reluble 0.1.0

Reliable deep network training, saving you hours.

Reluble

A thin PyTorch wrapper which performs static and dynamic analysis of the computational graph to prevent latent errors.

Install

pip install reluble

Tests

Install the testing dependencies:

• pytest, e.g. with pip install pytest.
• torchvision, e.g. with pip install torchvision. This is probably installed alongside torch.

Clone the repository and change into the root:

git clone [e-mail]:arcadelab/reluble.git
cd reluble

To run all the tests, simply run pytest. It is recommended to disable capturing, e.g. with:

pytest -s

These tests can take a long time, particularly if you are running on CPU. It is recommended to run on GPU. Some of the tests will produce output files in ./outputs/.

To show corruptions working for a good net, run:

pytest -s tests/test_training.py::TestCorruptions::test_good_Net

And for a bad net:

pytest -s tests/test_training.py::TestCorruptions::test_bad_Net

In this case, the LearningError is the desired behavior.