Rupesh K Srivastava

We agree, it does not look like we need full blown logging. @untom: There are three alternatives we can suggest monitoring large amount of logs, say from update-wise logs: 1....

The `InfoUpdater` hook actually submits the logs to the database at every call. We have used it this way and it was pretty handy. Currently it only supports being called...

In that case you would have verbosity for `Trainer` to be True (default), specify the filename to dump logs to, and set verbosity for per update hooks to False. You'll...

To be clear, the verbosity behavior is how the hooks work already. It does seem like a bit of flexibility that may not be needed. We have this so that...

@untom, any ideas for tackling this issue?

@ntfrgl, nice to see you paying attention to the issues already! Here's my current take on this issue. I don't feel that we gain much from incorporating full blown logging...

Our usage of `skcuda.sum()` was rather weird, due to a bug in scikit-cuda. The fix above will hopefully avoid such crashes. Let's keep this open for a while to see...

The reason for not using pickle is security, since saved networks will probably be shared by users with others. I recommend that we use hdf5 as much as possible. Brainstorm...

Indeed: http://www.benfrederickson.com/dont-pickle-your-data/ On 11 August 2015 at 22:52, untom [email protected] wrote: > Is pickle insecure? > > — > Reply to this email directly or view it on GitHub >...

JSON + array in HDF5 is good. Full state/gradients are probably not needed. What would actually help is to have helper functions (which can be used as hooks) which allow...