Where communities thrive


  • Join over 1.5M+ people
  • Join over 100K+ communities
  • Free without limits
  • Create your own community
People
Activity
David Buchaca Prats
@davidbp
for what? neuralnets¿
RMSProp/adam work quite well for big nnets
Lyndon White
@oxinabox
yes neural nets
I know it works well
But is there a theoretical equilelvence / relationship between the adaptive learning rates and the Nessian in traditional optimistion
I thought there was, but it isn't mentioned in AdaGrad, AdaDelta or Adam papers.
And from my recollection of the RMSProp lecture is inot there either
David Buchaca Prats
@davidbp
I don´t know about any theoretical work on the equivalence among those
pevnak
@pevnak
I would like to implement this method
https://github.com/ProbabilisticNumerics/cabs
which should automatically adjust the size of the minibatch in SGD. I am curious about the method, since in security the signals are typically very weak and therefore I would expect that large batch sizes would be needed. Looking at the source code
https://github.com/ProbabilisticNumerics/cabs/blob/master/gradient_moment.py
it seems like they need to identify consumers of a tensor (line 79 of gradient_moment.py). I wonder if it is possible to do so in TensorFlow.jl?
Lyndon White
@oxinabox
The information is there, but I can't think how to get it
I think that information is not nesc there in a nice way on the julia side in particular.
pevnak
@pevnak
That is frustrating. It just sound weird to me that you can extract it in python and not in Julia. I thought that the TensorFlow API is now fully exposed.
Lyndon White
@oxinabox
Oh I could 100% be wrong
sorry, I thought it was more clear that I was kinda speculating there.
I'm not super familar with that part of the code
pevnak
@pevnak
Should I post it as a bug to TensorFlow.jl?
Lyndon White
@oxinabox
Or a thread on Discourse
Malmaud checks it regularly
pevnak
@pevnak
Ok, will do
Asbjørn Nilsen Riseth
@anriseth
Is Clustering.jl the best source for kmeans in julia?
Andre Bieler
@abieler
Question to the heavy hitters in the field: I started converting the material from Andrew Ng's coursera course to Julia. Maybe we could use this to get people started with Julia/ML. You can see the current status on: https://github.com/abieler/mlng where only ex2 is "done". Good idea to keep going? If yes we can discuss about the format, etc. In the original course it is a .pdf file with instructions + a bunch of matlab/octave files. I condensed it into one jupyter notebook. Comments are welcome.
There are two versions of each notebook in the src folders. The one with suffix "solution" is just so people can take a look at it now and run the code, but should not be available at a later point
Lyndon White
@oxinabox
This seems like a good idea, (No further comment til I have time to look at them properly which will not tonight)
Andre Bieler
@abieler
Maybe a further comment: This is very close to the original material, just for people to be able to follow along the other course material. Implementing some julian extras could be fun too.
Vaibhav Kumar Dixit
@Vaibhavdixit02
Hello everyone, I am Vaibhav Dixit a sophomore in Mathematics and Computing from India. I wanted to get involved in JuliaQuant and needed help to get started. Can anybody help me out?
Patrick Kofod Mogensen
@pkofod
@abieler cool stuff. Just make sure you don't end up infringing on Ng or coursera's IP
Patrick Kofod Mogensen
@pkofod
Christof Stocker
@Evizero
thanks. glad people actually read it
really need to get back into it. I do have time again just need some discipline to get back into the habit
Patrick Kofod Mogensen
@pkofod
I really think some of the documentation you've written is top notch compared to many other projects (Optim very very much included in that lost of subpar docs)
Christof Stocker
@Evizero
thanks
Tom Breloff
@tbreloff
@MikeInnes (or anyone else that is interested). My colleague Stefano Pacifico is looking for presenters at the next CAIM event: http://thepostersession.com and asked if I knew of anyone in the Julia ML/AI community that wanted to do a poster. I thought presenting on Flux or maybe the GPU packages would be interesting. Let me know
Rainer Engelken
@RainerEngelken
@tbreloff Sounds interesting and not far away. Would you mind, If I send it around at my institute (https://ctn.zuckermaninstitute.columbia.edu/)?
Tom Breloff
@tbreloff
Please do @RainerEngelken !
datnamer
@datnamer
@datnamer
Anyone here know about Flux?
it seems like conv function isn't defined in both master and the latest release
FluxML/Flux.jl#218
Any help would be appreciated
Andrei Zhabinski
@dfdx
is Flux.jl even supposed to provide it? it provides types Conv and Conv2D, function conv is available in NNlib.jl which is a kind of functional layer of Flux.jl
also consider using Slack (channel #machine-learning) - seems like most folks have moved there
Christopher Rackauckas
@ChrisRackauckas
I pointed him to the Flux channel which is bridged.
but yes, the Slack is very helpful too
datnamer
@datnamer
@dfdx Yes, flux wraps the Conv function in nnlib
It's working now with a fresh base julia install
Ismael Venegas Castelló
@SalchiPapa
Hello everyone!

One of my coworkers is trying out Julia and he asked me about:

I am looking for fastest as possible (from calculation perspective):
Feature Selector and Hyper Parameters Tuner

I pointed him to all the Julia ML channels and resources I know about, but was left wondering if there is a curated list of ML packages (I haven't been actively reading Julia news lately), what would you guys recommend to him?
datnamer
@datnamer
@oxinabox or @Evizero Is there anything in Mlutils that can help me write a generator for batches from an image directory, where the target is a number (so a regression) instead of a label.
I have a dataframe of image locations and a integer y, and the manual approach would be to split that and write code against the dataframe shuffle and load the image and the target.
Christof Stocker
@Evizero
nothing aside the functionality discussed in the docs
Paulito Palmes
@ppalmes
he can use scikitlearn.jl for grid search and feature selection: https://github.com/cstjean/ScikitLearn.jl
Md. Tarikul Islam
@mti_tarik_twitter
Hello. I'm a new Julia user. I want to write a vehicle identification program that can classify vehicle from live video. Is it possible to implement it using Julia? if Yes, please suggest me which libraries should I use and relative links. Thank you
Lyndon White
@oxinabox
Take a look at MetalHead.jl