encoder_decoder_information

Specifying Encoder and Decoder Architectures for <code>VAE_train()</code>

Build and train a variational autoencoder (VAE) for mixed-type
tabular data (continuous, binary, categorical).
Models are implemented using 'TensorFlow' and 'Keras' via the 'reticulate'
interface, enabling reproducible VAE training for heterogeneous tabular
datasets.

Sarah Milligan

autotab

Variational Autoencoders for Heterogeneous Tabular Data

encoder_decoder_information function

The arguments <code>encoder_info</code> and <code>decoder_info</code> define the architecture of
the encoder and decoder networks used in <code>VAE_train()</code>. Each is a list in
which every element describes one layer in sequence.
AutoTab currently supports two layer types: <code>"dense"</code> and <code>"dropout"</code>.
Dense layers
When <code>input1 = "dense"</code>, the layer specification takes the form:
<ul>
<li><code>input2</code>: Numeric. Number of units (nodes).</li>
<li><code>input3</code>: Character. Activation function
(any TensorFlow/Keras activation name).</li>
<li><code>input4</code>: Integer (0/1). L2 regularization flag. Default: <code>0</code>.</li>
<li><code>input5</code>: Numeric. L2 regularization strength (<code>lambda</code>). Default: <code>1e-4</code>.</li>
<li><code>input6</code>: Logical. Apply batch normalization. Default: <code>FALSE</code>.</li>
<li><code>input7</code>: Numeric. Batch normalization momentum. Default: <code>0.99</code>.</li>
<li><code>input8</code>: Logical. Whether batch normalization scale and center
parameters are trainable. Default: <code>TRUE</code>.</li>
</ul>Dropout layers
When <code>input1 = "dropout"</code>, the layer specification is:
<ul>
<li><code>input2</code>: Numeric. Dropout rate.</li>
</ul>Together, these lists fully specify the encoder and decoder architectures
used during VAE training.

Encoder and Decoder configuration

Specifying Encoder and Decoder Architectures for VAE_train() — encoder_decoder_information

The arguments <code>encoder_info</code> and <code>decoder_info</code> define the architecture of
the encoder and decoder networks used in <code>VAE_train()</code>. Each is a list in
which every element describes one layer in sequence.
AutoTab currently supports two layer types: <code>"dense"</code> and <code>"dropout"</code>.
Dense layers
When <code>input1 = "dense"</code>, the layer specification takes the form:
<ul>
<li><code>input2</code>: Numeric. Number of units (nodes).</li>
<li><code>input3</code>: Character. Activation function
(any TensorFlow/Keras activation name).</li>
<li><code>input4</code>: Integer (0/1). L2 regularization flag. Default: <code>0</code>.</li>
<li><code>input5</code>: Numeric. L2 regularization strength (<code>lambda</code>). Default: <code>1e-4</code>.</li>
<li><code>input6</code>: Logical. Apply batch normalization. Default: <code>FALSE</code>.</li>
<li><code>input7</code>: Numeric. Batch normalization momentum. Default: <code>0.99</code>.</li>
<li><code>input8</code>: Logical. Whether batch normalization scale and center
parameters are trainable. Default: <code>TRUE</code>.</li>
</ul>

Dropout layers
When <code>input1 = "dropout"</code>, the layer specification is:
<ul>
<li><code>input2</code>: Numeric. Dropout rate.</li>
</ul>

Together, these lists fully specify the encoder and decoder architectures
used during VAE training.

encoder_decoder_information: Specifying Encoder and Decoder Architectures for `VAE_train()`

Description

Arguments

Encoder and Decoder configuration

See Also