mariusae / tune Goto Github PK

Hi,

I wrote a quick nightscout-to-tune input converter (https://github.com/trixing/tune). It produces json files like attached, which seems to conform to the specified format (for 1, 3, 7 days respectively in this case).

Unfortunately the produced basal schedule is rather meaningless: a factor of ~10x too high, and also the same result if I run on 3 or 7 days.

3-day data
{"version": 1, "timezone": "Europe/Berlin", "insulin_sensitivity_schedule": {"index": [0, 390], "values": [170.0000000000016, 169.99999999999957]}, "carb_ratio_schedule": {"index": [0, 360, 660, 1050], "values": [9.000000000000052, 9.000000000000027, 9.000000000000002, 9.000000000000005]}, "basal_rate_schedule": {"index": [0, 60, 120, 180, 240, 300, 360, 420, 480, 540, 600, 660, 720, 780, 840, 900, 960, 1020, 1080, 1140, 1200, 1260, 1320, 1380], "values": [6.060164554997598, 5.789143251467735, 6.669406893763722, 7.100211844444492, 7.206123554462863, 7.198766166916485, 7.2001180743763005, 7.200090756536767, 7.1998197971922835, 7.200449445844885, 7.198854832952293, 7.202448204774257, 7.1960631269896504, 7.202852101333228, 7.2101639269212, 7.1397788680270615, 7.410869150932365, 6.530692446003393, 6.099377983276996, 5.995011592147588, 5.998777410458621, 6.0038361020430315, 5.997054760167179, 5.98992515397004]}, "training_loss": NaN}

7-day-data
{"version": 1, "timezone": "Europe/Berlin", "insulin_sensitivity_schedule": {"index": [0, 390], "values": [170.0000000000016, 169.99999999999957]}, "carb_ratio_schedule": {"index": [0, 360, 660, 1050], "values": [9.000000000000052, 9.000000000000027, 9.000000000000002, 9.000000000000005]}, "basal_rate_schedule": {"index": [0, 60, 120, 180, 240, 300, 360, 420, 480, 540, 600, 660, 720, 780, 840, 900, 960, 1020, 1080, 1140, 1200, 1260, 1320, 1380], "values": [6.060164554997598, 5.789143251467735, 6.669406893763722, 7.100211844444492, 7.206123554462863, 7.198766166916485, 7.2001180743763005, 7.200090756536767, 7.1998197971922835, 7.200449445844885, 7.198854832952293, 7.202448204774257, 7.1960631269896504, 7.202852101333228, 7.2101639269212, 7.1397788680270615, 7.410869150932365, 6.530692446003393, 6.099377983276996, 5.995011592147588, 5.998777410458621, 6.0038361020430315, 5.997054760167179, 5.98992515397004]}, "training_loss": NaN}

I'm uncertain though if I'm inputing everything correctly. My understanding is that "tune" expects a delta amount from the scheduled basal rate for example? The delta modeling is also easy to get wrong, but it looks correct.

I'd be interested to get the model to work, just as a reference point to autotune and manual tuning.

There is also code in model.py which does a cumsum ("undelta") on the basal durations. I'm not sure how that is useful but uncommenting the line doesn't change the result, which is a bit surprising. In general the algorithm seems a bit too stable?

Archive.zip