Skip to content
Different optimizers tested in Tensorflow
Branch: master
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Permalink
Type Name Latest commit message Commit time
Failed to load latest commit information.
adalbert
cocob
configs
scinol
scripts
.gitignore
Dockerfile
LICENSE
README.md
artificial_new.csv
clean.sh
datasets.py
distributions.py
exp0_best_runs.txt
models.py
plot.py
plot_b128.py
plot_distributions.py
plot_linear.py
preprocess.py
requirements.txt
sc2_numpy.py
short_names.py
tb.sh
test.py
util_plot.py

README.md

Requiremets

To run this coe, a Linux machine with python3 is needed. There is a high chance it will also work on Mac but it wasn't tested.

Additionally python3 dependencies from requirements.txt (sudo pip3 install -r requirements.txt).

Internet connection is also needed (to download data).

Docker

If you wish so, you can run the scripts in docker, a Dockerfile and launch script is prepared (see: scripts/docker_build_n_run.sh). In this case no dependencies except for Docker are required

How to run

Just run:

./scripts.icml_reproduce.sh

You should see this output on the console:

# Optimizers: 26
# Models: 1
Tests in total: 26
Downloading bank-additional.zip 101.3%
Successfully downloaded bank-additional.zip 444572 bytes.
Running optimizers for dataset: 'UCI_Bank', model: 'LR{'init0': True}'
adagrad_l1.0:   0%|                                      | 0/10 [00:00<?, ?it/s]

Do not be bothered by the fact that more than 100% is downloaded. Our scripts work hard!

Unfortunately reproducing the whole experiment will take much time on a single machine (More than a day most likely) because the code was created with more focus on deep models and batchsize>1.

Output

The scripts create tb_logs_linear directory with summaries from tensorflow, however do not try to run it via tensorboard because so much data will clog your ram. This directory weighs ~2.7 GB because tensordflow apparently can't write data efficiently.

Additionally graphs_linear directory will be created with graphs just like those used in the paper and more (separate graphs for each algorithm and runs for learning rates not shown in the paper).

Artificial experiment

tb_logs_art are also created which contain logs from artificial experiment. To see them run:

tensorboard --logdir=tb_logs_art
You can’t perform that action at this time.