SENTRY: Selective Entropy Optimization via Committee Consistency for Unsupervised Domain Adaptation
PyTorch Code forViraj Prabhu, Shivam Khare, Deeksha Kartik, Judy Hoffman
Many existing approaches for unsupervised domain adaptation (UDA) focus on adapting under only data distribution shift and offer limited success under additional cross-domain label distribution shift. Recent work based on self-training using target pseudolabels has shown promise, but on challenging shifts pseudolabels may be highly unreliable and using them for self-training may cause error accumulation and domain misalignment. We propose Selective Entropy Optimization via Committee Consistency (SENTRY), a UDA algorithm that judges the reliability of a target instance based on its predictive consistency under a committee of random image transformations. Our algorithm then selectively minimizes predictive entropy to increase confidence on highly consistent target instances, while maximizing predictive entropy to reduce confidence on highly inconsistent ones. In combination with pseudolabel-based approximate target class balancing, our approach leads to significant improvements over the state-of-the-art on 27/31 domain shifts from standard UDA benchmarks as well as benchmarks designed to stress-test adaptation under label distribution shift.
Table of Contents
Setup and Dependencies
- Create an anaconda environment with Python 3.6:
conda create -n sentry python=3.6.8
and activate:conda activate sentry
- Navigate to the code directory:
cd code/
- Install dependencies:
pip install -r requirements.txt
And you're all set up!
Usage
Download data
Data for SVHN->MNIST is downloaded automatically via PyTorch. Data for other benchmarks can be downloaded from the following links. The splits used for our experiments are already included in the data/
folder):
- DomainNet
- OfficeHome
- VisDA2017 (only train and validation needed)
Pretrained checkpoints
To reproduce numbers reported in the paper, we include a a few pretrained checkpoints. We include checkpoints (source and adapted) for SVHN to MNIST (DIGITS) in the checkpoints
directory. Source and adapted checkpoints for Clipart to Sketch adaptation (from DomainNet) and Real_World to Product adaptation (from OfficeHome RS-UT) can be downloaded from this link, and should be saved to the checkpoints/source
and checkpoints/SENTRY
directory as appropriate.
Train and adapt model
- Natural label distribution shift: Adapt a model from to for a given (where benchmark may be DomainNet, OfficeHome, VisDA, or DIGITS), as follows:
python train.py --id <experiment_id> \
--source <source> \
--target <target> \
--img_dir <image_directory> \
--LDS_type <LDS_type> \
--load_from_cfg True \
--cfg_file 'config/<benchmark>/<cfg_file>.yml' \
--use_cuda True
SENTRY hyperparameters are provided via a sentry.yml
config file in the corresponding config/<benchmark>
folder (On DIGITS, we also provide a config for baseline adaptation via DANN). The list of valid source/target domains per-benchmark are:
- DomainNet: real, clipart, sketch, painting
- OfficeHome_RS_UT: Real_World, Clipart, Product
- OfficeHome: Real_World, Clipart, Product, Art
- VisDA2017: visda_train, visda_test
- DIGITS: Only svhn (source) to mnist (target) adaptation is currently supported.
Pass in the path to the parent folder containing dataset images via the --img_dir <name_of_directory>
flag (eg. --img_dir '~/data/DomainNet'
). Pass in the label distribution shift type via the --LDS_type
flag: For DomainNet, OfficeHome (standard), and VisDA2017, pass in --LDS_type 'natural'
(default). For OfficeHome RS-UT, pass in --LDS_type 'RS_UT'
. For DIGITS, pass in --LDS_type
as one of IF1
, IF20
, IF50
, or IF100
, to load a manually long-tailed target training split with a given imbalance factor (IF), as described in Table 4 of the paper.
To load a pretrained DA checkpoint instead of training your own, additionally pass --load_da True
and --id <benchmark_name>
to the script above. Finally, the training script will log performance metrics to the console (average and aggregate accuracy), and additionally plot and save some per-class performance statistics to the results/
folder.
Note: By default this code runs on GPU. To run on CPU pass: --use_cuda False
Reference
If you found this code useful, please consider citing:
@article{prabhu2020sentry
author = {Prabhu, Viraj and Khare, Shivam and Kartik, Deeksha and Hoffman, Judy},
title = {SENTRY: Selective Entropy Optimization via Committee Consistency for Unsupervised Domain Adaptation},
year = {2020},
journal = {arXiv preprint: 2012.11460},
}
Acknowledgements
We would like to thank the developers of PyTorch for building an excellent framework, in addition to the numerous contributors to all the open-source packages we use.
License
MIT