#Maté
Maté is a MATLAB library for ConvNets (Convolutional Neural Networks). Maté is derived from MatConvNet. Maté uses MatConvNet routines for several time-critical operations and can load and manipulate pretrained networks supplied with MatConvNet.
Maté is object-oriented and has the goal to simplify prototyping and experimentation, in particular to minimize the efforts needed to try new non-standard layers and new non-standard network graphs.
Maté continues the tradition of naming a ConvNet library after a caffeine-related drink. Other such libraries that also inspired Maté include Decaf (Python, CPU) and Caffe (C++ with interfaces, CPU/GPU).
##Functionality Maté provides functionality for:
- Easy definition of new layers (just define forward and, if needed, backward procedures; it often takes couple of lines in MATLAB).
- Easy definition and training of non-chain networks (any directed acyclic graph can be used) within MATLAB code.
- Easy loading and manipulation of pretrained networks.
- Easy GPU mode thanks to MATLAB Parallel Toolbox and MatConvNet routines. For many (most?) new layers same MATLAB code can be reused for CPU and GPU modes.
- Easy visualization of the internal state of a network during training.
##Installation Install MatConvNet from here, clone this repository, and run setup.m (which simply adds subfolders to MATLAB path).
##Howto
- Build a network (+other basics)
- Use a pre-trained network
- Train a network
- Layer catalog
- Define a new layer
##Citation If you use Maté to prepare your publication, please cite MatConvNet report as suggested here.
##Status The project is in a very early prerelease state, use at your own risk.