# -*- coding: utf-8 -*-
"""Using Adversarially Learned Anomaly Detection
"""
# Author: Michiel Bongaerts (but not author of the ALAD method)
# Pytorch version Author: Jiaqi Li <jli77629@usc.edu>
import numpy as np
import pandas as pd
try:
import torch
except ImportError:
print('please install torch first')
import torch
import torch.nn as nn
import torch.optim as optim
from matplotlib import pyplot as plt
from sklearn.preprocessing import StandardScaler
from sklearn.utils import check_array
from sklearn.utils.validation import check_is_fitted
from .base import BaseDetector
from ..utils.utility import check_parameter
[docs]
class ALAD(BaseDetector):
"""Adversarially Learned Anomaly Detection (ALAD).
Paper: https://arxiv.org/pdf/1812.02288.pdf
See :cite:`zenati2018adversarially` for details.
Parameters
----------
output_activation : str, optional (default=None)
Activation function to use for output layers for encoder and dector.
activation_hidden_disc : str, optional (default='tanh')
Activation function to use for hidden layers in discrimators.
activation_hidden_gen : str, optional (default='tanh')
Activation function to use for hidden layers in encoder and decoder
(i.e. generator).
epochs : int, optional (default=500)
Number of epochs to train the model.
batch_size : int, optional (default=32)
Number of samples per gradient update.
dropout_rate : float in (0., 1), optional (default=0.2)
The dropout to be used across all layers.
dec_layers : list, optional (default=[5,10,25])
List that indicates the number of nodes per hidden layer for the d
ecoder network.
Thus, [10,10] indicates 2 hidden layers having each 10 nodes.
enc_layers : list, optional (default=[25,10,5])
List that indicates the number of nodes per hidden layer for the
encoder network.
Thus, [10,10] indicates 2 hidden layers having each 10 nodes.
disc_xx_layers : list, optional (default=[25,10,5])
List that indicates the number of nodes per hidden layer for
discriminator_xx.
Thus, [10,10] indicates 2 hidden layers having each 10 nodes.
disc_zz_layers : list, optional (default=[25,10,5])
List that indicates the number of nodes per hidden layer for
discriminator_zz.
Thus, [10,10] indicates 2 hidden layers having each 10 nodes.
disc_xz_layers : list, optional (default=[25,10,5])
List that indicates the number of nodes per hidden layer for
discriminator_xz.
Thus, [10,10] indicates 2 hidden layers having each 10 nodes.
learning_rate_gen: float in (0., 1), optional (default=0.001)
learning rate of training the encoder and decoder
learning_rate_disc: float in (0., 1), optional (default=0.001)
learning rate of training the discriminators
add_recon_loss: bool optional (default=False)
add an extra loss for encoder and decoder based on the reconstruction
error
lambda_recon_loss: float in (0., 1), optional (default=0.1)
if ``add_recon_loss= True``, the reconstruction loss gets multiplied
by ``lambda_recon_loss`` and added to the total loss for the generator
(i.e. encoder and decoder).
preprocessing : bool, optional (default=True)
If True, apply standardization on the data.
verbose : int, optional (default=1)
Verbosity mode.
- 0 = silent
- 1 = progress bar
contamination : float in (0., 0.5), optional (default=0.1)
The amount of contamination of the data set, i.e.
the proportion of outliers in the data set. When fitting this is used
to define the threshold on the decision function.
device : str or None, optional (default=None)
The device to use for computation. If None, the default device will be used.
Possible values include 'cpu' or 'gpu'. This parameter allows the user
to specify the preferred device for running the model.
Attributes
----------
decision_scores_ : numpy array of shape (n_samples,)
The outlier scores of the training data [0,1].
The higher, the more abnormal. Outliers tend to have higher
scores. This value is available once the detector is
fitted.
threshold_ : float
The threshold is based on ``contamination``. It is the
``n_samples * contamination`` most abnormal samples in
``decision_scores_``. The threshold is calculated for generating
binary outlier labels.
labels_ : int, either 0 or 1
The binary labels of the training data. 0 stands for inliers
and 1 for outliers/anomalies. It is generated by applying
``threshold_`` on ``decision_scores_``.
"""
def __init__(self, activation_hidden_gen='tanh',
activation_hidden_disc='tanh',
output_activation=None,
dropout_rate=0.2,
latent_dim=2,
dec_layers=[5, 10, 25],
enc_layers=[25, 10, 5],
disc_xx_layers=[25, 10, 5],
disc_zz_layers=[25, 10, 5],
disc_xz_layers=[25, 10, 5],
learning_rate_gen=0.0001, learning_rate_disc=0.0001,
add_recon_loss=False, lambda_recon_loss=0.1,
epochs=200,
verbose=0,
preprocessing=False,
add_disc_zz_loss=True, spectral_normalization=False,
batch_size=32, contamination=0.1, device=None):
super(ALAD, self).__init__(contamination=contamination)
self.device = device if device else torch.device(
"cuda" if torch.cuda.is_available() else "cpu")
self.activation_hidden_disc = activation_hidden_disc
self.activation_hidden_gen = activation_hidden_gen
self.output_activation = output_activation
self.dropout_rate = dropout_rate
self.latent_dim = latent_dim
self.dec_layers = dec_layers
self.enc_layers = enc_layers
self.disc_xx_layers = disc_xx_layers
self.disc_zz_layers = disc_zz_layers
self.disc_xz_layers = disc_xz_layers
self.add_recon_loss = add_recon_loss
self.lambda_recon_loss = lambda_recon_loss
self.add_disc_zz_loss = add_disc_zz_loss
self.contamination = contamination
self.epochs = epochs
self.learning_rate_gen = learning_rate_gen
self.learning_rate_disc = learning_rate_disc
self.preprocessing = preprocessing
self.batch_size = batch_size
self.verbose = verbose
self.spectral_normalization = spectral_normalization
if self.spectral_normalization:
try:
import torch.nn.utils.spectral_norm as spectral_norm
self.spectral_norm = spectral_norm
except ImportError:
print('Spectral normalization not available. '
'Install torch>=1.0.0.')
self.spectral_normalization = False
check_parameter(dropout_rate, 0, 1, param_name='dropout_rate',
include_left=True)
def _build_model(self):
def get_activation(name):
if name == 'tanh':
return nn.Tanh()
elif name == 'sigmoid':
return nn.Sigmoid()
elif name == 'relu':
return nn.ReLU()
else:
raise ValueError(
"Unsupported activation function: {}".format(name))
# Create the decoder
dec_layers = []
input_dim = self.latent_dim
for l_dim in self.dec_layers:
dec_layers.append(nn.Linear(input_dim, l_dim))
dec_layers.append(nn.Dropout(self.dropout_rate))
dec_layers.append(get_activation(self.activation_hidden_gen))
input_dim = l_dim
dec_layers.append(nn.Linear(input_dim, self.n_features_))
if self.output_activation:
dec_layers.append(get_activation(self.output_activation))
self.dec = nn.Sequential(*dec_layers).to(self.device)
# Create the encoder
enc_layers = []
input_dim = self.n_features_
for l_dim in self.enc_layers:
enc_layers.append(nn.Linear(input_dim, l_dim))
enc_layers.append(nn.Dropout(self.dropout_rate))
enc_layers.append(get_activation(self.activation_hidden_gen))
input_dim = l_dim
enc_layers.append(nn.Linear(input_dim, self.latent_dim))
if self.output_activation:
enc_layers.append(get_activation(self.output_activation))
self.enc = nn.Sequential(*enc_layers).to(self.device)
# Create the discriminators
def create_discriminator(layers, input_dim):
disc_layers = []
for l_dim in layers:
disc_layers.append(nn.Linear(input_dim, l_dim))
if self.spectral_normalization:
disc_layers[-1] = nn.utils.spectral_norm(disc_layers[-1])
disc_layers.append(nn.Dropout(self.dropout_rate))
disc_layers.append(get_activation(self.activation_hidden_disc))
input_dim = l_dim
disc_layers.append(nn.Linear(input_dim, 1))
disc_layers.append(nn.Sigmoid())
return nn.Sequential(*disc_layers).to(self.device)
self.disc_xx = create_discriminator(self.disc_xx_layers,
2 * self.n_features_)
self.disc_zz = create_discriminator(self.disc_zz_layers,
2 * self.latent_dim)
self.disc_xz = create_discriminator(self.disc_xz_layers,
self.n_features_ + self.latent_dim)
# Optimizers
self.opt_gen = optim.Adam(
list(self.enc.parameters()) + list(self.dec.parameters()),
lr=self.learning_rate_gen)
self.opt_disc = optim.Adam(list(self.disc_xx.parameters()) + list(
self.disc_xz.parameters()) + list(self.disc_zz.parameters()),
lr=self.learning_rate_disc)
self.hist_loss_disc = []
self.hist_loss_gen = []
def train_step(self, data):
x_real, z_real = data
x_real = torch.FloatTensor(x_real).to(self.device)
z_real = torch.FloatTensor(z_real).to(self.device)
self.opt_disc.zero_grad()
x_gen = self.dec(z_real)
z_gen = self.enc(x_real)
out_true_xz = self.disc_xz(torch.cat((x_real, z_gen), dim=1))
out_fake_xz = self.disc_xz(torch.cat((x_gen, z_real), dim=1))
out_true_xx = self.disc_xx(torch.cat((x_real, x_real), dim=1))
out_fake_xx = self.disc_xx(torch.cat((x_real, x_gen), dim=1))
loss_dxz = nn.BCELoss()(out_true_xz,
torch.ones_like(out_true_xz)) + nn.BCELoss()(
out_fake_xz, torch.zeros_like(out_fake_xz))
loss_dxx = nn.BCELoss()(out_true_xx,
torch.ones_like(out_true_xx)) + nn.BCELoss()(
out_fake_xx, torch.zeros_like(out_fake_xx))
if self.add_disc_zz_loss:
out_true_zz = self.disc_zz(torch.cat((z_real, z_real), dim=1))
out_fake_zz = self.disc_zz(torch.cat((z_real, z_gen), dim=1))
loss_dzz = nn.BCELoss()(out_true_zz, torch.ones_like(
out_true_zz)) + nn.BCELoss()(out_fake_zz,
torch.zeros_like(out_fake_zz))
loss_disc = loss_dxz + loss_dzz + loss_dxx
else:
loss_disc = loss_dxz + loss_dxx
loss_disc.backward()
self.opt_disc.step()
self.opt_gen.zero_grad()
x_gen = self.dec(z_real)
z_gen = self.enc(x_real)
out_true_xz = self.disc_xz(torch.cat((x_real, z_gen), dim=1))
out_fake_xz = self.disc_xz(torch.cat((x_gen, z_real), dim=1))
out_true_xx = self.disc_xx(torch.cat((x_real, x_real), dim=1))
out_fake_xx = self.disc_xx(torch.cat((x_real, x_gen), dim=1))
loss_gexz = nn.BCELoss()(out_fake_xz,
torch.ones_like(out_fake_xz)) + nn.BCELoss()(
out_true_xz, torch.zeros_like(out_true_xz))
loss_gexx = nn.BCELoss()(out_fake_xx,
torch.ones_like(out_fake_xx)) + nn.BCELoss()(
out_true_xx, torch.zeros_like(out_true_xx))
if self.add_disc_zz_loss:
out_true_zz = self.disc_zz(torch.cat((z_real, z_real), dim=1))
out_fake_zz = self.disc_zz(torch.cat((z_real, z_gen), dim=1))
loss_gezz = nn.BCELoss()(out_fake_zz, torch.ones_like(
out_fake_zz)) + nn.BCELoss()(out_true_zz,
torch.zeros_like(out_true_zz))
cycle_consistency = loss_gezz + loss_gexx
loss_gen = loss_gexz + cycle_consistency
else:
cycle_consistency = loss_gexx
loss_gen = loss_gexz + cycle_consistency
if self.add_recon_loss:
x_recon = self.dec(self.enc(x_real))
loss_recon = torch.mean((x_real - x_recon) ** 2)
loss_gen += loss_recon * self.lambda_recon_loss
loss_gen.backward()
self.opt_gen.step()
self.hist_loss_disc.append(loss_disc.item())
self.hist_loss_gen.append(loss_gen.item())
[docs]
def fit(self, X, y=None, noise_std=0.1):
"""Fit detector. y is ignored in unsupervised methods.
Parameters
----------
X : numpy array of shape (n_samples, n_features)
The input samples.
y : Ignored
Not used, present for API consistency by convention.
Returns
-------
self : object
Fitted estimator.
"""
# validate inputs X and y (optional)
X = check_array(X)
self._set_n_classes(y)
# Get number of sampels and features from train set
self.n_samples_, self.n_features_ = X.shape[0], X.shape[1]
self._build_model()
# Apply data scaling or not
if self.preprocessing:
self.scaler_ = StandardScaler()
X_norm = self.scaler_.fit_transform(X)
else:
X_norm = np.copy(X)
for n in range(self.epochs):
if n % 50 == 0 and n != 0 and self.verbose == 1:
print(f'Train iter: {n}')
# Shuffle train
np.random.shuffle(X_norm)
X_train_sel = X_norm[
:min(self.batch_size, self.n_samples_)].astype(
np.float32)
latent_noise = np.random.normal(0, 1, (
X_train_sel.shape[0], self.latent_dim))
X_train_sel += np.random.normal(0, noise_std,
size=X_train_sel.shape)
self.train_step((X_train_sel, latent_noise))
if self.preprocessing:
X_norm = self.scaler_.transform(X)
else:
X_norm = np.copy(X)
pred_scores = self.get_outlier_scores(X_norm)
self.decision_scores_ = pred_scores
self._process_decision_scores()
return self
def train_more(self, X, epochs=100, noise_std=0.1):
"""This function allows the researcher to perform extra training
instead of the fixed number determined
by the fit() function.
"""
# fit() should have been called first
check_is_fitted(self, ['decision_scores_'])
# Apply data scaling or not
if self.preprocessing:
X_norm = self.scaler_.transform(X)
else:
X_norm = np.copy(X)
for n in range(epochs):
if n % 50 == 0 and n != 0 and self.verbose == 1:
print(f'Train iter: {n}')
# Shuffle train
np.random.shuffle(X_norm)
X_train_sel = X_norm[
:min(self.batch_size, self.n_samples_)].astype(
np.float32)
latent_noise = np.random.normal(0, 1, (
X_train_sel.shape[0], self.latent_dim))
X_train_sel += np.random.normal(0, noise_std,
size=X_train_sel.shape)
self.train_step((X_train_sel, latent_noise))
if self.preprocessing:
X_norm = self.scaler_.transform(X)
else:
X_norm = np.copy(X)
pred_scores = self.get_outlier_scores(X_norm)
self.decision_scores_ = pred_scores
self._process_decision_scores()
return self
def get_outlier_scores(self, X_norm):
X_norm = torch.FloatTensor(X_norm).to(self.device)
X_enc = self.enc(X_norm).detach().cpu().numpy()
X_enc_gen = self.dec(
torch.FloatTensor(X_enc).to(self.device)).detach().cpu().numpy()
out_true_xx = self.disc_xx(
torch.cat((X_norm, X_norm), dim=1)).detach().cpu().numpy()
out_fake_xx = self.disc_xx(
torch.cat((X_norm, torch.FloatTensor(X_enc_gen).to(self.device)),
dim=1)).detach().cpu().numpy()
outlier_scores = np.mean(np.abs((out_true_xx - out_fake_xx) ** 2),
axis=1)
return outlier_scores
[docs]
def decision_function(self, X):
"""Predict raw anomaly score of X using the fitted detector.
The anomaly score of an input sample is computed based on different
detector algorithms. For consistency, outliers are assigned with
larger anomaly scores.
Parameters
----------
X : numpy array of shape (n_samples, n_features)
The training input samples. Sparse matrices are accepted only
if they are supported by the base estimator.
Returns
-------
anomaly_scores : numpy array of shape (n_samples,)
The anomaly score of the input samples.
"""
check_is_fitted(self, ['decision_scores_'])
X = check_array(X)
if self.preprocessing:
X_norm = self.scaler_.transform(X)
else:
X_norm = np.copy(X)
X_norm = torch.FloatTensor(X_norm).to(self.device)
pred_scores = self.get_outlier_scores(X_norm.cpu().numpy())
return pred_scores
[docs]
def plot_learning_curves(self, start_ind=0, window_smoothening=10):
fig = plt.figure(figsize=(12, 5))
l_gen = pd.Series(self.hist_loss_gen[start_ind:]).rolling(
window=window_smoothening).mean()
l_disc = pd.Series(self.hist_loss_disc[start_ind:]).rolling(
window=window_smoothening).mean()
ax = fig.add_subplot(1, 2, 1)
ax.plot(range(len(l_gen)), l_gen)
ax.set_title('Generator')
ax.set_ylabel('Loss')
ax.set_xlabel('Iter')
ax = fig.add_subplot(1, 2, 2)
ax.plot(range(len(l_disc)), l_disc)
ax.set_title('Discriminator(s)')
ax.set_ylabel('Loss')
ax.set_xlabel('Iter')
plt.show()