tensorflow/sig-addons

where SIG Addons talks

Nanubala Gnana Sai

@jonpsy

Although I'm not sure if that file lies in your subdomain

Sean Morgan

@seanpmorgan

Yeah I’ve been swamped at work so have very little time to work on Addons will try my best

Nanubala Gnana Sai

@jonpsy

Thanks, appreciate it!

Nanubala Gnana Sai

@jonpsy

Hey, so I'm working on Matthew metric and I strongly believe the Multiclass implementation is wrong. More on #2339.

Matthews loss*

Suraj Upadhyay

@SurajUpadhyay0_twitter

Hii, where can I find design document for Tensorflow ? I just want to make a sense of the codebase

Harshith MohanKumar

@Harsh188

Kind of unrelated to Addons, but I wanted to ask if there's a significant industry-level difference between AWS and GCP for building and deploying models onto the internet. Should I prefer one over the other?
After reading several articles, I've noticed that there are similar services in both ends but GCP seems to have better integrations with TF. However, if a model doesn't currently exist on TF would that advantage be lost? For example we are trying to implement and deploy PEGASUS and we aren't sure which route to go down.

Sean Morgan

@seanpmorgan

I work for AWS, so I try to avoid conversations like this when its associated with Addons. I'll keep it short and say that both are perfectly viable options.

Jason Zaman

@perfinion

I dont work for either so here are my two cents: underneath both AWS and GCP are regular intel or AMD cpus with nvidia gpus running linux. they are extremely similar to each other so use whichever one you want. the only recommendation i'd have it to use the optimized VM images, iirc AWS calls them "DeepLearningAMIs" and GCP has "deep learning optimized images" and they both put a lot of work in to make sure TensorFlow is optimized (as opposed to using eg generic ubuntu)

Sean Morgan

@seanpmorgan

:thumbsup:

bhack

@bhack

More you go down to the bare metal more they are similar (if you exclude TPU).. more you are using managed services more they diverge

Harshith MohanKumar

@Harsh188

Thanks for the responses!

Nanubala Gnana Sai

@jonpsy

From what I understand, instead of letting users manually use the command export TF_ADDONS_PY_OPS=1, we would wrap it under a function?

Basically this, in options.py?

Nanubala Gnana Sai

@jonpsy

sorry, it was meant for @windqaq

Nanubala Gnana Sai

@jonpsy

hey @bhack , i wanted to discuss something. Lemme know if you're free

bhack

@bhack

hey @bhack , i wanted to discuss something. Lemme know if you're free

Hi, what is the topic?

Nanubala Gnana Sai

@jonpsy

Two things,
a) The LookAhead module, I thought of a stand-in solution by overriding __setattr__ of the LookAhead to set learning rate of the internal optimizer. Maybe we could do that for all the wrappers by introducing a Mixin of a sort and letting the wrappers inherit that?

b)The Matthew correlation implementation is wrong imo, there've been some discussion and I've made a PR. Might you be interested? tensorflow/addons#2406

good day ! :)

Harshith MohanKumar

@Harsh188

Hey @seanpmorgan @bhack,

As I was going through the "TensorFlow Project Ideas" document for google summer of code when I noticed that TF Addons was listed under the propose your own idea section.

I was wondering what projects/tasks I could propose that would be beneficial for TFAddons. I would love to hear your suggestions on possible projects that I could work on. I remember @seanpmorgan talking about custom ops during the last SIG meeting and how it could be an interesting task for a university student which fits my interests.

So far I have only made minor changes/fixes, however, I would be more than happy to tackle a challenging task :)

Kartikey Rawat

@carrycooldude

Hello everyone, I am Kartikey Rawat and I am Open Source and ML Enthusiast and I am really want to contribute in tensorflow.js and even I am also advocating tensorflow.js through my speaker session love to be a part of this community

Sean Morgan

@seanpmorgan

Hey @Harsh188, apologies for the slow answer. Unfortunetely I won’t have time right now to be a mentor. If there is any one else that is willing to sponsor a project though that may work. The custom op idea is not really enough of a GSoC project in as far as it’s more maintainence and reviewing code than anything

Not to say there couldn’t be a way to incorporate that into a project but nothing is coming to mind

Harshith MohanKumar

@Harsh188

Thanks for the reply! By any chance would you know any other members in TFAddons who would be willing to mentor for GSoC?

KiranKhanna721

@KiranKhanna721

Hello everyone , I am Kiran Khanna .

bhack

@bhack

Abhinav Prakash

@abhinavsp0730

Hi, @bhack @seanpmorgan do you know How to reach the mentors of this year's GSOC 2021.

Last year I've applied for GSOC but I didn't hear anything from TF team . It was written that kindly mail your drafted proposal so that we can assign you the Mentor to refine your proposal. But unluckily I've mailed them but I didn't got any reply. So, this year I want to discuss about my project Ideas with mentor. So that I'll include only those things which are relevant to TF ecosystem.

bhack

@bhack

Last year I've applied for GSOC but I didn't hear anything from TF team . It was written that kindly mail your drafted proposal so that we can assign you the Mentor to refine your proposal. But unluckily I've mailed them but I didn't got any reply. So, this year I want to discuss about my project Ideas with mentor. So that I'll include only those things which are relevant to TF ecosystem.

You can try to post in https://groups.google.com/a/tensorflow.org/g/developers

Sean Morgan

@seanpmorgan

Hmmm not able to join the meeting...

Is it just me?

Rehan Guha

@rehanguha

@seanpmorgan It will be great if we have a list of checks(all possible CI checks) with required package versions used in CI available in the repo for development purpose.
This will reduce the commit counts.
Eg.

black --check addons
flake8 addons

I am not sure whether it is already in place or not. :-)

bhack

@bhack

Is there the meeting?

I cannot access

@seanpmorgan ?

vulkomilev

@vulkomilev

does somebody needs a help on some issue/project ?

bhack

@bhack

@vulkomilev We are going to clean up open issues and PRs in the next days

vulkomilev

@vulkomilev

okay if you have something specific in mind let me know

vulkomilev

@vulkomilev

@bhack can you assist me about tensorflow/addons#2593 I have found the solution to the problem and I just need to merge it

I don't have a lot of experience in this

bhack

@bhack

@vulkomilev Ok the First step Is:

Do you have a test or can you extend an existing test to cover this case?

vulkomilev

@vulkomilev

Sorry for the late answer but the modification is for Triangular2CyclicalLearningRate and I am not aware of any test for this class

bhack

@bhack

Test https://github.com/tensorflow/addons/blob/e83e71cf07f65773d0f3ba02b6de66ec3b190db7/tensorflow_addons/optimizers/tests/cyclical_learning_rate_test.py#L57

vulkomilev

@vulkomilev

okay since he have test what is the next thing to do ?

to test it and after that?

bhack

@bhack

I think that we Need or to modify or to add a new test case with a PR. If tests are not failing now It mean that we don't have enoug test coverage for that case

Then you could add commit on that PR to let the new test to pass

Where communities thrive

People

Repo info

Activity