fix

scripts/question_answering/run_squad.py

@@ -374,8 +374,8 @@ def untune_params(model, untunable_depth, not_included=[]):
         A list or parameter names that not included in the untunable parameters
     """
     all_layers = model.backbone.encoder.all_encoder_layers
-    for _, v in model.collect_params('.*embed*').items():
-        model.grad_req = 'null'
+    for _, value in model.collect_params('.*embed*').items():
+        value.grad_req = 'null'
 
     for layer in all_layers[:untunable_depth]:
         for key, value in layer.collect_params().items():

src/gluonnlp/models/electra.py

@@ -392,9 +392,8 @@ def apply_layerwise_decay(self, layerwise_decay, not_included=[]):
         # consider the task specific finetuning layer as the last layer, following with pooler
         # In addition, the embedding parameters have the smaller learning rate based on this setting.
         max_depth = self.num_layers
-        for key, value in self.collect_params().items():
-            if 'embed' in key:
-                value.lr_mult = layerwise_decay**(max_depth + 1)
+        for _, value in self.collect_params('.*embed*').items():
+            value.lr_mult = layerwise_decay**(max_depth + 1)
 
         for (layer_depth, layer) in enumerate(self.encoder.all_encoder_layers):
             layer_params = layer.collect_params()