Autoencoder

Model

class ParticleAutoencoder(*args, **kwargs)

Bases: Model

Autoencoder model for dimensionality reduction.

input_shape

shape of input, default (100,3)

Type:: tuple, optional

latent_dim

size of the latent dimension, default 6

Type:: int, optional

x_mean_stdev

mean and standard deviation of inputs, default (0,1)

Type:: tuple, optional

kernel_n

number of kernels, default 16

Type:: int, optional

kernel_sz

kernel size, default 3

Type:: int, optional

activation

activation function, default “elu”

Type:: string, optional

activation_latent

activation before bottleneck, default :class:tensorflow.keras.activations.linear`

Type:: tensorflow.keras.activations, optional

build_encoder(mean, stdev)

Builds encoder model

Parameters:

mean (float) – Mean of data.
stdev (float) – Standard deviation of data.

Returns:

encoder

Return type:

tensorflow.keras.Model

build_decoder(mean, stdev)

Builds decoder model

Parameters:

mean (float) – Mean of data.
stdev (float) – Standard deviation of data.

Returns:

decoder

Return type:

tensorflow.keras.Model

classmethod load(path)

Loads autoencoder.

Parameters:: path (str) – Model path.
Returns:: autoencoder
Return type:: tensorflow.keras.Model

compile(optimizer, reco_loss)

Compiles the autoencoder - set model’s configuration.

Parameters:

optimizer (tensorflow.keras.Optimizer) – optimizer.
reco_loss (Callable) – reconstruction loss function

call(x)

Calls the model.

Parameters:: x (numpy.ndarray) – inputs
Returns:: outputs
Return type:: numpy.ndarray

train_step(x)

Training step.

Parameters:: x (numpy.ndarray) – inputs
Returns:: loss dictionary
Return type:: dict

test_step(x)

Inference step.

Parameters:: x (numpy.ndarray) – inputs
Returns:: loss dictionary
Return type:: dict

Layers

class Conv1DTranspose(*args, **kwargs)

Bases: Layer

Custom 1d transposed convolution that expands to 2d output for vae decoder

activation

activation

Type:: tensorflow.keras.activation

ConvTranspose

sublayer

Type:: tensorflow.keras.Layer

ExpandChannel

sublayer

Type:: tensorflow.keras.Lambda

filters

number kernels

Type:: int

kernel_initializer

kernel init

Type:: tensorflow.keras.Initializer

kernel_sz

kernel size

Type:: int

SqueezeChannel

sublayer

Type:: tensorflow.keras.Layer

__init__(filters, kernel_sz, activation, kernel_initializer, **kwargs)

Construtor.

Parameters:

filters (int) – number kernels
activation (tensorflow.keras.activation) – activation
kernel_initializer (tensorflow.keras.Initializer) – kernel init
kernel_sz (int) – kernel size

call(inputs)

Call.

Parameters:: inputs (tensorflow.Tensor) – call
Returns:: outputs
Return type:: tensorflow.Tensor

get_config()

Get configuration of layer.

Returns:: dictionary containing configuration used to initialize this layer
Return type:: dict

class StdNormalization(*args, **kwargs)

Bases: Layer

Normalizing input feature to std Gauss (mean 0, var 1).

mean_x

mean of inputs

Type:: float

std_x

standard deviation of inputs

Type:: float

__init__(mean_x, std_x, name='Std_Normalize', **kwargs)

Constructor.

Parameters:

mean_x (float) – mean of inputs
std_x (float) – standard deviation of inputs
name (str, optional) – name

get_config()

Get configuration for layer.

Returns:: dictionary containing configuration used to initialize this layer
Return type:: dict

call(x)

Call.

Parameters:: inputs (tensorflow.Tensor) – call
Returns:: outputs
Return type:: tensorflow.Tensor

class StdUnnormalization(*args, **kwargs)

Bases: StdNormalization

Rescaling feature to original domain

__init__(mean_x, std_x, name='Un_Normalize', **kwargs)

Constructor.

Parameters:

mean_x (float) – mean of inputs
std_x (float) – standard deviation of inputs
name (str, optional) – name

call(x)

Call.

Parameters:: inputs (tensorflow.Tensor) – call
Returns:: outputs
Return type:: tensorflow.Tensor

Train

train(data_sample, input_shape=(100, 3), read_n=10000, batch_size=256, latent_dim=6, epochs=10, act_latent=None)

Trains autoencoder

Parameters:

data_sample (numpy.ndarray) – inputs
input_shape (tuple, optional) – shape, default (100, 3)
batch_size (int, optional) – batch size, default 256
latent_dim (int, optional) – latent dim, default 6
epochs (int, optional) – number of epochs, default 10
act_latent (tensorflow.keras.Actication, optional) – latent activation, default None

Predict

map_to_latent_space(data_sample, model) → ndarray

Autoencoder mapping input space to latent representation.

Parameters:

data_sample (tensorflow.data.Dataset) – tensorflow.data.Dataset.from_tensor_slices(input_sample (:class:`numpy.ndarray)).batch(batch_size)`
model (tensorflow.keras.Model) – the autoencoder

Returns:

latent representation

Return type:

numpy.ndarray