3 ms·
Maybe we can repurpose the SETI@home infrastructure :)
by sceadu 4y ago
Maybe we can repurpose the SETI@home infrastructure :)
- kmeisthax 4y agoBOINC might be usable but the existing distributed training setups assume all nodes have very high speed I/O so they can trade gradients and model updates around quickly. The kind of setup that's feasible for BOINC is "here's a dataset shard, here's the last epoch, send me back gradients and I'll average them with the other ones I get to make the next epoch". This is quite a bit different from, say, the single-node case which is entirely serial and model updates happen every step rather than epoch.