linkedin / detext

DeText: A Deep Neural Text Understanding Framework for Ranking and Classification Tasks
BSD 2-Clause "Simplified" License
1.26k stars 134 forks source link

expose bert dropout option #30

Closed xwli-chelsea closed 4 years ago

xwli-chelsea commented 4 years ago

Description

This pr add a param to control whether to apply dropouts in the attention and hidden layers in Bert. Previously the dropouts are not applied.

Fixes # (issue)

Please delete options that are not relevant.

List all changes

Please list all changes in the commit.

Added param use_bert_dropout to control this behavior, defaults to False to keep the same default behavior Modified logic in bert_model Added a Bert test in test_run_detext.py

Testing

local test

Test Configuration:

Checklist