70a188776f7470c838dd22b1636462b75573a734,src/gluonnlp/models/bert.py,BertTransformer,__init__,#BertTransformer#Any#Any#Any#Any#Any#Any#Any#Any#Any#Any#Any#Any#Any#Any#Any#,110

Before Change

                                              weight_initializer=weight_initializer,
                                              bias_initializer=bias_initializer,
                                              activation=activation,
                                              prefix="{}_".format(layer_idx)))

    def hybrid_forward(self, F, data, valid_length):
        

After Change

        self.embed_layer_norm = nn.LayerNorm(epsilon=self.layer_norm_eps)
        self.embed_dropout = nn.Dropout(hidden_dropout_prob)
        // Construct token type embedding
        self.token_type_embed = nn.Embedding(input_dim=num_token_types,
                                             output_dim=units,
                                             weight_initializer=weight_initializer)
        self.token_pos_embed = PositionalEmbedding(units=units,
                                                   max_length=max_length,
                                                   dtype=self._dtype,
                                                   method=pos_embed_type)
        if self.use_pooler:
            // Construct pooler
            self.pooler = nn.Dense(units=units,
                                   in_units=units,
                                   flatten=False,
                                   activation="tanh",
                                   weight_initializer=weight_initializer,

Instances