SamsungLabs / pytorch-ensembles

The official PyTorch implementation of:
Pitfalls of In-Domain Uncertainty Estimation and Ensembling in Deep Learning, ICLR'20

OpenReview / arXiv / Poster video (5 mins) / Blog / bibtex

Poster video (5 mins)

Environment Setup

The following allows to create and to run a python environment with all required dependencies using miniconda:

conda env create -f condaenv.yml
conda activate megabayes

Logs, Plots, Tables, Pre-trained weights

At the notebooks folder we provide:

• Saved logs with all computed results
• Examples of ipython notebooks to reproduce plots, tables, and compute the deep ensemble equivalent (DEE) score

Pre-trained weights of of some models are available at link: Deep Ensembles on ImageNet (~10G) and Deep Ensembles on CIFARs (~62G), etc. The weights can be also download with a command line interface by yadisk-direct:

pip3 install wldhx.yadisk-direct

% ImageNet
curl -L $(yadisk-direct https://yadi.sk/d/rdk6ylF5mK8ptw?w=1) -o deepens_imagenet.zip
unzip deepens_imagenet.zip 

% CIFARs
curl -L $(yadisk-direct https://yadi.sk/d/8C5jBz-licWMqQ?w=1) -o deepens_cifars.zip
unzip deepens_cifars.zip 

Pre-trained weights for other models can be provided on the request---make an issue if you need some specific models.

Evaluation

The evaluation of ensembling methods can be done using the scripts from the ens folder, which contains a separate script for each ensembling method. The scrips have the following interface:

ipython -- ens/ens-<method>.py -h
usage: ens-onenet.py [-h] --dataset DATASET [--data_path PATH]
                     [--models_dir PATH] [--aug_test] [--batch_size N]
                     [--num_workers M] [--fname FNAME]

optional arguments:
  -h, --help         show this help message and exit
  --dataset DATASET  dataset name CIFAR10/CIFAR100/ImageNet (default: None)
  --data_path PATH   path to a data-folder (default: ~/data)
  --models_dir PATH  a dir that stores pre-trained models (default: ~/megares)
  --aug_test         enables test-time augmentation (default: False)
  --batch_size N     input batch size (default: 256)
  --num_workers M    number of workers (default: 10)
  --fname FNAME      comment to a log file name (default: unnamed)

• All scripts assume that pytorch-ensembles is the current working directory (cd pytorch-ensembles).
• The scripts will write .csv logs in pytorch-ensembles/logs in the following format rowid, dataset, architecture, ensemble_method, n_samples, metric, value, info.
• The notebooks folder contains ipython notebooks to reproduce the tables and plots using these logs
• The scripts will write final log-probs of every method in '.npy' format to pytorch-ensembles/.megacache folder.
• The interface for K-FAC-Laplace differs and is described below.

Examples:

ipython -- ens/ens-onenet.py  --dataset=CIFAR10/CIFAR100/ImageNet
ipython -- ens/ens-deepens.py --dataset=CIFAR10/CIFAR100/ImageNet
ipython -- ens/ens-sse.py     --dataset=CIFAR10/CIFAR100/ImageNet
ipython -- ens/ens-csgld.py   --dataset=CIFAR10/CIFAR100
ipython -- ens/ens-fge.py     --dataset=CIFAR10/CIFAR100/ImageNet
ipython -- ens/ens-dropout.py --dataset=CIFAR10/CIFAR100
ipython -- ens/ens-vi.py      --dataset=CIFAR10/CIFAR100/ImageNet
ipython -- ens/ens-swag.py    --dataset=CIFAR10/CIFAR100

Training models

Deep Ensemble members, Variational Inference (CIFAR-10/100)

All the models trained on CIFAR use a single GPU for training. Examples of training commands:

bash train/train_cifar.sh \
--dataset CIFAR10/CIFAR100 \
--arch VGG16BN/PreResNet110/PreResNet164/WideResNet28x10 \
--method regular/vi

Fast Geometric Ensembling, SWA-Gaussian, Snapshot Ensembles, and Cyclical SGLD (CIFAR-10/100)

Examples of training commands:

bash train/train_fge.sh CIFAR10 WideResNet28x10 1 ~/weights ~/datasets
bash train/train_swag.sh CIFAR10 WideResNet28x10 1 ~/weights ~/datasets
bash train/train_sse_mcmc.sh CIFAR10 WideResNet28x10 1 ~/weights ~/datasets SSE
bash train/train_sse_mcmc.sh CIFAR10 WideResNet28x10 1 ~/weights ~/datasets cSGLD

Script parameters: dataset, architecture name, training run id, root directory for saving snapshots (created automatically), root directory for datasets (downloaded automatically)

K-FAC Laplace (CIFAR-10/100, ImageNet)

Given a checkpoint, ens/ens-kfacl.py builds the Laplace approximation and produces the results of the approximate posterior averaging. Use keys --scale_search and --gs_low LOW --gs_high HIGH --gs_num NUM to find the optimal posterior noise scale on the validation set. We have used the following noise scales (also listed in Table 3, Appendix D in the paper):

For ResNet50 on ImageNet, the optimal scale found was 2.0 with test-time augmentation and 6.8 without test-time augmentation.

Refer to `ens/ens-kfacl.py' for the full list of arguments and default values. Example use:

ipython -- ens/ens-kfacl.py --file CHECKPOINT --data_path DATA --dataset CIFAR10 --model PreResNet110 --scale 0.213

Deep Ensemble members, Variational Inference, Snapshot Ensembles (ImageNet)

Examples of training commands:

bash train/train_imagenet.sh --method regular/sse/fge/vi

We strongly recommend using multi-gpu training for Snapshot Ensembles.

Attribution

Parts of this code are based on the following repositories:

• Stochastic Weight Averaging (SWA). Pavel Izmailov, Dmitrii Podoprikhin, Timur Garipov, Dmitry Vetrov, Andrew Gordon Wilson.
• A Simple Baseline for Bayesian Deep Learning. Wesley Maddox, Timur Garipov, Pavel Izmailov, Dmitry Vetrov, Andrew Gordon Wilson.
• Cyclical Stochastic Gradient MCMC for Bayesian Deep Learning. Ruqi Zhang, Chunyuan Li, Jianyi Zhang, Changyou Chen and Andrew Gordon Wilson.
• Loss Surfaces, Mode Connectivity, and Fast Ensembling of DNNs. Timur Garipov, Pavel Izmailov, Dmitrii Podoprikhin, Dmitry Vetrov and Andrew Gordon Wilson.
• PyTorch
• PyTorch Examples

Citation

If you found this code useful, please cite our paper

@article{ashukha2020pitfalls,
  title={Pitfalls of In-Domain Uncertainty Estimation and Ensembling in Deep Learning},
  author={Ashukha, Arsenii and Lyzhov, Alexander and Molchanov, Dmitry and Vetrov, Dmitry},
  journal={arXiv preprint arXiv:2002.06470},
  year={2020}
}