Cheatsheet

Abstraction: Comprehensive reference index of Torch7 packages, tutorials, and GPU setup

Key points:

• Torch7 is a Lua/LuaJIT-based scientific computing and deep learning framework; core packages include nn (neural networks), optim (SGD, Adagrad, LBFGS, etc.), and rnn for recurrent models.
• GPU support requires cutorch and cunn packages; CUDA double precision is not supported; multi-GPU parallel training is supported via cutorch.
• EC2 AMI (ami-b36981d8, g2.2xlarge) available pre-loaded with CUDA, cuDNN, iTorch, and Anaconda; Docker images available via kaixhin/torch.
• Notable demos include Karpathy's char-rnn (character-level LSTM language models), Neural Turing Machines, variational autoencoders, and DQN/A3C reinforcement learning.
• Torch's nn.LookupTable supports neural language models; word2vec and GloVe embeddings can be loaded as Tensors.
• Debugging via ZeroBrane Studio (standard Lua IDE) or Eclipse LDT; trepl provides an embedabble Lua REPL.

Connections: Torch · Lua · Deep Learning · Neural Networks · Machine Learning Frameworks

Source: https://github.com/torch/torch7/wiki/Cheatsheet#tutorials-demos-by-category