Platform Information for Fast AI Part 1 (2020)

Part 1 (2020)
1

What will be the recommend platforms for Fast AI Part 1 (2020).

For instance, for AWS is a p2.xlarge machine good? What are the other recommended platforms?

1 Like
2

+1.

Are there any recommendations for server setup like Colab, GCP, etc (same way it was done for v3), so we can set it up and use once the course starts on March 17?

3

For Colab, I know I’d recommend the following:

  • Open from GitHub
  • Navigate to the course repo
  • Open whatever notebook you want
  • Select the GPU option under Runtime
  • !pip install fastai2

That’s about all that should be needed. It used to be a bit more complex on the install but dependencies were updated on both fastai2 and what Colab natively runs :slight_smile: (Also for any future questions about Colab specifically, please feel free to @ me, as I run it exclusively)

9 Likes
4

Great, thank you, I will try it. Do you run your Colab with free or 9 usd plan? Which one will suffice?

5

I personally upgraded, but that’s simply because I love Colab and use it so much. The free version should be good enough though! (Just note that sometimes GPU’s may not be available depending on the current usage by others)

2 Likes
6

ok, perfect! I will try to set it up, thank you!

7

Colab is a great service, and I was not aware of the Pro service till now. Unfortunately Pro is currently only available in the US. Looking forward to it reaching the EU.

1 Like
8

I’ve heard of people getting access to it outside the states IIRC (don’t know how but it’s possible). But shouldn’t be too long :slight_smile:

1 Like
9

If you want to go the local development route, I think this is one of the best guides to building your own DL machine …

https://waydegg.com/making-a-dl-server.html

1 Like
10

Generally the best approach is to follow the “server setup” guides here:

https://course.fast.ai/

Then just add pip install fastai2 to the end of the process.

6 Likes
11

GCP is decent. When you sign up you get $300 credits which is more than enough for the course. Additionally, there are deep learning images with PyTorch and fastai v1 installed. Very easy to set up. With the in-browser SSH feature, all work can be done in the browser!

12

Sorry for asking a stupid question once again, but I am struggling with installing fastai2 on GCP. I’d like to keep my previous instance with previous fastai version since it is working well for me now.

So I created a new instance for fastai2, but will doing pip install fastai2 mess up my old instance since it is using old fastai version?

The instructions said that after creating new instance, do the following:

cd tutorials/fastai/course-v3
git checkout .
git pull

sudo /opt/anaconda3/bin/conda install -c fastai fastai

Will these install fastai 2 , all lessons and the ebook? Could someone please clarify this for me if possible? Thank you in advance.

13

No. IIRC that’s not even updated yet. You should install fastai2 alone, Jeremy hasn’t released any course notebooks (for this year yet), we just have the ones from last year living in course/ under fastai2 and the book draft

1 Like
14

oh, ok, thank you very much!

15

For AWS, I would recommend g4dn.2xlarge or any of g4dn series which come with T4 GPUs that are generally twice as faster as the K80 in p2.xlarge.

Also, the g4dn.2xlarge costs $0.752 - cheaper than the p2.xlarge which is $0.9.

4 Likes
16

I thought that T4s were meant for inference–are there any drawbacks to using them for training?

17

They work pretty well actually. With Fp16, it’s faster than K80 fp16 actually.