Closed nicholasturner1 closed 9 years ago
Added (optional) "logging" option to configuration file, and recording the configuration for a given run tagged with a timestamp.
I have plans for logging: (1) Data samples fed to the network, by sample id, offset, and transformations performed before the forward pass (2) Learning Curve statistics alongside the above information
Hopefully this combination, though slightly redundant, will make debugging easier
It might be helpful to have the training (and possibly the forward-pass) record a log of the samples passed to the network. That way, if we get an obscure error (segfault, etc.), we can more easily replicate the bug. Logging the samples could be as easy as identifying the sample filename and the deviance from the middle.