Providing a tl;dr: Google provides a jupyter / ipython research environment with some amount of integration to their existing services. This environment they call "Colaboratory". They have now started providing free access to GPU resources from within this environment. Details for using with tensorflow here: https://colab.research.google.com/notebook#fileId=/v2/extern...
Anybody know how to upload large data files into these environments? Possibly from google drive? I spent some time at it and the example loading techniques all run into download limits or IO fails. Anybody successful at loading large data files?
It's previous gen hardware, so presumably the only cost to google is energy usage and similar. They probably use the newer p100 and v100 cards internally which are significantly faster for deep learning.
Not only that, the hardware is already on, so it's only a matter of the cost difference between an idle GPU and a busy one. It's not inconceivable that if enough paying customers showed up with K80 workloads, these freebie users might have to wait longer.
Is it possible to start a training run, logout for a month-two and return back once training is complete? Or does it only allow a single browser session?
1 hour timeout on runs. I'm assuming that means from one command execution to another, not per batch but shrug. Once it times out it looks like they save state and you can reconnect but I've not checked very closely. Same for saved data but I assume there's some way to chuck it into drive which is the storage backend as far as I can tell.
17 comments
[ 3.1 ms ] story [ 52.5 ms ] thread1. Upload local files directly (Doesn't work on Firefox)
2. HTTP download
3. Google Drive, Google sheets, Google Cloud Storage..
Take a look at this for more info: https://colab.research.google.com/notebook#fileId=/v2/extern...
I just copy and pasted a keras MNIST demo and it seemed to work like a charm.
https://colab.research.google.com/notebook#fileId=1tD_viugd-...
(think you need a google account, fwiw)