[mlpack] GSoC '21

[CED18I048 SHAH ANWAAR KHALID]

Hello Marcus,

Thanks for the suggestion. Sorry for the late response, I was a little
caught up with my college exams.

I’ve already started working on the Dual Optimizer PR and hope to finish it
before GSoC. I got it to work for Standard GAN and DCGAN. I’ve opened a PR.
For the summer, I’d like to take up CycleGAN PR and implement StyleGAN.


I went through the PR for CycleGANs and as per my understanding, I have
made the following TODO’s list:


   1.

   The class has been implemented with an older mlpack version. A lot of
   refactoring of the ANN code has been done since then.
   1.

      Implement ResetData() function.
      2.

      Remove boost variant..
      3.

      Implement Dual Optimizer train function.
      4.

      Testing.



Style GAN consists of the following components:


   1.

   Noise Mapping Network and Progressive Growing Generator.
   2.

   Adaptive Instance Normalization.
   3.

   Discriminator Network.


I'll write a detailed explanation of each component in the proposal. Please
let me know what implementation details I should include.

Tentative Timeline:


   1.

   Before GSoC:
   1.

       Finish Dual Optimizer.
      2.

   May 17 - June 7 : Community Bonding Period
   1.

      Get more familiar with StyleGAN and CycleGAN.
      2.

      Look at the official implementation in pytorch and tensorflow.
      3.

   June 7 : Coding Period:
   1.

      Week 1 - 2
      1.

         I’m thinking of spending the first two weeks on finishing CycleGAN
         PR.
         2.

      Week 3-4:
      1.

         Implement Instance Normalization.
         2.

         Write tests..
         3.

      Week 5-8:
      1.

         Implement Generator and Discriminator Network.
         2.

         Write tests.
         4.

      Week 8- 10:
      1.

         I’m thinking of keeping this as a buffer to finish any pending
         work.


Please let me know your thoughts on this.

Thanks,
Anwaar





On Thu, Mar 11, 2021 at 10:13 PM Marcus Edel <marcus.edel at fu-berlin.de>
wrote:

> Hello Shah,
>
> happy to provide feedback; everything you proposed in the mail sounds
> great to me.
> My recommendation would be to focus on two tasks for the summer, that
> could be
> StyleGAN and Cycle GANs, or Dual Optimizer and StyleGAN, I guess you know
> what
> I mean.
>


>
> The program this year is going to be shorter, so we should make sure we
> don't end
> up with a lot of unfinished projects because we run out of time. That
> said, keep in mind
> testing and documentation need time as well, often more time than the
> actual implementation.
>
> Thanks,
> Marcus
>
>
> On 10. Mar 2021, at 19:37, CED18I048 SHAH ANWAAR KHALID <
> ced18i048 at iiitdm.ac.in> wrote:
>
> Hello mentors!
>
> I’m Shah  Anwaar Khalid, currently a 3rd year student at IIITDM, Chennai,
> India. I’ve been working with mlpack for quite some time, and have been
> familiarizing myself with the ANN codebase.
> I wish to spend this summer working with mlpack under GSoC ‘21. I’m
> particularly interested in working on improving the GAN class of mlpack and
> I hope to get some suggestions on improving my proposal through this
> thread. I apologize, in advance, in case this email gets too long.
>
> After exploring the GAN class for quite some time now, i’ve realized that
> there are still a lot of open pull requests on GANs that have some pending
> work. Furthermore, a lot of state of the art GAN models have not yet been
> implemented. So, in my project I want to finish the pending work in the
> open pull requests and extend the functionality of GAN class by
> implementing some of new models that have come up in the last couple of
> years.
>
> I’m interested in the following pull requests:
>
>    1. Dual Optimizer for GANs <https://github.com/mlpack/mlpack/pull/1888>
>    2. Label smoothing in GANs <https://github.com/mlpack/mlpack/pull/1915>
>    3. Frechet Inception Distance
>    <https://github.com/mlpack/mlpack/pull/1939>
>    4. Cycle GANs <https://github.com/mlpack/mlpack/pull/1497>
>
>
> I’m also interested in adding the following GAN models:
> 2. Pix2Pix
> 3. StyleGAN
>
>
> I’ve started working on dual optimizer PR and hope to finish it before
> GSoC. I’ll write a more detailed mail describing the components of Pix2Pix
> and StyleGAN once I get some feedback on this.
>
> Thanks and Regards
> Shah Anwaar Khalid
> GitHub: hello-fri-end
>
> _______________________________________________
> mlpack mailing list
> mlpack at lists.mlpack.org
> http://knife.lugatgt.org/cgi-bin/mailman/listinfo/mlpack
>
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://knife.lugatgt.org/pipermail/mlpack/attachments/20210316/648f8a66/attachment-0001.htm>