Poll: Which deep learning framework would you use?

Posted by cs702 on Tuesday August 30, 2016 @12:54PM
from the you-decide dept.

Say you are about to build a new product or service that uses or requires deep learning.

Which deep learning framework would you use?

If you vote on a framework, please also don't forget to upvote the poll itself, so we can keep it visible to others and collect more votes. Thank you!

( 12 comments | 103 points | on HN )

Poll

Caffe: https://github.com/BVLC/caffe 13 votes

CNTK: https://github.com/Microsoft/CNTK 7 votes

DL4J: https://github.com/deeplearning4j/deeplearning4j 11 votes

DSSTNE: https://github.com/amznlabs/amazon-dsstne 3 votes

H2O: https://github.com/h2oai/h2o-3/tree/master/h2o-algos/src/main/java/hex/deeplearning 2 votes

Marvin: https://github.com/PrincetonVision/marvin 1 vote

MXNet: https://github.com/dmlc/mxnet 5 votes

Neon: https://github.com/NervanaSystems/neon 3 votes

Paddle: https://github.com/baidu/paddle 4 votes

TensorFlow (including SkFlow): https://github.com/tensorflow/tensorflow / https://github.com/tensorflow/skflow 81 votes

Theano: https://github.com/Theano/Theano 20 votes

Torch: https://github.com/torch/torch7 17 votes

Veles: https://github.com/samsung/veles 1 vote

Other 4 votes

Keras (uses Theano or TensorFlow as backend): https://github.com/fchollet/keras 36 votes

Dali: https://github.com/dali-ml/Dali 1 vote

Lasagne (on top of Theano): https://github.com/Lasagne/Lasagne 2 votes

Blocks (on top of Theano): https://github.com/mila-udem/blocks 1 vote

Chainer: http://chainer.org/ 2 votes

Re:Poll: Which deep learning framework would you use? by cosmoharrigan on Tuesday August 30, 2016 @01:04PM (#12391855)

The deep learning frameworks landscape, August 2016: https://twitter.com/fchollet/status/765212287531495424

[ Parent | on HN ]

Re:Poll: Which deep learning framework would you use? by cs702 on Tuesday August 30, 2016 @01:15PM (#12391947)

I saw that post, but still think it makes to have this poll here because the HN community's preferences are probably a better indicator of what startup developers will do going forward than a ranking based on historical github figures accumulated in the past from a wider range of users.

[ Parent | on HN ]

Re:Poll: Which deep learning framework would you use? by minimaxir on Tuesday August 30, 2016 @01:32PM (#12392093)

That chart is slightly misleading in this context; it only measures GitHub repository activity (naturally skewed toward newer/less mature frameworks), not actual usage in real-world applications.

[ Parent | on HN ]

Re:Poll: Which deep learning framework would you use? by cs702 on Wednesday August 31, 2016 @08:53AM (#12397905)

Below is a bar chart of votes which I find helpful, constructed from data as I'm writing this. The figures are unlikely to change much at this point.

TensorFlow dominates, followed by Keras (which uses Theano or TensorFlow as its backend) and then Theano.

TensorFlow and Theano (on their own or via Keras) account for two thirds of all votes.

  TensorFlow (78) ******************************************************************************
       Keras (34) **********************************
      Theano (20) ********************
       Torch (16) ****************
       Caffe (13) *************
         DL4J (9) *********
         CNTK (5) *****
        MXNet (4) ****
       Paddle (4) ****
       DSSTNE (3) ***
        Other (3) ***
          H2O (2) **
      Lasagne (2) **
         Neon (2) **
        Block (1) *
      Chainer (1) *
         Dali (1) *
       Marvin (1) *
        Veles (1) *

[ Parent | on HN ]

Re:Poll: Which deep learning framework would you use? by jonathanraiman on Tuesday August 30, 2016 @01:18PM (#12391964)

Dali: http://www.dali.ml

[ Parent | on HN ]

Re:Poll: Which deep learning framework would you use? by cs702 on Tuesday August 30, 2016 @01:20PM (#12391984)

Thanks for pointing Dali out. I just added it to the list of choices.

[ Parent | on HN ]

Re:Poll: Which deep learning framework would you use? by cosmoharrigan on Tuesday August 30, 2016 @03:09PM (#12392972)

Note that SkFlow is now TF Learn and is part of the TensorFlow repository: https://github.com/tensorflow/tensorflow/tree/master/tensorf... The skflow repository linked to in the poll is being deprecated.

[ Parent | on HN ]

Re:Poll: Which deep learning framework would you use? by cs702 on Tuesday August 30, 2016 @05:03PM (#12393878)

Oh, thanks for pointing that out. I now remember reading it had been moved not long ago, but I still linked to the old repo without thinking, as I have never used this simpler scikit-learn-like interface to TensorFlow. Sorry for the oversight.

[ Parent | on HN ]

Re:Poll: Which deep learning framework would you use? by jmmcd on Tuesday August 30, 2016 @05:36PM (#12394127)

Lasagne on top of Theano https://github.com/Lasagne/Lasagne

[ Parent | on HN ]

Re:Poll: Which deep learning framework would you use? by cs702 on Tuesday August 30, 2016 @05:45PM (#12394186)

Thanks. I knew about but forgot to add Lasagne -- and Blocks (https://github.com/mila-udem/blocks) for that matter. It's been a while since I looked at either.

[ Parent | on HN ]

Re:Poll: Which deep learning framework would you use? by mbajkowski on Wednesday August 31, 2016 @10:11AM (#12398404)

Chainer: http://chainer.org/

[ Parent | on HN ]

Re:Poll: Which deep learning framework would you use? by cs702 on Wednesday August 31, 2016 @10:34AM (#12398585)

Thanks. Just added it.

[ Parent | on HN ]