[mlpack] Regarding GSOC Proposal for ANN Algorithms Implemented in mlpack

[Marcus Edel]

Hello Kartik,

I think I said this before, but welcome and thanks for all the contribution.

> I would love to dicuss the design related aspect in more detail as I am really
> excited about this project. Also is this the right approach to a GSOC proposal
> or Open Source in general?

I can't really speak about other projects but we like to discuss ideas on public
channels, so this is totally the right approach.

About the project or projects, I like the ideas: YOLOv3 and Feature Pyramid
Network. However, I think going for both might be too ambitious for the Summer,
I'd rather have a single model that we can merge in during the summer than two
models that can't be merged because we run out of time. So I would probably
focus on a single model instead. That said, why not include some of the tasks:

1. IoU metric for object localization.
2. Upsampling layer (I will hopefully open a PR for the same by Sunday).
3. Depthwise seperable Convolution (I have opened a PR for this, currently working on fixing some bugs).
4. AlexNet (I have to put it on training this week).

into a GSoC project, you would need some of the features anyway, so why not
include them into the project itself. Also, we can always include more if there
is some time at the end, but we should make it clear what the focus is. Let me
know what you think.

Thanks,
Marcus

> On 2. Mar 2020, at 10:12, Kartik Dutt <kartikdutt at live.in> wrote:
> 
> Hello everyone,
> I am Kartik Dutt, currently a 3rd year student at DTU, India. I have been getting familiar with mlpack for the last couple of months. I wanted to get some opinions regarding my project proposal to help me improve it.
> I apologize in adavance incase this email gets a bit too long.
> 
> I hope to add the following functionalities in the following months before GSOC so that I only have to work on implementing models during the summer.
> 
> 1. IoU metric for object localization.
> 2. Upsampling layer (I will hopefully open a PR for the same by Sunday).
> 3. Depthwise seperable Convolution (I have opened a PR for this, currently working on fixing some bugs).
> 4. AlexNet (I have to put it on training this week).
> 
> I hope to complete most of them this week after my exams end.
> 
> During GSOC, I wanted to implement the following models:
> 
> 1. Darknet (Backbone) and YOLOv3 (Performs better on devices like RPi than networks such as RetinaNet).
> 2. Feature Pyramid Network (Can be placed on object detection models to implement models such as FasterRCNN, RetinaNet in the future, hopefully just after GSOC).
> 
> The primary focus during the project would be defining API and tests for all future models that are added. I have opened a PR for AlexNet on models repo to dicuss the same.
> I think the above task can only be achieved is if I implement major chunk of DarkNet even before GSOC begins.
> If time allows me, I would also love to work on Spatio-Temporal Networks becuase I am doing my final year thiesis on it.
> I wanted to get your opinion on the project proposal. This seems like a lot of work but I think after implementing it, the models repository can open up to various new possibilities.
> I would love to dicuss the design related aspect in more detail as I am really excited about this project.
> Also is this the right approach to a GSOC proposal or Open Source in general?
> I would love to hear back from you.
> 
> Project: Application of ANN Algorithms Implemented in mlpack
> Mentors: Marcus Edel, Sumedh Ghaisas, Shikhar Jaiswal, Saksham Bansal
> 
> Regards,
> Kartik Dutt,
> Github ID: kartikdutt18
> _______________________________________________
> mlpack mailing list
> mlpack at lists.mlpack.org
> http://knife.lugatgt.org/cgi-bin/mailman/listinfo/mlpack