Saving multiple deep learning Models during training

User 1319 | 3/15/2015, 6:08:09 AM

I’m using <i class="Italic">GraphLab Create</i> deep learning for image classification problem with a relatively small number of labeled samples. So, I want to use all available samples for training (<i class="Italic">validation set = None</i>). Currently, <i class="Italic">gl.neuralnetclassifier.create()</i> saves the model after the specified <i class="Italic">modelcheckpointinterval</i> to the same directory in <i class="Italic">modelcheckpointpath</i>, which overwrites previous models.

Is it possible to have the ability to save multiple deep learning models (at specific iterations) during training by passing a dictionary to <i class="Italic">modelcheckpointpath</i>. Something like: <i class="Italic">modelcheckpointpath= {‘25’: WORKINGDIR + '/result/model25', ‘35’: WORKINGDIR + '/result/model35',‘45’: WORKINGDIR + '/result/model45'}</i>

By setting the <i class="Italic">max_iterations</i> to a reasonably high number, we can have several models at different iterations. This is useful for models comparison or the creation of a diverse ensemble of deep neural models.




User 940 | 3/16/2015, 5:02:02 AM

Hi Tarek,

Thank you for the feature request. We'll put it on our roadmap!

Cheers! -Piotr

User 1319 | 3/16/2015, 5:51:02 AM

Thank you Piotr for considering my feature request. Highly appreciated.

Cheers! Tarek