JOIN
Get Time
forums   
Search | Watch Thread  |  My Post History  |  My Watches  |  User Settings
View: Flat (oldest first)  | Threaded  | Tree
Previous Thread  |  Next Thread
Re: Final Testing clarification (response to post by Mloody2000) | Reply
The training process should do everything that makes it possible to run inference later. It should consume raw training data and should output a model that you'll use for testing. I suggest we return to this problem later when contestants have measurement data to prove that the given configuration is not enough to finish training. If this will be the case then we have several options, like verifying your training process only on a subset of data, or using stronger test machines. Unfortunately the 7 days training limit can't be extended, even with the 7 days it will be tough to run all tests between the end of submission phase and the time final results should be announced.
Final Testing clarification | Reply
Hi guys,

could you please clarify if the train.sh file should produce a trained model from scratch or just pre-process the data (if necessary) ?

In case it should do the full training, the given configuration seems a bit low:
- 7 days
- p2.xlarge

From what I've read, p2.xlarge are rather slow nowadays (btw, best article on GPUs for DL => http://timdettmers.com/2017/04/09/which-gpu-for-deep-learning/ ). This is roughly equivalent to 1-1.5 days for a 1080Ti or Titan. Given the dataset size, this looks low (I didn't start training yet, just noticed this point that may be a problem).
RSS