# Algorithms with constructible architecture

## Specify network structure

There are three ways to define relations between layers. We can define network’s architecture separately from the training algorithm.

from neupy import algorithms, layers

network = layers.join(
layers.Input(10),
layers.Sigmoid(40),
layers.Sigmoid(2),
)

network,
step=0.2,
shuffle_data=True
)


Or, we can set up a list of layers that define sequential relations between layers.

from neupy import algorithms, layers

[
layers.Input(10),
layers.Sigmoid(40),
layers.Sigmoid(2)
layers.Softmax(2),
],
step=0.2,
shuffle_data=True
)


This is just a syntax simplification that allows to avoid using layer.join function.

Small networks can be defined with a help of inline operator.

from neupy import algorithms
from neupy.layers import *

Input(10) > Sigmoid(40) > Sigmoid(2),
step=0.2,
shuffle_data=True
)


## Train networks with multiple inputs

NeuPy allows to train networks with multiple inputs.

from neupy import algorithms, layers

[
[[
# 3 categorical inputs
layers.Input(3),
layers.Embedding(n_unique_categories, 4),
layers.Reshape(),
], [
# 17 numerical inputs
layers.Input(17),
]],
layers.Concatenate(),
layers.Relu(16),
layers.Sigmoid(1)
],

step=0.5,
verbose=True,
error='binary_crossentropy',
)

# Categorical variable should be the first, because
# categorical input layer was defined first in the network
network.train([x_train_cat, x_train_num], y_train,
[x_test_cat, x_test_num], y_test,
epochs=180)
y_predicted = network.predict([x_test_cat, x_test_num])


From the example above, you can see that we specified first layer as a list of lists. Each list has small sequence of layers specified and each sequence starts with the Input layer. This list of lists is just simple syntax sugar around the parallel function. Exactly the same architecture can be rewritten in the following way.

gdnet = algorithms.GradientDescent(
[
layers.parallel([
# 3 categorical inputs
layers.Input(3),
layers.Embedding(n_unique_categories, 4),
layers.Reshape(),
], [
# 17 numerical inputs
layers.Input(17),
]),
layers.Concatenate(),
layers.Relu(16),
layers.Sigmoid(1)
]
)


The training and prediction looks slightly different as well.

network.train([x_train_cat, x_train_num], y_train,
[x_test_cat, x_test_num], y_test,
epochs=180)
y_predicted = network.predict([x_test_cat, x_test_num])


Input we specified as a list where number of values equal to the number of input layers in the network. The order in the list is also important. We defined first input layer for categorical variables and therefore we need to pass it as the first element to the input list. The same is true for the predict method.

## Algorithms

NeuPy supports lots of different training algorithms based on the backpropagation. You can check Cheat sheet if you want to learn more about them.

Before using these algorithms you must understand that not all of them are suitable for all problems. Some of the methods like Levenberg-Marquardt or Conjugate Gradient work better for small networks and they would be extremely slow for networks with millions parameters. In addition, it’s important to note that not all algorithms are possible to train with mini-batches. Algorithms like Conjugate Gradient don’t work with mini-batches.

## Loss functions

NeuPy has many different loss functions. These loss functions can be specified specified as a string.

from neupy import algorithms, layers

[
layers.Input(784),
layers.Relu(500),
layers.Relu(300),
layers.Softmax(10),
],
error='categorical_crossentropy',
)


Also, it’s possible to create custom loss functions. Loss function should have two mandatory arguments, namely expected and predicted values.

import tensorflow as tf
from neupy import algorithms, layers

def mean_absolute_error(expected, predicted):
abs_errors = tf.abs(expected - predicted)
return tf.reduce_mean(abs_errors)

[
layers.Input(784),
layers.Relu(500),
layers.Relu(300),
layers.Softmax(10),
],
error=mean_absolute_error,
)


Loss function should return a scalar, because during the training output from the loss function will be used as a variable with respect to which we are differentiating.

Algorithms with constructible architectures allow to use additional update rules for parameter regularization and learning rate updates. For instance, we want to add Weight Decay regularization and we want to minimize step monotonically after each epoch.

from neupy import algorithms, layers

[
layers.Input(784),
layers.Relu(500),
layers.Relu(300),
layers.Softmax(10),
],
step=0.1,
batch_size=16,

algorithms.StepDecay]
)


Both WeightDecay and StepDecay algorithms have additional parameters. In case if we need to modify them we can add them to the training algorithm.

from neupy import algorithms, layers

[
layers.Input(784),
layers.Relu(500),
layers.Relu(300),
layers.Softmax(10),
],

step=0.1,
batch_size=16,

# Parameters from StepDecay
reduction_freq=50,

# Parameters from WeightDecay
decay_rate=0.05,

algorithms.StepDecay]
)


NeuPy doesn’t allow to use multiple regularizations and step update add-ons for training algorithm.

>>> from neupy import algorithms, layers
>>>
...     [
...         layers.Input(784),
...         layers.Relu(500),
...         layers.Relu(300),
...         layers.Softmax(10),
...     ],