haeusser / Learning_by_association

This repository contains code for the paper Learning by Association - A versatile semi-supervised training method for neural networks (CVPR 2017) and the follow-up work Associative Domain Adaptation (ICCV 2017)

It is implemented with TensorFlow. Please refer to the TensorFlow documentation for further information.

The core functions are implemented in semisup/backend.py. The files train.py and eval.py demonstrate how to use them. A quick example is contained in mnist_train_eval.py.

In order to reproduce the results from the paper, please use the architectures and pipelines from the tools/{stl10,svhn,synth}.py. They are loaded automatically by setting the flag [target_]dataset in {train,eval}.py accordingly.

Before you get started, please make sure to add the following to your ~/.bashrc:

export PYTHONPATH=/path/to/learning_by_association:$PYTHONPATH

Copy the file semisup/tools/data_dirs.py.template to semisup/tools/data_dirs.py, adapt the paths and .gitignore this file.

Domain Adaptation Hyper parameters

Synth. Signs -> GTSRB

"target_dataset": "gtsrb",
"walker_weight_envelope_delay": "0",
"max_checkpoints": 5,
"dataset": "synth_signs",
"visit_weight": "0.1",
"sup_per_batch": 24,
"walker_weight_envelope_steps": 1,
"eval_batch_size": 24,
"walker_weight_envelope": "linear",
"unsup_batch_size": 1032,
"visit_weight_envelope": "linear",
"decay_steps": 9000,
"sup_per_class": -1,
"max_steps": 12000,
"architecture": "svhn_model"

MNIST -> MNIST-M

"target_dataset": "mnistm",
"walker_weight_envelope_delay": "500",
"max_checkpoints": 5,
"new_size": 32,
"dataset": "mnist3",
"visit_weight": "0.6",
"augmentation": true,
"walker_weight_envelope_steps": 1,
"walker_weight_envelope": "linear",
"unsup_batch_size": 1000,
"visit_weight_envelope": "linear",
"decay_steps": 9000,
"architecture": "svhn_model",
"sup_per_class": -1,
"sup_per_batch": 100,
"max_steps": "12000",

SVHN -> MNIST

"target_dataset": "mnist3",
"walker_weight_envelope_delay": "500",
"max_checkpoints": 5,
"new_size": 32,
"dataset": "svhn",
"sup_per_batch": 100,
"decay_steps": 9000,
"unsup_batch_size": 1000,
"sup_per_class": -1,
"walker_weight_envelope_steps": 1,
"walker_weight_envelope": "linear",
"visit_weight_envelope": "linear",
"architecture": "svhn_model",
"visit_weight": 0.2,
"max_steps": "12000"

Synth. Digits --> SVHN

"target_dataset": "svhn",
"walker_weight_envelope_delay": "2000",
"max_checkpoints": 5,
"dataset": "synth",
"sup_per_class": -1,
"sup_per_batch": 100,
"walker_weight_envelope_steps": 1,
"walker_weight_envelope": "linear",
"decay_steps": 9000,
"unsup_batch_size": 1000,
"visit_weight_envelope": "linear",
"architecture": "svhn_model",
"visit_weight": 0.2,
"max_steps": "20000",

If you use the code, please cite the paper "Learning by Association - A versatile semi-supervised training method for neural networks" or "Associative Domain Adaptation":

@string{cvpr="IEEE Conference on Computer Vision and Pattern Recognition (CVPR)"}
@InProceedings{haeusser-cvpr-17,
  author = 	 "P. Haeusser and A. Mordvintsev and D. Cremers",
  title = 	 "Learning by Association - A versatile semi-supervised training method for neural networks",
  booktitle = cvpr,
  year = 	 "2017",
}

@string{iccv="IEEE International Conference on Computer Vision (ICCV)"}
@InProceedings{haeusser-iccv-17,
  author = 	 "P. Haeusser and T. Frerix and A. Mordvintsev and D. Cremers",
  title = 	 "Associative Domain Adaptation",
  booktitle = iccv,
  year = 	 "2017",
}

For questions please contact Philip Haeusser ([email protected]).