Notes

transformertf

  • Fix RNN sequence padding with torch.compile: https://github.com/pytorch/pytorch/issues/155238

  • Implement masking of padded values during loss computation.

  • Finalize eddy current fitting

    • Use early stopping, and use best params for BFGS [completion:: 2025-08-08]
    • Consolidate plotting functions, and save plots to cernbox [completion:: 2025-08-08]
    • Use different loss weights for the convolution and orbit losses [completion:: 2025-08-08]
    • Develop converter publishing for calculation of eddy current decay. N.B. this converter must publish the real in case in DYNECO or other machine modes.. [completion:: 2025-08-10]
  • Validate fitted TransformerLSTM model for Dedicated MD 2025-08-13.

  • Regarding HOLEC issues from yesterday:

“Firstine has not yet been able to resolve the issue. They have observed that raising the earth current detection threshold makes a difference, but they do not yet know exactly why this detection is causing problems.”