https://github.com/1ytic/warp-rna

Recurrent Neural Aligner
https://github.com/1ytic/warp-rna

cuda forward-backward rna rnn-transducer

Last synced: about 2 months ago
JSON representation

Recurrent Neural Aligner

• Host: GitHub
• URL: https://github.com/1ytic/warp-rna
• Owner: 1ytic
• License: mit
• Created: 2019-08-11T08:12:41.000Z (over 5 years ago)
• Default Branch: master
• Last Pushed: 2020-04-14T10:48:16.000Z (almost 5 years ago)
• Last Synced: 2024-11-14T22:44:22.665Z (3 months ago)
• Topics: cuda, forward-backward, rna, rnn-transducer
• Language: Python
• Size: 71.3 KB
• Stars: 49
• Watchers: 7
• Forks: 7
• Open Issues: 2
• Metadata Files:
  • Readme: README.md
  • License: LICENSE

Awesome Lists containing this project

README

        
# Recurrent Neural Aligner

Recurrent Neural Aligner (RNA) is a restricted version of RNN-Transducer loss (RNN-T). It assumes that the length of input sequence is equal to or greater than the length of target sequence ([Sak, et al., 2017](https://www.isca-speech.org/archive/Interspeech_2017/pdfs/1705.PDF); [Dong, et al., 2018](https://arxiv.org/abs/1806.06342)).

In this repository you can find a reference implementation of RNA loss in python, as well as a fast GPU version on CUDA. In order to apply the same efficient procedure from [warp_rnnt](https://github.com/1ytic/warp-rnnt), the alphas/betas arrays are represented as shown below. Because RNA loss assumes that a model produces only one output at each input step, we can reduce the T dimension to S=T-U+2.

![](aligner.gif)

## Convergence

In the figure below shows a sanity check of the implementation for the speech recognition task with a small dataset. The decoding procedure was the same for both models. As you can see RNN-T loss more stable in this case.

![](check.png)

If you have a successful example of using RNA loss or if you find errors in this implementation, please make an issue for this repository.

## Install

Currently, there is only a binding for PyTorch 1.0 and higher.

```bash

pip install warp_rna

```

## Test

There is a unittest in pytorch_binding/warp_rna which includes tests for arguments and outputs as well.

```bash

cd ..

python -m warp_rna.test

```