batch bug fix of decoder_weight shape in transformers' model (#608)

* fix a bug of weight shape in bert model

* batch bug fix of decoder_weight shape

Author: yingyibiao

paddlenlp/transformers/bert/modeling.py

@@ -501,7 +501,7 @@ def __init__(self,
         self.activation = getattr(nn.functional, activation)
         self.layer_norm = nn.LayerNorm(hidden_size)
         self.decoder_weight = self.create_parameter(
-            shape=[hidden_size, vocab_size],
+            shape=[vocab_size, hidden_size],
             dtype=self.transform.weight.dtype,
             is_bias=False) if embedding_weights is None else embedding_weights
         self.decoder_bias = self.create_parameter(

paddlenlp/transformers/bigbird/modeling.py

@@ -569,7 +569,7 @@ def __init__(self,
         self.activation = getattr(nn.functional, activation)
         self.layer_norm = nn.LayerNorm(hidden_size, epsilon=1e-12)
         self.decoder_weight = self.create_parameter(
-            shape=[hidden_size, vocab_size],
+            shape=[vocab_size, hidden_size],
             dtype=self.transform.weight.dtype,
             is_bias=False) if embedding_weights is None else embedding_weights
         self.decoder_bias = self.create_parameter(

paddlenlp/transformers/ernie/modeling.py

@@ -645,7 +645,7 @@ def __init__(self,
         self.activation = getattr(nn.functional, activation)
         self.layer_norm = nn.LayerNorm(hidden_size)
         self.decoder_weight = self.create_parameter(
-            shape=[hidden_size, vocab_size],
+            shape=[vocab_size, hidden_size],
             dtype=self.transform.weight.dtype,
             is_bias=False) if embedding_weights is None else embedding_weights
         self.decoder_bias = self.create_parameter(

paddlenlp/transformers/gpt/modeling.py

@@ -858,7 +858,7 @@ class GPTLMHead(nn.Layer):
     def __init__(self, hidden_size, vocab_size, embedding_weights=None):
         super(GPTLMHead, self).__init__()
         self.decoder_weight = self.create_parameter(
-            shape=[hidden_size, vocab_size],
+            shape=[vocab_size, hidden_size],
             dtype=paddle.get_default_dtype(),
             is_bias=True) if embedding_weights is None else embedding_weights
 

paddlenlp/transformers/unified_transformer/modeling.py

@@ -372,7 +372,7 @@ def __init__(self,
         self.activation = getattr(nn.functional, activation)
         self.layer_norm = nn.LayerNorm(hidden_size)
         self.decoder_weight = self.create_parameter(
-            shape=[hidden_size, vocab_size],
+            shape=[vocab_size, hidden_size],
             dtype=self.transform.weight.dtype,
             is_bias=False) if embedding_weights is None else embedding_weights
         self.decoder_bias = self.create_parameter(