chore: Add whisper dropout to large turbo, up lr to 3e-5

alexandrainst · Oct 29, 2024 · 6443f2e · 6443f2e
1 parent 0594341
commit 6443f2e
Showing 1 changed file with 3 additions and 3 deletions.
diff --git a/config/model/whisper-large-turbo.yaml b/config/model/whisper-large-turbo.yaml
@@ -9,8 +9,8 @@ clean_text: true
 
 # Model hyperparameters
 sampling_rate: 16_000
-dropout: 0.0
-activation_dropout: 0.1
+dropout: 0.1
+activation_dropout: 0.0
 attention_dropout: 0.0
 mask_time_prob: 0.5
 mask_time_length: 10
@@ -20,4 +20,4 @@ layerdrop: 0.1  # NOTE: This will automatically be set to 0 in a multi-gpu setti
 max_length: 225
 
 # Model-specific optimisation hyperparameters
-learning_rate: 1e-5
+learning_rate: 3e-5