Hi, I recently came across Fast AI through Nvidia’s Deep Learning Website. I have watched Lesson 1 Overview and AWS Setup.
At this point I’m pursing bachelors and not possible for me to pay for p2.xlarge instances. Is there any alternative that lets you access GPU instances for free for a month or so?
If not, how can I setup environment locally using Docker on macOS? I want to start coding as soon as I can to implement Deep Learning Techniques.
If you’re enrolled at a Uni, you could get about 150$ in AWS credits under their awseducate program. Just apply via the GitHub student pack and you should be good.
I suggest you get familiar with Spot instances and then you could ‘boost’ your credits into giving you more usage time
But I’m also looking for a local setup to learn offline when possible. Wonder if you could help me out!
You would need a Nvidia GPU to do that.
Which would be a higher investment upfront.
Just FYI, AWS doesn’t require much data of your internet package while you’re using it, so it shouldn’t be a much of a problem.
As far as I understood from the video is that t2.micro is free to use and which has only CPUs. correct?
And if I am to use p2.xlarge that has GPU the cost would be $0.9/hr? Is it seconds based or I can use it for free for less than an hour?
It’s hours based, but no free lunch. You get charged for an hour whether you use 1 second or 59 minutes.
Docker on a Mac would work, but you won’t be able to use the GPU unless you have an NVIDIA card (and think nvidia-docker might be only Linux, but can’t remember). You can definitely use Keras and TensorFlow on a Mac, just again, maybe not with GPU.
Okay for the AWS info.
I definitely know that Nvidia-docker image is for computers with Nvidia chips built-in. I wonder if using Keras and Tensorflow locally could be enough to create projects or run fast-ai notebooks/projects as I will be using CPUs on Mac.
One last question - How long does take while running these codes on CPU? (as far as the first and second lessons are concerned) It wouldn’t take days, would it?
I wouldn’t suggest training them on your Mac. You might end up frying your motherboard in extreme cases (if you keep your MacBook overheated for a while, that’s the extreme end result)
Oh! Then I guess I wouldn’t try it…!
Keras and TensorFlow work on CPU (including Mac), but very slowly.
Can handle small networks on MNIST, for example, but epochs will take an order of magnitude or two longer.
In fact, the CPU is enough for your train and fine-tuning based on simple NN structure.
You can use the Keras Docker following the readme inside keras repo. Just modify the Dockerfile a little bit for using python 2.7 and Keras v1.