krishnakumarsekar/Adversarial-Domain-Adaptation-for-Identifying-Phase-Transitions

Python code for our paper "Adversarial Domain Adaptation for Identifying Phase Transitions"

This is the computational appendix for the following paper:

Patrick Huembeli, Alexandre Dauphin, Peter Wittek. Adversarial Domain Adaptation for Identifying Phase Transitions. arXiv:1710.xxxxx, 2017.

Installation

To train the DANN, you need a few changes to the Keras source files. We recommend to make a new environment for this. We assume that the Python distribution is Anaconda. Create and activate a new environment:

$ conda create -n dann anaconda
$ source activate dann

Files to Generate States

Bose_Hubbard_Gutzwiller_coefficients.py

• Code partially from https://github.com/tcompa/BoseHubbardGutzwiller, install these files to use our code.
• With this file we generated the Gutzwiller coefficients.

Bogoliubov_Kitaev.py

• Produces Kitaev states

SSH_states_and_Winding_Nr.py

• Produces SSH states for OBC and PBC and calculates also the winding number and gives a plot of it.

• To do the same for long range SSH, replace the Hamiltonian in this file with the Hamiltonian from SSH_Long_Range_Hamiltonian.py.

Ising

• Ising_energy_Gibbs_sampling.py generates the Ising configurations via Monte Carlo method. The code has been made faster by using the beginning of each Markov chain more than once.
• CNN_Ising.py is a normal convolutional neural network, that can give the same output as the DANN.

Files for the neural network

Gradient_Reverse_Layer.py

• Code from keras-team/keras#3119 (comment).
• Needs to be in same folder as DANN main file DANN_example.py.

DANN_helper_file.py

• Summerizes all the important building blocks from https://github.com/fchollet/keras/pull/4031/files.
• Needs to be in same folder as DANN main file DANN_example.py.

DANN_example.py

• Main file for the DANN, with all the parameters.
• First there is specified, which data has to be loaded. Then there is a section for the training, one for the preddiction / evaluation of the DANN, and the last part is to apply unsupervised techniques on the feature space directly.

training.py and training_old.py

• training.py is the new file that has to be placed in the keras backend.
• training_old.py is a backup of the original file.