DeepOBS

A Deep Learning Optimizer Benchmark Suite

About DeepOBS

DeepOBS is a benchmarking suite that drastically simplifies, automates and improves the evaluation of deep learning optimizers. It can evaluate the performance of new optimizers on a variety of real-world test problems and automatically compare them with realistic baselines.



Fast

You only have to code your new optimizer, we take care of the benchmarking for you!



Fair

Automatically use competitive baselines!



Comparable

The same test problems for all optimizers!

How to Install DeepOBS

You can install the latest stable release of DeepOBS using pip

pip install deepobs

You are ready to dive into the tutorials found in the DeepOBS documentation.

Leaderboard

Overview over the current optimizer leaderboard on the DeepOBS test problems. Click on See Full Results to see the plots and tables of the full benchmarking results.

Quadratic Deep

A 100-dimensional noisy quadratic problem with an eigenspectrum similar to the one reported for deep neural networks.

Optimizer Test Loss Speed
#1 Momentum 87.05 70.5
#2 Adam 87.11 39.9
#3 SGD 87.40 51.1

MNIST - VAE

A basic variational autoencoder for the MNIST data set with three convolutional and three deconvolutional layers.

Optimizer Test Loss Speed
#1 Adam 27.83 1.0
#2 SGD 38.46 1.0
#3 Momentum 52.93 1.0

F-MNIST - CNN

A simple convolutional network for the Fashion-MNIST data set, consisting of two conv and two fully-connected layers.

Optimizer Test Accuracy Speed
#1 Adam 92.34 % 40.1
#2 SGD 92.27 % 40.6
#3 Momentum 92.14 % 59.1

CIFAR-10 - CNN

A slightly larger convolutional network for the Cifar-10 data set, with three conv and three fully-connected layers.

Optimizer Test Accuracy Speed
#1 Adam 84.75 % 36.0
#2 Momentum 84.41 % 40.7
#3 SGD 83.71 % 42.5

F-MNIST - VAE

A basic variational autoencoder for the Fashion-MNIST data set with three convolutional and three deconvolutional layers.

Optimizer Test Loss Speed
#1 Adam 23.07 1.0
#2 SGD 23.80 1.0
#3 Momentum 59.23 1.0

CIFAR-100 - All CNN C

Variant C of the All Convolutional Network from Striving for Simplicity for the CIFAR-100 data set consisting solely of convolutional layers.

Optimizer Test Accuracy Speed
#1 Momentum 60.33 % 72.8
#2 SGD 57.06 % 128.7
#3 Adam 56.15 % 152.6

SVHN - Wide ResNet 16-4

The Wide ResNet 16-4 for the Street View House Numbers data set using the variant with 16 conv layers and a widening factor of 4.

Optimizer Test Accuracy Speed
#1 Momentum 95.53 % 10.8
#2 SGD 95.37 % 28.3
#3 Adam 95.25 % 12.1

Tolstoi - Char RNN

A recurrent neural network for character-level language modeling on the novel War and Peace by Leo Tolstoy using two LSTM layers.

Optimizer Test Accuracy Speed
#1 SGD 62.07 % 47.7
#2 Momentum 61.30 % 88.0
#3 Adam 61.23 % 62.8

DeepOBS is Open Source

If you use DeepOBS in your research, please cite our paper.

@InProceedings{ schneider2018deepobs,
Title = {Deep{OBS}: A Deep Learning Optimizer Benchmark Suite},
Author = {Frank Schneider and Lukas Balles and Philipp Hennig},
Booktitle = {International Conference on Learning Representations},
Year = {2019},
Url = {https://openreview.net/forum?id=rJg6ssC5Y7}
}