Floyd - alternative to AWS P2 instance


#61

I ran into a couple of issues regarding backward compatibility of Keras 2 which is the current default on Floyd.
My issues could be fixed by following the notes expressed in https://github.com/YuelongGuo/floydhub.fast.ai/issues/1#issuecomment-323515380

Thanks everyone! Jasper


(Fernando Santos) #62

Hey @todiketan and @joelg,

You can search for floyd public datasets here. Some datasets used in the course are already there, like catsvsdogs and catsvsdogs redux.

And I saw now there’s also a project there with the files for the first lesson of the course: https://www.floydhub.com/fastai/projects/lesson1_dogs_cats. I could clone it and create my own project with the exact same files. It was quite simple to do it…

Fernando


(Giulio Matteucci) #63

Hi @sai!

Is there a way to customize Jupyter running in Floyd using jupyterthemes (https://github.com/dunovank/jupyter-themes)?

I tried to do so installing the package via terminal and then typing “jt -t grade3” but the appeareance of my jupyter notbook did not change (also after a refresh)…


(Francisco Ingham) #64

Hi kijes if could you share

prepare.py

that would be very useful.

I am having some trouble building up the data structure from within Floydhub given that the /input directory where I have my data is read-only.

Thanks!


(Rafal Kijewski) #65

Hi,
You can check this gist: https://gist.github.com/kijes/163bbceca3b6fb25f210f1d2033a4f97
I’ve refactored the code since the last post. Generally the preparation is a separate step and then I use its output as input to training phase.
In the gist:

  1. I call “prepare_train_full_fh_cmd.sh” to run FH command to prepare the data
  2. The script calls another one “prepare_train_full_fh.sh” which just calls python script “prepare_train_full_fh.py”
  3. “prepare_train_full_fh.py” downloads and unpack data from Google Drive into “/output” (you need to provide file id that you can get when you share your file)
    I download everythin that’s is needed for the experiment: training data, test data or pretrained models.
    You will need to adjust the paths.

(Francisco Ingham) #66

Thanks kijes!


(Vivek Reddy) #67

Hi kijes,

Thank you for your post.
You mentioned earlier that you had a shareable google drive link to the lung data.
I was wondering by any chance if you would be willing to share the link? I have had trouble being able to download the data and no space on my personal computers.


(Rafal Kijewski) #68

Hi Vivek,

I don’t store the data permanently. I upload before doing some experiments on FloydHub.
It all depends what I’m working on.