{ "id": 13266, "title": "Tools / Language", "url": "/competitions/diabetic-retinopathy-detection/discussion/13266", "author_name": "", "post_date": "2015-04-07T17:01:21.710Z", "votes": 1, "comment_count": 3, "views": 1768, "content": "

Hey Guys just checking what language / tools are you using in this problem ?

\n\n

Almost every solution that I'm approaching is using or licensed products or open-source with limited comercial use.

\n\n

Thanks

", "messages": [ { "id": "69904", "postDate": "04/07/2015 17:01:21", "content": "

Hey Guys just checking what language / tools are you using in this problem ?

\n\n

Almost every solution that I'm approaching is using or licensed products or open-source with limited comercial use.

\n\n

Thanks

", "rawMarkdown": "", "votes": null }, { "id": "69932", "postDate": "04/07/2015 20:10:48", "content": "

http://caffe.berkeleyvision.org/
https://github.com/dmlc/cxxnet
http://deeplearning.net/software/theano/

\n\n

All of these was used for winning solutions and completely free.

", "rawMarkdown": "", "votes": null }, { "id": "69950", "postDate": "04/07/2015 22:51:40", "content": "

The short list of selected tools is very amusing.

\n

If these are the only methodologies and tools being use then this particular challenge will come down to who has the computing resources to run the deep learners...

\n

For the most part this particular problem is a gift... it solution is readily available requiring minimal fine tuning and hopefully win ..... if you are going to use those tools for this problem i would recommend you invest in some cuda hardware.... my personal cluster is rated at near 30 TeraFlops and still i am concerned that some of the other teams have more resources than me to play around with the layers of the deep learners 

\n

My real hope here is discover some other novel method to solve this problem, other than the standard deep learning ....

", "rawMarkdown": "", "votes": null }, { "id": "69980", "postDate": "04/08/2015 06:06:06", "content": "

@ Michael George Hart,

\n

1) it is just fact that it is VERY hard to beat deep learning in image classification problem. Just as fact that fastest way from Paris to New York is airplane, which is not cheap.

\n

2) But. There is still three months till end of competition - more than enough time to take some conventional (not deep learning) good working paper and reimplement it from scratch. I personally think that is a way of current leader, btw. But in this case question about tools is not a point, someone just need Python (or C++) and possibly OpenCV, hardly called "a tool".

\n

3)"novel method to solve this problem, other than the standard deep learning"

\n

Sorry, but "standard"??? All three (actually four, since 2nd place just merged results, not methods) winners of National Bowl developed a method at least worth ICLR/CVPR paper. Xudong Cao`s work on architecture, new layer from Deep Sea, Fractional pooling from Ben Graham (actually, last one is ICLR paper 2015 - to appear), new non-linearity from Happy Lantern Festival. Very, very far from just "competing with hardware".

\n

4) There are enough competitions in Kaggle with inhomogenous data (like from Otto) where CNNs perform worse that random forests. And RFs need no GPU :) 

", "rawMarkdown": "", "votes": null } ], "comments": [ { "id": 69932, "author_name": "oldufo", "author_url": "", "post_date": "04/07/2015 20:10:48", "content": "

http://caffe.berkeleyvision.org/
https://github.com/dmlc/cxxnet
http://deeplearning.net/software/theano/

\n\n

All of these was used for winning solutions and completely free.

", "votes": null, "replies": [] }, { "id": 69950, "author_name": "spaceman", "author_url": "", "post_date": "04/07/2015 22:51:40", "content": "

The short list of selected tools is very amusing.

\n

If these are the only methodologies and tools being use then this particular challenge will come down to who has the computing resources to run the deep learners...

\n

For the most part this particular problem is a gift... it solution is readily available requiring minimal fine tuning and hopefully win ..... if you are going to use those tools for this problem i would recommend you invest in some cuda hardware.... my personal cluster is rated at near 30 TeraFlops and still i am concerned that some of the other teams have more resources than me to play around with the layers of the deep learners 

\n

My real hope here is discover some other novel method to solve this problem, other than the standard deep learning ....

", "votes": null, "replies": [] }, { "id": 69980, "author_name": "oldufo", "author_url": "", "post_date": "04/08/2015 06:06:06", "content": "

@ Michael George Hart,

\n

1) it is just fact that it is VERY hard to beat deep learning in image classification problem. Just as fact that fastest way from Paris to New York is airplane, which is not cheap.

\n

2) But. There is still three months till end of competition - more than enough time to take some conventional (not deep learning) good working paper and reimplement it from scratch. I personally think that is a way of current leader, btw. But in this case question about tools is not a point, someone just need Python (or C++) and possibly OpenCV, hardly called "a tool".

\n

3)"novel method to solve this problem, other than the standard deep learning"

\n

Sorry, but "standard"??? All three (actually four, since 2nd place just merged results, not methods) winners of National Bowl developed a method at least worth ICLR/CVPR paper. Xudong Cao`s work on architecture, new layer from Deep Sea, Fractional pooling from Ben Graham (actually, last one is ICLR paper 2015 - to appear), new non-linearity from Happy Lantern Festival. Very, very far from just "competing with hardware".

\n

4) There are enough competitions in Kaggle with inhomogenous data (like from Otto) where CNNs perform worse that random forests. And RFs need no GPU :) 

", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "69904": "", "69932": "", "69950": "", "69980": "" }, "source": "meta" }