sciml 0.0.7__py3-none-any.whl → 0.0.8__py3-none-any.whl

sciml/__init__.py

@@ -1,2 +1,2 @@
 # coding: utf-8
-__all__ = ["utils", "pipelines"]
+__all__ = ["utils", "pipelines", "models"]

sciml/models.py

@@ -0,0 +1,269 @@
+import numpy as np
+import copy
+import itertools
+import warnings
+from xgboost import XGBRegressor
+from sklearn.metrics import mean_squared_error
+from sklearn.model_selection import train_test_split
+
+class SmartForest:
+    """
+    SmartForest: A deep, intelligent decision forest model for complex sequential and tabular data.
+
+    SmartForest blends ideas from deep forests (cascade forest structures), LSTM-style forget gates,
+    and ensemble learning using XGBoost. It is especially suited for time series or structured tabular data 
+    where layer-wise feature expansion and memory-inspired filtering can enhance performance.
+
+    Key Features:
+    -------------
+    - Deep cascade of XGBoost regressors
+    - Optional Multi-Grained Scanning (MGS) for local feature extraction
+    - Forget-gate-inspired mechanism to regulate information flow across layers
+    - Early stopping to prevent overfitting
+    - Full retention of best-performing model (lowest validation RMSE)
+
+    Parameters:
+    -----------
+    n_estimators_per_layer : int
+        Number of XGBoost regressors per layer.
+    
+    max_layers : int
+        Maximum number of layers (depth) in the model.
+
+    early_stopping_rounds : int
+        Number of layers with no improvement before early stopping is triggered.
+
+    param_grid : dict
+        Grid of XGBoost hyperparameters to search over.
+
+    use_gpu : bool
+        If True, use GPU-accelerated training (CUDA required).
+
+    gpu_id : int
+        ID of GPU to use (if use_gpu=True).
+
+    window_sizes : list of int
+        Enables Multi-Grained Scanning if non-empty, with specified sliding window sizes.
+
+    forget_factor : float in [0, 1]
+        Simulates LSTM-style forget gate; higher values forget more past information.
+
+    verbose : int
+        Verbosity level (0 = silent, 1 = progress updates).
+
+    Methods:
+    --------
+    fit(X, y, X_val=None, y_val=None):
+        Train the SmartForest model layer by layer, using optional validation for early stopping.
+
+    predict(X):
+        Make predictions on new data using the trained cascade structure.
+
+    get_best_model():
+        Returns a copy of the best model and the corresponding RMSE from validation.
+
+    Example:
+    --------
+    >>> model = SmartForest(n_estimators_per_layer=5, max_layers=10, window_sizes=[2, 3], forget_factor=0.2)
+    >>> model.fit(X_train, y_train, X_val, y_val)
+    >>> y_pred = model.predict(X_val)
+    >>> best_model, best_rmse = model.get_best_model()
+    """
+    def __init__(self, n_estimators_per_layer = 5, max_layers = 10, early_stopping_rounds = 3, param_grid = None,
+                 use_gpu = False, gpu_id = 0, window_sizes = [], forget_factor = 0, verbose = 1):
+        self.n_estimators_per_layer = n_estimators_per_layer
+        self.max_layers = max_layers
+        self.early_stopping_rounds = early_stopping_rounds
+        self.param_grid = param_grid or {
+            "objective": ["reg:squarederror"],
+            "random_state": [42],
+            'seed': [0],
+            'n_estimators': [100],
+            'max_depth': [6],
+            'min_child_weight': [4],
+            'subsample': [0.8],
+            'colsample_bytree': [0.8],
+            'gamma': [0],
+            'reg_alpha': [0],
+            'reg_lambda': [1],
+            'learning_rate': [0.05],
+        }
+        self.use_gpu = use_gpu
+        self.gpu_id = gpu_id
+        self.window_sizes = window_sizes
+        self.forget_factor = forget_factor
+        self.layers = []
+        self.best_model = None
+        self.best_rmse = float("inf")
+        self.verbose = verbose
+
+    def _get_param_combinations(self):
+        keys, values = zip(*self.param_grid.items())
+        return [dict(zip(keys, v)) for v in itertools.product(*values)]
+
+    def _multi_grained_scanning(self, X, y):
+        new_features = []
+        for window_size in self.window_sizes:
+            if X.shape[1] < window_size:
+                continue
+            for start in range(X.shape[1] - window_size + 1):
+                window = X[:, start:start + window_size]
+                if y is None:
+                    new_features.append(window)
+                    continue
+
+                param_combos = self._get_param_combinations()
+                for params in param_combos:
+                    if self.use_gpu:
+                        params['tree_method'] = 'hist'
+                        params['device'] = 'cuda'
+                    model = XGBRegressor(**params)
+                    model.fit(window, y)
+                    preds = model.predict(window).reshape(-1, 1)
+                    new_features.append(preds)
+        return np.hstack(new_features) if new_features else X
+
+    def _apply_forget_gate(self, X, layer_index):
+        forget_weights = np.random.rand(X.shape[1]) * self.forget_factor
+        return X * (1 - forget_weights)
+
+    def _fit_layer(self, X, y, X_val=None, y_val=None, layer_index=0):
+        layer = []
+        layer_outputs = []
+        param_combos = self._get_param_combinations()
+        X = self._apply_forget_gate(X, layer_index)
+
+        for i in range(self.n_estimators_per_layer):
+            best_rmse = float('inf')
+            best_model = None
+
+            for params in param_combos:
+                if self.use_gpu:
+                    params['tree_method'] = 'hist'
+                    params['device'] = 'cuda'
+
+                model = XGBRegressor(**params)
+                model.fit(X, y)
+
+                if X_val is not None:
+                    preds_val = model.predict(X_val)
+                    rmse = np.sqrt(mean_squared_error(y_val, preds_val))
+                    if rmse < best_rmse:
+                        best_rmse = rmse
+                        best_model = model
+                else:
+                    best_model = model
+
+            preds = best_model.predict(X).reshape(-1, 1)
+            layer.append(best_model)
+            layer_outputs.append(preds)
+
+        output = np.hstack(layer_outputs)
+        return layer, output
+
+    def fit(self, X, y, X_val=None, y_val=None):
+        X_current = self._multi_grained_scanning(X, y)
+        X_val_current = self._multi_grained_scanning(X_val, y_val) if X_val is not None else None
+        no_improve_rounds = 0
+
+        for layer_index in range(self.max_layers):
+            if self.verbose: print(f"Training Layer {layer_index + 1}")
+            layer, output = self._fit_layer(X_current, y, X_val_current, y_val, layer_index)
+            self.layers.append(layer)
+            X_current = np.hstack([X_current, output])
+
+            if X_val is not None:
+                val_outputs = []
+                for reg in layer:
+                    n_features = reg.n_features_in_
+                    preds = reg.predict(X_val_current[:, :n_features]).reshape(-1, 1)
+                    val_outputs.append(preds)
+                val_output = np.hstack(val_outputs)
+                X_val_current = np.hstack([X_val_current, val_output])
+
+                y_pred = self.predict(X_val)
+                rmse = np.sqrt(mean_squared_error(y_val, y_pred))
+                if self.verbose: print(f"Validation RMSE: {rmse:.4f}")
+
+                if rmse < self.best_rmse:
+                    self.best_rmse = rmse
+                    self.best_model = copy.deepcopy(self.layers)
+                    no_improve_rounds = 0
+                    if self.verbose: print(f"✅ New best RMSE: {self.best_rmse:.4f}")
+                else:
+                    no_improve_rounds += 1
+                    if no_improve_rounds >= self.early_stopping_rounds:
+                        if self.verbose: print("Early stopping triggered.")
+                        break
+
+    def predict(self, X):
+        X_current = self._multi_grained_scanning(X, None)
+        X_current = self._apply_forget_gate(X_current, layer_index=-1)
+
+        for layer in self.layers:
+            layer_outputs = []
+            for reg in layer:
+                n_features = reg.n_features_in_
+                preds = reg.predict(X_current[:, :n_features]).reshape(-1, 1)
+                layer_outputs.append(preds)
+            output = np.hstack(layer_outputs)
+            X_current = np.hstack([X_current, output])
+
+        final_outputs = []
+        for reg in self.layers[-1]:
+            n_features = reg.n_features_in_
+            final_outputs.append(reg.predict(X_current[:, :n_features]).reshape(-1, 1))
+        return np.mean(np.hstack(final_outputs), axis=1)
+
+    def get_best_model(self):
+        return self.best_model, self.best_rmse
+
+"""
+# ============================== Test Example ==============================
+from sklearn.datasets import load_diabetes
+
+warnings.simplefilter('ignore')
+
+X, y = load_diabetes(return_X_y=True)
+X_train, X_val, y_train, y_val = train_test_split(X, y, test_size=0.3, random_state=42)
+
+# Hyperparameter grid
+param_grid = {
+    "objective": ["reg:squarederror"],
+    "random_state": [42],
+    'seed': [0],
+    'n_estimators': [100],
+    'max_depth': [6],
+    'min_child_weight': [4],
+    'subsample': [0.8],
+    'colsample_bytree': [0.8],
+    'gamma': [0],
+    'reg_alpha': [0],
+    'reg_lambda': [1],
+    'learning_rate': [0.05],
+}
+
+# Create the model with Multi-Grained Scanning enabled (with window sizes 2 and 3)
+df_reg = SmartForest(
+    n_estimators_per_layer = 5,
+    max_layers = 10,
+    early_stopping_rounds = 5,
+    param_grid = param_grid,
+    use_gpu = False,
+    gpu_id = 0,
+    window_sizes = [],  # Enables MGS if e.g., [2, 3], else empty disables MGS.
+    forget_factor = 0.,  # Set forget factor to simulate forget gate behavior
+    verbose = 1
+)
+
+df_reg.fit(X_train, y_train, X_val, y_val)
+
+# Predict on validation set and evaluate
+y_pred = df_reg.predict(X_val)
+rmse = np.sqrt(mean_squared_error(y_val, y_pred))
+print("\nFinal RMSE:", rmse)
+
+# Output best model and RMSE
+best_model, best_rmse = df_reg.get_best_model()
+print("\nBest validation RMSE:", best_rmse)
+"""

sciml/pipelines.py

@@ -171,4 +171,265 @@ def train_lstm(X_train, y_train, nfeature, ntime, verbose = 2, epochs = 200, bat
     # es = EarlyStopping(monitor='loss', mode='min', verbose=1)
     # model.fit(X_train.reshape(-1, nsites, nfeats), y_train, epochs=100, batch_size=256, verbose=2, callbacks=[es])
     model.fit(X_train, y_train, epochs = epochs, batch_size = batch_size, verbose=verbose)
-    return model
+    return model
+
+
+'''
+# ========================================================================================================
+import numpy as np
+from xgboost import XGBRegressor
+from sklearn.metrics import mean_squared_error
+
+class XGBoostDeepForestRegressor:
+    def __init__(self, n_estimators_per_layer=2, max_layers=20, early_stopping_rounds=2):
+        self.n_estimators_per_layer = n_estimators_per_layer
+        self.max_layers = max_layers
+        self.early_stopping_rounds = early_stopping_rounds
+        self.layers = []
+
+    def _fit_layer(self, X, y):
+        layer = []
+        layer_outputs = []
+        for _ in range(self.n_estimators_per_layer):
+            reg = XGBRegressor()
+            reg.fit(X, y)
+            preds = reg.predict(X).reshape(-1, 1)
+            layer.append(reg)
+            layer_outputs.append(preds)
+        output = np.hstack(layer_outputs)
+        return layer, output
+
+    def fit(self, X, y, X_val=None, y_val=None):
+        X_current = X.copy()
+        best_rmse = float("inf")
+        no_improve_rounds = 0
+
+        for layer_index in range(self.max_layers):
+            print(f"Training Layer {layer_index + 1}")
+            layer, output = self._fit_layer(X_current, y)
+            self.layers.append(layer)
+            X_current = np.hstack([X_current, output])
+
+            if X_val is not None:
+                y_pred = self.predict(X_val)
+                # rmse = mean_squared_error(y_val, y_pred, squared=False)
+                rmse = np.sqrt(mean_squared_error(y_val, y_pred))
+                print(f"Validation RMSE: {rmse:.4f}")
+
+                if rmse < best_rmse:
+                    best_rmse = rmse
+                    no_improve_rounds = 0
+                else:
+                    no_improve_rounds += 1
+                    if no_improve_rounds >= self.early_stopping_rounds:
+                        print("Early stopping triggered.")
+                        break
+
+    def predict(self, X):
+        X_current = X.copy()
+        for layer in self.layers:
+            layer_outputs = []
+            for reg in layer:
+                n_features = reg.n_features_in_
+                preds = reg.predict(X_current[:, :n_features]).reshape(-1, 1)
+                layer_outputs.append(preds)
+            output = np.hstack(layer_outputs)
+            X_current = np.hstack([X_current, output])
+
+        # Final prediction = average of last layer regressors
+        final_outputs = []
+        for reg in self.layers[-1]:
+            n_features = reg.n_features_in_
+            final_outputs.append(reg.predict(X_current[:, :n_features]).reshape(-1, 1))
+        return np.mean(np.hstack(final_outputs), axis=1)
+
+
+from sklearn.datasets import load_diabetes
+from sklearn.model_selection import train_test_split
+from sklearn.metrics import mean_squared_error
+
+X, y = load_diabetes(return_X_y=True)
+X_train, X_val, y_train, y_val = train_test_split(X, y, test_size=0.3, random_state=42)
+
+df_reg = XGBoostDeepForestRegressor(n_estimators_per_layer=2, max_layers=5)
+df_reg.fit(X_train, y_train, X_val, y_val)
+
+y_pred = df_reg.predict(X_val)
+# rmse = mean_squared_error(y_val, y_pred, squared=False)
+rmse = np.sqrt(mean_squared_error(y_val, y_pred))
+print("Final RMSE:", rmse)
+
+# ----------------------------------------------------------------------------------------------------
+
+import numpy as np
+from xgboost import XGBRegressor
+from sklearn.metrics import mean_squared_error
+import itertools
+
+class XGBoostDeepForestRegressor:
+    def __init__(self, n_estimators_per_layer=2, max_layers=20, early_stopping_rounds=2, param_grid=None, use_gpu=True, gpu_id=0):
+        self.n_estimators_per_layer = n_estimators_per_layer
+        self.max_layers = max_layers
+        self.early_stopping_rounds = early_stopping_rounds
+        self.param_grid = param_grid or {
+            'max_depth': [3],
+            'learning_rate': [0.1],
+            'n_estimators': [100]
+        }
+        self.use_gpu = use_gpu
+        self.gpu_id = gpu_id
+        self.layers = []
+
+    def _get_param_combinations(self):
+        keys, values = zip(*self.param_grid.items())
+        return [dict(zip(keys, v)) for v in itertools.product(*values)]
+
+    def _fit_layer(self, X, y, X_val=None, y_val=None):
+        layer = []
+        layer_outputs = []
+        param_combos = self._get_param_combinations()
+
+        for i in range(self.n_estimators_per_layer):
+            best_rmse = float('inf')
+            best_model = None
+
+            for params in param_combos:
+                # Set GPU support parameters in XGBRegressor
+                if self.use_gpu:
+                    params['tree_method'] = 'hist'  # Use hist method
+                    params['device'] = 'cuda'  # Enable CUDA for GPU
+
+                model = XGBRegressor(**params)
+                model.fit(X, y)
+
+                if X_val is not None:
+                    preds_val = model.predict(X_val)
+                    rmse = np.sqrt(mean_squared_error(y_val, preds_val))
+                    if rmse < best_rmse:
+                        best_rmse = rmse
+                        best_model = model
+                else:
+                    best_model = model
+
+            final_model = best_model
+            preds = final_model.predict(X).reshape(-1, 1)
+            layer.append(final_model)
+            layer_outputs.append(preds)
+
+        output = np.hstack(layer_outputs)
+        return layer, output
+
+    def fit(self, X, y, X_val=None, y_val=None):
+        X_current = X.copy()
+        X_val_current = X_val.copy() if X_val is not None else None
+
+        best_rmse = float("inf")
+        no_improve_rounds = 0
+
+        for layer_index in range(self.max_layers):
+            print(f"Training Layer {layer_index + 1}")
+            layer, output = self._fit_layer(X_current, y, X_val_current, y_val)
+            self.layers.append(layer)
+            X_current = np.hstack([X_current, output])
+
+            if X_val is not None:
+                val_outputs = []
+                for reg in layer:
+                    n_features = reg.n_features_in_
+                    preds = reg.predict(X_val_current[:, :n_features]).reshape(-1, 1)
+                    val_outputs.append(preds)
+                val_output = np.hstack(val_outputs)
+                X_val_current = np.hstack([X_val_current, val_output])
+
+                y_pred = self.predict(X_val)
+                rmse = np.sqrt(mean_squared_error(y_val, y_pred))
+                print(f"Validation RMSE: {rmse:.4f}")
+
+                if rmse < best_rmse:
+                    best_rmse = rmse
+                    no_improve_rounds = 0
+                else:
+                    no_improve_rounds += 1
+                    if no_improve_rounds >= self.early_stopping_rounds:
+                        print("Early stopping triggered.")
+                        break
+
+    def predict(self, X):
+        X_current = X.copy()
+        for layer in self.layers:
+            layer_outputs = []
+            for reg in layer:
+                n_features = reg.n_features_in_
+                preds = reg.predict(X_current[:, :n_features]).reshape(-1, 1)
+                layer_outputs.append(preds)
+            output = np.hstack(layer_outputs)
+            X_current = np.hstack([X_current, output])
+
+        final_outputs = []
+        for reg in self.layers[-1]:
+            n_features = reg.n_features_in_
+            final_outputs.append(reg.predict(X_current[:, :n_features]).reshape(-1, 1))
+        return np.mean(np.hstack(final_outputs), axis=1)
+
+
+from sklearn.datasets import load_diabetes
+from sklearn.model_selection import train_test_split
+from sklearn.metrics import mean_squared_error
+
+# Load dataset
+X, y = load_diabetes(return_X_y=True)
+X_train, X_val, y_train, y_val = train_test_split(X, y, test_size=0.3, random_state=42)
+
+# Hyperparameter grid
+param_grid = {
+    'max_depth': [3, 4],
+    'learning_rate': [0.1, 0.05],
+    'n_estimators': [50, 100]
+}
+
+# Create and fit the model with GPU enabled
+df_reg = XGBoostDeepForestRegressor(
+    n_estimators_per_layer=2,
+    max_layers=5,
+    early_stopping_rounds=2,
+    param_grid=param_grid,
+    use_gpu=True,  # Enable GPU acceleration
+    gpu_id=0  # Default to the first GPU
+)
+
+df_reg.fit(X_train, y_train, X_val, y_val)
+
+# Final evaluation
+y_pred = df_reg.predict(X_val)
+rmse = np.sqrt(mean_squared_error(y_val, y_pred))
+print("Final RMSE:", rmse)
+
+# ----------------------------------------------------------------------------------------------------
+
+xgb_params = {
+    "objective": "reg:squarederror",
+    "random_state": 0,
+    'seed': 0,
+    'n_estimators': 100,
+    'max_depth': 6, 
+    'min_child_weight': 4, 
+    'subsample': 0.8, 
+    'colsample_bytree': 0.8, 
+    'gamma': 0, 
+    'reg_alpha': 0, 
+    'reg_lambda': 1,
+    'learning_rate': 0.05, 
+}
+
+from xgboost import XGBRegressor
+regr = XGBRegressor(**xgb_params)
+
+regr.fit(X_train, y_train)
+y_pred = regr.predict(X_val)
+
+
+from scipy import stats
+
+stats.linregress(y_val, y_pred)
+
+'''

{sciml-0.0.7.dist-info → sciml-0.0.8.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: sciml
-Version: 0.0.7
+Version: 0.0.8
 Summary: draw and basic calculations/conversions
 Home-page: https://github.com/soonyenju/sciml
 Author: Songyan Zhu

sciml-0.0.8.dist-info/RECORD

@@ -0,0 +1,9 @@
+sciml/__init__.py,sha256=6iQAGgCEMuw4yoLBzZDax46a45LZgzEeNSHQMdmcBSQ,58
+sciml/models.py,sha256=p6cw3SxTQaOtFhJx8KdW0Z2QtxBlSBlVPHETTNCjJ2w,9880
+sciml/pipelines.py,sha256=CJolleJakoEQc-EV-v6NovP3bDb1hif7SvObXdaLXdY,15268
+sciml/utils.py,sha256=u5DzQJV4aCZ-p7sY56Fxzj8WDGYOgn1rOTeGzAw0vwY,1831
+sciml-0.0.8.dist-info/LICENSE,sha256=dX4jBmkgQPWc_TfYkXtKQzVIgZQWFuHZ8vQjV4sEeV4,1060
+sciml-0.0.8.dist-info/METADATA,sha256=uMCtigVwS2e0abqbvfbLZca6iZnkdDTBXtbjdg34yIA,313
+sciml-0.0.8.dist-info/WHEEL,sha256=2wepM1nk4DS4eFpYrW1TTqPcoGNfHhhO_i5m4cOimbo,92
+sciml-0.0.8.dist-info/top_level.txt,sha256=dS_7aBCZFKQE3myPy5sh4USjQZCZyGg382-YxUUYcdw,6
+sciml-0.0.8.dist-info/RECORD,,

sciml-0.0.7.dist-info/RECORD

@@ -1,8 +0,0 @@
-sciml/__init__.py,sha256=Asqzx08kEOBLv_IRE20VlHxZu9XgydyrzIMUDRE-qiU,48
-sciml/pipelines.py,sha256=5qfeHdxGhF-GMu-rTiInPv5metXiT32uSENIDFd2Ths,6333
-sciml/utils.py,sha256=u5DzQJV4aCZ-p7sY56Fxzj8WDGYOgn1rOTeGzAw0vwY,1831
-sciml-0.0.7.dist-info/LICENSE,sha256=dX4jBmkgQPWc_TfYkXtKQzVIgZQWFuHZ8vQjV4sEeV4,1060
-sciml-0.0.7.dist-info/METADATA,sha256=363EbWoSVqR9qAdhOfeVD8RiP6DfcalvDiZECJ6LW3s,313
-sciml-0.0.7.dist-info/WHEEL,sha256=2wepM1nk4DS4eFpYrW1TTqPcoGNfHhhO_i5m4cOimbo,92
-sciml-0.0.7.dist-info/top_level.txt,sha256=dS_7aBCZFKQE3myPy5sh4USjQZCZyGg382-YxUUYcdw,6
-sciml-0.0.7.dist-info/RECORD,,