Machine Learning on AWS and not

So gonna give machine learning a try on AWS. Ultimately, I think this is not economically viable, but since I’m just screwing around, whatever.

I searched for a prebuilt AMI ami-180ad670 called gpu_theano or something. Picked the first one for no reason except it seemed reasonable.

I picked a spot instance with .30$ per hour as max rate. g2.2xlarge

ran some script inside called Theano appears to be working on the gpu. (theano )

Let’s see if we can get this puppy running

Error: Command ‘[‘/home/ubuntu/neural-doodle/pyvenv/bin/python3’, ‘-Im’, ‘ensurepip’, ‘–upgrade’, ‘–default-pip’]’ returned non-zero exit status 1

Okay. Forget doing the virtual env stuff. Not worth it on a burner computer.

sudo apt-get install python3-dev

sudo apt-get install python3-pip

sudo python3 -m pip install numpy

sudo python3 -m pip install scipy

Why aren’t these in requirements.txt

I wonder if I could use python2. scipy and numpy are probably already installed.

sudo python3 -m pip install –upgrade setuptools

sudo python3 -m pip install –upgrade cython

sudo apt-get update

sudo apt-get install libfreetype6-dev

sudo apt-get build-dep pillow

sudo python3 -m pip install –upgrade scikit-image

sudo python3 -m pip install  theano

sudo python3 -m pip install lasagne

Bad move DOn’t install lasagne and theano on their own

This is running slow as hell. What is up. Github page

Using screen so ssh failing won’t quit job

use screen

run your job

detach with ctrl-a ctrl-d

then you can reattach with screen -r,-CUDA-7,-cuDNN)

Interesting Link. Should try this next time.

Okay. I got a free nvidia graphics card (GTX 560 ti) from a bro. I set up my router to forward port 22 to my desktop so I can ssh in from anywhere. Installed cuda and cudnn. Tensorflow by default does support this old of a graphics card. Saw some rumblings

FInally got scipy to install once I download the blas and lapack libaryr prerequisites using apt-get.

I was having a lot of trouble installing scikit-image with some kind of error about pgen. Eventually I renamed /usr/local/bin/pgen which is not the porgram it is expecting to /usr/local/bin/pgentmp  and then it seems txo get past it fine.

sudo apt-get install libatlas-base-dev

Needed to install some matplotlib dependancies

Needed to set a .theanorc file. Change cude5.5 to the verison you have

Finally running.

I’d say the speed is on par with the AWS. It will take about an hour or two to finish the job at 40 iterations.

Phase 3 is the beast.

Failed at phase 3. My card has only 1GB of ram. Not enough I guess.

I’ll post this for now, but clearly a work in progress.



Leave a Reply

Your email address will not be published. Required fields are marked *