Hey everyone, I’m building Crestle, with the goal of making it effortless to train deep learning models in the cloud. It’s still in private beta, but you can sign up for an invite at https://www.crestle.com. I’ll be letting people in over the next few weeks as I build out GPU capacity.
12 Likes
Is this running on AWS or are you building your own infrastructure?
Thanks @anurag, I’m interested in checking it out. I’m curious what GPUs you’ll be using. And also your pricing of course.
@davecg The intent is to abstract over cloud and bare metal GPU servers; ideally, where it’s running shouldn’t matter to the end user. Having said that, AWS is certainly one of the commodity backends.
@Even more details to come.
Question for folks who’ve signed up for an invite: are most of you still using Python 2/Keras 1 for Part 1 of the course?
yes I use python2/theano for part1 though I’d be interested to learn how to convert it to python 3/tensorflow
1 Like
I moved to python3 and keras 2 and TF for backend
python 3 keras 1 on tensorflow backend
FloydHub is also a platform for training and deploying deep learning models in the cloud. We support all the popular DL frameworks. Your first 100 hours of GPU are on us 
Happy to answer questions!
Disclaimer: I’m one of the co-creators
1 Like
Are most of you developing with CPU-only notebooks and copying work over to run on GPUs?
I moved to Python 3 and keras 2 and tensorflow for backend.
Quick update: I’ve invited around half of the folks who signed up, and will be sending out invites to the rest over the next few days.
1 Like
Another update: pricing and other details are now live on the site (once you sign in). Everyone starts off with 25 GPU and CPU hours free. Feel free to ask me questions here or on email at anurag@crestle.com.
My quick review of Crestle:
It works very well with this course. The deep learning packages required for this course are pre-installed and configured. I got it working successfully for Lesson 1 and the Kaggle’s Dogs-Vs-Cats homework task.
The python workbooks that are required for this course have also been pre-loaded. I noticed that minor additions/fixes/enhancements were made on the Python notebooks (as these were originally designed for the “official” VM in AWS) so it was not necessary to fix the keras and theano configurations. This saved me a lot of time. There is also a console in the Crestle web site, so an external SSH client is not necessary. Sessions start almost immediately when the “Start Jupyter” button is clicked, I found it to be much faster than booting up an instance in AWS.
Thanks for this service, Anurag!
6 Likes
Thanks for the feedback CY! The next thing on my list is adding datasets used in part 1/2 of the course so students don’t have to spend time (and storage costs) downloading them.
As an update, I’ve uploaded all datasets for part 1 to a shared directory and edited all the notebooks to point to them. Updated notebooks here.
4 Likes
This looks just amazing and very user-friendly
1 Like
Hi, I’m from Good AI Lab, and I’d like to share with you TensorPort. It is is a cloud platform exclusively for TensorFlow. We provide an easy CLI and a browser based interface we call Matrix.
The CLI works in conjunction with git, you can create projects with ‘tport create project’ and upload a repository of models (or just one) with ‘git push tensorport master’.
Matrix is a command center, it’s an interface for training models, collecting output, adding to tensorBoard for side by side comparison, sharing, and collaborating. We really hope you like it.
Sign up and your first 100 GPU hours are free. We would love your feed back, and want to hear what you would like next!