configuration.md

Configuration File

Configuration files manage hyperparameters of trainings. It must provide at least the following information:

• batchsize,
• epoch,
• dataset (both train / test),
• model (network structure),
• optimizer (optimization algorithm like Adam),
• mode (list of configuration mode).

Additionally, we can provide the following information:

• hook (called before calculation by optimization algorithms at every iteration),
• extension (called by Trainer object of Chainer).

Structure of config/ directory

Sample configuration files can be found under config/ directory.

|- attack/
|  - configuration files for attack methods
|- outer_polytope/
|  - experiment 6.1
|- parseval_svhn/
|  - experiment 6.2
|- svhn_avg/
|  - experiment 6.2 (average-pooling ver.)

Configuration of attack methods

Configuration files for attack methods are different from those for training. They must provide the following information:

• attack method,
• args of the attack method,
• kwargs of the attack method.

They can be found under config/attack/ directory.

Dataset

Datasets are provided under src/dataset/ directory.

Note: in this public repository, they return a pair of training and test data, so please be careful if you start a new project using this repository.

Model

Models are provided under src/model/ directory.

Mode

Mode is a list of configurations. We provide the folloing modes:

• default (usual training)
• lmt (train using lmt (Prop.1 ver.))
• lmt_fc (train using lmt (Prop.2 ver.) activated only when lmt is also specified)
• parseval (Parseval networks)

Hook

Hooks are provided under src/hook/ directory.

Extension

Extensions are provided under src/extension/ directory.