From ab3785adc4d72179e0c59f6adfa51d4376a31c07 Mon Sep 17 00:00:00 2001
From: Nikolai Hartmann <Nikolai.Hartmann@physik.uni-muenchen.de>
Date: Wed, 19 Sep 2018 13:43:41 +0200
Subject: [PATCH] Switch to functional interface and make usage of bias vectors
configurable
---
toolkit.py | 40 +++++++++++++++++++++++++++-------------
1 file changed, 27 insertions(+), 13 deletions(-)
diff --git a/toolkit.py b/toolkit.py
index 65c526f..b224fed 100755
--- a/toolkit.py
+++ b/toolkit.py
@@ -124,6 +124,8 @@ class ClassificationProject(object):
:param dropout_input: dropout fraction for the input layer. Set to None for no Dropout.
+ :param use_bias: use bias constant for each neuron? (default: True). You can also pass a list of booleans for each layer.
+
:param batch_size: size of the training batches
:param validation_split: split off this fraction of training events for loss evaluation
@@ -229,6 +231,7 @@ class ClassificationProject(object):
nodes=64,
dropout=None,
dropout_input=None,
+ use_bias=True,
batch_size=128,
validation_split=0.33,
kfold_splits=None,
@@ -288,9 +291,14 @@ class ClassificationProject(object):
self.dropout = dropout
if not isinstance(self.dropout, list):
self.dropout = [self.dropout for i in range(self.layers)]
+ self.dropout_input = dropout_input
+ self.use_bias = use_bias
+ if not isinstance(self.use_bias, list):
+ self.use_bias = [self.use_bias for i in range(self.layers)]
if len(self.dropout) != self.layers:
raise ValueError("List of dropout fractions has to be of equal size as the number of layers!")
- self.dropout_input = dropout_input
+ if len(self.use_bias) != self.layers:
+ raise ValueError("List biases has to be of equal size as the number of layers!")
self.batch_size = batch_size
self.validation_split = validation_split
self.kfold_splits = kfold_splits
@@ -703,23 +711,29 @@ class ClassificationProject(object):
if self._model is None:
- self._model = Sequential()
+ # input
+ input_layer = Input((len(self.fields),))
+ # optional dropout on inputs
if self.dropout_input is None:
- self._model.add(Dense(self.nodes[0], input_dim=len(self.fields), activation=self.activation_function))
- # in case of no Dropout we already have the first hidden layer
- start_layer = 1
+ hidden_layer = input_layer
else:
- self._model.add(Dropout(rate=self.dropout_input, input_shape=(len(self.fields),)))
- start_layer = 0
- # the (other) hidden layers
- for node_count, dropout_fraction in zip(self.nodes[start_layer:], self.dropout[start_layer:]):
- self._model.add(Dense(node_count, activation=self.activation_function))
+ hidden_layer = Dropout(rate=self.dropout_input)(input_layer)
+
+ # densely connected hidden layers
+ for node_count, dropout_fraction, use_bias in zip(
+ self.nodes,
+ self.dropout,
+ self.use_bias,
+ ):
+ hidden_layer = Dense(node_count, activation=self.activation_function, use_bias=use_bias)(hidden_layer)
if (dropout_fraction is not None) and (dropout_fraction > 0):
- self._model.add(Dropout(rate=dropout_fraction))
- # last layer is one neuron (binary classification)
- self._model.add(Dense(1, activation=self.activation_function_output))
+ hidden_layer = Dropout(rate=dropout_fraction)(hidden_layer)
+
+ # one output node for binary classification
+ output_layer = Dense(1, activation=self.activation_function_output)(hidden_layer)
+ self._model = Model(inputs=[input_layer], outputs=[output_layer])
self._compile_or_load_model()
return self._model
--
GitLab