Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Package up useful changes from my experimentaiton branch #37

Merged
merged 3 commits into from
Nov 19, 2018
Merged

Conversation

nbren12
Copy link
Owner

@nbren12 nbren12 commented Nov 19, 2018

The last weeks experimentation branch got too crazy. I am repacking the useful parts to be merged into master.

I hypothesize that blow-up in SAM is connected to linearly unstable
modes of the neural network. This commit introduces code for computing
the jacobian and it's eigenvalues in an autodiff friendly way. Also,
the weak temperature gradient in the tropics significantly alters what
the unstable modes will look like, so I include code that penalizes
unstable WTG modes.

These changs required some additions to the Batch class in the
timestepper module.
Previously the dataset class defined a sample as the whole time series from an
indididual location, but this does not allow enough separate training steps per
batch. This code allows dividing the time series into equal non-overallaping
chunks.
All the training methods are now based timestepper.Batch objects rather
than having to create them on their own. This required some changes in
the loss functions.

I also added plots for the moisture budget imbalance.
@nbren12 nbren12 merged commit fd2bd2c into master Nov 19, 2018
@nbren12 nbren12 deleted the tmp branch November 19, 2018 23:37
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

Successfully merging this pull request may close these issues.

1 participant