search

dicarlolab / archconvnets

Architecturally optimized neural networks trained with regularized backpropagation
5 stars 5 forks source link

checkpoint pickle file -> database #26

Open daseibert opened 10 years ago

daseibert commented 10 years ago

Ideally we would have a function that takes as input the path of a pickle checkpoint file which the function then inserts into the database as a new model entry.

ardila commented 10 years ago

I am going to write some code to do this so I can run the caffe model (whose weights I now have). @yamins81 do you have any strong desires regarding required fields for storing models that have been trained with an outside procedure? I think I might just have a comment describing how I got the weights, and check the code that does so into some repo

yamins81 commented 10 years ago

Code to insert checkpoints in the standard format is one problem, and it is basically already solved by existing code. I will break this code out into a separate function this afternoon. The other question about models from a different format should be resolved separately. On Jul 24, 2014 3:17 PM, "Diego Ardila" notifications@github.com wrote:

I am going to write some code to do this so I can run the caffe model (whose weights I now have). @yamins81 https://github.com/yamins81 do you have any strong desires regarding required fields for storing models that have been trained with an outside procedure? I think I might just have a comment describing how I got the weights, and check the code that does so into some repo

— Reply to this email directly or view it on GitHub https://github.com/dicarlolab/archconvnets/issues/26#issuecomment-50064606 .

ardila commented 10 years ago

What I meant is that there are many fields in that are typically there in 'rec' that cannot really be filled if the training is an outside procedure. I guess your code just makes some assumptions about this, and I'm ok with that.