ImageNet: release configurations

diux-dev / cluster

train on AWS

75 stars 15 forks source link

ImageNet: release configurations #55

Open yaroslavvb opened 5 years ago

yaroslavvb commented 5 years ago

current best-known list of configurations to be used for release @bearpelican

1 machine: xar_args_pytorch
4 machines: x4ar_args_bench
8 machines: x8ar_args_benchmark
16 machines: x16ar_args_benchmark
32 machines: x32ar_throughput

yaroslavvb commented 5 years ago

x4ar_args_bench is broken because of missing sz/352 folder (it's missing on EFS, so it didn't make it into snapshots used by replicate_imagenet.py), so instead

4 machines: x4ar_args

bearpelican commented 5 years ago

@yaroslavvb I'll do some testing on this to see the performance difference. Perhaps it's not worth it to have that folder

The current pytorch_source ami has all the necessary resize folders. I can copy those over to our snapshots.

yaroslavvb commented 5 years ago

Dropping the 32 machine configuration for release, testing 1, 4, 8 and 16 on new account