Each team had access to whatever hardware they wished to use. It's a competition! It's hard to scale over GPUs with cifar10. If others had gotten it working they would have submitted that result.
-
-
-
টুইটটি অনুপলব্ধ
-
That's discussed in the DAWNBench paper. It's neither good nor bad - it has its uses, but its not useful for everything. Personally I think the 'cost' competition is in some ways more interesting - although a combination of time and cost would be better still.
-
Maybe having another leaderboard which is "time on AWS instance X" would be a good comparison point
-
Yes I'd like that too. Or 'best accuracy for <$1' or 'best accuracy in <5 mins on P2'
কথা-বার্তা শেষ
নতুন কথা-বার্তা -
-
-
Great work! Would you say your hyperparameters were already tuned for CIFAR-10 (giving an edge), or would this speed be transferable to another dataset with minimal tuning?
-
Bit of both. The basic ideas work on a range of datasets - although we haven't gotten great results on imagenet yet. We only spent ~1 week on this and generally only 1-2 people working on it at a time. So not heavily optimized yet.
-
Fair enough. It's still a 50% improvement in GPU time versus the ResNet50 which is impressive. Another reason for me to expedite learning PyTorch/fastai I guess :)
-
We'll have a single GPU result up soonish FYI. You really can't assume linear scaling with cifar10!
কথা-বার্তা শেষ
নতুন কথা-বার্তা -
-
-
Now go for Imagenet :)
-
Think we might miss the deadline for Imagenet. Still need to do a lot of work to figure out how to train it fast...
-
(I don't spend much time on imagenet so I'm less familiar with it - we're doing OK at the moment, but I think there's much room to improve)
কথা-বার্তা শেষ
নতুন কথা-বার্তা -
-
-
No doubt some Leslie Smith cyclical magic

-
Spot on.
কথা-বার্তা শেষ
নতুন কথা-বার্তা -
-
-
Congrats! I am wondering how you compute the train for each epoch. Did you include startup/initialization, data-preparation, and checkpoint saving time in your submission?
-
Yes it's end to end. See the DAWNBench paper.
কথা-বার্তা শেষ
নতুন কথা-বার্তা -
-
-
What was the final accuracy? Or does the benchmark just stop at 94%?
-
Yup we just train to 94%.
কথা-বার্তা শেষ
নতুন কথা-বার্তা -
-
-
Are you only training for 1 epoch, per the readme here? Amazing if so. https://github.com/fastai/imagenet-fast/tree/2748de3b0e28ecca1e02d90b2d5bf7d9b9e58e3d/cifar10/dawn_submission …
-
Poorly named param. It's one cycle, of 50 epochs
-
Ha ok -- for a moment, my mind was completely blown! Still great result though!
কথা-বার্তা শেষ
নতুন কথা-বার্তা -
লোড হতে বেশ কিছুক্ষণ সময় নিচ্ছে।
টুইটার তার ক্ষমতার বাইরে চলে গেছে বা কোনো সাময়িক সমস্যার সম্মুখীন হয়েছে আবার চেষ্টা করুন বা আরও তথ্যের জন্য টুইটারের স্থিতি দেখুন।