Notes
- Post-process yesterday’s MD. Will do cleaner postprocessing with Tirsi on Thursday.
- Validate online drift-correction on UCAP.
Critical bug in transformertf for LSTM-based models
- Found critical bug in transformertf where Claude implemented
encoder_alignment="left"as putting padding on the left and data on the right, which is not compatible with RNN Packed Sequences.- Affected TFLSTM-33, TFLSTM-34 and TFLSTM-38
- Resolved in MR 37 (Fix sequence alignment convention), to be released as v0.13.1
Issue with UCAP node for calculating BDOT
It seems the BDOT Converter is not publishing for user HIRADMT2 which is playing today, even though the converter is computing the BDOT properly and releasing it from the convert method with a correct FSPV. Contacted UCAP team
hi Marcin, I’m seeing a funky behavior on a UCAP node that neither me nor Michi understands. I have a device which has 3 transformations and they all publish first updates (around 20 each), but it seems that even though the (Python) converter is returning a correct FSPV they are being thrown away by UCAP. We’ve verified that the property name is correct (it’s taken from the context), and the values should be fine. it’s just a handful of users that are being muted.
Workaround is to manually restart MBI/IREF subscription in UCAP.SPSBEAM/BDOT to re-trigger the converter and publish.
Validation of sps-app-hysteresis for Dedicated MD
With BDOT mitigated, it seems the model is predicting accurately, down to about 0.2 G with TFLSTM-18. It’s important to use the online drift-correction for reliable measurements.
Interpolated predictors seem to work, but the reference must be reset after switching predictor.