Crestle [easier alternative to AWS]

Billing works through Stripe who accept debit cards as well. I’ll update the site to clarify.

@noskill 10 hours feels like enough time to try things out; the change helps keep Crestle sustainable long term. AWS doesn’t cover GPU instances under their free tier, so Crestle’s offer still works out to be better for users.

2 Likes

@anurag absolutely right!!. i will try Stripe for payments.
one quick question - how do i download my prediction files from directory??

The easiest way to do that is to upload the files to your own server (with scp) or use something like AWS S3.

@anurag i ran out of free offer. please update the payment details(stripe) on site so that i can continue using crestle!!.

I’m not sure I understand. Are you unable to enter your debit card into the payment form?

yes, i used mastercard’s debit card and it says you can’t use this type of card for purchase.

Oh, I see. It looks like Stripe does not support your particular card in India. Can you send me a DM with the type of card it is? I’ll get in touch with them to check.

Can you let me know if you will updating the github repo with the second course anytime soon? Also, the dogs-casts-redux dataset is in the /datasets folder as a zip file. Do we have to extract them to the user space to work on them or is it possible to extract them in the root directory and work (will reduce the usage space for user)?

Thanks a lot for Crestle!

If you look under /datasets/fast.ai/dogscats/ everything’s unzipped already and ready to use.

I haven’t looked into the changes needed for part 2 yet, but if you run into anything let me know.

Quick update: Crestle registrations are open now. No invites necessary.

Hi, @anurag :slight_smile:

I must say, this is one of the easiest workflows that I’ve experienced.

However, I cannot use it because the IO is a huge bottleneck (as it is with FloydHub; @sai, @narenst - you know this, I’m sure).
If you can somehow eliminate it, this has the potential to be the go to solution for cloud-based deep learning.

Best wishes!

P.S. Keep us updated. We’re rooting for you.

I must say, this is one of the easiest workflows that I’ve experienced.

Great to hear, thanks. And I hear you on the IO. It’s an NFS issue that only slows down writes (reads are fine) but it’s still annoying. On my list.

Besides the writes, reading/handling (eg.unzipping) a large number of files is a pain. I believe you referred to it earlier yourself:

Writing lots of small files is going to be slow over EFS (which Crestle uses under the hood)

So, for example, it’s slower to use a Keras ImageGenerator.

3 Likes

@anurag Any thoughts on teaming up with fast.ai when the next version of Part 1 starts in Oct? That would really make the setup easy.

Already working with @jeremy on it!

4 Likes

My experience, having started from nothing with fast.ai a few days ago, is that making everything work nicely on AWS took me about 10 hours, and even then I had poorly patched insecure systems. Making everything work up to the end of lesson 1 on crestle then took me about 20 minutes. including the compute time. Thank you very much, @anurag. I will persist with crestle for the next lessons.

4 Likes

Emrys: me too. From what I can tell, AWS no longer supports p2.xlarge, at least in my region. On to Crestle or TensorPort. I just hope the lesson instructions still work with these other resources.

Thanks for letting me know @Emrys !

This is great, thanks @anurag!

Quick question, is it possible to get the connection information for the Jupyter notebook to allow other Jupyter clients to connect? I do a fair amount of work on an iPad while traveling and unfortunately the keyboard I have doesn’t an escape key! I’ve started using the Kernels app to connect to Jupyter to compensate for this and would love to be able to connect the two.

Thank you!

The URL you see when you open the notebook in a new tab is the actual Jupyter connection URL. Something like https://s.users.crestle.com/bgg3yzmo. The port is 443 since SSL is enabled. Try it out?