Project Icon

returnn

Optimize Recurrent Neural Networks with Multi-GPU Support in RETURNN

Product DescriptionExplore the RETURNN framework, an extensible Theano/TensorFlow-based platform for training recurrent neural networks efficiently across multiple GPUs. The framework prioritizes simplicity, flexibility, and efficiency, streamlining setup and troubleshooting. RETURNN accommodates a range of experiments and neural network models while offering high-speed training and decoding essential for research and industry. Core features include mini-batch training, sequence chunking for RNNs, custom CUDA kernels for LSTM, and multi-device load balancing. Illustrative real-world cases and benchmarks demonstrate RETURNN's superior capacity and performance, particularly in speech recognition and encoder-decoder modeling.
Project Details