tag:blogger.com,1999:blog-8540876.post328231440874481473..comments2024-01-17T00:22:00.802+03:00Comments on Deniz Yuret's Homepage: Introducing Knet8: beginning deep learning with 100 lines of JuliaDeniz Yurethttp://www.blogger.com/profile/00578023665603100985noreply@blogger.comBlogger18125tag:blogger.com,1999:blog-8540876.post-82990213378610448202015-03-03T08:33:30.392+02:002015-03-03T08:33:30.392+02:00Matthew's parsing tutorial taught me how to do...Matthew's parsing tutorial taught me how to do parsing, yours would hopefully will teach me about this misterious thing called deep learning. However, would have loved it if it was in python. Thanks anyways!Anonymousnoreply@blogger.comtag:blogger.com,1999:blog-8540876.post-13899245244741190122015-03-03T08:14:03.390+02:002015-03-03T08:14:03.390+02:00Emrah: The repo uses an MIT license.Emrah: The repo uses an MIT license.Deniz Yurethttps://www.blogger.com/profile/00578023665603100985noreply@blogger.comtag:blogger.com,1999:blog-8540876.post-79216118509179088422015-03-02T22:09:09.985+02:002015-03-02T22:09:09.985+02:00Terrific post. Very timely and elegant indeed.
T...Terrific post. Very timely and elegant indeed. <br /><br />Thanks Deniz for sharing this with the rest of us. What kind of license do you have in mind for general use?Emrah Acarhttp://researcher.watson.ibm.com/researcher/view.php?person=us-emrahnoreply@blogger.comtag:blogger.com,1999:blog-8540876.post-11232628162655255242015-03-02T09:24:30.161+02:002015-03-02T09:24:30.161+02:00Andy: I just posted some benchmark results here.Andy: I just posted some benchmark results <a href="https://github.com/denizyuret/KUnet.jl/blob/master/docs/benchmark.md" rel="nofollow">here</a>.Deniz Yurethttps://www.blogger.com/profile/00578023665603100985noreply@blogger.comtag:blogger.com,1999:blog-8540876.post-44890527973893108342015-03-02T02:30:18.440+02:002015-03-02T02:30:18.440+02:00I find this code extremely succinct and easy to fo...I find this code extremely succinct and easy to follow. Very well done. Its great to see an example like this in Julia. Real boost for the language. <br /><br />Do you have any stats on the performance benefits you refer to at the top of the post?Andyhttps://www.blogger.com/profile/16882084691355634515noreply@blogger.comtag:blogger.com,1999:blog-8540876.post-89399317094641591382015-03-01T23:08:38.720+02:002015-03-01T23:08:38.720+02:00Alok,
You can try out the code at https://try.jup...Alok,<br /><br />You can try out the code at https://try.jupyter.org/. Create a new Julia notebook in the top right and go from there.Anonymousnoreply@blogger.comtag:blogger.com,1999:blog-8540876.post-50382056493960387952015-03-01T17:46:32.069+02:002015-03-01T17:46:32.069+02:00Hi,
I think it is clearer to write
for l in n
...Hi, <br /><br />I think it is clearer to write<br /><br />for l in n<br /><br />rather than <br /><br />for l=nUnknownhttps://www.blogger.com/profile/06544698063916882417noreply@blogger.comtag:blogger.com,1999:blog-8540876.post-26739929847893651342015-03-01T15:32:04.861+02:002015-03-01T15:32:04.861+02:00Thanks Dario: I have now fixed the v0.3 compatibil...Thanks Dario: I have now fixed the v0.3 compatibility issues, the latest from the repo should work out of the box.Deniz Yurethttps://www.blogger.com/profile/00578023665603100985noreply@blogger.comtag:blogger.com,1999:blog-8540876.post-3480740523201259972015-03-01T13:41:08.056+02:002015-03-01T13:41:08.056+02:00Hi Deniz, thanks for this very useful piece of wor...Hi Deniz, thanks for this very useful piece of work!<br /><br />I had some problems getting the library to work in Julia Version 0.3.6 (2015-02-17 22:12 UTC), Darwin build. In case anybody can find it useful, I had to change<br /><br />adagrad!(eps, dw2, dw)=(for i=1:length(dw) (dw2[i] += dw[i] * dw[i]; dw[i] /= (eps + sqrt(dw2[i]))) end)<br /><br />to <br /><br />adagrad!(eps, dw2, dw) = for i=1:length(dw) <br /> dw2[i] += dw[i] * dw[i] <br /> dw[i] /= (eps + sqrt(dw2[i])) <br />end<br /><br />in update.jl (line 17), and also rename the Net function in types.jl (line 16) so that it doesn't clash with the type name.DarĂohttps://www.blogger.com/profile/11516710513710645191noreply@blogger.comtag:blogger.com,1999:blog-8540876.post-43078928890326249362015-03-01T12:27:46.508+02:002015-03-01T12:27:46.508+02:00Check out the KUnet discussion on ycombinator and ...Check out <a href="https://news.ycombinator.com/item?id=9124176" rel="nofollow">the KUnet discussion on ycombinator</a> and the references therein: <a href="https://github.com/pluskid/Mocha.jl" rel="nofollow">Mocha.jl</a>, <a href="http://deeplearning4j.org/" rel="nofollow">deeplearning4j</a>, <a href="https://github.com/dfdx/Boltzmann.jl" rel="nofollow">Boltzmann.jl</a>.Deniz Yurethttps://www.blogger.com/profile/00578023665603100985noreply@blogger.comtag:blogger.com,1999:blog-8540876.post-63239219132649612192015-03-01T12:19:21.268+02:002015-03-01T12:19:21.268+02:00Jon: thanks for your kind comments. Trying new id...Jon: thanks for your kind comments. Trying new ideas with minimal coding was one of the main motivations for starting KUnet.jl, I hope it serves that purpose. I hadn't seen the infinite RBM paper, I will check it out (http://arxiv.org/abs/1502.02476).Deniz Yurethttps://www.blogger.com/profile/00578023665603100985noreply@blogger.comtag:blogger.com,1999:blog-8540876.post-68768445776971613692015-03-01T12:16:01.889+02:002015-03-01T12:16:01.889+02:00Alok: I haven't tested the code under Windows,...Alok: I haven't tested the code under Windows, but I know Julia and CUDArt support Windows so it should not be too difficult. Regarding ImageNet: as the size of the input grows, the number of weights for fully connected layers grow with it. So we would have to implement convolutional nets first to make large image processing feasible.Deniz Yurethttps://www.blogger.com/profile/00578023665603100985noreply@blogger.comtag:blogger.com,1999:blog-8540876.post-63607788971863782472015-03-01T08:55:14.599+02:002015-03-01T08:55:14.599+02:00Amazing, this is exactly what is needed, short rea...Amazing, this is exactly what is needed, short readable efficient code for learning and implementations. I hope this can implement the new ideas for people to try out as they are published. Many thanks. Anonymoushttps://www.blogger.com/profile/15967273973400871354noreply@blogger.comtag:blogger.com,1999:blog-8540876.post-88130596385447180662015-03-01T08:53:14.276+02:002015-03-01T08:53:14.276+02:00This is amazing, exactly what is needed and the id...This is amazing, exactly what is needed and the idea of readable and efficient code. It would deb amazing if we can try to implement the new ideas as they come out for people to play around with. Did you see the infinite RBM paper, about growing the layers adaptively?<br /><br />Many thanksAnonymoushttps://www.blogger.com/profile/15967273973400871354noreply@blogger.comtag:blogger.com,1999:blog-8540876.post-82557310097379159882015-03-01T05:25:04.926+02:002015-03-01T05:25:04.926+02:00Alok, not much work at all: https://www.virtualbox...Alok, not much work at all: https://www.virtualbox.org/Anonymousnoreply@blogger.comtag:blogger.com,1999:blog-8540876.post-14682772018002407652015-03-01T01:25:22.549+02:002015-03-01T01:25:22.549+02:00Thank you!Thank you!Anonymoushttps://www.blogger.com/profile/04457951892816469813noreply@blogger.comtag:blogger.com,1999:blog-8540876.post-76495639625472699152015-02-28T19:33:24.923+02:002015-02-28T19:33:24.923+02:00One of few approachable short tutorials I have bee...One of few approachable short tutorials I have been able to digest as a rank beginner, thanks!Berniehttps://www.blogger.com/profile/12629770310623405192noreply@blogger.comtag:blogger.com,1999:blog-8540876.post-68517384675367576162015-02-28T19:16:27.733+02:002015-02-28T19:16:27.733+02:00Thanks a lot for this.
Any idea how much effort m...Thanks a lot for this.<br /><br />Any idea how much effort may be involved in making this work under Windows?<br /><br />Also, have you tested with more a challenging dataset like ImageNet?Aloknoreply@blogger.com