Merge pull request #66 from matchms/fix_and_expand_architecture

Fix and expand architecture choices

Author: florian-huber

CHANGELOG.md

@@ -7,12 +7,16 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## [Unreleased]
 
+## Changed
+
+- Allow users to define L1 and L2 regularization of `SiameseModel` [#67](https://github.com/matchms/ms2deepscore/issues/67)
+- Allow users to define number and size of `SiameseModel` [#64](https://github.com/matchms/ms2deepscore/pull/64)
+
 ## [0.1.2] - 2021-03-05
 
 ## Added
 
 - `create_confusion_matrix_plot` in `plotting` [#58](https://github.com/matchms/ms2deepscore/pull/58)
-- Allow users to define number and size of `SiameseModel` [#64](https://github.com/matchms/ms2deepscore/pull/64)
 
 ## [0.1.1] - 2021-02-09
 
@@ -32,6 +36,6 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 - This is the initial version of MS2DeepScore
 
 [Unreleased]: https://github.com/matchms/ms2deepscore/compare/0.1.2...HEAD
-[0.1.2]: https://github.com/matchms/ms2deepscore/releases/tag/0.1.1...0.1.2
-[0.1.1]: https://github.com/matchms/ms2deepscore/releases/tag/0.1.0...0.1.1
+[0.1.2]: https://github.com/matchms/ms2deepscore/compare/0.1.1...0.1.2
+[0.1.1]: https://github.com/matchms/ms2deepscore/compare/0.1.0...0.1.1
 [0.1.0]: https://github.com/matchms/ms2deepscore/releases/tag/0.1.0

ms2deepscore/models/SiameseModel.py

@@ -45,6 +45,8 @@ def __init__(self,
                  base_dims: Tuple[int, ...] = (600, 500, 500),
                  embedding_dim: int = 400,
                  dropout_rate: float = 0.5,
+                 l1_reg: float = 1e-6,
+                 l2_reg: float = 1e-6,
                  keras_model: keras.Model = None):
         """
         Construct SiameseModel
@@ -59,7 +61,11 @@ def __init__(self,
         embedding_dim
             Dimension of the embedding (i.e. the output of the base model)
         dropout_rate
-            Dropout rate to be used in the base model
+            Dropout rate to be used in the base model.
+        l1_reg
+            L1 regularization rate. Default is 1e-6.
+        l2_reg
+            L2 regularization rate. Default is 1e-6.
         keras_model
             When provided, this keras model will be used to construct the SiameseModel instance.
             Default is None.
@@ -75,7 +81,9 @@ def __init__(self,
             self.base = self._get_base_model(input_dim=self.input_dim,
                                              dims=base_dims,
                                              embedding_dim=embedding_dim,
-                                             dropout_rate=dropout_rate)
+                                             dropout_rate=dropout_rate,
+                                             l1_reg=l1_reg,
+                                             l2_reg=l2_reg)
             # Create head model
             self.model = self._get_head_model(input_dim=self.input_dim,
                                               base_model=self.base)
@@ -100,22 +108,24 @@ def save(self, filename: Union[str, Path]):
     def _get_base_model(input_dim: int,
                         dims: Tuple[int, ...] = (600, 500, 500),
                         embedding_dim: int = 400,
-                        dropout_rate: float = 0.25):
+                        dropout_rate: float = 0.25,
+                        l1_reg: float = 1e-6,
+                        l2_reg: float = 1e-6):
+        # pylint: disable=too-many-arguments
         model_input = keras.layers.Input(shape=input_dim, name='base_input')
         for i, dim in enumerate(dims):
             if i == 0:
                 model_layer = keras.layers.Dense(dim, activation='relu', name='dense'+str(i+1),
-                                               kernel_regularizer=keras.regularizers.l1_l2(l1=1e-6, l2=1e-6))(
+                                               kernel_regularizer=keras.regularizers.l1_l2(l1=l1_reg, l2=l2_reg))(
                    model_input)
             else:
-                model_layer = keras.layers.Dense(dim, activation='relu', name='dense'+str(i+1),
-                                               kernel_regularizer=keras.regularizers.l1_l2(l1=1e-6, l2=1e-6))(
-                   model_layer)
+                model_layer = keras.layers.Dense(dim, activation='relu',
+                                                 name='dense'+str(i+1))(model_layer)
             model_layer = keras.layers.BatchNormalization(name='normalization'+str(i+1))(model_layer)
             model_layer = keras.layers.Dropout(dropout_rate, name='dropout'+str(i+1))(model_layer)
 
-        embedding = keras.layers.Dense(embedding_dim, activation='relu', name='embedding')(
-            model_layer)
+        embedding = keras.layers.Dense(embedding_dim, activation='relu',
+                                       name='embedding')(model_layer)
         return keras.Model(model_input, embedding, name='base')
 
     @staticmethod

tests/test_models.py

@@ -44,6 +44,10 @@ def test_siamese_model():
     assert len(model.model.layers[2].layers) == len(model.base.layers) == 11, \
         "Expected different number of layers"
     assert model.model.input_shape == [(None, 339), (None, 339)], "Expected different input shape"
+    np.testing.assert_array_almost_equal(model.base.layers[1].kernel_regularizer.l1, 1e-6), \
+        "Expected different L1 regularization rate"
+    np.testing.assert_array_almost_equal(model.base.layers[1].kernel_regularizer.l2, 1e-6), \
+        "Expected different L2 regularization rate"
 
     # Test base model inference
     X, y = test_generator.__getitem__(0)
@@ -67,6 +71,16 @@ def test_siamese_model_different_architecture():
     assert model.base.output_shape == (None, 100), "Expected different output shape of base model"
 
 
+def test_siamese_model_different_regularization_rates():
+    spectrum_binner, test_generator = get_test_binner_and_generator()
+    model = SiameseModel(spectrum_binner, base_dims=(200,),
+                         embedding_dim=100, l1_reg=1e-7, l2_reg=1e-5)
+    np.testing.assert_array_almost_equal(model.base.layers[1].kernel_regularizer.l1, 1e-7), \
+        "Expected different L1 regularization rate"
+    np.testing.assert_array_almost_equal(model.base.layers[1].kernel_regularizer.l2, 1e-5), \
+        "Expected different L2 regularization rate"
+
+
 def test_load_model():
     """Test loading a model from file."""
     spectrum_binner, test_generator = get_test_binner_and_generator()