From a sister group at msft: you can load & run the notebooks using https://notebooks.azure.com.

When you create a Library, select +New from GitHub and enter the above path. It’s free, but no gpu support :(.

Very cool! Though you might want to encourage whomever manages that page to open it on their phone. There is something wrong with the column width calculation which makes the tiles unreadable.

I believe we have a bug open on this. I'll check, if not, I'll file. Thanks!

..they should offer for the user to pay for gpu access..

Do you have Spark support? I'd love a managed Spark notebook that runs on Azure which wouldn't require any setup/manageability on my end.

Hi, not directly within the free Azure Notebooks service, however, we do have:

https://azure.microsoft.com/en-us/services/hdinsight/

which is a paid service with its own Jupyter notebook support. You can authenticate yourself in Azure Notebooks and use any Azure service as well, but it's a bit more complicated. We hope to improve this in the future.

Ah thanks, I've tried HDinsight in the past and it's a bit of a pain to set up and maintain. Also, I had to pay for the cluster reservation. I was hoping for a more pay as you go Server less service like what Databricks has.

We've had plans to provide a paid version for a while (per the faq), and once we do, hopefully it'll be straightforward and friction free. Ideas, usage scenarios welcome (pls file your magic-wand wish list on github!)

That installation tutorial is a far, far cry from installing Anaconda then running "conda install tensorflow-gpu". Microsoft has a simpler python-only installation process [1], but when I tried a Linux installation I encountered the same errors that other people are seeing [2]. I expected a Microsoft product at version 2.2 would have more polish than this. Hopefully they will improve the installation process soon.

[1] https://docs.microsoft.com/en-us/cognitive-toolkit/Setup-CNT...

[2] https://docs.microsoft.com/en-us/cognitive-toolkit/setup-lin...

For the Microsoft shops out there, our Deep Learning platform SignalBox (https://SignalBox.ai) supports CNTK and we support deploying to Azure, if you’re interested please reach out!

/shameless plug

I’ve been doing a lot of work with CNTK lately (and for the most part I quite like it compared to TensorFlow), but there are two things in particular that need to be addressed before I think it can gain wider traction: support for batch normalization on CPUs (right now bn is GPU-only, which means a lot of Microsoft’s examples only run on Windows/Linux machines with GPUs, not something everyone has immediate access to always), and macOS support (even CPU-only support). I think those things are table stakes for any deep learning framework.

That said, it’s great to see the author’s work here - CNTK documentation can be a bit hard to come by.

Thanks for the kind words and directions.

If you have this information, one thing that would be really helpful with this type of tutorial is to know how long it took to train the models when you tested them last (and what it cost).

On the install page https://github.com/astorfi/CNTK-World/tree/master/docs/tutor... it would be nice if it says which version(s) of Ubuntu are supported.

Thank you so much for your attention.

After looking at brainscript I'm more and more drawn to the idea of building the network outside the data wrangling environment.

Do you have any experience with brainscript and can you comment on developing with it compared to python api?

It looks really simple but I do not have any experience on that.

Not warranted. From the article: "CNTK is new and it's hard to find easy-to-use code examples and tutorials for learning by doing!"