ltfmselector 0.1.12__py3-none-any.whl → 0.2.0__py3-none-any.whl

ltfmselector/env.py

@@ -14,7 +14,10 @@ capLowerValues = lambda x: 0.0 if x < 0.0 else x
 
 class Environment:
     def __init__(
-            self, X, y, X_bg, fQueryCost, mQueryCost,
+            self, X, y, X_bg,
+            fQueryCost, fQueryFunction,
+            fThreshold, fCap, fRate,
+            mQueryCost,
             fRepeatQueryCost, p_wNoFCost, errorCost, pType,
             regression_tol, regression_error_rounding, pModels, device,
             sample_weight=None, **kwargs
@@ -42,6 +45,21 @@ class Environment:
         fQueryCost : float
             Cost of querying a feature
 
+        fQueryFunction : None or {'step', 'linear', 'quadratic', 'exponential'}
+            Function to progressively increase cost of recruiting a feature
+
+        fThreshold : None or int
+            If `fQueryFunction == {'step', 'linear', 'quadratic', 'exponential'}`
+            Threshold of number of features, before cost of recruiting 
+            increases
+            
+        fCap : None or float
+            If `fQueryFunction == {'step'}`, upper limit of penalty
+
+        fRate : None or float
+            If `fQueryFunction == {'linear', 'quadratic', 'exponential'}`, rate
+            individual cost functions
+
         mQueryCost : float
             Cost of querying a prediction model
 
@@ -87,6 +105,11 @@ class Environment:
 
         # Reward functions
         self.fQueryCost = fQueryCost
+        self.fQueryFunction = fQueryFunction
+        self.fThreshold = fThreshold
+        self.fCap = fCap
+        self.fRate = fRate
+
         self.mQueryCost = mQueryCost
         self.fRepeatQueryCost = fRepeatQueryCost
         self.p_wNoFCost = p_wNoFCost
@@ -204,7 +227,7 @@ class Environment:
                 self.state[action] = self.X_test.iloc[0, action]
 
                 # Punish for querying a feature
-                return [self.state, -self.fQueryCost, False]
+                return [self.state, -self.get_fQueryCost(), False]
 
             # Punish agent for attempting to query a feature already
             # previously selected
@@ -331,6 +354,44 @@ class Environment:
 
                 return [None, 0.0, True]
 
+    def get_fQueryCost(self):
+        '''
+        Get cost of querying a feature
+        '''
+        if self.fQueryFunction is None:
+            return self.fQueryCost
+
+        # Get number of total recruited features
+        nFSubset = (self.get_feature_mask()).sum()
+
+        if self.fQueryFunction == "step":
+            return self.get_fQueryCostStep(nFSubset)
+        elif self.fQueryFunction == "linear":
+            return self.get_fQueryCostLinear(nFSubset)
+        elif self.fQueryFunction == "quadratic":
+            return self.get_fQueryCostQuadratic(nFSubset)
+
+    def get_fQueryCostStep(self, _nFSubset):
+        '''Step function for querying feature'''
+        if _nFSubset > self.fThreshold:
+            return self.fCap
+        else:
+            return self.fQueryCost
+
+    def get_fQueryCostLinear(self, _nFSubset):
+        '''Linear function for querying feature'''
+        return max(
+            self.fQueryCost,
+            self.fQueryCost + self.fRate*(_nFSubset-self.fThreshold)
+        )
+
+    def get_fQueryCostQuadratic(self, _nFSubset):
+        '''Quadratic function for querying feature'''
+        if _nFSubset > self.fThreshold:
+            return self.fQueryCost + self.fRate*(_nFSubset-self.fThreshold)**2
+        else:
+            return self.fQueryCost
+
     def get_feature_mask(self):
         '''
         Get the (boolean) feature mask that indicates if a feature has

ltfmselector/ltfmselector.py

@@ -72,7 +72,9 @@ class LTFMSelector:
     def __init__(
             self, episodes, batch_size=256, tau=0.0005,
             eps_start=0.9, eps_end=0.05, eps_decay=1000,
-            fQueryCost=0.01, mQueryCost=0.01,
+            fQueryCost=0.01, fQueryFunction=None,
+            fThreshold=None, fCap=None, fRate=None,
+            mQueryCost=0.01,
             fRepeatQueryCost=1.0, p_wNoFCost=5.0, errorCost=1.0,
             pType="regression", regression_tol=0.5,
             regression_error_rounding=1,
@@ -105,7 +107,32 @@ class LTFMSelector:
             Rate of exponential decay
 
         fQueryCost : float
-            Cost of querying a feature
+            Cost of querying a feature.
+
+        fQueryFunction : None or {'step', 'linear', 'quadratic'}
+            User can also decide to progressively increase the cost of 
+            querying features in the following manner:
+            'step' : 
+                Every additional feature adds a fixed constant, determined 
+                by user.
+            'linear' : 
+                Cost of every additional feature linearly increases according    
+                to user-defined gradient
+            'quadratic' : 
+                Cost of every additional feature increases quadratically, 
+                according to a user-defined rate
+
+        fThreshold : None or int
+            If `fQueryFunction == {'step', 'linear', 'quadratic', 'exponential'}`
+            Threshold of number of features, before cost of recruiting 
+            increases
+            
+        fCap : None or float
+            If `fQueryFunction == {'step'}`, upper limit of penalty
+
+        fRate : None or float
+            If `fQueryFunction == {'linear', 'quadratic', 'exponential'}`, rate
+            individual cost functions
 
         mQueryCost : float
             Cost of querying a prediction model
@@ -196,6 +223,35 @@ class LTFMSelector:
 
         # Reward function
         self.fQueryCost = fQueryCost
+        self.fQueryFunction = fQueryFunction
+        self.fThreshold = fThreshold
+        self.fCap = fCap
+        self.fRate = fRate
+
+        # Options for progressive cost functions
+        if isinstance(self.fQueryFunction, str):
+            fQueryFunctions = ['step', 'linear', 'quadratic']
+            if not self.fQueryFunction in fQueryFunctions:
+                raise ValueError(
+                    f"{self.fQueryFunction} is not a valid option. Available " +
+                    f"options are {fQueryFunctions}"
+                )
+            else:
+                if not isinstance(fThreshold, int):
+                    raise ValueError("Parameter fThreshold must be an integer!")
+
+                if self.fQueryFunction == "step":
+                    if not (isinstance(fCap, float) or isinstance(fCap, int)):
+                        raise ValueError("Parameter fCap must be an int or float!")
+                    else:
+                        self.fCap = float(fCap)
+                else:
+                    if self.fQueryFunction in ["linear", "quadratic"]:
+                        if not (isinstance(fRate, float) or isinstance(fRate, int)):
+                            raise ValueError("Parameter fRate must be an int or float!")
+                        else:
+                            self.fRate = float(fRate)
+
         self.mQueryCost = mQueryCost
         self.fRepeatQueryCost = fRepeatQueryCost
         self.p_wNoFCost = p_wNoFCost
@@ -327,7 +383,7 @@ class LTFMSelector:
         # If user wants to save average computed action-value functions and
         # rewards of sampled batches
         if returnQ:
-            total_iterations = 10000000000
+            total_iterations = 16777216 # 2^24
             LearningValuesMatrix = np.zeros(
                 (total_iterations, 3), dtype=np.float32
             )
@@ -336,7 +392,9 @@ class LTFMSelector:
         # Initializing the environment
         env = Environment(
             self.X, self.y, self.background_dataset,
-            self.fQueryCost, self.mQueryCost,
+            self.fQueryCost, self.fQueryFunction,
+            self.fThreshold, self.fCap, self.fRate,
+            self.mQueryCost,
             self.fRepeatQueryCost, self.p_wNoFCost, self.errorCost,
             self.pType, self.regression_tol, self.regression_error_rounding,
             self.pModels, self.device, sample_weight=self.sample_weight,
@@ -484,6 +542,7 @@ class LTFMSelector:
             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.flush()
             writer.close()
 
         if returnQ:
@@ -517,7 +576,9 @@ class LTFMSelector:
         # Initializing the environment
         env = Environment(
             self.X, self.y, self.background_dataset,
-            self.fQueryCost, self.mQueryCost,
+            self.fQueryCost, self.fQueryFunction,
+            self.fThreshold, self.fCap, self.fRate,
+            self.mQueryCost,
             self.fRepeatQueryCost, self.p_wNoFCost, self.errorCost,
             self.pType, self.regression_tol, self.regression_error_rounding,
             self.pModels, self.device, **kwargs

{ltfmselector-0.1.12.dist-info → ltfmselector-0.2.0.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: ltfmselector
-Version: 0.1.12
+Version: 0.2.0
 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>

ltfmselector-0.2.0.dist-info/RECORD

@@ -0,0 +1,9 @@
+ltfmselector/__init__.py,sha256=lf3e90CNpEDvEmNZ-0iuoHOPsA7D-WN_opbBsTYLVEA,76
+ltfmselector/env.py,sha256=898o_g6-i0Rz5R-4WxZInf3xaxXHf58kPJId0KeewQM,16070
+ltfmselector/ltfmselector.py,sha256=zxGTLtuaoqdWbGxM8JmQES1_kGpNad1utRfDkepPoko,32329
+ltfmselector/py.typed,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
+ltfmselector/utils.py,sha256=VXYZSDm7x4s0p9F_58NLW8WQa3dxi0vHZewRy6miC2E,5438
+ltfmselector-0.2.0.dist-info/METADATA,sha256=76QDgOBLL81otMAwr9D-eNfviT5SY76Tf70-WNGIgyg,3020
+ltfmselector-0.2.0.dist-info/WHEEL,sha256=WLgqFyCfm_KASv4WHyYy0P3pM_m7J5L9k2skdKLirC8,87
+ltfmselector-0.2.0.dist-info/licenses/LICENSE,sha256=tmIDlkkp4a0EudXuGmeTdGjHjPhmmXkEMshACXLqX2w,1092
+ltfmselector-0.2.0.dist-info/RECORD,,

ltfmselector-0.1.12.dist-info/RECORD

@@ -1,9 +0,0 @@
-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,,