Skip to content

Commit 6534773

Browse files
authored
Dropout in DeLight Layers
1 parent f6df883 commit 6534773

File tree

1 file changed

+5
-1
lines changed

1 file changed

+5
-1
lines changed

fairseq/models/delight_transformer_lm.py

Lines changed: 5 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -398,11 +398,15 @@ def delight_transformer_lm_wiki103(args):
398398
scale_attn_drop = 0.1
399399
scale_attn_drop_d_m = round(scale_attn_drop / delta_model_dimension, 2)
400400
scale_attn_drop_d_m = bound_function(0, 0.1, scale_attn_drop_d_m)
401+
402+
scale_delight_drop = 0.1
403+
scale_delight_drop_d_m = round(scale_delight_drop / delta_model_dimension, 2)
404+
scale_delight_drop_d_m = bound_function(0, 0.1, scale_delight_drop_d_m)
401405

402406
args.dropout = getattr(args, "dropout", scale_dropout_d_m)
403407
args.delight_emb_dropout = getattr(args, "delight_emb_dropout", 0.1) # We used a fixed value
404408
args.attention_dropout = getattr(args, "attention_dropout", scale_attn_drop_d_m)
405-
args.delight_dropout = getattr(args, "delight_dropout", 0.0)
409+
args.delight_dropout = getattr(args, "delight_dropout", scale_delight_drop_d_m)
406410
args.pe_dropout = getattr(args, "pe_dropout", 0.1) # We used a fixed value
407411
args.activation_dropout = getattr(args, "activation_dropout", 0.0) # we didn't use it
408412
args.ffn_dropout = getattr(args, "ffn_dropout", scale_dropout_d_m)

0 commit comments

Comments
 (0)