default dropout to correct value for big transformer

facebookresearch · May 17, 2018 · e560a12 · e560a12
1 parent c903798
commit e560a12
Showing 1 changed file with 1 addition and 0 deletions.
diff --git a/fairseq/models/transformer.py b/fairseq/models/transformer.py
@@ -435,6 +435,7 @@ def transformer_vaswani_wmt_en_de_big(args):
     args.decoder_ffn_embed_dim = 4096
     args.decoder_layers = 6
     args.decoder_attention_heads = 16
+    args.dropout = 0.3
 
 
 @register_model_architecture('transformer', 'transformer_wmt_en_de_big')