Crestle [easier alternative to AWS]

joshfp · November 21, 2017, 12:10pm

I started using Crestle. First, I found that Theano was not installed. After installing it, when running “import theano” I get this error “You are tring to use the old GPU back-end. It was removed from Theano. Use device=cuda* now. See https://github.com/Theano/Theano/wiki/Converting-to-the-new-gpu-back-end(gpuarray) for more information.”. I tried to set device=cuda in $HOME/.theanorc, but it didn’t work. Hope you can help me.

sgaseretto · November 21, 2017, 2:11pm

I’m having the same exact problem, has any one solved it?

anurag · November 21, 2017, 9:38pm

Have you installed keras 1.2.2? It is needed for the course and automatically installs the supported Theano version.

I’d recommend uninstalling the Theano version you installed and following the instructions here:

github.com

anurag/crestle-fastai/blob/master/deeplearning1/nbs/lesson1.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Using Convolutional Neural Networks"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Welcome to the first week of the first deep learning certificate! We're going to use convolutional neural networks (CNNs) to allow our computer to see - something that is only possible thanks to deep learning."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [

This file has been truncated. show original

sgaseretto · November 22, 2017, 12:59am

I just did what you said, and also a friend did the same when trying for the first time Crestle, and we are still getting an error when running that notebook with enable GPU. The curious thing is that when running the instance without “Enable GPU” the import of utils works, but then it throws an error when trying to train the model.
The error I’m getting when running crestle with GPU is the following.

ValueError: You are tring to use the old GPU back-end. It was removed from Theano. 
Use device=cuda* now. See https://github.com/Theano/Theano/wiki/Converting-to-the-new-gpu-back-end%28gpuarray%29 for more information.

anurag · November 22, 2017, 2:49am

It should work if you install the previous version of Theano:

pip2 install theano==0.9.0

kvinicki · November 22, 2017, 9:41am

I got this error: Unreadable Notebook: /home/nbuser/courses/fastai/deeplearning1/nbs/lesson1.ipynb NotJSONError('Notebook does not appear to be JSON: ‘{\n “cells”: [\n {\n “cell_type”: "m…’,)

Can you please help? I can open other lessons. Problem is with just lesson1

sgaseretto · November 22, 2017, 12:07pm

Thanks a lot for the help Anurag!
Everything is working now. I’ll summary everything so that if anyone has the same problems will know how to solve it. For anyone who is trying to run the original notebooks of part 1 from the course, first you will need to run the following commands in the terminal

pip2 install keras==1.2.2 theano==0.9.0

and then

mkdir -p ~/.keras
echo '{"epsilon":1e-07,"floatx":"float32","backend":"theano", "image_dim_ordering": "th"}' > ~/.keras/keras.json

Also this can be included in the notebooks the same way anurag included it in the crestle-fastai notebooks here:

github.com

anurag/crestle-fastai/blob/master/deeplearning1/nbs/lesson1.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Using Convolutional Neural Networks"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Welcome to the first week of the first deep learning certificate! We're going to use convolutional neural networks (CNNs) to allow our computer to see - something that is only possible thanks to deep learning."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [

This file has been truncated. show original

If you have already installed theano before doing this steps, you will probably encounter an error like this:

ImportError: cannot import name inplace_increment

To solve this, uninstall that theano and delete the theano-cache running theano-cache purge or deleting it manually with rm -rf ~/.theano
and then install the theano version 0.9.0

sgaseretto · November 22, 2017, 12:22pm

I had the same error, what I did was to upload myself the notebook from lesson1 that Anurag posted, and it worked as intended. If you want to run the original notebooks maybe you will need to install the packages with the terminal before running any notebook

kvinicki · November 22, 2017, 1:45pm

tnx, that solved the problem

anurag · November 25, 2017, 5:51am

As an update, Crestle includes Theano 0.9 as the default for now. When the new version of the course is released in a few weeks we can stop using Theano/Py2 altogether.

GiulioM · November 25, 2017, 5:29pm

Hi @anurag!

I was wandering if you already managed to solve the issue with jupyter themes already reported by @mosessoh.
I tried to use jupyter themes on crestle and I encountered the same problem… Is there any workaround?

anurag · November 26, 2017, 8:52pm

I haven’t yet. Which theme are you looking to use?

zest · November 26, 2017, 9:06pm

Hello @anurag,

I am having problems downloading data from Kaggle (the carvana-image-masking-challenge). I have tried the kaggle-cli (install, kg download etc …). Is it not possible to use the kaggle-cli on Crestly?
I have also tried wget as recommended in your site. The wget works on the cifar example you provide in your FAQ. But I am not able to download the Carvana data. This is a sample command I use (I have tried other versions):

wget https://www.kaggle.com/c/carvana-image-masking-challenge/download/train.zip
wget https://www.kaggle.com/c/carvana-image-masking-challenge/data/train.zip

An empty file is downloaded in the first instance, the index page in the second.

I have written to support@crestle a few times so you will see my question there as well. I have found my deep learning setup on AWS to be very frustrating which was why I was trying out Crestle.

anurag · November 26, 2017, 9:34pm

I wouldn’t use wget for this because it does not send your Kaggle auth credentials by default. In the terminal, try upgrading your version of kaggle-cli:

pip3 install -U kaggle-cli

# followed by

kg download -u <username> -p <password> -c  carvana-image-masking-challenge

Replace <username> and <password> above with your details. If this doesn’t work, please post the output of the commands above.

GiulioM · November 26, 2017, 9:59pm

I like “grade3”

zest · November 26, 2017, 10:00pm

Thanks. Yes “pip3 install -U kaggle-cli” did the trick. I had tried “pip install -U kaggle-cli” before. Perhaps you can include information about kaggle downloads in your FAQ.

vateye · November 27, 2017, 9:15am

I want to use Caffe on Crestle, but I cannot use the ‘sudo’ command which shows that not permitted. How can I fixed this, And how can I use ‘Tensorboard’ on Crestle? Thx.

anurag · November 27, 2017, 5:54pm

Caffe can’t be installed manually on Crestle at the moment because it requires admin privileges. I’ll look into adding it to the base image soon. Tensorboard isn’t supported, since it’s a Jupyter-only environment for now.

anurag · November 28, 2017, 5:17am

I’ll plan on adding support for user-installable extensions and themes to Crestle.

GiulioM · November 28, 2017, 8:37pm

Thank you, it would be great!