ltfmselector 0.2.1__py3-none-any.whl → 0.2.2__py3-none-any.whl

ltfmselector/logger.py

@@ -0,0 +1,39 @@
+import numpy as np
+
+class Logger:
+    def __init__(self, env, max_steps=500):
+        self.state_dim = env.X.shape[1]*2 + 1
+        self.max_steps = max_steps
+        
+        # Pre-allocate the "scratchpad" for the current episode
+        self.s_buffer = np.zeros((self.max_steps, self.state_dim), dtype=np.float32)
+        self.a_buffer = np.zeros((self.max_steps,), dtype=np.int32)
+        self.ptr = 0
+        
+        # Final storage lists
+        self.all_states = []
+        self.all_actions = []
+
+    def log_step(self, obs, action):
+        if self.ptr < self.max_steps:
+            self.s_buffer[self.ptr] = obs
+            self.a_buffer[self.ptr] = action
+            self.ptr += 1
+
+    def log_episode(self):
+        """Internal helper to copy the current buffer into the main list."""
+        if self.ptr > 0:
+            self.all_states.append(self.s_buffer[:self.ptr].copy())
+            self.all_actions.append(self.a_buffer[:self.ptr].copy())
+            self.ptr = 0
+
+    def save_data(self, filename):
+        """Saves all completed and currently-running episodes to disk."""
+        # Capture the current episode if it's mid-run
+        self.log_episode()
+        np.savez_compressed(
+            filename, 
+            states=np.array(self.all_states, dtype=object),
+            actions=np.array(self.all_actions, dtype=object)
+        )
+        print(f"Saved {len(self.all_states)} episodes (states + actions) to {filename}")

ltfmselector/ltfmselector.py

@@ -11,6 +11,7 @@ import numpy as np
 import pandas as pd
 from collections import defaultdict
 
+from .logger import Logger
 from .env import Environment
 from .utils import ReplayMemory, DQN, Transition
 
@@ -110,25 +111,25 @@ class LTFMSelector:
             Cost of querying a feature.
 
         fQueryFunction : None or {'step', 'linear', 'quadratic'}
-            User can also decide to progressively increase the cost of 
+            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 
+            'step' :
+                Every additional feature adds a fixed constant, determined
                 by user.
-            'linear' : 
-                Cost of every additional feature linearly increases according    
+            'linear' :
+                Cost of every additional feature linearly increases according
                 to user-defined gradient
-            'quadratic' : 
-                Cost of every additional feature increases quadratically, 
+            '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 
+            Threshold of number of features, before cost of recruiting
             increases
-            
+
         fCap : None or float
-            If `fQueryFunction == {'step', 'linear', 'quadratic'}`, upper 
+            If `fQueryFunction == {'step', 'linear', 'quadratic'}`, upper
             limit of penalty
 
         fRate : None or float
@@ -284,7 +285,7 @@ class LTFMSelector:
     def fit(
             self, X, y, loss_function='mse', sample_weight=None,
             agent_neuralnetwork=None, lr=1e-5, returnQ=False,
-            monitor=False, background_dataset=None, **kwargs
+            monitor=False, log=False, background_dataset=None, **kwargs
     ):
         '''
         Initializes the environment and agent, then trains the agent to select
@@ -323,15 +324,18 @@ class LTFMSelector:
 
         returnQ : bool
             Return average computed action-value functions and rewards of
-            the sampled batches, as a (<total_iterations>, 3) matrix. The 
+            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
+            Run `tensorboard --logdir=runs` in the terminal to monitor the
             progression of the action-value function.
 
+        log : bool
+            Log states and actions for research purposes
+
         background_dataset : None or pd.DataFrame
             If None, numerical features will be assumed when computing the
             background dataset.
@@ -403,6 +407,9 @@ class LTFMSelector:
         )
         env.reset()
 
+        if log:
+            logger = Logger(env, max_steps=self.max_timesteps)
+
         # Initializing length of state and actions as public fields for
         # loading the model later
         self.state_length = len(env.state)
@@ -465,6 +472,9 @@ class LTFMSelector:
                 # - reward
                 observation, reward, terminated = env.step(action.item())
 
+                if log:
+                    logger.log_step(observation, action)
+
                 if terminated:
                     next_state = None
                 else:
@@ -529,6 +539,9 @@ class LTFMSelector:
                     )
                     break
 
+            if log:
+                logger.log_episode()
+
             # Saving trained policy network intermediately
             if not self.checkpoint_interval is None:
                 if (i_episode + 1) % self.checkpoint_interval == 0:
@@ -539,6 +552,9 @@ class LTFMSelector:
                 self.policy_network_checkpoints[self.episodes] =\
                     self.policy_net.state_dict()
 
+        if log:
+            logger.save_data("ActionStates_fromFit.npz")
+
         if monitor:
             writer.add_scalar("Metrics/Average_QValue", _res[0], monitor_count)
             writer.add_scalar("Metrics/Average_Reward", _res[1], monitor_count)
@@ -556,7 +572,7 @@ class LTFMSelector:
         else:
             return doc
 
-    def predict(self, X_test, **kwargs):
+    def predict(self, X_test, log=False, **kwargs):
         '''
         Use trained agent to select features and a suitable prediction model
         to predict the target/class, given X_test.
@@ -566,6 +582,9 @@ class LTFMSelector:
         X_test : pd.DataFrame
             Test samples
 
+        log : bool
+            Log states and actions for research purposes
+
         Returns
         -------
         y_pred : array
@@ -584,6 +603,8 @@ class LTFMSelector:
             self.pType, self.regression_tol, self.regression_error_rounding,
             self.pModels, self.device, **kwargs
         )
+        if log:
+            logger = Logger(env, max_steps=self.max_timesteps)
 
         # Create dictionary to save information per episode
         doc_test = defaultdict(dict)
@@ -608,6 +629,9 @@ class LTFMSelector:
 
                 observation, reward, terminated = env.step(action.item())
 
+                if log:
+                    logger.log_step(observation, action)
+
                 if terminated:
                     next_state = None
                 else:
@@ -637,6 +661,12 @@ class LTFMSelector:
                     y_pred[i] = env.y_pred
                     break
 
+            if log:
+                logger.log_episode()
+
+        if log:
+            logger.save_data("ActionStates_fromPredict.npz")
+
         return y_pred, doc_test
 
     def select_action(self, state, env):

{ltfmselector-0.2.1.dist-info → ltfmselector-0.2.2.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: ltfmselector
-Version: 0.2.1
+Version: 0.2.2
 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>
@@ -29,9 +29,11 @@ License-File: LICENSE
 Requires-Python: >=3.12
 Requires-Dist: gymnasium>=1.1.1
 Requires-Dist: matplotlib>=3.10.1
+Requires-Dist: moviepy>=2.2.1
 Requires-Dist: numpy>=2.2.4
 Requires-Dist: openpyxl>=3.1.5
 Requires-Dist: pandas>=2.2.3
+Requires-Dist: pygame>=2.6.1
 Requires-Dist: scikit-learn<1.6
 Requires-Dist: seaborn>=0.13.2
 Requires-Dist: tensorboard>=2.20.0

ltfmselector-0.2.2.dist-info/RECORD

@@ -0,0 +1,10 @@
+ltfmselector/__init__.py,sha256=lf3e90CNpEDvEmNZ-0iuoHOPsA7D-WN_opbBsTYLVEA,76
+ltfmselector/env.py,sha256=F0NycqUkNn-p2zC1EPdds73__G8yyMW5f9F93yPDHTA,16371
+ltfmselector/logger.py,sha256=of5fgVmh1CctRE3ckjO0R_Wo6WFg8Mg1RxAI84oPKVA,1448
+ltfmselector/ltfmselector.py,sha256=jBd7dj8_l7PdTtkm1lALTzc6JX9Q954eKZrSk59EfII,33151
+ltfmselector/py.typed,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
+ltfmselector/utils.py,sha256=VXYZSDm7x4s0p9F_58NLW8WQa3dxi0vHZewRy6miC2E,5438
+ltfmselector-0.2.2.dist-info/METADATA,sha256=8AFbJExeTNTjSUvNbuw_wDo4NKfWkj0huNY599I4i20,3079
+ltfmselector-0.2.2.dist-info/WHEEL,sha256=WLgqFyCfm_KASv4WHyYy0P3pM_m7J5L9k2skdKLirC8,87
+ltfmselector-0.2.2.dist-info/licenses/LICENSE,sha256=tmIDlkkp4a0EudXuGmeTdGjHjPhmmXkEMshACXLqX2w,1092
+ltfmselector-0.2.2.dist-info/RECORD,,

ltfmselector-0.2.1.dist-info/RECORD

@@ -1,9 +0,0 @@
-ltfmselector/__init__.py,sha256=lf3e90CNpEDvEmNZ-0iuoHOPsA7D-WN_opbBsTYLVEA,76
-ltfmselector/env.py,sha256=F0NycqUkNn-p2zC1EPdds73__G8yyMW5f9F93yPDHTA,16371
-ltfmselector/ltfmselector.py,sha256=GdxazN6JG_ELZ7a7x6bbBCVDsgmdRWOsbi-TnRFNk8Y,32354
-ltfmselector/py.typed,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
-ltfmselector/utils.py,sha256=VXYZSDm7x4s0p9F_58NLW8WQa3dxi0vHZewRy6miC2E,5438
-ltfmselector-0.2.1.dist-info/METADATA,sha256=mHEsAKWtYYsGOQyRxHX5f8fAaSXePDDklL_CLzMay9A,3020
-ltfmselector-0.2.1.dist-info/WHEEL,sha256=WLgqFyCfm_KASv4WHyYy0P3pM_m7J5L9k2skdKLirC8,87
-ltfmselector-0.2.1.dist-info/licenses/LICENSE,sha256=tmIDlkkp4a0EudXuGmeTdGjHjPhmmXkEMshACXLqX2w,1092
-ltfmselector-0.2.1.dist-info/RECORD,,