Library of Pre-Trained Networks

The page contains a library of pre-trained networks (i.e. weights) that can be simply loaded as models in tensorflow. To do this, please see the Apply to Data Cube example on the examples page. Additionally, I will report the results of each pre-trained network.

What were these networks trained on?

For each set of data in the library, we created 30,000 synthetic SITELLE spectra using the formulation described in the Create Synthetic Data example of varying resolutions. The velocity was uniformly sampled from a distribution ranging from -500 km/s to 500 km/s. The broadening values were uniformly sampled from a distribution ranging from 10 km/s to 200 km/s. Each section contains a pickle file containing the spectra used, the saved CNN model, an image of the reference spectrum, and statistics on the accuracy of the method on the test set. Click on the icon to download the pre-trained CNN. If you would like a pre-trained library with different specifications, please feel free to contact us at carter.rhea@umontreal.ca.
Here are the files I used to create these pre-trained networks: Generate.py & Train.py