fix masked_lm for loading in pytext

Summary: lm_output_learned_bias doesn't exist when loading the model for fine-tuning

Reviewed By: jingfeidu

Differential Revision: D15579190

fbshipit-source-id: 45e8e193399943c89b77cc553d3d6d49b056e55a

fairseq/models/masked_lm.py

@@ -174,6 +174,7 @@ class MaskedLMEncoder(FairseqEncoder):
         self.activation_fn = utils.get_activation_fn(args.activation_fn)
         self.layer_norm = LayerNorm(args.encoder_embed_dim)
 
+        self.lm_output_learned_bias = None
         if self.load_softmax:
             self.lm_output_learned_bias = nn.Parameter(torch.zeros(self.vocab_size))
 
@@ -229,6 +230,7 @@ class MaskedLMEncoder(FairseqEncoder):
         elif self.embed_out is not None:
             x = self.embed_out(x)
 
+        if self.lm_output_learned_bias is not None:
             x = x + self.lm_output_learned_bias
         sentence_logits = None
         if self.sentence_projection_layer: