search

mtewes / tenbilac

Neural network for inverse regression problems
1 stars 2 forks source link

By default, an existing Normer should be reused instead of making a new one #42

Closed mtewes closed 7 years ago

mtewes commented 7 years ago

... when restarting an existing training. Add a setting in the config file to control this.

mtewes commented 7 years ago

I'll keep working in the post-xmas-fixes branch for this. Any comments welcome! I'm also wondering if the committe-member directories should not be better just have always the same names, say "member001", instead of using the config name. OK if I change this (breaks backwards compatibility with existing trainings), @kuntzer ?

kuntzer commented 7 years ago

Seems a good idea! I can change the names manually anyhow

mtewes commented 7 years ago

This is implemented, but I'm busy with the "minimize". It turns out that we also kept carrying around the "cache" of each layer, dealing with this right now. Will push asap.

kuntzer commented 7 years ago

Oh, you mean we could be freeing a lot of memory?

mtewes commented 7 years ago

Not while running (except if we don't use the cache, which maybe should also be a configurable option). But when saving to disk, yes.

mtewes commented 7 years ago

All implemented and merged, closing.