Hello Rajiv, > I did a bit of research and read a few papers related to CNNs. I found the > following to be very interesting:
Yes, these are some really neat ideas, especially Spatial Transformer Networks would be a great addition. > For this, following improvements can also be done: 1. Adding image training > support in MLPack(If it already exists, I would love to work to simplify the > process of training images by adding features like Command Line Interface, > train.txt, test.txt parsing). There is an open issue https://github.com/mlpack/mlpack/issues/1254 that is directly related to the CLI idea. > 2. Also, for working on images, GPU support is needed. Hence, I can work on > adding GPU support for training CNNs. There is some work in this direction; https://github.com/conradsnicta/bandicoot- code is a GPU accelerator add-on for the Armadillo. In the meantime, we could use NVBLAS which is a GPU-accelerated implementation of BLAS. > I also read a few papers on RBFNs which were really interesting. If possible, > I > would love to work on it as well. My recommendation is to focus on one or two ideas, each takes time to implement and writing tests often takes more time as you might anticipate. I hope this is helpful, let me know if I should clarify anything. Thanks, Marcus > On 25. Feb 2018, at 19:06, Rajiv Vaidyanathan <[email protected]> > wrote: > > Hi Marcus, > > I did a bit of research and read a few papers related to CNNs. I found the > following to be very interesting: > 1. R-CNN > 2. Spatial Transformer Networks > > They improve the CNN results without directly affecting the core > architecture. They act as an additional layer before the CNN input layer and > I felt it would be great addition to the MLPack library which already has the > core architecture > ready(https://github.com/mlpack/mlpack/tree/master/src/mlpack/methods/ann > <https://github.com/mlpack/mlpack/tree/master/src/mlpack/methods/ann>). > > Also, I looked into the implementation of Digit > Recogniser(https://github.com/mlpack/models/ > <https://github.com/mlpack/models/>). The main application of CNNs is in the > domain of images and thus I want to work on improving the process of working > with images and that's why I decided to go with R-CNN and Spacial Transformet > Networks. > > For this, following improvements can also be done: > 1. Adding image training support in MLPack(If it already exists, I would love > to work to simplify the process of training images by adding features like > Command Line Interface, train.txt, test.txt parsing). > 2. Also, for working on images, GPU support is needed. Hence, I can work on > adding GPU support for training CNNs. > > I also read a few papers on RBFNs which were really interesting. If possible, > I would love to work on it as well. > > Please let me know if anything more could be added to this or if any of this > is feasible. It would be great if you could give your suggestions. > > Regards, > Rajiv > > > On 19 February 2018 at 00:44, Rajiv Vaidyanathan > <[email protected] <mailto:[email protected]>> wrote: > Hi Marcus, > > The paper "Learning Methods for radial basis function networks" is already > present in the ideas page(There is where I found it initially :p). Has RBFN > already been implemented in MLPack? Also, I have a small query. In the ideas > page, it has been said that we have to implement deep learning modules. Does > this necessarily mean more than one? For example, I want to implement RBFN as > well something in CNN. But, I am not sure about its feasibility. > > As you said, I'll look into the CNN implementation in MLPack. I'll research > for ideas related to it. If possible, could you give me an example of idea > which is feasible to work on? As CNN is a vast domain, I would be > specifically able to search for ideas in those lines and then start a > discussion based on it. > > Regards, > Rajiv > > On 18 February 2018 at 21:29, Marcus Edel <[email protected] > <mailto:[email protected]>> wrote: > Hello Rajiv, > > thanks for the references, I will add the "Learning methods for radial basis > function networks" paper to the ideas page. > >> Also, I have worked before on CNNs using AlexNet have also worked on a few >> Caffe >> Models like SegNet, etc. I found CNNs to be amazingly creative in the way >> they >> work. Is there a possibility of working on CNN? Has it already been >> implemented? >> Convolutional Neural Networks would be a great addition to MLPack and maybe >> if >> we have a simple interface like Caffe(maybe a simpler one with great >> tutorials >> :p) for using it with image dataset(and for tuning the parameters), then >> MLPack >> would be a really popular choice for learning images. > > Yes, mlpack provides an CNN implementation. Eugene Freyman wrote a really nice > implementation of a network that learns the MNIST digits for a Kaggle > competition in the models repository: > > https://github.com/mlpack/models/ <https://github.com/mlpack/models/> > > You could play with that, as well as the tests in src/mlpack/tests/, to learn > how the framework works and get some experience with it. > > However, there are a bunch of CNN related ideas that that could be worked on, > so > if you have something interesting in mind, please start a discussion. > > Best, > Marcus > >> On 16. Feb 2018, at 14:43, Rajiv Vaidyanathan <[email protected] >> <mailto:[email protected]>> wrote: >> >> Hi Marcus, >> >> Sorry for the late reply... As we had discussed before, I am interested in >> contributing to the Deep Learning Modules in MLPack. I was going through the >> list of papers and the links provided by you in the Ideas page. I found the >> idea of RBFN to be very interesting. I read the paper >> https://www.sciencedirect.com/science/article/pii/S0167739X04000378#bib2 >> <https://www.sciencedirect.com/science/article/pii/S0167739X04000378#bib2> >> and this concept looks very exciting and promising. I also found the >> following material which explained the learning methods of RBFN in a very >> simple manner with some great illustrations: >> http://home.iitk.ac.in/~lbehera/Files/Lecture5_RBFN.pdf >> <http://home.iitk.ac.in/~lbehera/Files/Lecture5_RBFN.pdf>. If you like it, >> you can add it to the list of papers in the ideas page :) >> >> Also, I have worked before on CNNs using AlexNet have also worked on a few >> Caffe Models like SegNet, etc. I found CNNs to be amazingly creative in the >> way they work. Is there a possibility of working on CNN? Has it already been >> implemented? Convolutional Neural Networks would be a great addition to >> MLPack and maybe if we have a simple interface like Caffe(maybe a simpler >> one with great tutorials :p) for using it with image dataset(and for tuning >> the parameters), then MLPack would be a really popular choice for learning >> images. >> >> Regards, >> Rajiv >> >> On 14 February 2018 at 15:46, Marcus Edel <[email protected] >> <mailto:[email protected]>> wrote: >> Hello Rajiv, >> >> We don't track the time, in the community bonding time, we will work with >> every >> student on milestones that ideally should be met over the summer. But keep in >> mind students who are accepted into the Google Summer of Code program are >> expected to work full-time hours for the 12 weeks of the program -- that's 40 >> hours a week. >> >> Best, >> Marcus >> >>> On 13. Feb 2018, at 21:37, Rajiv Vaidyanathan >>> <[email protected] <mailto:[email protected]>> >>> wrote: >>> >>> Hi Marcus, >>> >>> Yeah, please let me know once you finish the PR. I thought that I'll do >>> some research on the algorithms to be implemented first and make an >>> abstract of it first. Based on it, I'll decide the time line of work. >>> Roughly, how long per day should I spend on the project? I might be working >>> on an intern during the day time... So is it an issue? I am sure that I >>> will be able to provide 40 hours per week as given in the Requirements page. >>> >>> Regards, >>> Rajiv >>> >>> On 13 February 2018 at 21:38, Marcus Edel <[email protected] >>> <mailto:[email protected]>> wrote: >>> Hello Rajiv, >>> >>>> I initially had 3 topics of choice(in the previous mail), but after going >>>> through the mailing list archive, mlpack blog and other links provided by >>>> you, I >>>> have decided to go ahead with "ESSENTIAL DEEP LEARNING MODULES". I like >>>> the list >>>> of suggested algorithms. I'll also see if I can think of anything. Please >>>> let me >>>> know if you have any suggestions in mind. >>> >>> Sounds good, we will update the project description in the next days and >>> probably add more suggestions. >>> >>>> So, how should I go about from here? Is there any open ticket to work on? >>>> Should >>>> I start thinking about the proposal? >>> >>> If you don't mind I'd like to finish the open PR first, before opening >>> another >>> PR, how does this sound? About the proposal, if you like you can start >>> working >>> on it, the most important part is the timeline. >>> >>> I hope anything I said is helpful. >>> >>> Thanks, >>> Marcus >>> >>>> On 13. Feb 2018, at 08:07, Rajiv Vaidyanathan >>>> <[email protected] <mailto:[email protected]>> >>>> wrote: >>>> >>>> Hi Marcus, >>>> >>>> Firstly, congratulations for making into GSoC 2018! >>>> >>>> I initially had 3 topics of choice(in the previous mail), but after going >>>> through the mailing list archive, mlpack blog and other links provided by >>>> you, I have decided to go ahead with "ESSENTIAL DEEP LEARNING MODULES". I >>>> like the list of suggested algorithms. I'll also see if I can think of >>>> anything. Please let me know if you have any suggestions in mind. >>>> >>>> So, how should I go about from here? Is there any open ticket to work on? >>>> Should I start thinking about the proposal? >>>> >>>> Also, I thought I should to finish the SPSA >>>> optimizer(https://github.com/mlpack/mlpack/pull/1153 >>>> <https://github.com/mlpack/mlpack/pull/1153>) before I start working on a >>>> new issue. I am stuck up in the implementation of the test code. I have >>>> posted the latest error in the comments. It would be great if you could >>>> help me out. >>>> >>>> Regards, >>>> Rajiv >>>> >>>> On 3 February 2018 at 20:37, Rajiv Vaidyanathan >>>> <[email protected] <mailto:[email protected]>> >>>> wrote: >>>> >>>> >>>> Hi Marcus, >>>> >>>> As of now, I am working on writing a test for SPSA optimizer which I have >>>> implemented... I'll try to finish it ASAP. As of now, I cannot think of >>>> any DL model good enough to replace the suggested ones... Also, as you >>>> said, lets wait until Google officials confirms :) >>>> >>>> Regards, >>>> Rajiv >>>> >>>> On 3 February 2018 at 03:22, Marcus Edel <[email protected] >>>> <mailto:[email protected]>> wrote: >>>> Hello Rajiv, >>>> >>>> Nice to hear from you again how are things going? >>>> >>>> > I am interested in the following topics(listed in the order of interest): >>>> > 1. Reinforcement Learning >>>> > 2. Essential Deep Learning Modules >>>> > 3. Particle Swarm Optimization >>>> > >>>> > How should I go about it? I read a few Mailing List archives for 1 and >>>> > 3. What >>>> > should I do after that? Can I start working on the proposal submission? >>>> >>>> >>>> Going through the mailing list archive is definitely a good starting >>>> point, also >>>> the weekly updates from Kris and Shangtong >>>> (http://www.mlpack.org/gsocblog/index.html >>>> <http://www.mlpack.org/gsocblog/index.html>) could be interesting too. >>>> >>>> The models listed for the Essential Deep Learning Modules idea are just >>>> suggestions, if you like to work on an interesting network model over the >>>> summer >>>> please feel free to start a discussion. >>>> >>>> A good starting point, in general, is to get familiar with the codebase, >>>> if you >>>> have any question please don't hesitate to ask. About the submission >>>> that's up >>>> to you, but note Google hasn't announced the accepted organizations yet and >>>> there is plenty of time to prepare the proposal and get feedback. >>>> >>>> I hope this is helpful, let us know if we should clarify anything. >>>> >>>> Thanks, >>>> Marcus >>>> >>>> > On 2. Feb 2018, at 16:19, Rajiv Vaidyanathan >>>> > <[email protected] <mailto:[email protected]>> >>>> > wrote: >>>> > >>>> > Hi Marcus, >>>> > >>>> > I'm N Rajiv Vaidyanathan(github handle: rajiv2605). I have contributed >>>> > to mlpack in the past and I really like this organisation. I am very >>>> > interested in participating in GSoC. >>>> > >>>> > I am interested in the following topics(listed in the order of interest): >>>> > 1. Reinforcement Learning >>>> > 2. Essential Deep Learning Modules >>>> > 3. Particle Swarm Optimization >>>> > >>>> > How should I go about it? I read a few Mailing List archives for 1 and >>>> > 3. What should I do after that? Can I start working on the proposal >>>> > submission? >>>> > >>>> > Thanking you. >>>> > >>>> > Regards, >>>> > Rajiv >>>> > >>>> > _______________________________________________ >>>> > mlpack mailing list >>>> > [email protected] <mailto:[email protected]> >>>> > http://knife.lugatgt.org/cgi-bin/mailman/listinfo/mlpack >>>> > <http://knife.lugatgt.org/cgi-bin/mailman/listinfo/mlpack> >>>> >>>> >>>> >>>> >>>> _______________________________________________ >>>> mlpack mailing list >>>> [email protected] <mailto:[email protected]> >>>> http://knife.lugatgt.org/cgi-bin/mailman/listinfo/mlpack >>>> <http://knife.lugatgt.org/cgi-bin/mailman/listinfo/mlpack> >>> >> >> > > >
_______________________________________________ mlpack mailing list [email protected] http://knife.lugatgt.org/cgi-bin/mailman/listinfo/mlpack
