Commit e7b494f8 authored by alexeib's avatar alexeib Committed by Myle Ott
Browse files

default decoder_learned_pos for lm

parent 67ee6d1f

fairseq/models/transformer.py

+1 −0
Changes for fairseq/models/transformer.py: 1 added line, 0 removed lines.
Original line number Diff line number Diff line
@@ -533,6 +533,7 @@ def base_lm_architecture(args): 
    args.decoder_layers = getattr(args, 'decoder_layers', 6)

    args.decoder_attention_heads = getattr(args, 'decoder_attention_heads', 8)

    args.adaptive_softmax_cutoff = getattr(args, 'adaptive_softmax_cutoff', None)

    args.decoder_learned_pos = getattr(args, 'decoder_learned_pos', False)



    # The model training is not stable without this

    args.decoder_normalize_before = True