HackerReposA FIELD GUIDE TO OPEN SOURCE
ProSponsor the shelf♧ 0
← Back to top all-time
salesforce
pytorch-qrnn
Python2017–2020
salesforce / pytorch-qrnn · Archived

pytorch-qrnn

PyTorch implementation of the Quasi-Recurrent Neural Network - up to 16 times faster than NVIDIA's cuDNN LSTM

AI & Machine LearningSystems & InfrastructureNetworking
Explore on GitHubRead Hacker News discussions
4Distinct people
7Mentions
0Explicit recommenders
1.3KGitHub stars
REPOSITORY AT A GLANCE

What to know before exploring

What it is

PyTorch implementation of the Quasi-Recurrent Neural Network - up to 16 times faster than NVIDIA's cuDNN LSTM

GitHub

Where it is used

No documented product use has been verified yet.

What people are saying

No recommendation phrase was found in the indexed comments.

No classified warning is attached to this page. Read the discussions before treating silence as approval.

Also in these discussions

No other repository is repeatedly linked in the same discussions.

This week on Hacker News

Nov 7–Nov 14 UTC · Direct repository links and separately labeled related project stories. Indexed through 11/14/2026, 5:04:57 PM UTC.

No direct repository links were found in the indexed HN discussions over the past seven days. Older discussions are available below.

Mentions over time

HN mentions by month7 dated mentions

Observed mentions: 2017-10-09 21:43:11 UTC to 2020-04-08 19:41:45 UTC.

0122017-10-01 00:00:00 UTC to 2017-11-01 00:00:00 UTC (exclusive): 2 mentions; 1 submissions, 1 commentsOct 20172017-11-01 00:00:00 UTC to 2017-12-01 00:00:00 UTC (exclusive): 0 mentions; 0 submissions, 0 comments2017-12-01 00:00:00 UTC to 2018-01-01 00:00:00 UTC (exclusive): 0 mentions; 0 submissions, 0 comments2018-01-01 00:00:00 UTC to 2018-02-01 00:00:00 UTC (exclusive): 0 mentions; 0 submissions, 0 comments2018-02-01 00:00:00 UTC to 2018-03-01 00:00:00 UTC (exclusive): 0 mentions; 0 submissions, 0 comments2018-03-01 00:00:00 UTC to 2018-04-01 00:00:00 UTC (exclusive): 0 mentions; 0 submissions, 0 comments2018-04-01 00:00:00 UTC to 2018-05-01 00:00:00 UTC (exclusive): 0 mentions; 0 submissions, 0 comments2018-05-01 00:00:00 UTC to 2018-06-01 00:00:00 UTC (exclusive): 1 mentions; 0 submissions, 1 comments2018-06-01 00:00:00 UTC to 2018-07-01 00:00:00 UTC (exclusive): 2 mentions; 0 submissions, 2 comments2018-07-01 00:00:00 UTC to 2018-08-01 00:00:00 UTC (exclusive): 0 mentions; 0 submissions, 0 comments2018-08-01 00:00:00 UTC to 2018-09-01 00:00:00 UTC (exclusive): 0 mentions; 0 submissions, 0 comments2018-09-01 00:00:00 UTC to 2018-10-01 00:00:00 UTC (exclusive): 0 mentions; 0 submissions, 0 comments2018-10-01 00:00:00 UTC to 2018-11-01 00:00:00 UTC (exclusive): 0 mentions; 0 submissions, 0 comments2018-11-01 00:00:00 UTC to 2018-12-01 00:00:00 UTC (exclusive): 0 mentions; 0 submissions, 0 comments2018-12-01 00:00:00 UTC to 2019-01-01 00:00:00 UTC (exclusive): 0 mentions; 0 submissions, 0 comments2019-01-01 00:00:00 UTC to 2019-02-01 00:00:00 UTC (exclusive): 0 mentions; 0 submissions, 0 commentsJan 20192019-02-01 00:00:00 UTC to 2019-03-01 00:00:00 UTC (exclusive): 0 mentions; 0 submissions, 0 comments2019-03-01 00:00:00 UTC to 2019-04-01 00:00:00 UTC (exclusive): 0 mentions; 0 submissions, 0 comments2019-04-01 00:00:00 UTC to 2019-05-01 00:00:00 UTC (exclusive): 1 mentions; 0 submissions, 1 comments2019-05-01 00:00:00 UTC to 2019-06-01 00:00:00 UTC (exclusive): 0 mentions; 0 submissions, 0 comments2019-06-01 00:00:00 UTC to 2019-07-01 00:00:00 UTC (exclusive): 0 mentions; 0 submissions, 0 comments2019-07-01 00:00:00 UTC to 2019-08-01 00:00:00 UTC (exclusive): 0 mentions; 0 submissions, 0 comments2019-08-01 00:00:00 UTC to 2019-09-01 00:00:00 UTC (exclusive): 0 mentions; 0 submissions, 0 comments2019-09-01 00:00:00 UTC to 2019-10-01 00:00:00 UTC (exclusive): 0 mentions; 0 submissions, 0 comments2019-10-01 00:00:00 UTC to 2019-11-01 00:00:00 UTC (exclusive): 0 mentions; 0 submissions, 0 comments2019-11-01 00:00:00 UTC to 2019-12-01 00:00:00 UTC (exclusive): 0 mentions; 0 submissions, 0 comments2019-12-01 00:00:00 UTC to 2020-01-01 00:00:00 UTC (exclusive): 0 mentions; 0 submissions, 0 comments2020-01-01 00:00:00 UTC to 2020-02-01 00:00:00 UTC (exclusive): 0 mentions; 0 submissions, 0 comments2020-02-01 00:00:00 UTC to 2020-03-01 00:00:00 UTC (exclusive): 0 mentions; 0 submissions, 0 comments2020-03-01 00:00:00 UTC to 2020-04-01 00:00:00 UTC (exclusive): 0 mentions; 0 submissions, 0 comments2020-04-01 00:00:00 UTC to 2020-05-01 00:00:00 UTC (exclusive): 1 mentions; 0 submissions, 1 commentsApr 2020

Counts come from the repository’s dated source evidence. Index cutoff: 2026-11-02 08:16:59 UTC.

View mention counts and date ranges
UTC ranges include their start and exclude their end. Select a period to filter the source evidence.
PeriodStart (UTC)End (exclusive, UTC)MentionsSubmissionsComments
Oct 20172017-10-01 00:00:00 UTC2017-11-01 00:00:00 UTC211
Nov 20172017-11-01 00:00:00 UTC2017-12-01 00:00:00 UTC000
Dec 20172017-12-01 00:00:00 UTC2018-01-01 00:00:00 UTC000
Jan 20182018-01-01 00:00:00 UTC2018-02-01 00:00:00 UTC000
Feb 20182018-02-01 00:00:00 UTC2018-03-01 00:00:00 UTC000
Mar 20182018-03-01 00:00:00 UTC2018-04-01 00:00:00 UTC000
Apr 20182018-04-01 00:00:00 UTC2018-05-01 00:00:00 UTC000
May 20182018-05-01 00:00:00 UTC2018-06-01 00:00:00 UTC101
Jun 20182018-06-01 00:00:00 UTC2018-07-01 00:00:00 UTC202
Jul 20182018-07-01 00:00:00 UTC2018-08-01 00:00:00 UTC000
Aug 20182018-08-01 00:00:00 UTC2018-09-01 00:00:00 UTC000
Sep 20182018-09-01 00:00:00 UTC2018-10-01 00:00:00 UTC000
Oct 20182018-10-01 00:00:00 UTC2018-11-01 00:00:00 UTC000
Nov 20182018-11-01 00:00:00 UTC2018-12-01 00:00:00 UTC000
Dec 20182018-12-01 00:00:00 UTC2019-01-01 00:00:00 UTC000
Jan 20192019-01-01 00:00:00 UTC2019-02-01 00:00:00 UTC000
Feb 20192019-02-01 00:00:00 UTC2019-03-01 00:00:00 UTC000
Mar 20192019-03-01 00:00:00 UTC2019-04-01 00:00:00 UTC000
Apr 20192019-04-01 00:00:00 UTC2019-05-01 00:00:00 UTC101
May 20192019-05-01 00:00:00 UTC2019-06-01 00:00:00 UTC000
Jun 20192019-06-01 00:00:00 UTC2019-07-01 00:00:00 UTC000
Jul 20192019-07-01 00:00:00 UTC2019-08-01 00:00:00 UTC000
Aug 20192019-08-01 00:00:00 UTC2019-09-01 00:00:00 UTC000
Sep 20192019-09-01 00:00:00 UTC2019-10-01 00:00:00 UTC000
Oct 20192019-10-01 00:00:00 UTC2019-11-01 00:00:00 UTC000
Nov 20192019-11-01 00:00:00 UTC2019-12-01 00:00:00 UTC000
Dec 20192019-12-01 00:00:00 UTC2020-01-01 00:00:00 UTC000
Jan 20202020-01-01 00:00:00 UTC2020-02-01 00:00:00 UTC000
Feb 20202020-02-01 00:00:00 UTC2020-03-01 00:00:00 UTC000
Mar 20202020-03-01 00:00:00 UTC2020-04-01 00:00:00 UTC000
Apr 20202020-04-01 00:00:00 UTC2020-05-01 00:00:00 UTC101

What Hacker News says

7 sources
Smerity·Apr 8, 2020
For high performance CUDA kernels people still need to write derivatives by hand. I know this as for my own research, and for many production systems, I'd still need to write it myself. Many of my architectures wouldn't have been possible without writing the CUDA myself (Quasi-Recurrent Neural Network[1]) or using optimized hand written black boxes (cuDNN RNN). The lack of open optimized hand written CUDA kernels has actually been an impediment to progress in the field. Automatic differentiation allows for great flexibility and composability but the performance is still far from good, even with the various JITs available. Jax seems to be one of the most flexible and optimized for many use cases for now however. [1]: https://github.com/salesforce/pytorch-qrnn
In Jax – Composable transformations of Python and NumPy programs
Smerity·Apr 16, 2019
Chainer and potentially CuPy (which was extracted from Chainer to be independent) were around before PyTorch as it served as inspiration for PyTorch. I feel like that's a good motivation for diversity in packages and ecosystems regardless of your feelings otherwise. Along with a colleague I used CuPy in first Chainer and then PyTorch for implementing the Quasi-Recurrent Neural Network (QRNN) which at the time was far faster than even NVIDIA's optimized cuDNN LSTM whilst getting the same (or better) performance for many tasks. CuPy at the time was both the easiest and most Pythonic of potential solutions for that problem - even if it did involve writing CUDA in Python strings =] n.b. Our use case was literally pushing state of the art in research - CuPy is even more Pythonic if you're hitting more standard use cases. [1]: https://github.co
Read the rest
m/salesforce/pytorch-qrnn
In CuPy – NumPy-compatible matrix library accelerated by CUDA
Smerity·Jun 28, 2018
Actually, yes, the QRNN has all of those features. First figure from our paper: how the LSTM with a twist allows for the equivalent speed of a plain convnet by running efficiently in parallel on GPUs, like image processing convents.[1] Best of all, as it's only an "LSTM with (these) twists", it's drop-in compatible with existing LSTMs but can get you a 2-17 times speed-up over NVIDIA's cuDNN LSTM - essentially speed equivalent to the TCN or WaveNet speed-up. That's why Baidu implemented QRNN in their production Deep Voice 2 neural text-to-speech (TTS) system[3]. This isn't to say TCN or QRNN is better, simply that it's dangerous to flat out say _no_ if you're not actually certain or don't correctly recall the underlying information. Disclaimer: I'm the co-author of the QRNN. Double disclaimer: The TCN paper cites the QRNN but decides not
Read the rest
to test against it. They also show results over one of my datasets. [1]: https://www.semanticscholar.org/paper/Quasi-Recurrent-Neural... [2]: https://github.com/salesforce/pytorch-qrnn [3]: https://www.semanticscholar.org/paper/Deep-Voice-2%3A-Multi-...
In Keras vs PyTorch
jph00·May 15, 2018
fasttext is just an encoding of the first layer of a model (the word embeddings - or subword embeddings). Full multi-layer pre-trained models are able to do a lot more. For instance, on IMDb sentiment our method is about twice as accurate as fasttext. As to whether CNNs can replace RNNs in general, the jury is still out. Over the last couple of years there have been some sequence tasks where CNNs are state of the art, some where RNNs are. Note that with stuff like QRNNs the assumption that CNNs are less computationally expensive is no longer necessarily true: https://github.com/salesforce/pytorch-qrnn I'd be surprised if for tasks that require long-term state (like sentiment analysis on large docs) whether CNNs will win out in the end, since RNNs are specifically designed to be stateful - especially with the addition of an attention layer
Read the rest
.
In State-of-the-art text classification with universal language models
Smerity·Oct 26, 2017
The P3 instances are the first widely and easily accessible machines that use the NVIDIA Tesla V100 GPUs. These GPUs are straight up scary in terms of firepower. To give an understanding of the speed-up compared to the P2 instances for a research project of mine: + P2 (K80) with single GPU: ~95 seconds per epoch + P3 (V100) with single GPU: ~20 seconds per epoch Admittedly this isn't exactly fair for either GPU - the K80 cards are straight up ancient now and the Volta isn't sitting at 100% GPU utilization as it burns through the data too quickly ([CUDA kernel, Python] overhead suddenly become major bottlenecks). This gives you an indication of what a leap this is if you're using GPUs on AWS however. Oh, and the V100 comes with 16GB of (faster) RAM compared to the K80's 12GB of RAM, so you win there too. For anyone using the standard set o
Read the rest
f frameworks (Tensorflow, Keras, PyTorch, Chainer, MXNet, DyNet, DeepLearning4j, ...) this type of speed-up will likely require you to do nothing - except throw more money at the P3 instance :) If you really want to get into the black magic of speed-ups, these cards also feature full FP16 support, which means you can double your TFLOPS by dropping to FP16 from FP32. You'll run into a million problems during training due to the lower precision but these aren't insurmountable and may well be worth the pain for the additional speed-up / better RAM usage. - Good overview of Volta's advantages compared to event the recent P100: https://devblogs.nvidia.com/parallelforall/inside-volta/ - Simple table comparing V100 / P100 / K40 / M40: https://www.anandtech.com/show/11367/nvidia-volta-unveiled-g... - NVIDIA's V100 GPU architecture white paper: http://www.nvidia.com/object/volta-architecture-whitepaper.h... - The numbers above were using my PyTorch code at https://github.com/salesforce/awd-lstm-lm and the Quasi-Recurrent Neural Network (QRNN) at https://github.com/salesforce/pytorch-qrnn which features a custom CUDA kernel for speed
In Introducing Amazon EC2 P3 Instances

MAKE YOURSELF AT HOME

A few shortcuts

/
Search the collection
?
Open this guide
Esc
Close dialogs and navigation

Your saved repositories and interests stay in this browser.