changing l2_regularization_coeff to weight_decay

matsengrp · Jun 13, 2024 · ecb19fc · ecb19fc
1 parent 7e806b4
commit ecb19fc
Show file tree

Hide file tree

Showing 4 changed files with 12 additions and 12 deletions.
diff --git a/data/cnn_joi_sml-shmoof_small.yml b/data/cnn_joi_sml-shmoof_small.yml
@@ -11,6 +11,6 @@ model_hyperparameters:
   kmer_length: 3
 serialization_version: 0
 training_hyperparameters:
-  l2_regularization_coeff: 1.0e-06
+  weight_decay: 1.0e-06
   learning_rate: 0.1
   min_learning_rate: 1.0e-06
diff --git a/netam/dnsm.py b/netam/dnsm.py
@@ -424,7 +424,7 @@ def to_crepe(self):
                 "batch_size",
                 "learning_rate",
                 "min_learning_rate",
-                "l2_regularization_coeff",
+                "weight_decay",
             ]
         }
         encoder = framework.PlaceholderEncoder()
@@ -448,7 +448,7 @@ def burrito_of_model(
         batch_size=1024,
         learning_rate=0.1,
         min_learning_rate=1e-4,
-        l2_regularization_coeff=1e-6,
+        weight_decay=1e-6,
     ):
         model.to(device)
         burrito = DNSMBurrito(
@@ -458,6 +458,6 @@ def burrito_of_model(
             batch_size=batch_size,
             learning_rate=learning_rate,
             min_learning_rate=min_learning_rate,
-            l2_regularization_coeff=l2_regularization_coeff,
+            weight_decay=weight_decay,
         )
         return burrito
diff --git a/netam/framework.py b/netam/framework.py
@@ -376,7 +376,7 @@ def __init__(
         batch_size=1024,
         learning_rate=0.1,
         min_learning_rate=1e-4,
-        l2_regularization_coeff=1e-6,
+        weight_decay=1e-6,
         name="",
     ):
         """
@@ -394,7 +394,7 @@ def __init__(
         self.batch_size = batch_size
         self.learning_rate = learning_rate
         self.min_learning_rate = min_learning_rate
-        self.l2_regularization_coeff = l2_regularization_coeff
+        self.weight_decay = weight_decay
         self.name = name
         self.reset_optimization()
         self.bce_loss = nn.BCELoss()
@@ -425,7 +425,7 @@ def reset_optimization(self, learning_rate=None):
             self.optimizer_name,
             self.model.parameters(),
             lr=learning_rate,
-            weight_decay=self.l2_regularization_coeff,
+            weight_decay=self.weight_decay,
         )
         self.scheduler = ReduceLROnPlateau(
             self.optimizer, mode="min", factor=0.5, patience=10
@@ -763,7 +763,7 @@ def __init__(
         batch_size=1024,
         learning_rate=0.1,
         min_learning_rate=1e-4,
-        l2_regularization_coeff=1e-6,
+        weight_decay=1e-6,
         name="",
     ):
         super().__init__(
@@ -774,7 +774,7 @@ def __init__(
             batch_size=batch_size,
             learning_rate=learning_rate,
             min_learning_rate=min_learning_rate,
-            l2_regularization_coeff=l2_regularization_coeff,
+            weight_decay=weight_decay,
             name=name,
         )
 
@@ -829,7 +829,7 @@ def to_crepe(self):
             for key in [
                 "learning_rate",
                 "min_learning_rate",
-                "l2_regularization_coeff",
+                "weight_decay",
             ]
         }
         encoder = KmerSequenceEncoder(

diff --git a/netam/hyper_burrito.py b/netam/hyper_burrito.py
@@ -181,7 +181,7 @@ def burrito_of_model(
         batch_size=1024,
         learning_rate=0.1,
         min_learning_rate=1e-4,
-        l2_regularization_coeff=1e-6,
+        weight_decay=1e-6,
     ):
         burrito = SHMBurrito(
             self.train_dataset,
@@ -190,6 +190,6 @@ def burrito_of_model(
             batch_size=batch_size,
             learning_rate=learning_rate,
             min_learning_rate=min_learning_rate,
-            l2_regularization_coeff=l2_regularization_coeff,
+            weight_decay=weight_decay,
         )
         return burrito