Gluon - a new deep learning library


(Mourad Hamoud) #1

In case you missed this, AWS and Microsoft announced Gluon, a new deep learning library. Read the full article https://aws.amazon.com/about-aws/whats-new/2017/10/aws-and-microsoft-announce-gluon-to-simplify-deep-learning-for-developers/

The github repo for the project: https://github.com/gluon-api/gluon-api/


Gluon :- Checkout the documentation
(Vikram Kalabi) #2

Looks like a crossover between keras and pytorch!


(Maxim Pechyonkin) #3

As is there aren’t enough libraries already. :joy:


(Maxim Pechyonkin) #4

Look here – they even have name scopes. Jeremy is going to love it!


(Mourad Hamoud) #5

Yes, that’s true :slight_smile:


(David Gutman) #6


(stijn) #7

Did anyone try it yet?


#8

I really like this gluon library. I have been through fast.ai’s part one dealing with theano/TF, and started Part 2 but never finished it due to time constraints. Now that I have more time, I was hoping that the new Part 1 would be available to all, but it seems we will have to wait until next year once the in person class is done. I didn’t want to go through the old Part 1 as Theano is dead. Stumbling upon this thread, I have focused my free learning time on it. The documents that support gluon are real easy to understand and read like a textbook. The philosophy behind it is similar to Jeremy’s. Setup of the library is simple and runs quick and easy, even on win10. It shows how to build things from scratch as well as using the API. It covers VGG, RNN, GANs, etc. I have just finished the CNN portion and can’t wait to get to get to topics such as training on multiple GPUs. It looks promising and I suggest you check it out if you have the time.


(Dominique Gilleman) #9

I tried Gluon
The API is awesome, tutorial is very clear
For now, it doesn’t need much custom code to do things, Jeremy would love it.

The big drawback, CNN are so slow (on cpu)
PyTorch: 1mn45
Gluon: 16mn
for same network, Hybridize doesn’t change anything, cpu is 100% used

It exist Mxnet Mkl, and Mxnet nnpack, with a speed up between 2x and 7x
but unfortunately, it doesn’t install on anaconda windows