tht106/dan

Domain Agnostic Normalization layer for Unsupervised Domain Adaptation

A Domain Agnostic Normalization Layer for Unsupervised Adversarial Domain Adaptation

R. Romijnders, P.Meletis, G. Dubbelman Eindhoven, University of technology TU/e-SPS (VCA): Mobile Perception Systems Delivery date: July 3rd 2018

Upon acceptance, we will publish the code in this repository to reproduce all our experiments and results.

Note of July 17: this code will be made public after first review round of IEEE WACV (something like October 2018)

Notes for instruction:

• Best starting point is to change scripts/set_env_names.sh
  • Then run a script like train_many.sh or evaluate_many.sh
• The scripts in train.py and evaluate.py can be run from command line

Notes on versions

• doc/requirements.txt contains the output of pip freeze run on July 17, 2018
• This code base started as fork from the semantic-segmentation/v0.7 code by Panos Meletis. Last fork on last week October, 2017

Repository structure

• Estimator: all code necessary to run via tf.estimator API
• Input: all code related to the input pipeline of the data. We use the tf.data API
• Misc: Miscellaneous code. Mainly contains code for plotting
• Model: the actual model for the representation learner, segmenter and domain classifier
• Scripts: all code for shell scripting. Note, I learned bash along this project, so this code probably has some beginner-mistakes
• Utils: all kinds of utility functions