rwth-i6/returnn

What it solves

RETURNN provides a flexible and efficient framework for training modern recurrent neural network (RNN) architectures, specifically addressing the need for fast training and decoding speeds in research and production environments.

How it works

Built on PyTorch and TensorFlow, the framework optimizes the training of RNNs in multi-GPU environments. It utilizes sequence-chunking for batch training and includes a custom fast CUDA kernel for Long Short-Term Memory (LSTM) networks, as well as support for multidimensional LSTMs.

Who it’s for

It is designed for researchers and developers working with recurrent neural networks, particularly those implementing encoder-attention-decoder models for tasks like automatic speech recognition (ASR) and machine translation (MT).

Highlights

• Multi-GPU Optimization: Optimized for reliable and fast training across multiple devices.
• Custom CUDA Kernels: Includes a specialized fast CUDA kernel for LSTMs to improve performance.
• Flexible Architectures: Supports a wide variety of encoder-attention-decoder models.
• Memory Management: Includes built-in memory management for handling large datasets.