Crestle [easier alternative to AWS]

(Rengarajan Bashyam) #21

@anurag, Crestle is just awesome. Though I love the rigmarole of AWS, Crestle is a real boon for time constrained people like me. Great stuff.


Thanks Anurag. Even though you couldn’t have made the layout any simpler, in my excitement I missed the front page sign up field.

I have another naive question that I haven’t been able to answer and I’m just curious - where are the lecture datasets stored?

For instance, in the lesson 1’s (CatVsDogs) lecture notes the dataset path has been specified as "path = “/datasets/” which contains the train, test and validation sets etc. I wasn’t able to find the datasets folder in the terminal.

Crestle is just perfect, thank you so much once again for creating this once again.


Hi Anurag,

This looks great and it couldn’t be easier to get started. However, I jumped on and gave it a quick trial run and was running into package version issues. For instance, the original scripts that Jeremy wrote for this course use Keras 1.1.0, I believe. It appears as if the instances Crestle provides come equipped with Keras 2.0.5 at the moment. I don’t want to have a change a bunch of the Notebooks for this course to accommodate Keras 2, so I tried to install Keras 1.1.0, and it appeared as if it installed when I ran pip2 install keras=1.1.0. Unfortunately, when I went to import the various components in the Keras module it was still loading Keras 2.0.5.

Can you advise on how overcome these versioning issues?


(Anurag Goel) #24

The datasets are in a read-only folder at the root of the filesystem: /datasets/.

(Anurag Goel) #25 notebooks preinstalled in your account are already set up to work with keras 1.2.2, and you install it as part of lesson 1:

Note that you need == instead of = for pip to work, and the kernel needs to be restarted to reflect the new version. Let me know if you run into any other issues.

(Erik) #26

Hi Anurag,
I just emailed you about this. I’m on lesson1 dogs-cats and was trying to figure out how to upload the entire dataset. Any suggestions you can offer would be appreciated. I am concerned it will take a very long time and I didn’t see a way to upload the entire folder into Jupyter.


(Anurag Goel) #27

@erikvp the best way to get data into Crestle is to use wget or curl from the terminal. For an example, see the last question here:

(Erik) #28

Thanks Anurag, that worked!

(satish) #29

@anurag Somehow Copying folders is very slow . It is wasting most of the time . I am copying the files from /datasets to my own directory

(Anurag Goel) #30

Writing lots of small files is going to be slow over EFS (which Crestle uses under the hood). I’d avoid copying files from datasets if possible. Alternatively, you could run the operation overnight in a CPU-only terminal.

(satish) #31

Thanks @anurag . Only problem is cost associated with it cost. If we can do a SSH to the terminal , we can have a script written to do that and close it at the end of the day.

By the way you service works like charm . Great job


@anurag somehow I could not type anything under crestle terminal :cold_sweat::cold_sweat::cold_sweat: it was okay under notebook

(Anurag Goel) #34

@cesco Try restarting Jupyter? I wasn’t able to reproduce the issue and haven’t heard similar reports from other users.


@anurag it worked under chrome, could be problems of safari?

(Anurag Goel) #36

Works for me in Safari too.

(Jaime) #37

This looks great, will give a try!

(Anurag Goel) #38

Thanks. LMK what you think!

(Bhabani) #39

I am using Keras 2.0 python 3.5.x

(noskill) #40

@anurag now you are offering 10hrs instead of 25 hours??? please provide a debit card option as here in india , credit cards are not used very often…

and thanks for providing GPUs at such low cost. I used AWS and google but after using the crestle, i feel like its worth paying for crestle GPUs than aws and google.

(Simon) #41

Same goes for Germany/Europe, in regards to credit cards. Paypal option would be great as well!