Welcome to mirror list, hosted at ThFree Co, Russian Federation.

github.com/stanfordnlp/stanza.git - Unnamed repository; edit this file 'description' to name the repository.
summaryrefslogtreecommitdiff
diff options
context:
space:
mode:
authorJohn Bauer <horatio@gmail.com>2022-08-31 10:36:54 +0300
committerJohn Bauer <horatio@gmail.com>2022-08-31 10:36:54 +0300
commitf257cd4bdb841aa5e69545040bdea8f8a10286c5 (patch)
tree982806f13c6bfc28d32a6dd66e9d7772cc605294
parent1169177e62acb17e362b864f914d58ee85ca5664 (diff)
Update to 0.0005 - less likely to go completely bad
-rw-r--r--stanza/models/classifier.py3
1 files changed, 2 insertions, 1 deletions
diff --git a/stanza/models/classifier.py b/stanza/models/classifier.py
index 137e3212..28ab9287 100644
--- a/stanza/models/classifier.py
+++ b/stanza/models/classifier.py
@@ -118,7 +118,8 @@ def convert_fc_shapes(arg):
# For the most part, these values are for the constituency parser.
# Only the WD for adadelta is originally for sentiment
-DEFAULT_LEARNING_RATES = { "adamw": 0.0002, "adadelta": 1.0, "sgd": 0.001, "adabelief": 0.00005, "madgrad": 0.001, "sgd": 0.001 }
+# Also LR for adadelta and madgrad
+DEFAULT_LEARNING_RATES = { "adamw": 0.0002, "adadelta": 1.0, "sgd": 0.001, "adabelief": 0.00005, "madgrad": 0.0005, "sgd": 0.001 }
DEFAULT_LEARNING_EPS = { "adabelief": 1e-12, "adadelta": 1e-6, "adamw": 1e-8 }
DEFAULT_LEARNING_RHO = 0.9
DEFAULT_MOMENTUM = { "madgrad": 0.9, "sgd": 0.9 }