Still WIP. Bugs to correct and changing API. Stable version will be pushed soon (~ 12/2015)
##What
Prototype for a visual framework for Deep Learning. Deep Learning has shown exciting results in several domains of machine learning. Training a deep learning network required a lot of expertise and understanding of the underlying concepts. For python, some libraries as Theano and then Lasagne have enabled people who design those models to spend more time on understanding what they rather than developing their models. By enabling people to have a framework to manage models and training sessions through usefull visualization, this project aims to push trough this direction.
Lasagne (https://github.com/Lasagne/Lasagne) developers have made an awesome job to leverage the power of Theano. We can now be efficient when defining a new network model. However, when your models are defined, you still need to manage their training, and their evolution by yourself. LasagneEngine is about making you efficient to iterate when you build new models, providing the following features :
- Run several INDEPENDANT models on the same machine when several GPUs are available.
- Manage theano variables allocation in an easy wayand thus GPUs memory allocation
- Save and load a models and parameters (weights in the network)
- Manage live training sessions
- Track the evolution of training metrics (training loss, validation loss, ..) on a running session
- Visualize interesting metrics on your parameters
- Engine is the main class. It manage your models and your different training sessions
- Displayer is a class that contains the visualisation bits for the different classes. It's the step 0 through a future UI for Deep Learning
- Model is just a container for a model that you define using Lasagne
- DataManager allow you to associate numpy arrays in memory with name and take care to load this data in theano variables (aka in GPUs) when needed by your sessions
- Session handles the training of your network
- Params define a specific set of parameters for a specific models. Each models can evolve through different Params that you want to keep track of.