just so that you appreciate this outstanding gift: the equivalent compute time would quickly be worth several dozen of dollars if not hundreds of dollars on Amazon (plus setup time).
this is truly one of the most remarkable giveaways since … the Internet. of course by doing this they want to find talent and hook you. but I see no harm in this bait and switch scheme.
This is great - especially for small projects and learning. However, there are a few limitations to be aware of if you’re thinking about using it for real projects.
First, is memory. I’ve seen reports of practical limitations being around 500mb. Call me a spoiled millennial noob but the vast majority of my real-world ML projects use a lot more - especially when working with image data.
Second, is getting data into your notebook. Most ML is data heavy and you want a fast way of working with it. For example, the Google Landmark challenge on Kaggle has nearly half a terabyte of unaugmented images just for the test set. You could easily push several terabytes for that one-off challenge alone if you were not careful.
Note: it is possible to work with much less data in many cases by resizing and preprocessing the images as you pull them down but that has problems too. Also, you can use Google Drive but it’s not ideal for large datasets.
15 comments
[ 1.6 ms ] story [ 42.8 ms ] threadQuick summary: Google Colaboratory (Colab) provides a Jupyter notebook environment with up to 12 hours continuous access to an Nvidia K80 for free.
https://research.google.com/colaboratory/faq.html
this is truly one of the most remarkable giveaways since … the Internet. of course by doing this they want to find talent and hook you. but I see no harm in this bait and switch scheme.
TL;DR: wow! free supercomputer time!
First, is memory. I’ve seen reports of practical limitations being around 500mb. Call me a spoiled millennial noob but the vast majority of my real-world ML projects use a lot more - especially when working with image data.
Second, is getting data into your notebook. Most ML is data heavy and you want a fast way of working with it. For example, the Google Landmark challenge on Kaggle has nearly half a terabyte of unaugmented images just for the test set. You could easily push several terabytes for that one-off challenge alone if you were not careful.
Note: it is possible to work with much less data in many cases by resizing and preprocessing the images as you pull them down but that has problems too. Also, you can use Google Drive but it’s not ideal for large datasets.
Related:
https://www.reddit.com/r/MachineLearning/comments/84532y/n_g...
https://stackoverflow.com/questions/48750199/google-colabora...
The 500mb is simply the host ram for the python runtime. As long as you aren't converting everything to numpy arrays, you'll be fine.
You can mount a remote ssh disk as a working directory, and for your actual data stream it in from cloud storage direct from tensorflow.
Nothing is ever free forever.
And they already include opencv
So I guess this is an incredibly awesome gift for many more than just ML devsAfter playing some random python module bingo here are some of the most useful libraries