From e560a12a64d46e0a65bea4224e255d171aadbeb5 Mon Sep 17 00:00:00 2001 From: alexeib Date: Thu, 17 May 2018 14:17:39 +0100 Subject: [PATCH] default dropout to correct value for big transformer --- fairseq/models/transformer.py | 1 + 1 file changed, 1 insertion(+) diff --git a/fairseq/models/transformer.py b/fairseq/models/transformer.py index bc8f2b1253..221405f5ce 100644 --- 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')