blasscoc / linkedinarticles Goto Github PK

This is public now and github is pulling this for the website. Change to 'our' or perhaps leave out entirely.

Hello Ben,

Thank you for publishing this code, was interesting to read. Maybe we can coolaborate at some point. About your LSTM implementation:

• It would seem like you are feeding your LSTM with features and time_steps flipped? I understood that your aim is to treat depth as time, and pass on the cell state from one batch to the next. Then it would seem like your features should be the different tool readings and depth your time_step (i.e I would not have included the line X_train = X_train.transpose(0,2,1)). I am very curious on your thoughts about that, even more curious why is it working in its current form.

• The padding solution may cause some overfitting problems, this padding could be quite large compared to a single well data. I would probably have decided to go with a smaller batch size and drop the now smaller remainder. I wonder if you gave this a thought/optimization.

• You do not need to OneHotEncode the facies if you switch your loss to 'sparse_categorial_crossentropy'

• Did you optmize de window size to 7?, The decision to pick the center depth as the representative facies value is interesting, I wonder if picking the most frequent value within the window causes more/less bias.