AbstractIntegratedModule 0.8.6__tar.gz → 0.8.7__tar.gz

{abstractintegratedmodule-0.8.6 → abstractintegratedmodule-0.8.7}/AbstractIntegratedModule.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: AbstractIntegratedModule
-Version: 0.8.6
+Version: 0.8.7
 Summary: Library for Advanced Integrated Non-LLM AI Models - Optimized Backend Framework For Non-LLM AI Agent
 Author: Micro-Novelty
 Author-email: hernikpuspita5@gmail.com
@@ -42,7 +42,7 @@ https://github.com/Micro-Novelty/IntegratedPipeline-Specialized-Non-LLM-AI-Agent
 #### Note: The README here you are reading is a direct copy from my README Repository, to download the necessary files, you can visit my Repository with the provided link above.
 
 ### Library Short Description:
-- Development Stage: 0.8.6 Official Release.
+- Development Stage: 0.8.7 Official Release.
 - Maintainer: Micro-Novelty.
 - library Source-Code is Open-sourced on github.
 - Purpose: Specifically Designed for providing Non-LLM AI Agent Framework for edge Devices, Optimized for ARM64 architecture.
@@ -80,10 +80,10 @@ https://github.com/Micro-Novelty/IntegratedPipeline-Specialized-Non-LLM-AI-Agent
    - Transformer Optimized using Cython, to reduce Memory overhead and Reduce CPU Usage, With Reduced Training Time.
 ______________
   - Changelog:
-     - v0.8.6:
+     - v0.8.7:
         - [=] New features: 
-        - Added new Optimization module using Cython maths for dynamic weighted ensemble function.
-        - Added Penalty for wrong prediction in both single advanced prediction and advanced batch prediction to reduce repetitive errors.
+        - Added new continuos predictive correction for MLP class to reduce inflated confidence despite outputting wrong answer.
+        - Reduced the frequency of penalty based calibration in advanced prediction method to prevent the models from being too conservative.
         - Fixed P2P Vulnerabilities:
            - Empty allowed_ips in PRODUCTION/HARDENED mode now DENIES external IPs instead of allowing all
            - Startup validation warns about dangerous config combinations

{abstractintegratedmodule-0.8.6 → abstractintegratedmodule-0.8.7}/AbstractIntegratedModule.py

@@ -1716,7 +1716,11 @@ class MLP:
         self.layers2 = []
         self.lr = 0.1
         self.feed_layers = []
-    
+
+        self.error_counts = None
+        self.pred_counts = None
+        self.error_decay = None
+
         self.softmax = SoftmaxOutput()
       
 
@@ -1735,7 +1739,7 @@ class MLP:
     def forward(self, x):
         for layer in self.layers:
             x = layer.forward(x)
-            
+   
         return self.softmax.forward(x)
 
 
@@ -1757,7 +1761,40 @@ class MLP:
 
         return calibration
 
+    def continuos_predictive_correction(self, manager, prob, predicted_index):
+        eps = 1e-5
+
+        error_counts = manager.error_counts
+        pred_counts = manager.pred_counts
+        decay = manager.decay
+        label_map = manager.label_map
+
+        self.error_counts = error_counts
+        self.pred_counts = pred_counts
+        self.error_decay = decay
+
+        try:
+            self.pred_counts[predicted_index] += 1.0
+            n_classes = len(label_map)
+
+            for c in range(n_classes):
+                if self.pred_counts[c] > 0:
+                    error_rate    = self.error_counts[c] / (self.pred_counts[c] + 1e-8)
+                    # sigmoid-shaped dampening — never goes negative
+                    # error_rate=0.0 → multiplier=1.0 (no change)
+                    # error_rate=0.5 → multiplier≈0.67
+                    # error_rate=1.0 → multiplier≈0.5
+                    reputation    = 1.0 / (1.0 + error_rate)
+                    prob[c]     *= reputation  
+                    
+            prob_sum = prob.sum()
+            if prob_sum > 1e-8:
+                prob /= prob_sum
 
+        except Exception as e:
+            print(f'[!] Cant check and calibrate probs based on penalty due to: {e}')    
+        
+        return prob 
 
 
     def focused_backward(self, grad, lr, AME, anisotropy):
@@ -2893,7 +2930,7 @@ class WeightedEnsemblePredictor:
         AME = self.pipeline.model2.AME_Encoder(input_ids)
 
         trans_probs, attn_weights = self.pipeline.model2.forward(input_ids, AME=AME, embedded=embedded)
-        mlp_probs = self.pipeline.mlp.forward(X_mlp)
+        mlp_probs = self.pipeline.model3.forward(X_mlp)
         lstm_probs, lstm_weight_hint = self._get_lstm_probs(input_ids, X_mlp, label_bins=label_bins)
 
         established_agreement = self.query_node._establish_node_connection("PredictEnsemble")
@@ -13970,8 +14007,6 @@ class PipelinePredictionManager:
         4. User home directory
         5. Common data folders (Downloads, Documents, Desktop)
         """
-        import os
-        import sys
 
         # build candidate paths
         candidates = []
@@ -14701,9 +14736,10 @@ class PipelinePredictionManager:
                     print(f'[!] Similarity: {cached['similarity']} is low, Cannot pick label due to low certainty, Initiating advanced prediction...')
             else:
                 print('[=] No verified output from cache available that matched samples, starting advanced prediction...')
-
-            target_probs = self.pipeline.predict_proba(input_ids, X, type='Hybrid', embedded=True)
-            target_probs = target_probs[:mlp_probs.shape[0], :mlp_probs.shape[1]] 
+                target_probs = self.pipeline.predict_proba(input_ids, X, type='Hybrid', embedded=True)
+                target_probs = target_probs[:mlp_probs.shape[0], :mlp_probs.shape[1]] 
+                target_probs = self.pipeline.model3.continuos_predictive_correction(self, target_probs, mlp_pred_indices)  
+            
             target_pred_indices = np.argmax(target_probs, axis=1)    
 
             if self.pipeline.cache and 'label_bins' in self.pipeline.cache:
@@ -14716,13 +14752,6 @@ class PipelinePredictionManager:
             results = []
             attention_data = [] if return_attention else None
 
-            target_probs = self.pipeline.predict_proba(input_ids, X, type='Hybrid', embedded=True)
-            target_probs = target_probs[:mlp_probs.shape[0], :mlp_probs.shape[1]] 
-            target_pred_indices = np.argmax(target_probs, axis=1)          
-
-            results = []
-            attention_data = [] if return_attention else None
-
             for i, title in enumerate(titles):
                 # Parse input
                 if isinstance(title, tuple):
@@ -14911,7 +14940,6 @@ class PipelinePredictionManager:
                     print('|| Uncertain advanced prediction, requesting peer assistance if allowed...')
                     final_probs = self.pipeline._handle_distributed_connections(final_probs, attn_weights, input_ids, agreement) 
 
-                    final_probs = self.calibration_penalized_check(final_probs, final_class_idx) 
                     final_idx = final_probs[0].argmax()
                     original_idx = final_idx
 
@@ -14934,7 +14962,6 @@ class PipelinePredictionManager:
                     print('[||] Iniating local peer output search in database for best output...')
                     final_probs = self.pipeline.distribution._handle_peer_agent_request(final_probs, attn_weights, input_ids, type='DevicePeer', agreement=agreement)
 
-                    final_probs = self.calibration_penalized_check(final_probs, final_class_idx)
                     final_idx = final_probs[0].argmax()
                     original_idx = final_idx              
 
@@ -14955,7 +14982,6 @@ class PipelinePredictionManager:
                         print("\n[⚠️] Low confidence detected, but both models don't agree. Using calibrated probabilities for final decision to ensure robustness.")
                         final_probs = self.pipeline.hybrid_prediction(rules, input_ids, dataset)
 
-                        final_probs = self.calibration_penalized_check(final_probs, final_class_idx)
                         final_idx = final_probs[0].argmax()
                         original_idx = final_idx
                         
@@ -14967,7 +14993,6 @@ class PipelinePredictionManager:
                         print('[🎯] Stable confidence established, But both Models doesnt Agree, Re-evaluating...')   
                         final_probs = self.pipeline.hybrid_prediction(rules, input_ids, dataset)
 
-                        final_probs = self.calibration_penalized_check(final_probs, final_class_idx)
                         final_idx = final_probs[0].argmax()
                         original_idx = final_idx
                         if final_idx > len(reverse_map):
@@ -14986,7 +15011,6 @@ class PipelinePredictionManager:
                         input_forward = sequence_ids if sequence_ids is not None else input_ids
                         final_probs, details = self.pipeline.ensemble.predict_ensemble(input_forward, X, y, method='dynamic', embedded=True)
 
-                        final_probs = self.calibration_penalized_check(final_probs, final_class_idx)
                         final_idx = final_probs[0].argmax()
                         original_idx = final_idx 
 
@@ -15002,7 +15026,6 @@ class PipelinePredictionManager:
                             confidence = float(final_probs[0][len(reverse_map)-1]) if isinstance(final_probs[0], (float, int)) else 0.0             
                     else:
                         input_forward = sequence_ids if sequence_ids is not None else input_ids
-                        final_probs = self.calibration_penalized_check(final_probs, final_class_idx)
                         final_idx = final_probs[0].argmax() if final_probs is not None else target_probs[0].argmax()
                         
                         original_idx = final_idx 
@@ -15028,7 +15051,6 @@ class PipelinePredictionManager:
                                 confidence = self.pipeline.confidence_threshold
 
             elif confidence < self.pipeline.confidence_threshold and not self.pipeline.agreement and not results[0].get('models_agree', True):
-                final_probs = self.calibration_penalized_check(final_probs, final_class_idx)
                 if trans_probs is not None:
                     prob_entropy = self.calculate_entropy(final_probs)
                     normalized_entropy = prob_entropy / np.log(prob_entropy.shape[-1]) if prob_entropy.shape[-1] > 1 else 0
@@ -15075,7 +15097,6 @@ class PipelinePredictionManager:
                         final_probs = target_probs.copy() 
             else:
                 print("\n[🎯] Using initial Regular final prediction as final decision.")
-                final_probs = self.calibration_penalized_check(final_probs, final_class_idx)
                 final_idx = final_probs[0].argmax()
 
                 if final_idx > len(reverse_map):
@@ -15101,7 +15122,6 @@ class PipelinePredictionManager:
 
             try:
                 consecutive_probs = self.pipeline.distribution._handle_peer_agent_request(target_probs, attn_weights, input_ids, type='DevicePeer', agreement=agreement)
-                consecutive_probs = self.calibration_penalized_check(consecutive_probs, final_class_idx)
                 sec_final_idx = consecutive_probs[0].argmax()
 
                 if sec_final_idx > len(reverse_map):
@@ -15252,6 +15272,7 @@ class PipelinePredictionManager:
                     correct += 1
                 else:
                     self.error_counts[predicted_index] += 1.0
+    
             
             # Agreement indicator
             agree_symbol = "✓" if result.get('models_agree', True) else "⚠️"