Reference trained models.

[ardila]

I saved the model/checkpoint I used for comparison with humans on munctional0:

/home/ardila/ConvNet__2013-11-22_15.25.16 @yamins81 is adding it to the database

[yamins81]

@ardila @cadieu @daseibert

This has been done. I've put this model in the "convnet_checkpoint_db" database, in the gridfs filesystem named "reference_models".

E.g. you can get this fs by running

import gpumodel
fs = gpumodel.get_checkpoint_fs(host, port, 'convnet_checkpoint_db', 'reference_models')

I inserted the model with the following experiment_data key: {'experiment_id': 'imagenet_training_reference_0'}

where host/port is where you have your database connection (forwarding to dicarlo5.mit.edu:27017)

[ardila]

@yamins81 If I want to further train a reference model, what is the correct way to do this? I imagine that if I follow what is done in the examples, new checkpoints will be saved to the same database.

[yamins81]

You follow the existing example; but you'll want to change the "experiment_id" value in the experiment_data dictionary. You'll also want to change the filesystem collection in this case; you don't want to save random models into the "reference_models" collection, but instead into some other collection (e.g., the default one, "convnet_checkpoint_fs").

[ardila]

I understand the changes you think should be made, and I can try them, but I don't understand why what you are saying doesn't load an untrained model. What about the call to convnet.py is telling it to load the trained model?

[ardila]

--load-query is used