testadmin
/
Parakeet
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

Merge pull request #75 from PaddlePaddle/revert-74-reborn 

Revert "bug fix: apply dropout to logits before softmax"
This commit is contained in:
Feiyu Chan 2021-01-07 15:23:50 +08:00 committed by GitHub
parent 91c54575fe e88cbace1c
commit 61ac117df5
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
1 changed files with 1 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
parakeet/modules/attention.py
View File

@ -61,8 +61,8 @@ def scaled_dot_product_attention(q,
if mask is not None:
scaled_logit += paddle.scale((1.0 - mask), -1e9) # hard coded here
scaled_logit = F.dropout(scaled_logit, dropout, training=training)
attn_weights = F.softmax(scaled_logit, axis=-1)
attn_weights = F.dropout(attn_weights, dropout, training=training)
out = paddle.matmul(attn_weights, v)
return out, attn_weights