+    sigma=5, # Standard deviation of Gaussian noise for data augmentation
     top_k=5, # Number of top features to select (Symbolic regression)
     epochs=100, # Number of training epochs
     lr=0.001, # Learning rate
@@ -177,6 +173,9 @@ loaded_model.load("outputs/model.json")
 
 
 ### Classification Example
+
+> **Suggestion:** Please ensure that the data is normalized using standard scaling (or another suitable normalization method), as Elastic Net assumes that the model intercept has already been accounted for.
+
 ```python
 from oikan import OIKANClassifier
 from sklearn.metrics import accuracy_score
@@ -186,8 +185,9 @@ model = OIKANClassifier(
     hidden_sizes=[32, 32], # Hidden layer sizes
     activation='relu', # Activation function (other options: 'tanh', 'leaky_relu', 'elu', 'swish', 'gelu')
     augmentation_factor=10, # Augmentation factor for data generation
-    alpha=0.1, # L1 regularization strength (Symbolic regression)
-    sigma=0.1, # Standard deviation of Gaussian noise for data augmentation
+    alpha=1.0, # ElasticNet regularization strength (Symbolic regression)
+    l1_rate=0.5, # ElasticNet mixing parameter (0 <= l1_ratio <= 1). 0 is equivalent to Ridge regression, 1 is equivalent to Lasso (Symbolic regression)
+    sigma=5, # Standard deviation of Gaussian noise for data augmentation
     top_k=5, # Number of top features to select (Symbolic regression)
     epochs=100, # # Number of training epochs
     lr=0.001, # Learning rate

oikan-0.0.3.9.dist-info/RECORD

@@ -0,0 +1,11 @@
+oikan/__init__.py,sha256=Dh1Rf9ONRdm75B6tFiv9Y9P6NNiHAiKPCGDMuag6TTE,724
+oikan/elasticnet.py,sha256=yByuG9KCFQ4PpT2ze6oTSDy0DxvdF5MAJoegUGEipSA,2614
+oikan/exceptions.py,sha256=GhHWqy2Q5LVBcteTy4ngnqxr7FOoLNyD8dNt1kfRXyw,901
+oikan/model.py,sha256=UAjRYwb-kEap4AJkJ3OVmpNWpun0qgcad5m6x-mUbN8,26237
+oikan/neural.py,sha256=PZjaffSuABuCNxu-7PinU1GR6ji0Y6xRgSQ3n5HRDxI,1572
+oikan/utils.py,sha256=7UCm9obO-8Q2zhetdAkukMDOZvGSBWUL_dSF04XqM7k,8808
+oikan-0.0.3.9.dist-info/licenses/LICENSE,sha256=75ASVmU-XIpN-M4LbVmJ_ibgbzbvRLVti8FhnR0BTf8,1096
+oikan-0.0.3.9.dist-info/METADATA,sha256=KCUIcXDdneq4MTgnIKmOHf0OPjWUzmQSm21ayOIYQZs,13124
+oikan-0.0.3.9.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
+oikan-0.0.3.9.dist-info/top_level.txt,sha256=XwnwKwTJddZwIvtrUsAz-l-58BJRj6HjAGWrfYi_3QY,6
+oikan-0.0.3.9.dist-info/RECORD,,

oikan-0.0.3.7.dist-info/RECORD

@@ -1,10 +0,0 @@
-oikan/__init__.py,sha256=zEzhm1GYLT4vNaIQ4CgZcNpUk3uo8SWnoaHYtHW_XSQ,628
-oikan/exceptions.py,sha256=GhHWqy2Q5LVBcteTy4ngnqxr7FOoLNyD8dNt1kfRXyw,901
-oikan/model.py,sha256=TC2-R00GOjFb7ePzKTqeYkOiVlqUK7KP0mXsnJhg9ik,24736
-oikan/neural.py,sha256=PZjaffSuABuCNxu-7PinU1GR6ji0Y6xRgSQ3n5HRDxI,1572
-oikan/utils.py,sha256=7UCm9obO-8Q2zhetdAkukMDOZvGSBWUL_dSF04XqM7k,8808
-oikan-0.0.3.7.dist-info/licenses/LICENSE,sha256=75ASVmU-XIpN-M4LbVmJ_ibgbzbvRLVti8FhnR0BTf8,1096
-oikan-0.0.3.7.dist-info/METADATA,sha256=nrel6O7TXdbtJHSNCzvqPq_IELeQWx0azfrU4Jq6sps,12749
-oikan-0.0.3.7.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
-oikan-0.0.3.7.dist-info/top_level.txt,sha256=XwnwKwTJddZwIvtrUsAz-l-58BJRj6HjAGWrfYi_3QY,6
-oikan-0.0.3.7.dist-info/RECORD,,