Update hyperparameters for larger MS2DeepScore model

iomega · Aug 17, 2023 · ec60c3a · ec60c3a
1 parent 34ebede
commit ec60c3a
Showing 1 changed file with 3 additions and 3 deletions.
diff --git a/ms2query/create_new_library/train_ms2deepscore.py b/ms2query/create_new_library/train_ms2deepscore.py
@@ -43,7 +43,7 @@ def train_ms2ds_model(training_spectra,
         reference_scores_df=tanimoto_df,
         dim=len(spectrum_binner.known_bins), # The number of bins created
         same_prob_bins=same_prob_bins,
-        num_turns=2,
+        num_turns=1,
         augment_noise_max=10,
         augment_noise_intensity=0.01)
 
@@ -58,9 +58,9 @@ def train_ms2ds_model(training_spectra,
         augment_removal_max=0, augment_removal_intensity=0, augment_intensity=0, augment_noise_max=0, use_fixed_set=True
     )
 
-    model = SiameseModel(spectrum_binner, base_dims=(500, 500), embedding_dim=200, dropout_rate=0.2)
+    model = SiameseModel(spectrum_binner, base_dims=(1000, 1000, 1000), embedding_dim=500, dropout_rate=0.2)
 
-    model.compile(loss='mse', optimizer=Adam(lr=0.001), metrics=["mae", tf.keras.metrics.RootMeanSquaredError()])
+    model.compile(loss='mse', optimizer=Adam(lr=0.0005), metrics=["mae", tf.keras.metrics.RootMeanSquaredError()])
 
     # Save best model and include early stopping
     checkpointer = ModelCheckpoint(filepath=output_model_file_name, monitor='val_loss', mode="min", verbose=1, save_best_only=True)