This repo contains my personal implementation of Variational autoencoder
in tensorflow for anomaly detection, that follows
Variational Autoencoder based Anomaly Detection using Reconstruction Probability by Jinwon An, Sungzoon Cho
In order to make work the variational autoencoder for anomaly detection i've to change the last layer of the decoder
from a simple fully connected layer to two layers that estimate mean and variance
of x~ ( p(x|z) )
NOTE: this works with tensorflow 1.x
- My implementation try to follow scikit-learn api, so it has a fit method to train the model
- The constructor of VAE define the architecture of the VAE but some things are fixed like the type of layers, fully connected layers with relu activation function. You can specify the size of each layer and the size of the latent space, optionally the decoder layer sizes but by default are the specular of the encoder layer sizes and mu and sigma of the prior p(z) that it is a normal distribution by default and cannot be changed
- It has also different methods like generate that does a sampling from the latent distribution, and run the decoder in order to generate new data like the X fitted
- reconstruct receive in input X then fit the encoder, calculate mean and variance of the latent space, then decode the sampled values from the latent distribution
- For anomaly detection there are two methods reconstructed_probability and is_outlier: the first estimate the reconstructed_probability like name guesses while the second use them to tell than an observation is an outlier if the rec. prob. is less than alpha, an input parameter such that by default if 0.05
latent_size = 32
encoders_sizes = np.linspace(X_train.shape[1], latent_size, 7).astype('int')[1:-1]
vae = VAE((X_train.shape[1],), encode_sizes=encoders_sizes, latent_size=latent_size, lr=0.00001)
vae.fit(X_train, epochs=200, batch_size=256)
p_x = vae.reconstructed_probability(X_train)
np.save('P_x', p_x)In my implementation i used InteractiveSession instead Session to ease the experiments in jupyter notebook but if you want to auto-close the session after getting the reconstructed probability you can use with keyword
latent_size = 32
encoders_sizes = np.linspace(X_train.shape[1], latent_size, 7).astype('int')[1:-1]
with VAE((X_train.shape[1],), encode_sizes=encoders_sizes, latent_size=latent_size, lr=0.00001) as vae:
vae.fit(X_train, epochs=200, batch_size=256)
p_x = vae.reconstructed_probability(X_train)
np.save('P_x', p_x)