HyperStreamOnlineLearning

Online Learning for HyperStream.

Table of Contents

• Requirements
• Unittest
• Run a simple classification test
• Example with Keras
• Example of Anomaly detection
• Example of Incremental PCA
• Example of an Autoencoder with Keras

Requirements

Running any example of HyperStream requires a MongoDB server. The configuration of the host and ports of the MongoDB server can be changed in the file hyperstream_config.json if needed.

To download the code and install the requirements ussing virtualenvironment just do the following:

git clone git@github.com:IRC-SPHERE/HyperStreamOnlineLearning.git
cd HyperStreamOnlineLearning
virtualenv venv
. venv/bin/activate
pip install -r requirements.txt

If you want to run the examples with Keras you will need to install some additional dependences. Once the previous requirements have been installed you can install the rest with the following command:

pip install -r keras_requirements.txt

Once the MongoDB server is up and running and you have installed all the Python requirements you can try to run the some fo the following examples:

Unittest

To test if the code is working you can run a nosetest

nosetest

You should get something of the style

.
----------------------------------------------------------------------
Ran 1 testin 0.585s
OK

Run a simple classification test

The following code should finish with the scores of a model trained on Iris dataset.

python example_classifier.py --dataset iris --classifier SGDClassifier --epochs 20 --seed 42

At the end of the training you should see the test scores during the training for each epoch

Testscoresperepochduringtraining
[ 0.590.650.680.710.720.720.650.730.690.720.710.730.720.720.720.720.720.760.750.77]

Example with Keras

There is an additional example using Keras to specify Multilayer Perceptrons or Logistic regression. To run the example use these steps

python example_classifier_keras.py --dataset digits --classifier mlp30ds40m --epochs 20 --seed 42

At the end of the training you should see the test scores during the training for each epoch

[ 0.480.750.840.890.890.90.910.910.920.910.920.940.930.940.930.930.940.940.950.94]

Example of Anomaly detection

Example of a model that trains only with the input space and predicts if the test data has been drawn from the same distribution. In this case using a Multivariate-Gaussian to estimate the density of the training data.

python example_anomalies.py --dataset iris --model Gaussian --epochs 1 --seed 42 -b 2

At the end of the training you should see the test scores during the training for each epoch

[ nannannan0.920.930.570.520.910.810.410.840.530.770.880.950.790.580.840.690.820.630.550.990.90.890.650.610.830.870.860.780.750.740.660.840.920.810.81]

Example of Incremental PCA

python example_incremental_pca.py --dataset digits --components 10 --epochs 10 --seed 42 -b 100

At the end of the training you should see the mean squared error of the reconstruction using the principal components of the PCA.

Testscoresperminibatch (cyclic)
[[ 5.85.115.545.335.595.085.355.145.045.184.795.425.245.485.065.335.135.015.24.765.435.25.455.095.365.064.995.264.735.415.225.45.125.355.084.935.294.735.445.155.375.155.355.084.945.274.775.435.135.375.265.235.074.945.294.785.415.135.355.275.245.064.935.284.835.395.185.285.275.285.034.945.284.845.435.145.255.35.265.064.895.34.855.45.175.215.325.235.085.04]]

Example of an Autoencoder with Keras

python example_autoencoder_keras.py --dataset digits --architecture auto30ns10ns2ns_10ns30ns --epochs 100 --seed 42 -b 10 --learning-rate 0.1

Will output the following

_________________________________________________________________Layer (type) OutputShapeParam# [20/9638]=================================================================dense_1 (Dense) (None, 30) 1950_________________________________________________________________batch_normalization_1 (Batch (None, 30) 120_________________________________________________________________activation_1 (Activation) (None, 30) 0_________________________________________________________________dense_2 (Dense) (None, 10) 310_________________________________________________________________batch_normalization_2 (Batch (None, 10) 40_________________________________________________________________activation_2 (Activation) (None, 10) 0_________________________________________________________________dense_3 (Dense) (None, 2) 22_________________________________________________________________batch_normalization_3 (Batch (None, 2) 8_________________________________________________________________activation_3 (Activation) (None, 2) 0=================================================================Totalparams: 2,450Trainableparams: 2,366Non-trainableparams: 84_________________________________________________________________None_________________________________________________________________Layer (type) OutputShapeParam#=================================================================dense_4 (Dense) (None, 10) 30_________________________________________________________________batch_normalization_4 (Batch (None, 10) 40_________________________________________________________________activation_4 (Activation) (None, 10) 0_________________________________________________________________dense_5 (Dense) (None, 30) 330_________________________________________________________________batch_normalization_5 (Batch (None, 30) 120_________________________________________________________________activation_5 (Activation) (None, 30) 0_________________________________________________________________dense_6 (Dense) (None, 64) 1984=================================================================Totalparams: 2,504Trainableparams: 2,424Non-trainableparams: 80_________________________________________________________________None_________________________________________________________________Layer (type) OutputShapeParam# =================================================================sequential_1 (Sequential) (None, 2) 2450_________________________________________________________________sequential_2 (Sequential) (None, 64) 2504=================================================================Totalparams: 4,954Trainableparams: 4,790Non-trainableparams: 164_________________________________________________________________None2017-12-0415:49:37.732891: Itensorflow/core/platform/cpu_feature_guard.cc:137] YourCPUsupportsinstructionsthatthisTensorFlowbinarywasnotcompiledtouse: SSE4.1SSE4.2AVXTestscoresperminibatch (cyclic)
[[ 65.9658.5660.33 ..., 13.5410.8912.94]]