avishkar58 / irdm-collab-filtering Goto Github PK

Not sure yet, but I assume we won't be able to run grid searches on our laptops, so we need to get the code running on AWS or Legion/Emerald cluster to make it feasible

Currently the batch size is set to one, which results in the model learning very slowly, add batching support - but be careful to consider that some playlists may not be of the same length. See the Spotify blog post for information on how they dealt with this

Add the ability to save and restore a trained model (so that training and evaluation can be separated)

Hi,
DO you have an updated version for this? I am trying to run on google colab. It's giving lots of dependency error.
Thanks,
Suti

Determine the hyperparameters that give the best evaluation metrics:

• Sequence Length
• Hidden size
• Number of layers
• Number of epochs
• Data splitting methodology

Currently using a slightly adapted version of Tensorflow PTB example

Currently grouping by User. But we may want to incorporate some logic on top of the timestamp to determine distinct playlists for the same user (eg. split it by day or something more clever)

What evaluation metrics will we use.
Given a model, compute these metrics.

Currently just using the top 10000 lines as a single training set. Need to split all of the data into a train, validation and test set (in appropriate proportions)