ltfmselector 0.1.10__py3-none-any.whl → 0.1.12__py3-none-any.whl

ltfmselector/env.py

@@ -16,7 +16,8 @@ class Environment:
     def __init__(
             self, X, y, X_bg, fQueryCost, mQueryCost,
             fRepeatQueryCost, p_wNoFCost, errorCost, pType,
-            regression_tol, regression_error_rounding, pModels, device
+            regression_tol, regression_error_rounding, pModels, device,
+            sample_weight=None, **kwargs
     ):
         '''
         The environment with which the agent interacts, including the actions
@@ -34,7 +35,7 @@ class Environment:
         X_bg : pd.DataFrame
             Background dataaset, pandas dataframe with the shape:
             (n_samples+1, n_features)
-            
+
             An extra row for 'Total', average feature values for all training
             samples
 
@@ -67,7 +68,7 @@ class Environment:
             error is bigger than regression_tol
 
         regression_error_rounding : int
-            Only applicable for regression models. The error between the 
+            Only applicable for regression models. The error between the
             prediction and true value is rounded to the input decimal place.
 
         pModels : None or ``list of prediction models``
@@ -75,6 +76,9 @@ class Environment:
 
         device : ``CPU`` or ``GPU``
             Computation device
+
+        sample_weight : list or array or None
+            Per-sample weights
         '''
         # Datasets
         self.X = X
@@ -91,6 +95,7 @@ class Environment:
         self.regression_error_rounding = regression_error_rounding
 
         self.device = device
+        self.sample_weight = sample_weight
 
         # Available prediction models
         self.pType = pType
@@ -106,6 +111,14 @@ class Environment:
         # Counter for prediction model change
         self.pm_nChange = 0
 
+        ### Special-tailored implementation ###
+        if "smsproject" in list(kwargs.keys()):
+            self.smsproject = True
+        else:
+            self.smsproject = False
+
+        self.y_pred_bg = self.get_bgPrediction()
+
         self.state = None
 
     def reset(self, sample=None):
@@ -167,7 +180,7 @@ class Environment:
 
         return self.state
 
-    def step(self, action, sample_weight=None, **kwargs):
+    def step(self, action):
         '''
         Agent carries out an action.
 
@@ -177,9 +190,6 @@ class Environment:
             = -1 (make a prediction with selected features and prediction model)
             = int : [0, n_features] (query a feature)
             = int : [n_features, n_features + n_model] (query a prediction model)
-
-        sample_weight : list or array or None
-            Per-sample weights
         '''
         # === === === ===
         # Query a feature
@@ -229,13 +239,14 @@ class Environment:
             # Punish agent if it decides to predict without selecting any
             # features
             if len(col_to_retain) == 0:
+                self.y_pred = self.y_pred_bg[int(self.state[-1])]
                 return [None, -self.p_wNoFCost, True]
 
             # === === === ===
             # Make a prediction with selected features and prediction model
 
             ### Special-tailored implementation ###
-            if "smsproject" in list(kwargs.keys()):
+            if self.smsproject:
                 testpatientID = getPatientID(X_test.index[0])
                 otherSP_of_testPatient = [
                     sp for sp in X_train.index if getPatientID(sp) == testpatientID
@@ -254,13 +265,15 @@ class Environment:
             selected_predModel = self.pModels[int(self.state[-1])]
 
             ### Special-tailored implementation ###
-            if "smsproject" in list(kwargs.keys()):
+            if self.smsproject:
                 X_train_wLabel = X_train.copy()
                 X_train_wLabel["Target"] = self.y.loc[X_train_wLabel.index]
 
-                sample_weight = balance_classDistribution_patient(
+                _weights = balance_classDistribution_patient(
                     X_train_wLabel, "Target"
                 ).to_numpy(dtype=np.float32)[:,0]
+            else:
+                _weights = self.sample_weight
 
             # Convert X_train and y_train into numpy arrays if they are Pandas
             # DataFrame or Series
@@ -273,16 +286,16 @@ class Environment:
             if isinstance(X_test, pd.DataFrame):
                 X_test = X_test.values
 
-            if sample_weight is None:
+            if _weights is None:
                 selected_predModel.fit(X_train, y_train)
             else:
                 selected_predModel.fit(
-                    X_train, y_train, sample_weight=sample_weight
+                    X_train, y_train, sample_weight=_weights
                 )
-                
+
             self.y_pred = selected_predModel.predict(X_test)[0]
 
-            if "smsproject" in list(kwargs.keys()):
+            if self.smsproject:
                 # Capping values between 0 and 3
                 self.y_pred = capUpperValues(self.y_pred)
                 self.y_pred = capLowerValues(self.y_pred)
@@ -337,6 +350,54 @@ class Environment:
         '''
         return int(self.state[-1])
 
+    def get_bgPrediction(self):
+        '''
+        Get prediction based on background dataset for each type of
+        prediction model, fitted with the training samples, to be used
+        for the case that the agent decides to make a prediction without
+        any recruited features.
+        '''
+        # Initialize map between model type with background prediction
+        yBg_Model = []
+
+        ### Special-tailored implementation ###
+        if self.smsproject:
+            X_train_wLabel = self.X.copy()
+            X_train_wLabel["Target"] = self.y.loc[X_train_wLabel.index]
+
+            _weights = balance_classDistribution_patient(
+                X_train_wLabel, "Target"
+            ).to_numpy(dtype=np.float32)[:,0]
+        else:
+            _weights = self.sample_weight
+
+        # DataFrame or Series -> convert to numpy arrays
+        if isinstance(self.X, pd.DataFrame):
+            _X = self.X.values
+
+        if isinstance(self.y, pd.Series):
+            _y = self.y.values
+
+        for m in self.pModels:
+            # Fit each prediction model with the entire dataset
+            if _weights is None:
+                m.fit(_X, _y)
+            else:
+                m.fit(_X, _y, sample_weight=_weights)
+
+            # Use fitted model to make a prediction based on background
+            # dataset
+            yBg_Model.append(
+                m.predict(self.X_bg.loc[["Total"]])[0]
+            )
+
+            # Capping values between 0 and 3
+            if self.smsproject:
+                yBg_Model[-1] = capUpperValues(yBg_Model[-1])
+                yBg_Model[-1] = capLowerValues(yBg_Model[-1])
+
+        return yBg_Model
+
     def __getstate__(self):
         state = self.__dict__.copy()
         print(state.keys())

ltfmselector/ltfmselector.py

@@ -1,6 +1,7 @@
 import torch
 import torch.nn as nn
 import torch.optim as optim
+from torch.utils.tensorboard import SummaryWriter
 
 import os
 import random
@@ -120,7 +121,7 @@ class LTFMSelector:
 
             If pType == 'regression', then
             Agent is punished -errorCost*abs(``prediction`` - ``target``)
-            
+
             If pType == 'classification', then
             Agent is punished -errorCost
 
@@ -132,7 +133,7 @@ class LTFMSelector:
             error is bigger than regression_tol
 
         regression_error_rounding : int (default = 1)
-            Only applicable for regression models. The error between the 
+            Only applicable for regression models. The error between the
             prediction and true value is rounded to the input decimal place.
 
         pModels : None or ``list of prediction models``
@@ -142,7 +143,7 @@ class LTFMSelector:
             1. Support Vector Machine
             2. Random Forest
             3. Gaussian Naive Bayes
-            
+
             For regression:
             1. Support Vector Machine
             2. Random Forest
@@ -160,7 +161,7 @@ class LTFMSelector:
             Maximum number of time-steps per episode. Agent will be forced to
             make a prediction with the selected features and prediction model,
             if max_timesteps is reached
-            
+
             If None, max_timesteps will be set to 3 x number_of_features
 
         checkpoint_interval : int or None
@@ -226,7 +227,7 @@ class LTFMSelector:
     def fit(
             self, X, y, loss_function='mse', sample_weight=None,
             agent_neuralnetwork=None, lr=1e-5, returnQ=False,
-            background_dataset=None, **kwargs
+            monitor=False, background_dataset=None, **kwargs
     ):
         '''
         Initializes the environment and agent, then trains the agent to select
@@ -255,7 +256,7 @@ class LTFMSelector:
             integer elements (n1, n2). n1 and n2 pertains to the number of units
             in the first and second layer of a multilayer-perceptron,
             implemented in PyTorch.
-            
+
             If None, a default multilayer-perceptron of two hidden layers, each
             with 1024 units is used.
 
@@ -265,7 +266,14 @@ class LTFMSelector:
 
         returnQ : bool
             Return average computed action-value functions and rewards of
-            the sampled batches, for debugging purposes.
+            the sampled batches, as a (<total_iterations>, 3) matrix. The 
+            columns correspond to the averaged Q, reward, and target functions.
+
+        monitor : bool
+            Monitor training process using a TensorBoard.
+
+            Run `tensorboard --logdir=runs` in the terminal to monitor the p
+            progression of the action-value function.
 
         background_dataset : None or pd.DataFrame
             If None, numerical features will be assumed when computing the
@@ -284,10 +292,10 @@ class LTFMSelector:
                 List of policy network's action-value function, Q(s,a),
                 averaged over the sampled batch during training, per iteration
             r_avr_list : list
-                List of rewards, r, averaged over the sampled batch during 
+                List of rewards, r, averaged over the sampled batch during
                 training, per iteration
             V_avr_list : list
-                List of max action-value function for the next state (s'), 
+                List of max action-value function for the next state (s'),
                 max{a} Q(s', a), averaged over the sampled batch during
                 training, per iteration
         '''
@@ -312,10 +320,18 @@ class LTFMSelector:
         self.sample_weight = sample_weight
 
         # If user wants to monitor progression of terms in the loss function
+        if monitor:
+            writer = SummaryWriter()
+            monitor_count = 1
+
+        # If user wants to save average computed action-value functions and
+        # rewards of sampled batches
         if returnQ:
-            Q_avr_list = []
-            r_avr_list = []
-            V_avr_list = []
+            total_iterations = 10000000000
+            LearningValuesMatrix = np.zeros(
+                (total_iterations, 3), dtype=np.float32
+            )
+            Q_count = 1
 
         # Initializing the environment
         env = Environment(
@@ -323,7 +339,8 @@ class LTFMSelector:
             self.fQueryCost, self.mQueryCost,
             self.fRepeatQueryCost, self.p_wNoFCost, self.errorCost,
             self.pType, self.regression_tol, self.regression_error_rounding,
-            self.pModels, self.device
+            self.pModels, self.device, sample_weight=self.sample_weight,
+            **kwargs
         )
         env.reset()
 
@@ -387,9 +404,7 @@ class LTFMSelector:
                 # Agent carries out action on the environment and returns:
                 # - observation (state in next time-step)
                 # - reward
-                observation, reward, terminated = env.step(
-                    action.item(), sample_weight=self.sample_weight, **kwargs
-                )
+                observation, reward, terminated = env.step(action.item())
 
                 if terminated:
                     next_state = None
@@ -407,13 +422,21 @@ class LTFMSelector:
                 state = next_state
 
                 # Optimize the model
-                _res = self.optimize_model(optimizer, loss_function, returnQ)
+                _res = self.optimize_model(optimizer, loss_function, monitor, returnQ)
+
+                if monitor:
+                    if not _res is None:
+                        writer.add_scalar("Metrics/Average_QValue", _res[0], monitor_count)
+                        writer.add_scalar("Metrics/Average_Reward", _res[1], monitor_count)
+                        writer.add_scalar("Metrics/Average_Target", _res[2], monitor_count)
+                        monitor_count += 1
 
                 if returnQ:
                     if not _res is None:
-                        Q_avr_list.append(_res[0])
-                        r_avr_list.append(_res[1])
-                        V_avr_list.append(_res[2])
+                        LearningValuesMatrix[Q_count, 0] = _res[0]
+                        LearningValuesMatrix[Q_count, 1] = _res[1]
+                        LearningValuesMatrix[Q_count, 2] = _res[2]
+                        Q_count += 1
 
                 # Apply soft update to target network's weights
                 targetParameters = self.target_net.state_dict()
@@ -457,11 +480,19 @@ class LTFMSelector:
                 self.policy_network_checkpoints[self.episodes] =\
                     self.policy_net.state_dict()
 
+        if monitor:
+            writer.add_scalar("Metrics/Average_QValue", _res[0], monitor_count)
+            writer.add_scalar("Metrics/Average_Reward", _res[1], monitor_count)
+            writer.add_scalar("Metrics/Average_Target", _res[2], monitor_count)
+            writer.close()
+
         if returnQ:
-            Q_avr_list.append(_res[0])
-            r_avr_list.append(_res[1])
-            V_avr_list.append(_res[2])
-            return doc, (Q_avr_list, r_avr_list, V_avr_list)
+            LearningValuesMatrix[Q_count, 0] = _res[0]
+            LearningValuesMatrix[Q_count, 1] = _res[1]
+            LearningValuesMatrix[Q_count, 2] = _res[2]
+
+        if (monitor or returnQ):
+            return doc, LearningValuesMatrix[0:Q_count+1, :]
         else:
             return doc
 
@@ -489,7 +520,7 @@ class LTFMSelector:
             self.fQueryCost, self.mQueryCost,
             self.fRepeatQueryCost, self.p_wNoFCost, self.errorCost,
             self.pType, self.regression_tol, self.regression_error_rounding,
-            self.pModels, self.device
+            self.pModels, self.device, **kwargs
         )
 
         # Create dictionary to save information per episode
@@ -513,9 +544,7 @@ class LTFMSelector:
                 if t > self.max_timesteps:
                     action = torch.tensor([[-1]], device=self.device)
 
-                observation, reward, terminated = env.step(
-                    action.item(), sample_weight=self.sample_weight, **kwargs
-                )
+                observation, reward, terminated = env.step(action.item())
 
                 if terminated:
                     next_state = None
@@ -575,7 +604,7 @@ class LTFMSelector:
             with torch.no_grad():
                 return (self.policy_net(state).max(1)[1].view(1, 1) - 1)
 
-    def optimize_model(self, optimizer, loss_function, returnQ):
+    def optimize_model(self, optimizer, loss_function, monitor, returnQ):
         '''
         Optimize the policy network.
 
@@ -603,8 +632,8 @@ class LTFMSelector:
         # 1. Draw a random batch of experiences
         experiences = self.ReplayMemory.sample(self.batch_size)
         # [
-        #    Experience #1: (state, action, next_state, reward), 
-        #    Experience #2: (state, action, next_state, reward), 
+        #    Experience #1: (state, action, next_state, reward),
+        #    Experience #2: (state, action, next_state, reward),
         #    ...
         # ]
 
@@ -688,7 +717,7 @@ class LTFMSelector:
                 criterion = nn.SmoothL1Loss()
         else:
             criterion = loss_function
-            
+
         loss = criterion(
             state_action_values, expected_state_action_values.unsqueeze(1)
         )
@@ -701,7 +730,7 @@ class LTFMSelector:
         # Optimize the model (policy network)
         optimizer.step()
 
-        if returnQ:
+        if (monitor or returnQ):
             Q_avr = state_action_values.detach().numpy().mean()
             r_avr = reward_batch.unsqueeze(1).numpy().mean()
             V_avr = expected_state_action_values.unsqueeze(1).numpy().mean()

{ltfmselector-0.1.10.dist-info → ltfmselector-0.1.12.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: ltfmselector
-Version: 0.1.10
+Version: 0.1.12
 Summary: Locally-Tailored Feature and Model Selector with Deep Q-Learning
 Project-URL: GitHub, https://github.com/RenZhen95/ltfmselector/
 Author-email: RenZhen95 <j-liaw@hotmail.com>
@@ -32,8 +32,9 @@ Requires-Dist: matplotlib>=3.10.1
 Requires-Dist: numpy>=2.2.4
 Requires-Dist: openpyxl>=3.1.5
 Requires-Dist: pandas>=2.2.3
-Requires-Dist: scikit-learn>=1.6.1
+Requires-Dist: scikit-learn<1.6
 Requires-Dist: seaborn>=0.13.2
+Requires-Dist: tensorboard>=2.20.0
 Requires-Dist: torch>=2.6.0
 Description-Content-Type: text/markdown
 

ltfmselector-0.1.12.dist-info/RECORD

@@ -0,0 +1,9 @@
+ltfmselector/__init__.py,sha256=lf3e90CNpEDvEmNZ-0iuoHOPsA7D-WN_opbBsTYLVEA,76
+ltfmselector/env.py,sha256=vizWGqDSc_2Zfs9aXjFARanIAz6PTKwUHu2_Lew9s3Y,13878
+ltfmselector/ltfmselector.py,sha256=vs9unOmoDKq1piV6t87GC1wdy7kP8ucKHihw6i0F4KI,29567
+ltfmselector/py.typed,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
+ltfmselector/utils.py,sha256=VXYZSDm7x4s0p9F_58NLW8WQa3dxi0vHZewRy6miC2E,5438
+ltfmselector-0.1.12.dist-info/METADATA,sha256=QaUPeSx9NlZx0ZUbkEPRyFS-8nfJz9Y8yV5TXXPc7fA,3021
+ltfmselector-0.1.12.dist-info/WHEEL,sha256=WLgqFyCfm_KASv4WHyYy0P3pM_m7J5L9k2skdKLirC8,87
+ltfmselector-0.1.12.dist-info/licenses/LICENSE,sha256=tmIDlkkp4a0EudXuGmeTdGjHjPhmmXkEMshACXLqX2w,1092
+ltfmselector-0.1.12.dist-info/RECORD,,

{ltfmselector-0.1.10.dist-info → ltfmselector-0.1.12.dist-info}/WHEEL

@@ -1,4 +1,4 @@
 Wheel-Version: 1.0
-Generator: hatchling 1.27.0
+Generator: hatchling 1.28.0
 Root-Is-Purelib: true
 Tag: py3-none-any

ltfmselector-0.1.10.dist-info/RECORD

@@ -1,9 +0,0 @@
-ltfmselector/__init__.py,sha256=lf3e90CNpEDvEmNZ-0iuoHOPsA7D-WN_opbBsTYLVEA,76
-ltfmselector/env.py,sha256=mHa6l7mWE5mZGFTGA7sqr2xbGLAuE1ll0c5Lh8Ju5Gw,11854
-ltfmselector/ltfmselector.py,sha256=JX3jtlRE2KRUssH-LGwcrvw0y9HALPNQutete6PI09c,28150
-ltfmselector/py.typed,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
-ltfmselector/utils.py,sha256=VXYZSDm7x4s0p9F_58NLW8WQa3dxi0vHZewRy6miC2E,5438
-ltfmselector-0.1.10.dist-info/METADATA,sha256=TjeFKEBs09qrB3cbDRMXVCJJ-mcE5-CDJ2nju5qoc6w,2989
-ltfmselector-0.1.10.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
-ltfmselector-0.1.10.dist-info/licenses/LICENSE,sha256=tmIDlkkp4a0EudXuGmeTdGjHjPhmmXkEMshACXLqX2w,1092
-ltfmselector-0.1.10.dist-info/RECORD,,