nicolasderumigny/ml-proj

Projet of the ENS M1 Machine Learning course

ML-Proj : Deep Structured Energy Based Models for Anomaly Detection

===================================================================== http://proceedings.mlr.press/v48/zhai16.pdf

Emma Kerinec and Nicolas Derumigny

Definitions

EBM = density estimatior tool based on a parametrisation of the negative log probability, which is called energy, and then computing the density with a proper normalization

DSEBM = EBM + the energy is learnt by a deep-structured NN (Fully connected for static data, RNN for sequential, CNN for images), output is decided either on energy (DSEBM-e) or reconstruction error (DSEBM-r)

Training by score matching method (intead of maximum likelyhood) by standard stochastic gradient descent (SGD) as a deep denoising autoencoder.

============================================================================= Use of tensorflow-opencl (https://github.com/benoitsteiner/tensorflow-opencl) Use of yadlt (http://deep-learning-tensorflow.readthedocs.io)

ml-proj's Issues

How to implement RNN-EBM?

When I use RNN or LSTM layer, it has a bug "Second-order gradient for while loops not supported."
Do you have any solution?

Is this working?

Hi,
is this code working. I had to fix a few things. But when I run it on my data. I get a reconstruction loss, that becomes bigger and bigger with every iteration?!

lg