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>2022-03-18 03:07:07 +0300
committerJohn Bauer <horatio@gmail.com>2022-06-21 09:37:59 +0300
commita59ce17e72682d577f0928343a95997eee7e6662 (patch)
treeec6cc260ef60d5552ccfd59f5c1ba9570beaa3c6
parent5e9a9b21ae3f896c744b928b4d919999d732dbe7 (diff)
Experiment with not doing weight decay at all for pattncon_lattn
-rw-r--r--stanza/models/constituency/lstm_model.py4
1 files changed, 2 insertions, 2 deletions
diff --git a/stanza/models/constituency/lstm_model.py b/stanza/models/constituency/lstm_model.py
index 7b79d478..8ef718c9 100644
--- a/stanza/models/constituency/lstm_model.py
+++ b/stanza/models/constituency/lstm_model.py
@@ -460,8 +460,8 @@ class LSTMModel(BaseModel, nn.Module):
def is_low_decay_parameter(self, name):
#if name.find("_embedding") >= 0:
# return True
- #if name.startswith("partitioned_transformer_module"):
- # return True
+ if name.startswith("partitioned_transformer_module"):
+ return True
#if name.find(".bias") >= 0:
# return True
if name.find(".norm") >= 0 or name.find(".layer_norm") >= 0: