bilstm trying to deal with NaN by creating a shallower network

config.yml

@@ -58,12 +58,12 @@ LinearRegression:
   Initialization: init_he
 
 BidirectionalLSTM:
-  HiddenSize: 32
-  NumLayers: 4
+  HiddenSize: 16
+  NumLayers: 2
   LSTMDropout: 0
   FFNDropout: 0.2
-  NumFFN: 8
-  Initialization: None
+  NumFFN: 3
+  Initialization: init_he
 
 RNN:
   HiddenSize: 32