Eddie - Githubissues

The code as it is on Eddie. Plenty of Easter eggs.

To do:

[x] Get rid of stuff that's not relevant for phi four
[x] Get benchmark test working
[x] Improve the model actions
- [x] Remove unnecessary parameters (sigma, interval, s_final_activation...)
- [x] Add learnable global rescaling layer
[x] Tidy up training
- [x] optimizer and scheduler initialisation in checkpoints.py
[x] Tidy up observables
- [x] Remove fields.py
- [x] Make sure multiprocessing still works
[x] Tidy up neural neural network class, removing final activation function altogether
[x] Update sampling as done in #60
- [x] Acceptance and integrated autocorrelation time should be accessible by plot/table actions
- Ideally replace the collect over sampling action (see #46 for why this would be nice), but only if the replacement is actually better (discuss in future)
[x] Update layers as done in #60 but double check training still works
[x] Use more recent PyTorch version
- [x] Update to use PyTorch searchsorted
- [x] Update free theory test to use torch.fft.fft2
Option to specify scale in GlobalRescaling instead of letting it be a learnable parameter, or otherwise stop it from varying after some number of iterations. Generally look at rescaling in more detail (separate PR)
Implement layer-by-layer plotting (separate PR)
Bootstrap the cosh fit (separate PR)
[ ] Add docstrings
[x] Add to set of example runcards
Add functionality for aggregating over many models and plotting results (separate PR)

wilsonmr / anvil