The main purpose of this package is to easily create/run/replicate deep learning experiments using Torch with minimal code.

Initially inspired by ImageNet multi-GPU.

Similar frameworks:

• dp
• torchnet

• Multi-GPU implementation with automatic saving/loading of models.
• Data iterator and loaders with multi-threading support.
• Multiple types of training (simple, GAN, WGAN, BEGAN), with automatic checkpointing and logging of training loss.
• Easy experiment creation and dispatching. Experiments are transferable accross machines (e.g. can start training on a GPU machine and finish on a non-GPU machine).
• Slurm scheduler support for usage on HPC facilities.
• Settings parsing, optimizer, logging, colorspaces (XYZ, IPT, LMS, Lαβ). More info here.
• Data loader interfaces for MNIST, CIFAR, CelebA, Places, pix2pix.
• Implementations of some (standard) models, including LeNet5, VGG, AlexNet, Squeezenet, Colornet, UNET.

Make sure you have Torch installed.

To install use:

git clone https://github.com/dmarnerides/dlt.git
cd dlt
./install.sh

I created this toolbox for my PhD, mostly to learn Lua, understand Torch in depth, and have a consistent workflow accross multiple machines and HPC facilities.

Only tested on Ubuntu and CentOS.

If you use this package you will probably encounter bugs. If so please let me know!

Use at your own risk.

If you use this code please cite the repo.

[email protected]