AbstractIntegratedModule 1.0.6__tar.gz → 1.0.7__tar.gz

{abstractintegratedmodule-1.0.6 → abstractintegratedmodule-1.0.7}/AbstractIntegratedModule.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: AbstractIntegratedModule
-Version: 1.0.6
+Version: 1.0.7
 Summary: Integrated Pipeline with Specialized Non-LLM AI Agent Framework for ARM64 architecture
 Author: Micro-Novelty
 Author-email: Micro-Novelty <hernikpuspita5@gmail.com>

{abstractintegratedmodule-1.0.6 → abstractintegratedmodule-1.0.7}/AbstractIntegratedModule.py

@@ -2122,12 +2122,13 @@ class LSTMCell:
         scale = np.sqrt(2.0 / (input_size + hidden_size))
         self.W = np.random.randn(4 * hidden_size, input_size + hidden_size) * scale
         self.b = np.zeros((4 * hidden_size,))
+        self.b[:hidden_size] = 1.0  # forget gate bias init to 1.0
 
         # Output projection: hidden → output
         self.Wy = np.random.randn(hidden_size, hidden_size) * scale
         self.by = np.zeros((hidden_size,))
 
-    # ── slicing helpers ──────────────────────
+    # ── slicing helpers utility functions ──────────────────────
     def _f(self, v): return v[:self.hidden_size]
     def _i(self, v): return v[self.hidden_size:2*self.hidden_size]
     def _g(self, v): return v[2*self.hidden_size:3*self.hidden_size]
@@ -2184,6 +2185,8 @@ class LSTMCell:
             o      = sigmoid(z[H3:])
 
             c_new  = f * c + i * g
+            c_new = np.clip(c_new, -10.0, 10.0)  # prevent overflow in tanh
+
             tanh_c = np.tanh(c_new)
             h_new  = o * tanh_c
 
@@ -2268,6 +2271,8 @@ class LSTMNetwork:
         self.pipeline     = pipeline
         self._trained     = False
 
+        self._grad_norm_history = []  # for logging gradient health
+
     # forward method to calculate proper weight for prediction and training.
     def forward(self, x_seq):
         # also Wy init only here, removed from train_step
@@ -2329,7 +2334,7 @@ class LSTMNetwork:
         self.by     -= lr * out_grads["dby"]
 
     # train step for each LSTM fitting method 
-    def train_step(self, x_seq, targets, lr=1e-3, AMR=None):
+    def train_step(self, x_seq, targets, lr=1e-3, AMR=None, log_grad_health=True):
         # accept precomputed AMR 
         if AMR is None:
             AME = self.pipeline.AME_Encoder(x_seq)
@@ -2338,6 +2343,23 @@ class LSTMNetwork:
         preds, hs, cs, cache = self.forward(x_seq)
         loss, dloss          = self.loss_mse(preds, targets, AMR)
         cell_grads, out_grads, _ = self.backward(dloss, hs, cache)
+
+        if log_grad_health:
+            grad_norm = np.sqrt(sum(np.sum(g**2) for g in {**cell_grads, **out_grads}.values()))
+            if not hasattr(self, '_grad_norm_history'):
+                self._grad_norm_history = []
+            self._grad_norm_history.append(float(grad_norm))
+
+            # flag genuinely pathological training, not just noise
+            if len(self._grad_norm_history) >= 10:
+                recent = self._grad_norm_history[-10:]
+                if np.mean(recent) < 1e-6:
+                    print('[⚠️] LSTM gradient norm near-zero for 10 steps — '
+                        'possible vanishing gradient, training may have stalled!')
+                elif np.mean(recent) > 100.0:
+                    print('[⚠️] LSTM gradient norm consistently large — '
+                        'clipping is doing heavy lifting, consider lowering learning rate!')
+
         self.update(cell_grads, out_grads, lr)
         return loss, preds
 
@@ -9303,7 +9325,7 @@ class IntegratedPipeline:
         self.transformer_d_model = 32
         self.transformer_heads = 4
         self.transformer_training_epochs = 100
-        self.lstm_training_epochs = 50
+        self.lstm_training_epochs = 100
         self.lstm_lr = 5e-2
         self.lstm_hidden_dim = 64