j30

nmcardoso · Aug 23, 2023 · 0137656 · 0137656
1 parent 48b834a
commit 0137656
Show file tree

Hide file tree

Showing 2 changed files with 6 additions and 6 deletions.
diff --git a/mergernet/estimators/parametric.py b/mergernet/estimators/parametric.py
@@ -56,7 +56,7 @@ def build(self, freeze_conv: bool = False) -> tf.keras.Model:
     x = conv_block(x)
 
     # Representation layer
-    representation_mode = self.hp.get('features_reduction', default='flatten')
+    representation_mode = self.hp.get('features_reduction', default='avg_pooling')
     if representation_mode == 'flatten':
       x = tf.keras.layers.Flatten()(x)
     elif representation_mode == 'avg_pooling':
@@ -73,14 +73,14 @@ def build(self, freeze_conv: bool = False) -> tf.keras.Model:
       units = self.hp.get(f'dense_{i}_units')
       bn = self.hp.get(f'batch_norm_{i}')
       activation = self.hp.get(f'activation_{i}', default='relu')
-      dropout_rate = self.hp.get(f'dropout_{i}_rate')
+      dropout_rate = self.hp.get(f'dropout_{i}_rate', default=0.0)
       if units:
         x = tf.keras.layers.Dense(units, use_bias=not bn)(x)
         if bn:
           x = tf.keras.layers.BatchNormalization()(x)
         if activation == 'relu':
           x = tf.keras.layers.Activation('relu')(x)
-        if dropout_rate:
+        if dropout_rate > 1e-7:
           x = tf.keras.layers.Dropout(dropout_rate)(x)
 
     # Classifications

diff --git a/mergernet/jobs/j030.py b/mergernet/jobs/j030.py
@@ -46,13 +46,13 @@ def call(self):
       HP.num('batch_size', low=64, high=256, step=64, dtype=int),
       HP.const('features_reduction', 'avg_pooling'),
       HP.num('dense_1_units', low=32, high=1024, step=1, dtype=int),
-      HP.cat('dropout_1_rate', [None, 0.1, 0.2, 0.3, 0.4]),
+      HP.num('dropout_1_rate', low=0.0, high=0.4, step=0.1, dtype=float),
       HP.cat('batch_norm_1', [True, False]),
       HP.num('dense_2_units', low=32, high=1024, step=1, dtype=int),
-      HP.cat('dropout_2_rate', [None, 0.1, 0.2, 0.3, 0.4]),
+      HP.num('dropout_2_rate', low=0.0, high=0.4, step=0.1, dtype=float),
       HP.cat('batch_norm_2', [True, False]),
       HP.num('dense_3_units', low=32, high=1024, step=1, dtype=int),
-      HP.cat('dropout_3_rate', [None, 0.1, 0.2, 0.3, 0.4]),
+      HP.num('dropout_3_rate', low=0.0, high=0.4, step=0.1, dtype=float),
       HP.cat('batch_norm_3', [True, False]),
     )
     ds = Dataset(config=Dataset.registry.LS10_TRAIN_224_PNG)