{ "id": 7210, "title": "Using R", "url": "/competitions/cifar-10/discussion/7210", "author_name": "Ryan Jones", "post_date": "2014-02-26T15:55:58.470000", "votes": 1, "comment_count": 8, "views": 5582, "content": "

Has anybody successfully trained a model that got good results using R?  So far, a number of the methods seem to be focused on python implementations.

\n

What data processing techniques/ packages have been useful so far?

", "messages": [ { "id": 219682, "postDate": "2017-09-09T04:01:14.573Z", "content": "

The team at RStudio (mainly JJ Allaire, AFAICT) have released a package called Keras, which interfaces to three backends: TensorFlow, CNTK and Theano. I've been using that on some toy examples and it seems to work well. You'll need a working installation of one or both of the backends, and an NVidia GPU, ideally at GTX 1060 with 6GB RAM or better. It runs on slower GPUs fine (I have a GTX 750ti with 2GB) and CPUs, which is OK for small models. But most real deep learning researchers recommend at least a 1060 these days.

\n\n

HTH,

\n\n

John

", "rawMarkdown": "The team at RStudio (mainly JJ Allaire, AFAICT) have released a [package][1] called Keras, which interfaces to three backends: TensorFlow, CNTK and Theano. I've been using that on some toy examples and it seems to work well. You'll need a working installation of one or both of the backends, and an NVidia GPU, ideally at GTX 1060 with 6GB RAM or better. It runs on slower GPUs fine (I have a GTX 750ti with 2GB) and CPUs, which is OK for small models. But most real deep learning researchers recommend at least a 1060 these days.\n\nHTH,\n\nJohn\n [1]: https://keras.rstudio.com/", "votes": 1, "replies": [ { "id": 329303, "postDate": "2018-05-16T07:30:46.350Z", "content": "

Will it work with AMD GPU?

", "rawMarkdown": "Will it work with AMD GPU?" }, { "id": 353986, "postDate": "2018-07-08T12:53:02.090Z", "content": "

No.

", "rawMarkdown": "No." } ] }, { "id": 51316, "postDate": "2014-07-29T18:26:07.107Z", "content": "

This toy example shows how to access pixel data in R:

\n

library('png')
download.file('http://i.stack.imgur.com/hakvE.png', 'sample.png')
img = readPNG('sample.png')
pix.top.left = img[1,1,] # row 1, column 1
pix.bottom.left = img[3,1,] # row 3, column 1
pix.top.right = img[1,3,] # row 1, column 3

\n

Once you have the matrix 'img' for your pics you can embark on whatever data reduction seems reasonable.  I don't think it's going to be slow, particularly.

", "votes": 1 }, { "id": 39516, "postDate": "2014-02-26T15:55:58.470Z", "content": "

Has anybody successfully trained a model that got good results using R?  So far, a number of the methods seem to be focused on python implementations.

\n

What data processing techniques/ packages have been useful so far?

", "votes": 1 }, { "id": 51802, "postDate": "2014-08-09T16:54:59.977Z", "content": "

You can read convert images to single csv and read into R.

\n

Christopher Beckham posted his sample python script on following forum article :

\n

https://www.kaggle.com/c/cifar-10/forums/t/8274/extracting-pixel-data-from-images/50915#post50915

\n\n

Hope this help.

" }, { "id": 40207, "postDate": "2014-03-12T10:59:02.100Z", "content": "

R is going to be damn slow on images. Atleast that's what I infer from other competitions. Python has a deep learning module. Should be a good place to learn that.

" }, { "id": 136754, "postDate": "2016-09-24T20:02:07.143Z", "rawMarkdown": "", "votes": 2, "isDeleted": true, "replies": [ { "id": 253352, "postDate": "2017-12-04T20:40:32.823Z", "content": "

This was supremely helpful, thanks!

", "rawMarkdown": "This was supremely helpful, thanks! " } ] } ], "comments": [ { "id": 219682, "author_name": "John Williams", "author_url": "", "post_date": "2017-09-09T04:01:14.573000", "content": "

The team at RStudio (mainly JJ Allaire, AFAICT) have released a package called Keras, which interfaces to three backends: TensorFlow, CNTK and Theano. I've been using that on some toy examples and it seems to work well. You'll need a working installation of one or both of the backends, and an NVidia GPU, ideally at GTX 1060 with 6GB RAM or better. It runs on slower GPUs fine (I have a GTX 750ti with 2GB) and CPUs, which is OK for small models. But most real deep learning researchers recommend at least a 1060 these days.

\n\n

HTH,

\n\n

John

", "votes": 1, "replies": [ { "id": 329303, "author_name": "RudraFCB", "author_url": "", "post_date": "2018-05-16T07:30:46.350000", "content": "

Will it work with AMD GPU?

", "votes": 0, "replies": [] }, { "id": 353986, "author_name": "Steve Devney", "author_url": "", "post_date": "2018-07-08T12:53:02.090000", "content": "

No.

", "votes": 0, "replies": [] } ] }, { "id": 51316, "author_name": "Jay Moore", "author_url": "", "post_date": "2014-07-29T18:26:07.107000", "content": "

This toy example shows how to access pixel data in R:

\n

library('png')
download.file('http://i.stack.imgur.com/hakvE.png', 'sample.png')
img = readPNG('sample.png')
pix.top.left = img[1,1,] # row 1, column 1
pix.bottom.left = img[3,1,] # row 3, column 1
pix.top.right = img[1,3,] # row 1, column 3

\n

Once you have the matrix 'img' for your pics you can embark on whatever data reduction seems reasonable.  I don't think it's going to be slow, particularly.

", "votes": 1, "replies": [] }, { "id": 51802, "author_name": "Yong Hee Park", "author_url": "", "post_date": "2014-08-09T16:54:59.977000", "content": "

You can read convert images to single csv and read into R.

\n

Christopher Beckham posted his sample python script on following forum article :

\n

https://www.kaggle.com/c/cifar-10/forums/t/8274/extracting-pixel-data-from-images/50915#post50915

\n\n

Hope this help.

", "votes": 0, "replies": [] }, { "id": 40207, "author_name": "Bargava", "author_url": "", "post_date": "2014-03-12T10:59:02.100000", "content": "

R is going to be damn slow on images. Atleast that's what I infer from other competitions. Python has a deep learning module. Should be a good place to learn that.

", "votes": 0, "replies": [] }, { "id": 136754, "author_name": "", "author_url": "", "post_date": "2016-09-24T20:02:07.143000", "content": "", "votes": 2, "replies": [ { "id": 253352, "author_name": "Emmanuel Perry", "author_url": "", "post_date": "2017-12-04T20:40:32.823000", "content": "

This was supremely helpful, thanks!

", "votes": 0, "replies": [] } ] } ], "raw_markdown_by_id": { "219682": "The team at RStudio (mainly JJ Allaire, AFAICT) have released a [package][1] called Keras, which interfaces to three backends: TensorFlow, CNTK and Theano. I've been using that on some toy examples and it seems to work well. You'll need a working installation of one or both of the backends, and an NVidia GPU, ideally at GTX 1060 with 6GB RAM or better. It runs on slower GPUs fine (I have a GTX 750ti with 2GB) and CPUs, which is OK for small models. But most real deep learning researchers recommend at least a 1060 these days.\n\nHTH,\n\nJohn\n [1]: https://keras.rstudio.com/", "51316": "", "39516": "", "51802": "", "40207": "", "136754": "" } }