Update modelA.py

modelA.py

@@ -32,15 +32,7 @@ dtype = torch.float32
 warnings.filterwarnings("ignore")
 logging.basicConfig(level=logging.ERROR)
 
-PATH = 'E:/hf'
-os.environ['HF_HOME'] = PATH
-os.environ['HF_DATASETS_CACHE'] = PATH
-os.environ['TORCH_HOME'] = PATH
-os.environ['TF_ENABLE_ONEDNN_OPTS'] = '0'
-os.environ["PYTORCH_CUDA_ALLOC_CONF"] = "expandable_segments:True"
-
 def get_activation(act: str) -> nn.Module:
-    """Get activation function by name."""
     act_map = {
         "gelu": nn.GELU(), 
         "relu": nn.ReLU(), 
@@ -205,7 +197,6 @@ def plot_waveform(x=None, w=None, p=None, per=None, sample_idx=0, sr=16000, hop_
     return fig
 
 def valid(default_value, *items):
-    """Get first non-None item"""
     for item in items:
         if item is not None:
             return item
@@ -715,8 +706,6 @@ class FocusWindow(nn.Module):
         return head_importance
 
     def forward(self, x, xa=None, mask=None, enc=None, layer=None, return_bias=False, return_head_weights=False, learn_lr=False):
-
-        print(f"🎯 FocusWindow running! Input: {x.shape}, Feature: {xa.shape if xa is not None else None}")
         
         q = self.q_proj(x)
         k = self.k_proj(x if xa is None else xa)
@@ -817,11 +806,8 @@ class AdaptiveAttentionLR(nn.Module):
         
     def forward(self, x, feature_data=None, mask=None):
         quality = self.quality_estimator(x.mean(dim=1))
-        
-        lr_factor = self.lr_predictor(x.mean(dim=1))
-        
+        lr_factor = self.lr_predictor(x.mean(dim=1)
         adaptive_lr = quality * lr_factor
-        
         return adaptive_lr, adaptive_lr
 
 class SmartSensorResidual(nn.Module):